Developmental Psychopathology (4 Volume Set) [1-4, 3 ed.] 9781118120873, 9781118120910, 9781118120927, 9781118120934, 9781119125556

Citation preview

Trim Size: 8.5in x 11in

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

12:39pm

Page ii

Trim Size: 8.5in x 11in

DEVELOPMENTAL PSYCHOPATHOLOGY

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

12:39pm

Page i

Trim Size: 8.5in x 11in

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

12:39pm

Page ii

Trim Size: 8.5in x 11in

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

DEVELOPMENTAL PSYCHOPATHOLOGY THIRD EDITION

Volume One: Theory and Method

Editor

DANTE CICCHETTI

12:39pm

Page iii

Trim Size: 8.5in x 11in

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

This book is printed on acid-free paper. Copyright © 2016 by John Wiley & Sons, Inc. All rights reserved. Published by John Wiley & Sons, Inc., Hoboken, New Jersey. Published simultaneously in Canada. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 646-8600, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008. Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold with the understanding that the publisher is not engaged in rendering professional services. If legal, accounting, medical, psychological or any other expert assistance is required, the services of a competent professional person should be sought. Designations used by companies to distinguish their products are often claimed as trademarks. In all instances where John Wiley & Sons, Inc. is aware of a claim, the product names appear in initial capital or all capital letters. Readers, however, should contact the appropriate companies for more complete information regarding trademarks and registration. For general information on our other products and services please contact our Customer Care Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002. Wiley publishes in a variety of print and electronic formats and by print-on-demand. Some material included with standard print versions of this book may not be included in e-books or in print-on-demand. If this book refers to media such as a CD or DVD that is not included in the version you purchased, you may download this material at http://booksupport.wiley.com. For more information about Wiley products, visit www.wiley.com. Library of Congress Cataloging-in-Publication Data: Developmental psychopathology / editor, Dante Cicchetti. – Third edition. pages cm Includes index. ISBN 978-1-118-12087-3 (volume 1 : cloth : alk. paper) – ISBN 978-1-118-12091-0 (volume 2 : alk. paper) – ISBN 978-1-118-12092-7 (volume 3 : alk. paper) – ISBN 978-1-118-12093-4 (volume 4 : alk. paper) 1. Mental illness–Etiology. 2. Developmental psychology. 3. Mental illness–Risk factors. 4. Adjustment (Psychology) I. Cicchetti, Dante. RC454.4.D483 2016 616.89–dc23 2015018216 Printed in the United States of America 10 9 8 7 6 5 4 3 2 1

12:39pm

Page iv

Trim Size: 8.5in x 11in

Cicchetti ffirs.tex V2 - Volume I - 12/10/2015

These volumes are dedicated to Marianne Gerschel in recognition of her great vision and staunch support of the field of developmental psychopathology.

12:39pm

Page v

Trim Size: 8.5in x 11in

Cicchetti

ffirs.tex V2 - Volume I - 12/10/2015

12:39pm

Page vi

Trim Size: 8.5in x 11in

Cicchetti

ftoc.tex V2 - Volume I - 12/10/2015

12:39pm

Contents

Preface to Developmental Psychopathology, Third Edition

xi

Dante Cicchetti

Contributors 1

xiii

ASSESSMENT OF PSYCHOPATHOLOGY IN YOUNG CHILDREN 1 Margaret J. Briggs-Gowan, Leandra Godoy, Amy Heberle, and Alice S. Carter

2

DEVELOPMENTAL ISSUES IN ASSESSMENT, TAXONOMY, AND DIAGNOSIS OF PSYCHOPATHOLOGY: LIFE SPAN AND MULTICULTURAL PERSPECTIVES 46 Thomas M. Achenbach and Leslie A. Rescorla

3

DEVELOPMENTAL EPIDEMIOLOGY 94 E. Jane Costello and Adrian Angold

4

USING NATURAL EXPERIMENTS TO TEST ENVIRONMENTAL MEDIATION HYPOTHESES 129 Michael L. Rutter and Anita Thapar

5

DEVELOPMENTAL MODELS AND MECHANISMS FOR UNDERSTANDING THE EFFECTS OF EARLY EXPERIENCES ON PSYCHOLOGICAL DEVELOPMENT 156 Thomas G. O’Connor

6

EMOTIONAL SECURITY THEORY AND DEVELOPMENTAL PSYCHOPATHOLOGY

199

Patrick T. Davies, Meredith J. Martin, and Melissa L. Sturge-Apple

7

EMOTION AND THE DEVELOPMENT OF PSYCHOPATHOLOGY

265

Pamela M. Cole

8

ATTACHMENT AND DEVELOPMENTAL PSYCHOPATHOLOGY

325

R.M. Pasco Fearon, Ashley M. Groh, Marian J. Bakermans-Kranenburg, Marinus H. van IJzendoorn, and Glenn I. Roisman

vii

Page vii

Trim Size: 8.5in x 11in

viii

9

Cicchetti ftoc.tex V2 - Volume I - 12/10/2015

Contents

AUTONOMY AND AUTONOMY DISTURBANCES IN SELF-DEVELOPMENT AND PSYCHOPATHOLOGY: RESEARCH ON MOTIVATION, ATTACHMENT, AND CLINICAL PROCESS 385 Richard M. Ryan, Edward L. Deci, and Maarten Vansteenkiste

10

ROOTS OF TYPICAL CONSCIOUSNESS: IMPLICATIONS FOR DEVELOPMENTAL PSYCHOPATHOLOGY 439 Philippe Rochat

11

I-SELF AND ME-SELF PROCESSES AFFECTING DEVELOPMENTAL PSYCHOPATHOLOGY AND MENTAL HEALTH 470 Susan Harter

12

PEER RELATIONS AND DEVELOPMENTAL PSYCHOPATHOLOGY

527

Mitchell J. Prinstein and Matteo Giletta

13

FAMILY SYSTEMS FROM A DEVELOPMENTAL PSYCHOPATHOLOGY PERSPECTIVE

580

Patricia K. Kerig

14

ADOLESCENT/YOUNG ADULT ROMANTIC RELATIONSHIPS AND PSYCHOPATHOLOGY 631 Joanne Davila, Deborah M. Capaldi, and Annette M. La Greca

15

WHAT CAN DYNAMIC SYSTEMS MODELS OF DEVELOPMENT OFFER TO THE STUDY OF DEVELOPMENTAL PSYCHOPATHOLOGY? 665 Michael F. Mascolo, Paul Van Geert, Henderien Steenbeek, and Kurt W. Fischer

16

A SURVEY OF DYNAMIC SYSTEMS METHODS FOR DEVELOPMENTAL PSYCHOPATHOLOGY 717 Isabela Granic, Tom Hollenstein, and Anna Lichtwarck-Aschoff

17

MISSING DATA 760 Todd D. Little, Kyle M. Lang, Wei Wu, and Mijke Rhemtulla

18

PERSON-ORIENTED APPROACHES

797

G. Anne Bogat, Alexander von Eye, and Lars R. Bergman

19

PERSON-SPECIFIC APPROACHES TO THE MODELING OF INTRAINDIVIDUAL VARIATION IN DEVELOPMENTAL PSYCHOPATHOLOGY 846 Michael J. Rovine and Peter C. M. Molenaar

20

CONFIGURAL FREQUENCY ANALYSIS FOR RESEARCH ON DEVELOPMENTAL PROCESSES 866 Alexander von Eye and Eun-Young Mun

12:39pm

Page viii

Trim Size: 8.5in x 11in

Cicchetti ftoc.tex V2 - Volume I - 12/10/2015

Contents

21

MODERATION AND MEDIATION IN INTERINDIVIDUAL LONGITUDINAL ANALYSIS Jennifer L. Krull, JeeWon Cheong, Matthew S. Fritz, and David P. MacKinnon

22

LATENT GROWTH MODELING AND DEVELOPMENTAL PSYCHOPATHOLOGY Jungmeen Kim-Spoon and Kevin J. Grimm

23

INTEGRATIVE DATA ANALYSIS FOR RESEARCH IN DEVELOPMENTAL PSYCHOPATHOLOGY 1042 Eun-Young Mun, Yang Jiao, and Minge Xie

Author Index

1089

Subject Index

1125

986

ix

922

12:39pm

Page ix

Trim Size: 8.5in x 11in

Cicchetti

ftoc.tex V2 - Volume I - 12/10/2015

12:39pm

Page x

Trim Size: 8.5in x 11in

Cicchetti

fpref.tex V3 - Volume I - 12/10/2015

Preface to Developmental Psychopathology, Third Edition

encourage interdisciplinary and translational research (Cicchetti & Gunnar, 2009; Cicchetti & Toth, 2006). In keeping with its integrative focus, contributions to developmental psychopathology have come from many disciplines of the biological and social sciences. A wide array of content areas, scientific disciplines, and methodologies has been germane. Risk and protective factors and processes have been identified and validated at multiple levels of analysis and in multiple domains. The increased emphasis on a multilevel, dynamic systems approach to psychopathology and resilience, the increased attention paid to gene–environment interplay in the development of psychopathology and resilience, and the application of a multiple levels of analysis developmental perspective to mental illnesses that have traditionally been examined nondevelopmentally (e.g., bipolar disorder, schizophrenia, and the personality disorders) not only have contributed to a deeper understanding of the dysfunctions but also have educated the public about the causes and consequences of mental disorder (see Cicchetti & Cannon, 1999; Cicchetti & Crick, 2009a, 2009b; Miklowitz & Cicchetti, 2006, 2010; Tackett & Sharp, 2014). Advances in genomics, GxE interactions, and epigenetics; growth in our understanding of neurobiology, neural plasticity, and resilience; and progress in the development of methodological and technological tools, including brain imaging, neural circuitry, hormone assays, immunology, social and environmental influences on brain development, and statistical analysis of developmental change, pave the way for interdisciplinary and for multiple levels of analysis research programs that will significantly increase the knowledge base of the development and course of maladaptation, psychopathology, and resilience. Moreover, randomized control prevention and intervention trials are being conducted based on theoretical models and efforts to elucidate the mechanisms and processes contributing to developmental change at both the biological and psychological levels (Belsky & van IJzendoorn, 2015; Cicchetti & Gunnar, 2008).

A decade has passed since the second edition of Developmental Psychopathology was published. The two prior editions (Cicchetti & Cohen, 1995, 2006) have been very influential in the growth of the field of developmental psychopathology. The volumes have been highly cited in the literature and have served as an important resource for developmental scientists and prevention and intervention researchers alike. In the present third edition, we have expanded from the three volumes contained in the second edition to four volumes. The increased number of volumes in this current edition reflects the continued knowledge gains that have occurred in the field over the past decade. A not insignificant contributor to this growth can be found in the very principles of the discipline (Cicchetti, 1984, 1990, 1993; Cicchetti & Sroufe, 2000; Cicchetti & Toth, 1991, 2009; Rutter & Sroufe, 2000; Sroufe & Rutter, 1984). Theorists, researchers, and prevention scientists in the field of developmental psychopathology adhere to a life span framework to elucidate the numerous processes and mechanisms that can contribute to the development of mental disorders in high-risk individuals as well as those operative in individuals who already have manifested psychological disturbances or who have averted such disorders despite their high-risk status (Cicchetti, 1993; Masten, 2014; Rutter, 1986, 1987, 2012). Not only is knowledge of normal genetic, neurobiological, physiological, hormonal, psychological, and social processes very helpful for understanding, preventing, and treating psychopathology, but also deviations from and distortions of normal development that are seen in pathological processes indicate in innovative ways how normal development may be better investigated and understood. Similarly, information obtained from investigations of experiments of nature, high-risk conditions, and psychopathology can augment the comprehension of normal development (Cicchetti, 1984, 1990, 1993; Rutter, 1986; Rutter & Garmezy, 1983; Sroufe, 1990; Weiss, 1969). Another factor that has expedited growth within the field of developmental psychopathology has been its ability to incorporate knowledge from diverse disciplines and to xi

12:39pm

Page xi

Trim Size: 8.5in x 11in

xii

Cicchetti fpref.tex V3 - Volume I - 12/10/2015

Preface to Developmental Psychopathology, Third Edition

Despite the significant advances that have occurred in the field of developmental psychopathology, much important work lies ahead. Undoubtedly these future developments will build on the venerable contributions of the past; however, as work in the field becomes increasingly interdisciplinary, multilevel, and technologically sophisticated, it is essential that even more emphasis be directed toward the process of development (Harter, 2006; Sroufe, 2007, 2013). It is not only genes and environments but also the cumulative developmental history of the individual that influences how future development will unfold (Sroufe, 2007, 2013). Developmental psychopathologists have incorporated concepts and methods derived from other disciplinary endeavors that are too often isolated from each other, thereby generating advances in knowledge that might have been missed in the absence of cross-disciplinary dialogue. The continuation and elaboration of the mutually enriching interchanges that have occurred within and across disciplines interested in normal and abnormal development not only will enhance the science of developmental psychopathology but also will increase the benefits to be derived for individuals with high-risk conditions or mental disorders, families, and society as a whole. Dante Cicchetti, Ph.D. Minneapolis, MN January 2015 REFERENCES Belsky, J., & van IJzendoorn, M. (2015). What works for whom? Genetic moderation of intervention efficacy. [Special Section]. Development and Psychopathology, 27, 1–6. Cicchetti, D. (1984). The emergence of developmental psychopathology. Child Development, 55(1), 1–7. Cicchetti, D. (1990). A historical perspective on the discipline of developmental psychopathology. In J. Rolf, A. Masten, D. Cicchetti, K. Nuechterlein, & S. Weintraub (Eds.), Risk and protective factors in the development of psychopathology (pp. 2–28). New York, NY: Cambridge University Press. Cicchetti, D. (1993). Developmental psychopathology: Reactions, reflections, projections. Developmental Review, 13, 471–502.

Cicchetti, D., & Crick, N. R. (Eds.). (2009b). Precursors of and diverse pathways to personality disorder in children and adolescents. [Special Issue, Part 2]. Development and Psychopathology, 21(4), 1031–1381. Cicchetti, D., & Gunnar, M. R. (2008). Integrating biological processes into the design and evaluation of preventive interventions. Development and Psychopathology, 20, 737–743. Cicchetti, D., & Gunnar, M. R. (Eds.). (2009). Meeting the challenge of translational research in child psychology: Minnesota symposia on child psychology (Vol. 35). New York, NY: Wiley. Cicchetti, D., & Sroufe, L. A. (2000). The past as prologue to the future: The times they’ve been a changin’. Development and Psychopathology, 12, 255–264. Cicchetti, D., & Toth, S. L. (1991). The making of a developmental psychopathologist. In J. Cantor, C. Spiker, & L. Lipsitt (Eds.), Child behavior and development: Training for diversity (pp. 34–72). Norwood, NJ: Ablex. Cicchetti, D., & Toth, S. L. (Eds.). (2006). Translational research in developmental psychopathology. [Special Issue]. Development and Psychopathology, 18(3), 619–933. Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50, 16–25. Harter, S. (2006). Self-processes and developmental psychopathology. In D. Cicchetti & D. Cohen (Eds.), Developmental psychopathology (2nd ed., 370–418). New York, NY: Wiley. Masten, A. S. (2014). Ordinary magic: Resilience in development. New York, NY: Guilford Publications, Inc. Miklowitz, D. J., & Cicchetti, D. (2006). Toward a life span developmental psychopathology perspective on bipolar disorder. Development and Psychopathology, 18, 935–938. Miklowitz, D. J., & Cicchetti, D. (Eds.). (2010). Bipolar disorder: A developmental psychopathology approach. New York, NY: Guilford. Rutter, M. (1986). Child psychiatry: The interface between clinical and developmental research. Psychological Medicine, 16, 151–169. Rutter, M. (1987). Psychosocial resilience and protective mechanisms. American Journal of Orthopsychiatry, 57, 316–331. Rutter, M. (2012). Resilience as a dynamic concept. Development and Psychopathology, 24, 335–344. Rutter, M., & Garmezy, N. (1983). Developmental psychopathology. In E. M. Hetherington (Ed.), Handbook of child psychology (pp. 774–911). New York, NY: Wiley. Rutter, M., & Sroufe, L. A. (2000). Developmental psychopathology: Concepts and challenges. Development and Psychopathology, 12, 265–296.

Cicchetti, D., & Cannon, T. (1999). Neurodevelopmental processes in the ontogenesis and epigenesis of psychopathology. Development and Psychopathology, 11, 375–393. Cicchetti, D., & Cohen, D. (Eds.). (1995). Developmental psychopathology (Vols. 1–2). New York, NY: Wiley. Cicchetti, D., & Cohen, D. (Eds.). (2006). Developmental psychopathology (2nd ed., Vols. 1–3). New York, NY: Wiley.

Sroufe, L. A. (1990). Considering normal and abnormal together: The essence of developmental psychopathology. Development and Psychopathology, 2, 335–347. Sroufe, L. A. (2007). The place of development in developmental psychopathology. In A. Masten (Ed.), Multilevel dynamics in developmental psychopathology pathways to the future: The Minnesota symposia on child psychology (pp. 285–299). Mahwah, NJ: Erlbaum. Sroufe, L. A. (2013). The promise of developmental psychopathology. Development and Psychopathology, 25, 1215–1224. Sroufe, L. A., & Rutter, M. (1984). The domain of developmental psychopathology. Child Development, 55, 17–29.

Cicchetti, D., & Crick, N. R. (Eds.) (2009a). Precursors of and diverse pathways to personality disorder in children and adolescents. [Special Issue, Part 1]. Development and Psychopathology, 21(3), 683–1030.

Tackett, J. L., & Sharp, C. (2014). A developmental psychopathology perspective on personality disorder. [Special Issue]. Journal of Personality Disorders, 28, 1–179. Weiss, P. (1969). Principles of development. New York, NY: Hafner.

12:39pm

Page xii

Trim Size: 8.5in x 11in

Cicchetti flast.tex V3 - Volume I - 12/10/2015

Contributors

Thomas M. Achenbach, PhD University of Vermont Burlington, VT

E. Jane Costello, PhD Duke University Medical Center Durham, North Carolina

Adrian Angold, MRCPsych Duke University Medical Center Durham, North Carolina

Patrick T. Davies, PhD University of Rochester Rochester, New York

Marian J. Bakermans-Kranenburg, PhD Leiden University Leiden, Netherlands

Joanne Davila, PhD Stony Brook University Stony Brook, New York

Lars R. Bergman, PhD Stockholm University Stockholm, Sweden

Edward L. Deci, PhD University of Rochester Rochester, New York

G. Anne Bogat, PhD Michigan State University East Lansing, Michigan

R. M. Pasco Fearon, PhD, DClinPsy University College London London, United Kingdom

Margaret J. Briggs-Gowan, PhD University of Connecticut Farmington, Connecticut

Kurt W. Fischer, PhD Harvard University Cambridge, Massachusetts

Deborah M. Capaldi, PhD Oregon Social Learning Center Eugene, Oregon

Matthew S. Fritz, PhD University of Nebraska Lincoln, Nebraska

Alice S. Carter, PhD University of Massachusetts Boston, Massachusetts

Matteo Giletta, PhD Tilburg University Tilburg, Netherlands

JeeWon Cheong, PhD University of Florida Gainesville, Florida

Leandra Godoy, PhD Children’s National Medical Center Washington, District of Columbia

Dante Cicchetti, PhD Institute of Child Development University of Minnesota Minneapolis, Minnesota

Isabela Granic, PhD Radboud University Nijmegen Nijmegen, Netherlands

Pamela M. Cole, PhD Pennsylvania State University University Park, Pennsylvania

Kevin J. Grimm, PhD Arizona State University Tempe, Arizona xiii

12:39pm

Page xiii

Trim Size: 8.5in x 11in

xiv

Cicchetti

flast.tex V3 - Volume I - 12/10/2015

Contributors

Ashley M. Groh, PhD University of Missouri Columbia, Missouri

Michael F. Mascolo, PhD Merrimack College North Andover, Massachusetts

Susan Harter, PhD University of Denver Denver, Colorado

Peter C. M. Molenaar, PhD Pennsylvania State University University Park, Pennsylvania

Amy Heberle, MS University of Massachusetts Boston, Massachusetts

Eun-Young Mun, PhD Rutgers, the State University of New Jersey Piscataway, New Jersey

Tom Hollenstein, PhD Queen’s University Kingston, Canada

Thomas G. O’Connor, PhD University of Rochester Medical Center Rochester, New York

Yang Jiao, MS Rutgers, the State University of New Jersey Piscataway, New Jersey

Mitchell J. Prinstein, PhD, ABPP University of North Carolina Chapel Hill, North Carolina

Patricia K. Kerig, PhD University of Utah Salt Lake City, Utah

Leslie A. Rescorla, PhD Bryn Mawr College Bryn Mawr, Pennsylvania

Jungmeen Kim-Spoon, PhD Virginia Tech Blacksburg, Virginia Jennifer L. Krull, PhD University of California Los Angeles, California Annette M. La Greca, PhD, ABPP University of Miami Coral Gables, Florida Kyle M. Lang, PhD Texas Tech University Lubbock, Texas

Mijke Rhemtulla, PhD University of Amsterdam Amsterdam, Netherlands Philippe Rochat, PhD Emory University Atlanta, Georgia Glenn I. Roisman, PhD University of Minnesota Minneapolis, Minnesota Michael J. Rovine, PhD Pennsylvania State University University Park, Pennsyvania

Anna Lichtwarck-Aschoff, PhD Radboud University Nijmegen Nijmegen, Netherlands

Michael L. Rutter, MD, FRS, FRC Psych., FBA, FAC Med Sci King’s College London London, United Kingdom

Todd D. Little, PhD Texas Tech University Lubbock, Texas

Richard M. Ryan, PhD Australian Catholic University Strathfield, Australia

David P. MacKinnon, PhD Arizona State University Tempe, Arizona

Henderien Steenbeek, PhD University of Groningen Groningen, Netherlands

Meredith J. Martin, PhD University of Rochester Rochester, New York

Melissa L. Sturge-Apple, PhD University of Rochester Rochester, New York

12:39pm

Page xiv

Trim Size: 8.5in x 11in

Cicchetti flast.tex V3 - Volume I - 12/10/2015

Contributors

Anita Thapar, MD, PhD Cardiff University Cardiff, United Kingdom

Alexander von Eye, PhD Michigan State University East Lansing, Michigan

Paul van Geert, PhD University of Groningen Groningen, Netherlands

Wei Wu, PhD University of Kansas Lawrence, Kansas

Marinus H. van IJzendoorn, PhD Leiden University Leiden, Netherlands

Minge Xie, PhD Rutgers, the State University of New Jersey Piscataway, New Jersey

Maarten Vansteenkiste, PhD University of Ghent Ghent, Belgium

12:39pm

xv

Page xv

Trim Size: 8.5in x 11in

Cicchetti

flast.tex V3 - Volume I - 12/10/2015

12:39pm

Page xvi

CHAPTER 1

Assessment of Psychopathology in Young Children MARGARET J. BRIGGS-GOWAN, LEANDRA GODOY, AMY HEBERLE, and ALICE S. CARTER

INTRODUCTION 1 EARLY PROBLEMS MATTER 2 Progress in Psychiatric Diagnosis in Young Children 3 IMPORTANT CONSIDERATIONS IN YOUNG CHILD ASSESSMENT 4 Reliance on Caregivers for Information 4 Sensitivity to Contextual Influences, Including Caregiving Contexts 6 DOMAINS OF DEVELOPMENT 9 SELECTING AN ASSESSMENT APPROACH AND TOOL 11 Types of Tools 11 Understanding Psychometric Properties 12 Reliability 12 Validity 13 Validity of Classification 14 Normatization 15 Cultural Validity and Cultural Norms 16 Knowing What Problems Are Really Being Assessed 16

Response Formats 17 Summary 17 ASSESSMENT TOOLS 17 Screening Methods 18 Screening Methods Characteristics of Screening Tools 19 Selected Screening Tools 20 Comprehensive Dimensional Tools for Assessing Social-Emotional/Behavioral Problems 23 Selected Dimensional Checklists 24 Variation in Emphasis of the Domains That Are Assessed 29 Diagnostic Approaches 29 Selected Diagnostic Interviews 30 Psychometric Properties of Diagnostic Interviews 32 Observational Assessment 35 Assessing Impairment 35 CONCLUSIONS AND DIRECTIONS FOR FUTURE RESEARCH 37 REFERENCES 38

INTRODUCTION

2006). Indeed, recent research has indicated that psychiatric disorders are just as prevalent in early childhood as they are in school-age children (Egger & Angold, 2006). Moreover, when young children manifest psychopathology that is impairing, it is often persistent and predicts later difficulties once they become of school age. Equally important, there is increasing awareness that these problems can interfere with learning within early childhood and may set in motion a developmental cascade that likely predicts challenges to lifespan functioning in multiple domains. Research focused on specific disorders has driven discovery of neurobiologic substrates, which has further validated the relevance and reality of early life psychopathology (Luby, Belden, Pautsch, Si, & Spitznagel, 2009; Luby, Si, Belden, Tandon, & Spitznagel, 2009; Stalets & Luby, 2006). These advances in our understanding of young child psychopathology are paralleled by, and one might argue largely driven by, an explosion in reliable, valid, developmentally sensitive measures for assessing a full range of self-regulation, social-emotional development in young

The past 20 years have witnessed a sea change for young children’s mental health. It is now recognized that early childhood (0–5 years) is a crucial period for the development of self-regulation, a critical set of competencies that have implications for adaptive functioning in school and through the life span. Early childhood is also recognized as a time when psychopathology may begin to emerge and disrupt young children’s developmental progress. In addition, enormous progress has been made in demonstrating that, when psychiatric disorders are defined in a manner that is developmentally meaningful, even very young children suffer from psychiatric disorders that are valid, impairing, and clinically very similar to those experienced by older children (Egger & Angold, 2006; Egger et al.,

1

Color versions of Figure 1.1 are available at http://onlinelibrary .wiley.com/book/10.1002/9781118963418 1

2

Assessment of Psychopathology in Young Children

children. Specifically, over the past 15-plus years, a number of instruments have been developed to assess parent and other caregiver appraisals of social-emotional functioning utilizing both questionnaire and interview methods. There have also been advances in observational tools to assess clinically significant emotional and behavior problems. With greater acceptance and building on advances in measurement, we are poised to evaluate the benefits of a broad range of prevention and intervention efforts and see increasing discovery of biological and environmental influences on young children’s mental health. As the field presses forward to address the mental health needs of young children both efficiently and effectively, success will be optimized by a well-informed approach to assessment that (1) acknowledges contextual factors, including the caregiving environments at home and in other settings, such as child care and early education environments, caregiver influences on social-emotional functioning and assessment, recent changes in family structure or contextual stressors, and sociocultural factors; (2) is framed within the context of a child’s functioning in other developmental domains, such as language, cognition, adaptive functioning, health, and sensory; (3) is tailored to the goals and purposes of the assessment and evaluation setting (e.g., pediatric clinic, day care center, mental health clinic, or private practice); (4) utilizes reliable, valid, developmentally sensitive tools; and (5) employs an approach to interpretation that views the whole child in relation to contextual and developmental factors and evaluates his or her capacities and participation in developmentally appropriate activities and settings (i.e., impairment). A primary goal of this chapter is to help clinicians and researchers determine the most suitable measures to use from a wide array of parent and other caregiver report, observational, and direct assessment measures that are now available. Rather than trying to offer an exhaustive list of all existing measures of social-emotional functioning and psychopathology appropriate for young children, we highlight some of the most widely employed and promising tools and approaches, including those that reflect advances in screening, comprehensive dimensional parent- and other caregiver-report instruments, and diagnostic approaches to young child evaluation. These assessment tools can be categorized as follows: (1) parent and other caregiver report instruments that focus on general problem behaviors; (2) parent and other caregiver report instruments that focus on specific problem areas or disorders (e.g., anxiety, disruptive behavior); (3) parent and other caregiver report instruments designed to assess both problem behaviors and competencies; (4) comprehensive diagnostic interviews for parents of young children; and (5) observational tools

and methods. Within the first three categories, measures can be further divided according to whether they are brief tools appropriate for screening or longer checklist tools or diagnostic interviews that provide more detailed information. Finally, we will close the chapter with a discussion of ongoing challenges, future directions, and opportunities in research on and clinical applications with assessment of young child psychopathology. Although many researchers and clinicians continue to express discomfort about pathologizing, or labeling, young children, our focus is on assessment tools that enhance the recognition and detection of early emerging psychopathology to address mental health needs in an effort to minimize adverse developmental cascades. Moreover, we argue optimistically that by labeling systematic behavioral patterns observed within young children (rather than labeling individual children) we create the potential to develop and disseminate guidance regarding appropriate contextual supports and specific behavioral interventions that are tailored to the needs of children with different behavioral profiles; these early prevention and targeted interventions can be designed to support family beliefs, values, and goals and children’s developmental progress while minimizing child and family distress. EARLY PROBLEMS MATTER There is now a consensus among child clinicians that children as young as 2 years of age can suffer from significant social-emotional and behavior problems, or psychopathology. Prevalence estimates of clinically significant problems in nonreferred samples have ranged considerably, from as low as 7% to as high as 26%, depending on whether problems are defined in terms of meeting criteria for psychiatric diagnosis or by exceeding a clinical cutoff on a checklist measure (Briggs-Gowan, Carter, Skuban, & Horwitz, 2001; Egger & Angold, 2006; Gleason et al., 2011; Karabekiroglu et al., 2013; Keenan et al., 1997; Lavigne, Lebailly, Hopkins, Gouze, & Binns, 2009; Wichstrom et al., 2012). Rates also tend to be higher among young children exposed to poverty and other psychosocial risk factors (McCue Horwitz et al., 2012; Qi & Kaiser, 2003; Weitzman, Edmonds, Davagnino, & Briggs-Gowan, 2014). Early social-emotional and behavioral problems are linked with impairment in child and family functioning as well as increased parenting stress and worry (Briggs-Gowan & Carter, 2008a; Briggs-Gowan, Carter, Bosson-Heenan, Guyer, & Horwitz, 2006; Briggs-Gowan et al., 2001; Egger & Angold, 2006; Fuchs, Klein, Otto, & von Klitzing, 2013; Keenan et al., 2007; Lavigne et al., 1996; Luby, Belden, Pautsch, et al., 2009). Moreover, social-emotional/behavior

Early Problems Matter

problems in young children have been associated with concomitant delays in child social-emotional competence (Briggs-Gowan & Carter, 2008a; Briggs-Gowan et al., 2001) and shown to predict poorer social competence in elementary school (Briggs-Gowan, Carter, & Ford, 2011). Intervening in preschool to address both social emotional problems and competencies is associated with greater improvements in both areas, including on experimental tasks of emotion knowledge and social problem-solving strategies (Stefan ¸ & Miclea, 2013). Furthermore, contrasting with the historical belief that young children’s difficult behavior is just a phase (Keenan & Wakschlag, 2000), there is consistent evidence that for some children these early emergent social-emotional and behavioral problems are persistent and predict poorer functioning at later ages (Briggs-Gowan & Carter, 2008b; Briggs-Gowan et al., 2006; Kim-Cohen et al., 2005; Lavigne et al., 1998; Mathiesen & Sanson, 2000; O’Neill, Schneiderman, Rajendran, Marks, & Halperin, 2014; Shaw, Lacourse, & Nagin, 2005; Speltz, McClellan, DeKlyen, & Jones, 1999; Spence, Najman, Bor, O’Callaghan, & Williams, 2002; Stalets & Luby, 2006; Wakschlag, Briggs-Gowan, et al., 2008). Notably, persistence has been documented for a wide range of problems, including anxiety, depression, attention-deficit hyperactivity, and disruptive behaviors. Thus, consistent and convincing evidence indicates that young children can and do suffer from a wide spectrum of social-emotional and behavioral problems that are often persistent—the presence of impairment further underscores the importance of early identification and prevention in this developmental period. Progress in Psychiatric Diagnosis in Young Children Empirical and conceptual work in the areas of posttraumatic stress disorder (PTSD), depression, and disruptive behavior disorders illustrates the role that developmental factors can play in how psychopathology manifests and how several groups have endeavored to establish the relevance of these psychopathologies in young children. PTSD Considerable work by Michael Scheeringa and colleagues documents that children as young as 9 months of age can and do suffer from PTSD (Scheeringa, 2007, 2008; Scheeringa, Myers, Putnam, & Zeanah, 2012; Scheeringa, Peebles, Cook, & Zeanah, 2001). This work highlights the importance of considering the developmental capacities of young children when determining the appropriateness of criteria employed for older children and illustrates the utility of adopting a multi-informant, multimethod approach to assessment (Hunsley & Mash, 2007). Scheeringa et al.’s work in this area has driven important recognition of

3

developmental factors that affect how PTSD presents in young children. For example, many avoidance and numbing symptoms are either developmentally implausible (e.g., sense of a foreshortened future) or internal in quality (e.g., avoidance of internal thoughts, feelings, or reminders of the event), making them very difficult to identify in young children who have limited verbal skills (Scheeringa, 2008). Scheeringa also noted that these types of symptoms may manifest differently in young children. For example, “markedly diminished interest in significant activities” is often observed as constriction of play, and “feeling of detachment or estrangement from others” is often observed as social withdrawal. Scheeringa and colleagues further documented the central role that parental reactions play in the emergence, promotion, and maintenance of symptoms in young children (Scheeringa & Zeanah, 2001), highlighting the importance of assessing young children’s possible PTSD symptoms in the context of parent– or caregiver–child relationships. Depression Luby and colleagues’ work has established the presence of early manifestations of clinically significant signs and symptoms of depressive disorders in young children (Luby, Belden, Pausch, et al., 2009; Luby & Navsaria, 2010; Luby, Si, et al., 2009; Stalets & Luby, 2006). Their research addresses the complicated developmental question of whether young children’s emotional repertoire is itself sufficiently differentiated to encompass true depressive or elated affect and how to differentiate atypical from normative developmental manifestations. They have further noted that greater variability in young children’s mood states calls into question the relevance of duration criteria employed by diagnostic systems developed for older children (Gaffrey, Belden, & Luby, 2011). Luby and colleagues utilized a multimethod, multi-informant approach that included (1) parent reports on both dimensional ratings scales and in an age-appropriate diagnostic interview; (2) comprehensive observation of the young child’s affective range via a laboratory based temperament assessment, thematic play, and parent–child interaction across structured and unstructured conditions; (3) a developmentally sensitive direct interview for preschoolers; (4) cognitive assessment of the child; and (5) neurocognitive assessment of the child. By employing these techniques, they were able to identify a group of young preschool-age children who met modified depression criteria developed for young children. Further, Luby and colleagues’ empirical data have provided crucial insights into how depression manifests in young children (Luby, Belden, Sullivan, et al., 2009;

4

Assessment of Psychopathology in Young Children

Luby, Belden, Pausch, et al., 2009; Luby, Si, et al., 2009). For example, their work has shown that withdrawal and vegetative symptoms are less consistently evident in young children with Major Depressive Disorder relative to older children. They have also shown that sadness, irritability, and thoughts of death, while evident in young children with depression, also are often present in children with anxiety disorders and attention-deficit/hyperactivity disorder (ADHD). In contrast, anhedonia, feelings of guilt, and psychomotor agitation appear to be fairly characteristic of young children with depression. Changes in activity, appetite, and sleep also have been noted. Anhedonia appears to be a very specific marker of depression in young children: it is commonly present in young children with depression and rarely seen in other psychiatric groups or typically developing children (Luby et al., 2002). A subtype of preschool depression characterized by anhedonia also appears to be particularly severe (Luby, Belden, Pausch, et al., 2009). The work of Luby and her colleagues is notable for inclusion of biological correlates, such as cortisol reactivity and addressing family genetic risk to document the validity of the revised young child criteria (Luby, Belden, Pausch, et al., 2009; Luby, Si, et al., 2009; Stalets & Luby, 2006).

impairment in preschool children (Egger & Angold, 2006; Egger et al., 2006; Keenan & Wakschlag, 2000, 2002; Keenan et al., 2007; Wakschlag et al., 2007). Through a longitudinal study, this group further demonstrated persistence over time in early disruptive behavior disorders. In one study, 55% of those with disorders at baseline continued to meet criteria for disorder one year later (Wakschlag, Briggs-Gowan, et al., 2008). These rates are strikingly similar to levels of persistence observed in school-age children (Briggs-Gowan et al., 2003). Summary As our understanding of specific early emerging disorders evolves, it is quite likely that additional diagnosis-specific instruments will be developed, allowing clinicians and researchers to elicit information about behaviors that assist in making differential diagnoses (i.e., are specific to diagnostic conditions). Clearly, multimethod, multi-informant assessment is integral to understanding young children’s development. Moreover, observations of the parent– child or caregiver–child interactions across several contexts must be considered a critical component of any young child assessment of social-emotional problems or psychopathology.

Disruptive Behavior Disorders The study of disruptive behavior disorders highlights the complexity of distinguishing normative from clinically problematic behavior during early childhood when some “problem” behaviors are normative (Wakschlag et al., 2007; Wakschlag, Tolan, & Leventhal, 2010). Historically, concern was raised about whether disruptive behavior problems could be reliably assessed during this period (Campbell, 1990) and whether it may be premature to diagnose children whose challenging behaviors might be transient and diminish over time. At the same time, disruptive behavior is the most common reason for referral of young children to mental health clinics (Keenan & Wakschlag, 2000; Thomas & Guskin, 2001). Wakschlag and her colleagues have pressed for the refinement of diagnostic approaches in a manner that describes early manifestations of disruptive behavior in a developmentally appropriate and meaningful fashion and distinguishes between clinically significant disruptive behaviors and expectable variation in children’s capacity to regulate emotions and behavior as they consolidate their self-regulatory skills (Gray & Wakschlag, in press; Wakschlag et al., 2007). This work and the work of others has demonstrated that disruptive behavior disorders can be reliably and validly diagnosed and are associated with marked

IMPORTANT CONSIDERATIONS IN YOUNG CHILD ASSESSMENT Though challenges exist when conducting any mental health assessment, several are unique to or require greater consideration when assessing young children, including the following: (1) young children’s limited communication skills, which results in greater reliance on caregivers for information about the child’s functioning; (2) young children’s sensitivity to contextual influences on their functioning and the resulting importance of assessing context when working with this population; and (3) the influence of sociocultural factors, which affect the meaning of the child’s behaviors to caregivers and other family members, the way that caregivers report on the child’s behavior, caregivers’ level of worry about the child overall and in relation to specific behaviors, caregivers’ decisions to seek or not seek services, and the assessor’s interpretation of findings. Understanding these challenges can enhance the evaluation process. Reliance on Caregivers for Information The fact that infants and young children have no or limited verbal abilities and metacognitive capacities makes it

Important Considerations in Young Child Assessment

difficult or impossible to directly collect information from them about their thoughts and feelings about presenting complaints, historical experiences and behaviors, and contextual events. Thus, in contrast to older children, youth, and adults, caregivers play a central role in providing information about the young child’s behaviors across multiple settings and contexts, including crucial insight into when behaviors may have emerged and changed over time and contextual factors that may have exacerbated or attenuated symptoms. Collecting information from caregivers brings unique considerations to the assessment process when evaluating questionnaire and interview assessment tools. First, it is important to establish the best informant in the child’s life to best speak for him or her about a given topic. This is usually determined by asking who has a parental or caregiving role for the child in relation to the area or domain of inquiry, that is, the person who assumes responsibility for meeting his or her physical and emotional needs. In this chapter, the term parent refers to biological, adoptive, or foster parents or guardians or extended family members who care for the child on a regular basis. The term caregiver encompasses parents as well as individuals who routinely spend enough time with the child to be knowledgeable about multiple aspects of the child’s social-emotional functioning and day-to-day behavior. Thus, caregiver may refer to extended family members and child-care or day-care providers who have cared for the child for at least one month and who care for the child on a regular basis. (Note, however, that some measures may list different criteria for classifying someone as a suitable informant on the child’s functioning.) Typically, the person who brings a child to his or her evaluation appointment and who consents to the evaluation is a parental figure and should be included in the evaluation process; however, it is also important to assess others who may function as caregivers—particularly when a child experiences care in multiple contexts. These additional individuals should also be included as informants. Moreover, it is important to invite relevant caregivers to participate in any child evaluation, as young children’s behavior is more variable across contexts (Clark, Tluczek, & Gallagher, 2004; Tronick, 1989). An evaluation of the child’s larger family context facilitates the process of identifying caregivers. It also provides important cues regarding factors such as recent stressors or strained dyadic and triadic family relationships that are currently influencing the child’s functioning or that may have had a role in shaping the child’s functioning over time (Carlson, 1990; Hayden et al., 1998). Information

5

about the larger family context also enhances the assessor’s understanding of a given caregiver’s perspective on the child (e.g., how long the caregiver has known the child, who else is present in typical interactions between the caregiver and the child, who supports or influences the caregiver’s beliefs about the child), potentially illuminating sources of bias and avenues for intervention. For example, if in the process of assessing the family context it comes to light that a young mother has received criticism about her child’s behavior and her own parenting from her own mother, who cares for the child on a regular basis, this information would point to the need for further assessment of the mother–grandmother relationship. This evaluation could focus on the potential utility of intervention to improve communication between and address inconsistencies in the disciplinary styles, beliefs and expectations of the two caregivers. It is important to note that in using the word family we refer to any person, whether biologically related to the child or not, who is considered by the child or other adults close to the child to be a family member. The assessor should consider the total makeup of the family group (e.g., how many members), who is considered a family member, how family members define relationships with one another, where family members live in relation to one another, and what roles family members hold from the perspective of the target child (Carlson, 1990). In addition to evaluating the current family context, a history of the family makeup should be obtained, including ruptures in family relationships, new additions to the family, and changes in the way family members are distributed across households. In the process of assessing the family and caregiving context and history, assessors must be alert to their own biases about what constitutes a family and how families function. Accurate assessment of family functioning will be facilitated by knowledge of cultural norms regarding family relationships in the groups to which a child belongs. In contrast to the popular notion of the two-parent household and caregiving system, many children are cared for in part or whole by relatives such as aunts, uncles, and grandparents (Dressler, 1985; Wilson & Tolson, 1990). In fact, as of 2012, 10% of all American children reside with a grandparent (with or without a parent also in the home); this number is even greater for Black (14%), Latino (12%), and Asian (14%) children (Ellis & Simmons, 2014). Twenty-six percent of children under the age of 5 whose mothers are employed have a grandparent or other relative as their primary child-care provider (Laughlin, 2013). Many more children may have close but informal caregiving relationships with relatives other than their parents.

6

Assessment of Psychopathology in Young Children

Thus, many children have nonparental caregivers who should be included in the evaluation process to elicit an accurate assessment of the child’s functioning. The utility of including multiple caregivers’ reports in the evaluation process is further discussed in the next section. Sensitivity to Contextual Influences, Including Caregiving Contexts Young children’s behavior is highly sensitive to contextual and relational influences (Clark et al., 2004; Dirks, De Los Reyes, Briggs-Gowan, Cella, & Wakschlag, 2012; Gray et al., 2012; Tronick, 1989), and it is therefore critical to include an assessment of relevant contextual factors in any evaluation involving a young child. When working with young children, social-emotional and cognitive functioning and impairment should always be assessed within the context of caregiving relationships and settings, with assessors incorporating what they are able to observe and measure about the various aspects of caregiver functioning and setting characteristics into their interpretation of findings from assessment instruments designed to measure an individual’s functioning. In many cases, the family history will reveal recent or chronic stressors to which the child may be responding. Where major events have occurred in the child’s history—whether these are identified as stressors by the reporter or not—it may be useful to explicitly inquire about the child’s functioning before and after the event as well as changes since the event. As an example, moving houses is a stressful—and common—transition for young children (Stoneman, Brody, Churchill, & Winn, 1999). In some cases, the family history may reveal a recent change in the child’s caregiving relationships, as in the case of a child who has recently transitioned to the care of a new foster parent, or in the case of a child whose parents have recently moved the family into their own apartment after having lived with the child’s grandparents for the majority of her life. In cases such as these, repeat assessment after one or two months may be warranted to assess the child’s functioning after a period of adjustment to the new caregiving environment. Understanding the makeup of a particular child’s family is an essential assessment task because young children’s development, behavior, and functioning are embedded within their caregiving relationships. While it is also true that older children are influenced by their caregiving contexts, young children may be especially sensitive to contextual influences due to their dependence on others for their basic needs, often limited exposure to outside institutions such as schools, limited relationships with peers,

and their unlikelihood of having developed independent emotion-regulation or coping strategies that they can utilize in the face of environmental stressors (Compas, 1987). Caregivers provide structure to the physical, emotional, and behavioral aspects of the child’s environment; for example, infants learn to regulate their own affect through exchanges with their caregivers and practice with sensitive, engaged caregivers to repair dysynchronous interactions, to reengage after a rupture, to recover behaviorally and physiologically after exposure to a stressor, and to convey information through affective expression (Bridgett, Burt, Laake, & Oddi, 2013; Haley & Stansbury, 2003; Kogan & Carter, 1996; Martinez-Torteya et al., 2014; Tronick, 1989). As another example, in the behavioral domain, children whose caregivers provide encouragement for independent efforts to achieve goals or learn new tasks may develop a sense of competence and intrinsic motivation that facilitates effortful engagement in challenging tasks in settings outside the home (Ryan & Deci, 2000). Relatedly, consistent with Vygotsky’s (1978) theory of proximal development, children whose parents engage in scaffolding behavior (including teaching, praise, encouragement, and positive affective engagement) to enable them to perform within their zone of proximal development— the space between what they have already mastered developmentally and what they are capable of doing with parental support—show superior academic and social selfregulation abilities compared with children whose parents’ style of instruction is less contingent on their current level of achievement (Clarke, Kelleher, Clancy, & Cannon, 2012; Neitzel & Stright, 2003). Caregivers’ structuring of the physical environment also may be influential; as an example, low-income children on average have less access to cognitively stimulating books and games (Evans, 2004), which may both impact their actual academic attainment and deflate their scores on assessments due to a lack of familiarity with the types of tasks commonly used in intellectual and academic assessment of young children. A striking percentage of low-income children are in child-care settings that provide inadequate (24%) or minimally adequate (36%) support for their development; in turn, these children show higher levels of internalizing and externalizing behavior problems and demonstrate lower levels of positive behavior compared with their peers in higher quality child-care settings (Votruba-Drzal, Coley, & Chase-Lansdale, 2004). Within all of the areas in which caregivers provide structure for their children, the impact of a given factor is dependent on the match or mismatch between the two; children’s temperamental and developmental characteristics may strongly

Important Considerations in Young Child Assessment

influence how they respond to a situation. Similarly, caregivers’ individual characteristics, including expectations regarding child development and availability of personal psychological resources with which to meet the children’s needs for whom they care, will influence whether a set of behaviors is viewed as problematic (Seifer, 2000). Characteristics of the setting (home, day care, school, and any other place the child spends substantial amounts of time), including both the overall quality of the setting and the match between the child and the setting, may also impact the child’s functioning. As noted already, young children’s behavior is highly sensitive to contextual and relational influences (Clark et al., 2004; Tronick, 1989), and it may therefore be more variable across settings than older children’s behavior. It is common for parents and teachers, for example, to provide very different perspectives, with often very low correlations between their reports (Achenbach, McConaughy, & Howell, 1987). Though this may be due to differences in caregiver behavior or attitudes, it may also, in many cases, be due to differences in the setting, including the physical structure of the setting, the way time and activity is structured in the setting, adult-to-child ratio, number of children, behavior of other children in the setting, level of sensory stimulation, and adequacy of sleep support. Inconsistency in rules between home and school or day care (e.g., requirements for where eating and sleeping take place) may also be a source of differences between caregivers’ reports. Given children’s sensitivity to contextual influences such as these, the evaluation process would ideally include information gathering not only from multiple caregivers but also across multiple settings and within multiple types of interactions (e.g., unstructured play, challenging goal-directed tasks, shared reading). However, as this goal is often impractical, a compromise is to ensure that data are collected from multiple informants who are familiar with the child’s behavior in different settings and in different types of interactions. When gathering information as part of an evaluation, reports from multiple caregivers are likely to yield a more complete picture of the child’s functioning than reports from the child’s parents alone (Dirks et al., 2012). It is especially important to gather information from multiple caregivers in cases where a child regularly spends time in different contexts, with different primary caregivers in each (e.g., a child who is in full-time day care with a neighbor, is cared for by his mother at home during the week, and spends weekends in the care of his father). As discussed, where the reports of different caregivers diverge—for example, where a child’s father reports problematic levels of disruptive behavior and her

7

teacher reports no issues with disruptive behavior—the discrepancy should typically be viewed not as evidence of error on the part of one reporter but rather as evidence that contextual, cultural, interpersonal, or other influences may be contributing to the perceived problem. The review of assessment data should then include an analysis of the contexts in which a child is reported to have difficulties, with an eye toward identifying clues for intervention strategies to recommend for the child (e.g., creating consistency between home and day care with respect to the child’s nap schedule). In the case in which the child’s difficulties occur across multiple caregiving relationships or contexts, they may be more severe than the case in which the difficulties are limited to one relationship or one caregiving context. Lastly, and briefly, it is also important to consider aspects of the child’s context outside of his or her caregiving environment, such as neighborhood and community relationships and functioning. As an example, acts of violence occur considerably more frequently in high-poverty areas compared with more advantaged areas (Briggs-Gowan, Ford, Fraleigh, McCarthy, & Carter, 2010; Friday, 1995; Sampson, Raudenbush, & Earls, 1997), and some studies have reported extremely high rates of community violence exposure among children in high-risk urban areas, with estimates ranging from 42 to 78% (Schechter & Willheim, 2009). Based on these figures, the possibility that a child’s behavioral, academic, or emotional difficulties are trauma related should become particularly salient when a child is known to live in a high-poverty/high-risk urban neighborhood; assessors will be even better served by more specific and nuanced knowledge of neighborhood characteristics in the communities in which they work. Even without identifying specific trauma exposure, living in underresourced neighborhoods is associated with elevated disruptive behavior in toddlers (Heberle, Thomas, Wagmiller, Briggs-Gowan, & Carter, 2014). Sociocultural Factors In addition to identifying the child’s caregiving relationships, mapping out the larger web of family relationships in which the child is developing, and considering the ways the child’s presentation may be influenced by the caregiving context and the setting in which the child is cared for, assessors must also consider the ways sociocultural factors may affect caregiver reports, children’s performance on assessment instruments, the meaning of a child’s behaviors to the caregiver and other family members, whether family members are worried or concerned about the child, family members’ service seeking behavior and openness to referrals, and the assessor’s own interpretation of the child

8

Assessment of Psychopathology in Young Children

TABLE 1.1 Illustrative Examples of Sociocultural Factors that Impact Assessment Findings1 Example of norm

Explanation of relevance in the assessment context

Communication style

Expectations for appropriate eye contact vary across cultural groups; for example, Asian American children may avoid eye contact with authority figures as a symbol of respect and deference (Thaler, Allen, & Scott, 2014).

Avoidance of eye contact may be interpreted as symptomatic if an assessor from a group in which sustained eye contact is normative (e.g., European Americans) is unaware of the variability in cultural norms for these behaviors.

Caregiving norms

Middle-class Puerto Rican mothers have been found to engage primarily in mother-led feeding practices (in which the mother feeds her infant with a spoon or bottle, in contrast to the infant using a spoon or bottle on his or her own) with their 12-month-old infants and to expect that their infants will not yet be able to self-feed at this age (Schulze, Harwood, & Schoelmerich, 2001).

Instruments designed to measure adaptive behavior through the lens of middle-class Euro American culture, in which early attainment of self-feeding skills is emphasized, may falsely identify delays in children whose caregivers have simply not yet organized their interactions with their children to require the development of these skills.

Priorities for child behavior

Endorsement of conformity as a child-rearing goal among mothers is negatively correlated with income, educational attainment, and occupational prestige; thus, overall, as socioeconomic status increases, mothers are less likely to see conformity to externally imposed norms as a model for their children’s ideal behavior (Luster, Rhoades, & Haas, 1989).

An assessor primed to assess a given caregiver’s attitude regarding conformity versus self-direction is likely to have greater success communicating with the caregiver about his or her child’s strengths and weaknesses than an assessor who assumes that the caregiver shares his or her own values regarding conformity versus self-direction.

Beliefs about mental health services

Latinos, Asian Americans, and African Americans have been found to report high levels of perceived stigma regarding mental health problems in their communities (Alvidrez, 1999; Brown, Marshall, Bower, Woodham, & Waheed, 2014).

An assessor who is familiar with how perceived stigma might function as a barrier to service engagement is likely to be more effective than one who is not at addressing caregivers’ concerns about confidentiality and encouraging the family’s continued engagement in services (if needed).

1 Endorsement

of cultural norms may vary depending on ethnicity, race, socioeconomic status, immigration status, level of acculturation (for immigrant families), geographic location, and other aspects of identity, both separately and in intersection with one another.

and family’s presentation (Table 1.1). Early childhood social-emotional and behavioral problems are identified when a set of behaviors appears with heightened or reduced frequency and intensity or when a child shows developmental deviance through the presence of unusual behaviors or the quality of typical behaviors. While diagnostic guidelines sometimes specify the frequency with which a behavior must occur to be consistent with a clinical diagnosis, it is often the case that caregivers are asked to determine whether a behavior occurs frequently or not, a determination that will necessarily be influenced by the caregivers’ expectations for child behavior. Such expectations will vary across cultures and communities. In addition, the determination of whether a behavior (e.g., arguing with adults) has occurred at all is culturally mediated. Thus, the identification of social-emotional and behavioral problems is far from culturally neutral. In addition to problems related to the frequency, intensity, or quality of behavior, children may also show delays in the acquisition of social-emotional competencies (Briggs-Gowan & Carter, 1998). The identification of delays in competencies, like the identification of problem behaviors, is dependent on the caregivers’ expectations for the child as well as the child’s access to opportunities to develop and demonstrate competencies.

For these reasons, sociocultural considerations must inform the entire assessment process, from instrument selection through interpretation of findings. In selecting instruments for assessment, assessors should consider the following (note that this is not an exhaustive list): (1) Are the items on the instrument culturally relevant for this parent–child dyad? (2) Does the sample on which this instrument was normed represent this parent–child dyad? (3) Is the language in which this instrument is written the optimal language in which to assess this parent–child dyad? (4) If an instrument is designed to use familiar tasks as part of the assessment process (e.g., asking a child to complete visual puzzles), are these tasks actually familiar for this parent–child dyad? Unfortunately, in many cases socioculturally and linguistically appropriate instruments may not be available. The use of culturally inappropriate instruments may have striking effects and is therefore no small problem; one example comes from a study of Australian Aboriginal children, in which all 124 children who participated in the study screened below the cutoff for detecting developmental disabilities or academic delays on a normed and validated screening tool (the Brigance screen) (D’Aprano, Carapetis, & Andrews, 2011). Such results highlight the damage that may occur when culturally invalid instruments are used

Domains of Development

with young children—though, unfortunately, the failure of the instrument to work as intended will likely be far less obvious in the context of one-on-one assessment with a child (as opposed to a large research study or a large-scale universal screening effort). In work with young children and their families, several strategies may be used to enhance the overall cultural appropriateness of an assessment, even if culturally appropriate standardized instruments are not available. For example, it may be helpful to review items with a caregiver, discussing the caregiver’s understanding of the meaning of both symptoms that have been endorsed and those that have not. Caregivers can be asked to give examples of a time when they observed a behavior that they have endorsed on a symptom checklist. (Note that this strategy is likely to yield valuable information from parents belonging to a cultural group on which an instrument was normed as well as from parents for whom an instrument may not be culturally valid.) In addition, it is important to ascertain whether the behaviors assessed by a given instrument—whether observed, assessed in an interview, or reported on a questionnaire—are culturally relevant for the target child. For example, if a child is from a cultural group in which nonverbal expression is valued, the examiner should be careful to look for signs that a child is responding nonverbally to a prompt meant to elicit a verbal response in the dominant culture but designed to fit a different cultural group (McLaughlin, Gesi Blanchard, & Osanai, 1995). (A nonverbal response in this case may indicate not a deficit in the skills needed to respond verbally but rather an incongruence between the forms of communication expected by the test designer and the forms of communication familiar to the child.) In addition to attempting to compensate for inaccuracies in a given standardized instrument (or set of instruments) that are not culturally valid for a given client, assessors should also consider whether there is important information about the child’s development that is missed entirely by standard measures. For example, standard psychological assessment batteries typically do not include measures of ethnic and racial identity development, although these are essential aspects of the overall development of racial and ethnic minority children (Yasui & Dishion, 2007). Thus, overall, in cases in which the set of available instruments contains no culturally valid tools, assessors must be resourceful in identifying means for gathering supplemental information and extremely cautious about interpreting any standardized results obtained by calibrating a child’s functioning against a normative group that does not represent their experience. In addition to instrument selection, assessors must also consider cultural factors throughout the process of

9

interviewing parents and children and interpreting the overall findings of the assessment (Kleinman, Eisenberg, & Good, 1978). We suggest that assessors consider the following questions as a starting point for framing a culturally informed evaluation: (1) What does the caregiver see as the problem? (2) What are the caregiver’s expectations with regard to treatment? (3) How does the problem affect the child, caregivers, and other family members? (4) What does the caregiver see as optimal behavior for this child? (5) What are the caregiver’s beliefs about child development? What does the caregiver see as the key tasks of development at this stage in the child’s life? (6) How have community members responded to the child and caregiver? (7) What is the predominant attitude regarding help seeking in the caregiver’s community? If known, how do community members see the organization with which the assessor is affiliated? Does the caregiver share these beliefs? Culturally competent assessment requires that these and related questions be addressed for each family individually; however, familiarity with the cultural groups with which a family identifies is likely to enhance the assessment process by informing the questions asked by the assessor, the way questions are asked, and the way the caregiver’s responses to questions are understood. To this end, consultation with colleagues who share an identity status with the caregiver may be invaluable. In many cases, assessors may have only limited access to information about a child’s caregiving context, family relationships, and cultural affiliations. For example, screening instruments may collect only minimal demographic information about a family and are often administered in contexts in which an assessor has little or no direct contact with caregivers. In these cases, the selection of instruments that are appropriate for the target population is particularly important, as is additional, culturally competent assessment of any problems identified by a screening instrument. Care should be taken to ensure that screening results are delivered in a culturally competent manner, with consideration given to the linguistic needs of the family receiving feedback as well as to beliefs known to be widely held in the community that may impact caregivers’ understanding of screening results and their actions following the delivery of these results.

DOMAINS OF DEVELOPMENT One of the myths about early emerging psychopathology that has been debunked in the past two decades of research on young children’s mental health is that early emerging

10

Assessment of Psychopathology in Young Children

psychopathology is undifferentiated. Indeed, whether validating the presence of differentiation across broad band domains of psychopathology, such as distinctions between disruptive or externalizing problems (which typically include problems in aggression, hyperactivity, inattention, and defiance), mood or internalizing problems (which typically include problems in social withdrawal, depression, and anxiety), and regulatory problems (which typically include sleep, eating, negative emotionality, and sensory symptoms) (Achenbach, 1966; Achenbach, Edelbrock, & Howell, 1987; Carter, Briggs-Gowan, Jones, & Little, 2003) or validating distinctions between different types of symptoms within a dimension of psychopathology, such as anxiety (Mian, Carter, Pine, Wakschlag, & Briggs-Gowan, in press; Mian, Godoy, Briggs-Gowan, & Carter, 2012; Spence, Rapee, McDonald, & Ingram, 2001) or disruptive behavior (Wakschlag et al., 2014; Wakschlag, Henry, et al., 2012; Wakschlag, Briggs-Gowan, et al., 2008; Wakschlag, Hill, et al., 2008), both parent reports and observations support differentiation of psychopathological symptoms in the toddler/preschool period. Based on this relatively new recognition of differentiation in the early emergence of psychopathology (and as reviewed later in this chapter), screening and assessment tools that focus on differentiated aspects of psychopathology have been developed. Whether used in clinical or research settings, tools that assess social-emotional behaviors and psychopathology need to be interpreted in relation to knowledge about children’s relational, family, and cultural contexts and in relation to the children’s developmental context, or developmental functioning. We consider the child’s developmental context quite broadly, as multiple developmental domains transact with normative social-emotional development and psychopathological trajectories. Risks for or the occurrence of psychopathology intersect with delays, deficits, or atypical functioning across many other developmental domains. Indeed, intersections in developmental risk across psychopathology and other developmental domains is almost always observed; children with intellectual and developmental disabilities (Einfeld et al., 2006; Emerson, 2003), language delays and specific language disorders (Henrichs et al., 2013; Ross & Weinberg, 2006), learning disabilities (Morgan, Farkas, Tufis, & Sperling, 2008; Yu, Buka, McCormick, Fitzmaurice, & Indurkhya, 2006), or sensory processing disorders (Ben-Sasson, Carter, & Briggs-Gowan, 2009; Ben-Sasson, Soto, Heberle, Carter, & Briggs-Gowan, 2014) show increased rates of psychopathology. Importantly, the nature of these intersections varies from one condition to another, so knowledge

of neurocognitive and linguistic profiles or identified intellectual and developmental disabilities can inform which aspects of psychopathology are assessed in greater depth. Thus, if a child has a known neurodevelopmental condition, such as autism or fragile X syndrome, a literature review should help to identify common areas of difficulty in the social-emotional domain (e.g., anxiety) that should be given additional attention in the assessment process. Conversely, if a social-emotional problem such as ADHD is under consideration, problems in executive functioning should be assessed as these two often co-occur. Direct assessment of multiple developmental domains is often essential to understanding the whole child. Further, given that many young children will not be in environments in which informants will have the knowledge to report on the child’s functioning across all of the relevant developmental domains, we strongly encourage including direct, norm-referenced assessment of cognitive and linguistic functioning prior to interpreting the results of social-emotional assessments, particularly since some assessments of social-emotional and behavioral functioning assume normative competencies in other developmental domains. single-informant screeners are not adequate for this purpose. For example, within assessments of attention, questions may include ratings of the child’s ability to follow multistep instructions, which may be better explained by a receptive language deficit than inattention. Similarly, a young child with an intellectual disability may not be engaging in pretend play, but this may reflect a general cognitive delay or deficit rather than a specific delay or deficit in the socialemotional domain. In general, we do not expect gains in social-emotional development to exceed those observed in language or cognitive domains. Therefore, these two developmental lines are particularly crucial for contextualizing interpretation of findings in the socialemotional domain. In clinical settings, assessment of other developmental domains may aid in motivating parents to participate in preventive and targeted interventions designed to reduce psychopathology. In early childhood parents are often more concerned about and more attuned to delays in language development than to atypical emotional, social, or behavioral development (Godoy, Carter, Silver, Dickstein, & Seifer, 2014). Presenting to parents the ways symptoms may be interfering with language (or other) cognitive learning can aid in building a shared view of the child that will motivate parents to participate in interventions focused on reducing symptoms.

Selecting an Assessment Approach and Tool

Explaining the intersections of developmental lines in relation to establishing risk for psychopathology may also aid in freeing parents from the unfair burden of responsibility or feeling blamed for causing their child’s social-emotional or behavioral problems. For example, explaining the ways challenges in processes such as inhibitory control, executive functioning, language processing, and emotional reactivity can intersect with and heighten risk for psychopathology—particularly given that the explanation can emphasize temperamental and neurocognitive vulnerabilities or adverse environmental exposures—can often facilitate greater empathy for a child who has been engaging in problem behaviors that cause burden to the family system. Therefore, knowledge of multiple developmental domains and their intersections, many of which are well documented in the literature, is extremely important in assessment domain and measurement selection. Moreover, knowledge of children’s functioning in multiple developmental domains is often essential to appropriate interpretation of findings in research and clinical settings, and can be used to promote a stronger alliance between clinicians and parents when giving feedback and offering recommendations about intervention to parents whose children evidence elevated symptoms and disorders in clinical settings.

11

SELECTING AN ASSESSMENT APPROACH AND TOOL When selecting tools to assess social-emotional problems in young children, a number of factors should be considered: (1) making sure the tool will best suit the needs and goals and purposes of the assessment and will fit within any setting constraints; (2) reviewing psychometric properties, such as reliability and validity; (3) evaluating the appropriateness of the tool; and (4) reviewing the appropriateness of the tool for children of a particular age or cultural background. Types of Tools When selecting a tool, it is important to determine what type of tool will provide the level of information required given the goals of the assessment. We discuss three types of tools for collecting information from parents and other caregivers: brief screeners, more comprehensive in-depth checklists, and psychiatric interviews (Table 1.2). These different types of tools vary tremendously in their format, the level of information that they provide, and the purposes for which they are most appropriate. Screeners typically yield scores that indicate whether a child may have

TABLE 1.2 General Characteristics of Different Types of Parent and Other Caregiver Report Tools Type

General format

Type of information obtained

Screeners

5–10 minutes to complete Completed by parents or other caregivers Minimal training to administer or interpret Many assess a wide array of problems, such as behavior problems and emotional problems Some focus on a specific problem area, such as behavior problems only

Screeners usually provide one or two scores that indicate whether a child falls into a risk category suggesting that the child may have problems without providing detail into the specific nature of the problem. Results often trigger a more in-depth assessment to determine specific areas that are problematic, to identify treatment needs, and to inform planning.

Comprehensive checklists

15–30 minutes to complete Completed by parents or other caregivers Minimal training required to administer Interpretation typically requires training in psychological assessment Provide 2+ broad domain level scores (e.g., externalizing, internalizing) May provide narrowband scores about specific areas (e.g., anxiety)

Comprehensive checklists provide in-depth information about a child’s functioning. Most cover multiple areas, but some focus on specific types of difficulties, such as executive functioning problems. Scoring typically provides a profile of functioning, showing strengths and weaknesses in different areas. These results can be used to establish eligibility for services, generating treatment plans, and documenting change over time.

Diagnostic interviews

1–2 hours to complete at a minimum Parent or other caregiver is interviewed by a trained individual All interviewers must be trained in administration Some may be administered by nonclinicians; others require clinical interviewers Interviews probe for information about whether symptoms are present, duration and onset of symptoms, and impairment

The detailed information obtained from these interviews allows determination of whether a child meets psychiatric criteria for a range of disorders. Scoring also can yield symptom counts within disorders, such as how many depressive symptoms were reported.

12

Assessment of Psychopathology in Young Children

social-emotional problems but do not provide in-depth information about specific areas that are problematic. Longer, comprehensive checklists provide more in-depth assessment of problems within a specific area or areas. Many of these checklists provide detailed profiles that illustrate a child’s strengths and weaknesses relative to children of a similar age in multiple areas, such as depression, anxiety, impulsivity, aggression, and social competencies. Such profiles and score details can be especially helpful for treatment planning, and documenting change over time or with treatment. However, checklist tools are not appropriate for establishing whether a child meets criteria for a psychiatric disorder. Instead, psychiatric interviews collect very detailed information necessary to establish whether a child meets diagnostic criteria for a range of psychiatric disorders (e.g., symptom presence, onset, duration, and impairment), and can be important for treatment planning. Observational methods also can supplement more traditional informant report methods. In the next section of this chapter, we will review each type of measure in detail and provide a review of selected measures available for assessing social-emotional functioning in young children. Understanding Psychometric Properties The first step in reviewing an instrument should be to examine its psychometric properties. Unfortunately, in both clinical and research settings, practical considerations such as a tool’s ease of use; the time required for training administration, scoring, and interpretation; costs associated with use; the level of professional training required to administer, score, and interpret the instrument; and the influence of common regional practices often outweigh consideration of psychometric properties. The following section outlines psychometric issues most relevant to parent–caregiver reports in young child assessment: reliability, validity, sensitivity/specificity, positive and negative

predictive value, and normatization. Other factors such as cultural considerations, developmental appropriateness, and response formats that may influence instrument adoption also are discussed. Reliability Reliability refers broadly to the consistency or stability of a measure. For measures that have scale scores, the consistency of a measure is typically evaluated by examining two or three aspects of reliability: internal consistency, test–retest reliability, and interrater reliability. Table 1.3 provides an overview of these types of reliability and the types of tools to which they generally apply. Rules of thumb for interpreting whether a tool has good enough reliability for different types of measures and related statistics are presented in Table 1.4. Internal consistency is the extent to which individual items on the scale hang together and reflect the same construct. Usually, this is assessed with Cronbach’s alpha (Cronbach, 1951). Scales with Cronbach’s alpha of .70 or greater are usually considered to have adequate internal consistency, whereas those with alphas between .60 and .69 are considered marginal and those falling below .60 are considered unacceptable (Cicchetti & Sparrow, 1981; Nunnally, 1978). However, some clinically informative young child measures may not have acceptable internal consistency. This can occur if a measure includes behaviors that rarely occur in the population (e.g., atypical behaviors related to autism spectrum disorder; ASD) or sets of behaviors that are clinically concerning but not likely to co-occur (Achenbach, Edelbrock & Howell, 1987; Briggs-Gowan & Carter, 1998). For example, the ITSEA Atypical Behaviors Index and Maladaptive Index each concern a set of behaviors that are rare but clinically important (e.g., rocking, spinning, PICA, head banging). These indices are expected to have low internal consistency in a normative population (Carter & Briggs-Gowan, 2006),

TABLE 1.3 Types of Reliability Type

Definition

Measures to which this is most relevant

Statistic used

Internal consistency

How well do the items in a scale hang together?

Measures with scale scores, such as screeners or checklists

Cronbach’s alpha (𝛼)

Test–retest reliability

How consistently does the tool provide the same results over time?

All measures

Correlation coefficients (r) or intraclass correlation coefficient (ICC) (Bartko, 1976)

Interrater reliability

How consistently does the tool provide the same results when different individuals rate it?

All measures, especially those where judgment is involved, such as semistructured interviews or observational systems

Intraclass correlations, kappa

Selecting an Assessment Approach and Tool

13

TABLE 1.4 Reliability Statistics and Rules of Thumb for Interpreting Reliability Score types

Statistic

Criteria1

Dimensional scores Test–retest (about the same person at different times) Interrater (about the same sample of information about the same person by different raters)

Symptom counts Coding composite scores Scale scores

Intraclass correlation (ICC)

MENICE

(15.1)

In abusive families, parental upset often leads to abuse, and children watch vigilantly for a hint of negativity and then try to shift to the positive, such as MENICE , and so avoid the negative, MEBAD or MEMEAN . Through emotional splitting they try to avoid the bad.

not MEBAD > MEGOOD

(15.2)

In this first crude dissociation skill a child tries to suppress MEBAD and works unconsciously to shift to MEGOOD , provided that the context supports it. If the context supports the opposite representation, MEBAD , children will have great difficulty sustaining the shift at this level, but they will gain more control as they develop. Children of this age also start to distinguish between how to act in private (e.g., dealing with their family at home) and in public (e.g., climbing on the town’s play structure). In this way, they shift between representations of private and public and thus begin to internalize differences in their public and private worlds. Children in public work to keep a focus on MEGOOD , reserving MEBAD mostly for private settings (as parents can attest). That is, they avoid MEBAD in public.

ME-PRIVATEBAD > ME-PUBLICGOOD

(15.3)

When maltreated children differentiate public and private, they try to dissociate the two. On the playground, they strive to keep in mind their private world, acting good and proper. They avoid acting out the private MEs that connect to their abuse at home. We represent this affective splitting with a bar through the symbol for shift of focus.

not ME-PRIVATEBAD > ME-PUBLICGOOD

(15.4)

Isolating dissociation allows ME-PUBLIC to suppress and control ME-PRIVATE in contexts that support focus

12:37pm

681

on the public sphere. The same limit was present for the isolating shift from MEBAD to MEGOOD . Single representations have little power for overcoming context effects and so are not generally effective. The ability to relate representations using representational mappings (Level Rp2) gives children a stronger capacity to direct their own activities. That means that more stable dissociation requires the stronger structures that derive from coordination of representations into mappings, because they provide a mechanism for children to control that mapping, shifting MOM from mean to nice.

MOMMEAN

MENICE

(15.5)

Normally the capacity to sustain mappings develops at about 4 years and creates many new skills involving relations with other people, such as mother to father or mother to child (Fischer & Pipp, 1984). Children living in a situation of hidden family violence manipulate these skills frequently in their life in private. When 4-year-old Jason’s mom scolded his father sitting at dinner, Jason drew attention intensely to the drawing he did earlier that day (taped on the wall near the table). This abrupt switch led his parents to recognize immediately that he was trying to change their interaction. Jason was attempting to direct his mother’s attention to something happy. At this age children show a surge in understanding role relationships, which gives children in abusive families a capacity to separate public and private roles, thus facilitating strong isolating dissociation of public and private. For example, a child can maintain a public pattern of interaction, such as acting out ME-PUBLIC playing with their father, even though the primary private relationship with him is abusive boss interacting with bad child:

not private public MEBAD DADBOSS > MEGOOD DADGOOD (15.6) By age 6 or 7 children can build representational systems (Level Rp3), giving them more reliable control over their public and private worlds. At school or on the playground they can shift from public to private modes depending on who is more powerful. Roger shows this movement to a higher developmental step, sharply distinguishing public versus private, as for example when he tried to impose his private mode on other children whom he thought he could dominate. At age 9 Roger made a conscious effort to please his teacher and earn her approval. Being watched by adults Roger was a model student, even helping teachers to

Page 681

Trim Size: 8.5in x 11in

682

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

monitor class when they had to leave the room. However, out of sight of the teacher he was a ruthless bully, threatening other children to take their possessions or make them help him with homework, based on his assessment of their power. Approached by an older, stronger boy, he would yield. Threatened by another boy, he would cry and beg for mercy. Roger bragged that he had a good family, including his father’s business, money, and community respect. He did not say that his father, Mr. S, was strict and unpredictable and his mother was visibly sad and passive. Mr. S expected the children to respond quickly when he made a demand, and he often grew angry. In private, he told his children that they were worthless misfits, and he routinely degraded his wife, even physically abusing her. She often cried in silence. At home, Roger was subjected to the hidden family violence pattern and dissociated public and private, as evident in his relationships at school. In public (with teachers watching) he was a good boy, seeking the teacher’s approval. At home, he was submissive (a follower or victim) and obeyed his father’s commands and was frequently punished. In school with no teachers watching, the private domain dominated: Roger would take on the role of either follower or bully, based on his judgment of the other child’s power. Now social relationships were becoming increasingly complex involving (1) both public and private representational systems (Level Rp3) with multiple roles and (2) an isolating shift of focus between public and private systems: OBEY private

ME-FOLL’R

BAD

GOOD

ME-STUD’T

OBEY

DEMAND

YOU-BOSS

PUNISH

public

>

COMMAND

LIE

(15.8)

ROGER HELP

JAMES

COLIN STOLE

(15.9)

Second Tier of Dissociation: Isolating Personalities As Roger grew cognitively he transitioned to the next developmental tier, moving from representational systems to abstractions for people as personalities. As a result, his isolating dissociation became more skilled, as he improved his capacity for deceiving people and for skillfully separating public and private. By coordinating representational systems into abstractions (Level A1), he could use abstractions to coordinate concrete interactions across his public versus private domains. In his private world he coordinated two representational systems for abusive bossiness, one as boss to some other victim and a second one as victim to some other boss. We name the general personality characteristic “tyranny” to reflect his goal of being a tyrant in his family:

BAD

(15.7)

lie

BLAME COLIN

OBEY

Here Roger acted out the roles of follower and student, ME-FOLLOWER and ME-STUDENT. In appropriate contexts he could switch his role, acting as boss or teacher (ME-BOSS or ME-TEACHER) to someone else as follower or student (YOU-FOLLOWER or YOUSTUDENT). He could shift between follower or boss, student or teacher based on his judgment of context. By Roger’s age, children work hard to skillfully deceive and lie (Lamborn et al., 1994; Saarni & Lewis, 1993). Deception is at the center of hidden family violence, with individuals lying to both others and themselves. Detecting lies requires relating representations to each other, what is true and what is a lie. Understanding a lie requires coordinating two representations in at least a mapping:

TRUTH

STOLE

ROGER

ME-FOLL’R

LIKE

YOU-TEAC’R

Children develop and detect lies with representational mappings and they master them with the next level of skills, representational systems. (Consolidation of skills often requires at least two skill levels.) For instance, Roger stole a key from James and then “helped” James by blaming the robbery on the least popular child in the class, Colin. To skillfully lie required coordinating several states of James’ knowledge (Roger “helping” and Colin “stealing”) with claims or actions by Roger (stealing the toy and blaming Colin) in a representational system:

DEMAND

ME-BOSS

PUNISH

DEMAND

YOU1-BOSS

PUNISH OBEY

YOU2-FOLL’R BAD

ROGER

TYRANT

(15.10)

Roger elaborated these personalities using isolating dissociation to keep them separate as in this formula: ROGER-PRIVTYRANT > ROGER-PUBLCOMPETENT

(15.11) By age 16 he had firmly established his capacity for single abstractions and started to construct abstract mappings (Level A2), establishing his public personality clearly at school, and using it skillfully: At 16 years Roger was a successful sophomore in high school—on the student council and the track team. But he did not have any close friends and spent lots of time seeking attention from older student council members. He helped

Page 682

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

The Development of Affective Dissociation in Maltreated Children

the vice principal, whom he believed had the power to set the rules. Roger was influential in school politics, presenting himself as knowing the answers and handling issues better than others. He was directly cruel to younger students, and especially condescending to girls. Roger was quiet and withdrawn at home with his father, who remained rigid and cold. His father often confronted Roger, who whined and hung his head. With his mother he insulted her when his father was not at home, and one time he pushed her into a door. He bullied his younger brother to do many of his chores.

Roger is building abstract mappings that connect the characteristics of victim and tyrant in the private domain, and in the public he is building the image of competent student leader and satisfied authority, as in the following isolating dissociation skill:

ROGERVICTIM ROGERCOMPETENT

private

public

FATHERTYRANT > AUTH’TYIMPRESSED (15.12)

With this skill Roger focuses on roles of both victim and competent leader. At the same time he can reverse his stance to take the role of tyrant to someone else’s victim, as shown by his treatment of girls, his mother, and younger students:

ROGERTYRANT

private

MOTHERVICTIM

(15.13)

As Roger became an adult, he used his developing skills to create the isolating dissociative roles that he has sought all along. As he searched for a wife, he soon showed affective splitting in his treatment of the young woman he fancied.

12:37pm

683

Roger was constructing more complex skills to control June in both private and public domains, continuing to dissociate his positive public competence and negative private tyranny. He was attempting to construct his own family with an unfortunate joining of love with tyranny. He strived to bring June into the vice of his isolating dissociation, letting her be part of his public one only as a perfect stereotype. In public, in contrast, he connected his general competence with doing good for his community, limited by his rigid control of his private tyranny. Thus, he built abstract systems (Level A3) to continue his isolating dissociation of the two worlds called Public and Private: LOVER

ROGER

private

TYRANT

PROSOCIAL

ROGER

COMPETENT

LOVED

JUNE

VICTIM

public

>

GRATIFIED

AUTH’Y

IMPRESSED

(15.14)

Over time, Roger will build ever more complex skills of isolating dissociation that separate his public and private worlds. These are complex, sophisticated skills, and at the same time they are clearly dangerous and pathological. Unfortunately, Roger is creating hidden violence in his family as evidenced by his actions with June. For hidden family violence, the pattern starts with a maltreating family that dissociates their proper public world from their private, tyrannical one. Children growing up in this family are abused physically and emotionally, and sometimes sexually. From this contextual and affective dissociation, the children isolate public from private, and eventually their maltreatment promotes active dissociation, separating the public and private worlds. Psychopathology Is Not Developmental Immaturity

At age 24 Roger demonstrated his competence and apparent leadership in his public self. He joined the local church and became a member of the Young Republican Club. He offered to volunteer to promote his town’s business for the Chamber of Commerce, and he founded his own real-estate business. He started to date a 17-year-old named June, who had recently finished high school, having met her when she worked for him as a temporary secretary. Roger was highly attentive to her, buying expensive gifts and sending flowers almost daily. He said repeatedly how special she was and that she did not have a single fault. Roger grew increasingly controlling and possessive. He accompanied her to the hair stylist to tell the stylist how to do her hair. Telling June to throw out all the clothes in her closet, he insisted on picking her new clothes. He began to complain that June was a poor cook, had no sense of style, and was too naive to go out alone.

This dynamic analysis of development of isolating dissociation questions the traditional view that psychopathology comes from developmental immaturity or regression. Whereas traditional approaches tend to presume that psychopathology reflects a relative lack of skills or some form of deficiency, a developmental perspective brings forth the possibility that maladaptive behavioral development involves advances in skill development, not simply immature development. In the case of maltreated children, such advances take the form of increasingly powerful and more complex types of dissociation. Most scholars and practitioners who study maladaptive development have not detected this complexity because they have not taken the viewpoint of their patients. Moreover, patients are

Page 683

Trim Size: 8.5in x 11in

684

Cicchetti

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

commonly observed under minimizing conditions because of absence of contextual support and presence of stress. People subjected to stress and abuse follow strange pathways that organize their worlds in radically different ways, and which are characterized by increasing sophistication and complexity—and not regression or simplicity. Modeling the Growth of Dissociation and Coordination Dynamic mathematical modeling, based on general principles of dynamic interactions between variables, can provide deductive explanations of empirical observations. They are deductive explanations in the sense that they are derived from basic principles of developmental change and are expressed as mathematical rules generating descriptions of developmental trajectories. The strength of dynamic modeling lies in their power to deductively explain—on the basis of a small set of general dynamic principles of change—a broad range of qualitative properties that are typical of the phenomena that we can empirically observe in a particular domain such as developmental psychopathology. According to Weisstein (1991), a dynamic system is a formal way of describing how one state (of a system) develops or changes into another state over the course of time. A highly simplified mathematical representation of a dynamic system is as follows xt+ 1 = f (xt ) which can be read as: the next state—at time t +1—of a particular variable x is a mathematical function, f, of the current state, namely the state at time t.1 If the function f is a multiplication by a fraction such as 0.02, the dynamic model would represent stepwise exponential growth, such as money in a savings account. We have seen that most dynamic systems take the form of interactions among many components or variables. Such interactions can be modeled through the coupling of change equations. In order to dynamically model the developmental process occurring under traumatic conditions, we shall assume that skills may vary between 0, meaning that the skill is totally absent, and some value around 1, meaning that the skill is optimally or maximally developed. Hence, 1

c15.tex V3 - Volume I - 12/10/2015

The step size +1 can be given any temporal value, such as a second, a day, or a month, but it can also be given a value that approaches an infinitely small time step, in which case we shall call the above equation a differential equation, instead of a difference equation.

for a psychologically healthy person, the value of the coordination skill may be about 1, and that of the dissociation skill about 0, meaning that there are hardly any contexts that cause dissociation in this healthy person. For young children, coordination and dissociation may still be a developmental issue, and they may still be sensitive to transient (nontraumatic) contexts in which they tend to solve the situation by dissociation (e.g., by lying about things, while suppressing the fact that they know they are lying2 ). In the discussion on dissociation, we have seen that people who suffered from serious traumas may develop sophisticated and adaptive dissociation skills. In principle, these people may still have reasonable coordination skills, which means that there are still contexts in which the person’s personality is not dissociated. However, it is likely that such skills would not reach the higher levels of abstractions or principles that are typical of coordination in psychologically healthy people. We can mathematically represent this situation by values of skill levels between 0 and 1, for example, around 0.5 (for the modeling, the exact details of these numerical values do not really matter as long as they express observable differences in skill levels). In previous work, the authors have shown that there exists a powerful and yet relatively simple mathematical expression for defining the growth of dynamic skills (Fischer & Bidell, 2006; Van Geert, 1991, 1994; Van Geert & Steenbeek, 2005). This equation is the coupled logistic growth equation, which takes the following form. Let the value C represent the level of the coordination skill of a particular person at a particular moment in the development, and D the level of the dissociation skill. For simplicity, let us focus on the growth of coordination, which can be expressed as follows: Ct+1 = Ct + Ct ∗ rC ∗ (1 –Ct ∕KC ) − a ∗ Dt ∗ Ct The equation says that the next level of coordination (e.g., after a day or after a week, dependent on the chosen step size) is equal to the current level of coordination plus a certain amount of change. This amount of change is the result of the growth of coordination on the one hand (the underlined part of the equation), and the competition with other skills, in this particular case the competition with the dissociation skill (the double underlined part of the equation). The growth part of the equation specifies 2

See figure 15.8, illustrating the changing levels of coordination and dissociation as a consequence of childhood trauma.

Page 684

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

The Development of Affective Dissociation in Maltreated Children

change in coordination as a function of the current level of the coordination, a growth rate r and a carrying capacity K (Van Geert, 1991, 1994). Both the growth rate and the carrying capacity are an expression of the total underlying dynamic system, consisting of internal properties such as memory capacity, intelligence, information processing skills and external properties such as the family environment and experienced events. The carrying capacity is an important parameter since it is the summary of the entire underlying system, in terms of the total support it can give to this particular skill, namely coordination. In a well-functioning and caring family, this total support and hence the carrying capacity for personality coordination is likely to be much higher than in a badly functioning family characterized by many coercive interactions. The competition part of the equation describes the change of coordination as the multiplication of a competition parameter, a, the level of the competing skill, which in this particular case is the dissociation skill, and the level of the coordination skill itself. The equation says that the higher the level of the dissociation skill, the more the coordination skill will suffer from it, implying that level of the coordination skill will diminish over time, unless the competition can be counteracted by the growth of coordination, as it is influenced by all the personal and family factors that support it. We can now present a similar equation for the growth of the dissociation skill, but with parameter values that relate to dissociation this time (for instance, in a well-functioning family the total of factors supporting the growth of dissociation will be very low compared with the total of factors supporting the growth of coordination). Dt+1 = Dt + Dt ∗ rD ∗ (1 –Dt ∕KD ) –b ∗ Ct ∗ Dt We shall now assume that if a serious trauma occurs in the life of a particular young person—for instance sexual abuse by the father as described in the earlier example—the total set of factors supporting dissociation as an adaptive skill is considerably increased. Hence, the carrying capacity for the dissociation skill, KD , strongly increases as a consequence of traumatic events. In addition, the growth rate rD of dissociation skill is increasing, because the child now feels the need for developing dissociation as a way of coping with the traumatic situation. A very simple way of mathematically expressing the effect of trauma is the following: KDt+1 = KDt + d ∗ T rDt+1 = rDt ∗ e ∗ T

12:37pm

685

The factor T, which stands for the occurrence of traumatic events, is a parameter that is either 0 (when no traumatic events occur) or 1 (when traumatic events occur). The parameters d and e moderate the effect of trauma, and, like all other parameters, are specific for a particular individual child in a particular environment. For instance, children with a high level of resilience will be less affected by a particular trauma than children with a low level. In a family, where one of the parents is highly supportive of the child, a trauma caused by the other parent might eventually have less effect on the growth of dissociation of the child in a family where the mother for instance tacitly supports the abuse by the father. Since all parameters are individual and environment specific, we obtain an entire space of possible parameter values that represent possible individual cases. We can now use this dynamic systems model of coordination and dissociation skill growth to study the variety of individual trajectories that results from various combinations of parameters, including the various family support levels, growth rates that are dependent on individual personality characteristics, competition factors and factors influencing the effect of a particular traumatic experience. Figure 15.8 shows the simulation of two trajectories of coordination and dissociation skills in a child who suffers from serious traumatic abuse from the age of about eight years on (the models step size is one month). The top graph shows a trajectory of dissociation following trauma at around the age of eight years which stabilizes well into adulthood and becomes the sole way of coping for the person in question. The bottom graph shows a trajectory after a similar trauma that shows a pattern of stabilization between coordination and dissociation, with coordination still active, but with dissociation as the dominant mode of adaptation. This latter pattern implies that there still exist enough contexts in which the person can use personality coordination skills, for instance in the context of therapy where the person is able to deal with the consequences of trauma in a healing way. Surprisingly enough, the only difference between these two scenarios is the value of the growth rate parameter for coordination in the second case, which is slightly bigger than in the first case. That is to say, the person’s slightly bigger ability to build up coordination skills—which is represented by the value of the growth rate parameter—functions as a strongly protective factor, in that the situation is still maladaptive, but less so than the complete dissociation of the first example. This provides but one of the many illustrations of the nonlinear character of dynamic systems, where the effect

Page 685

Trim Size: 8.5in x 11in

686

Cicchetti c15.tex

in children who exhibit attention and emotional-behavioral difficulties (EBD). In this section, we examine how dynamic systems models can be used to understand the processes by which learning fostered and thwarted in children with attention and emotional-behavioral difficulties.

Skill level

1 0.8 0.6 0.4

Context Sensitivity of Attention Problems

0.2 0 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 Age in years Coordination

Dissociation

1.2 1 Skill level

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

1.2

0.8 0.6 0.4 0.2 0 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 Age in years Coordination

Dissociation

Figure 15.8 Modeling of two scenarios of dissociationcoordination after trauma occurring at 8 years of age; trajectories are determined by dynamic interactions between coordination and dissociation.

of changes in certain parameters strongly depends on the total constellation of the system including the values of the parameters.3 SELF-SUSTAINING LEARNING TRAJECTORIES IN CHILDREN WITH ATTENTION AND EMOTIONAL-BEHAVIORAL DIFFICULTIES What is true for maltreated children is also true for other children who exhibit maladaptive patterns of thinking, feeling, and acting. Both adaptive and maladaptive patterns of psychological activity undergo development and assume different forms over time. Dynamic systems models contribute to understanding the ways adaptive and maladaptive patterns of thinking, feeling, and acting develop 3

V3 - Volume I - 12/10/2015

To enable the reader to experiment with the complex interactions between the parameter values and the resulting effects on the developmental pattern, we shall provide a simple Excel model that can be downloaded at (http://www.paulvangeert.nl/articles _appendices.htm).

Drawing upon the medical model, the concept of ADHD is typically understood as a fixed and trait-like property of individuals. Traditional models maintain that attention difficulties are products of primary deficits in the capacity to sustain attention, inhibit behavior, executive functioning, regulation of motor behavior, and related central deficits (Nigg, 2006). However, from a dynamic systems perspective (see Figure 15.2), all behavior must be analyzed in the context in which it occurs. Behavior—whether adaptive or maladaptive—is not a thing that a person has, but instead is an ongoing process that emerges from interactions between individuals and their contexts. Further, research supporting the idea that ADHD classified behavior is the direct product of central cognitive deficits has been mixed. Research suggests that primary deficits in the capacity to sustain attention, inhibit impulses, regulate motor action, etc., apply only to subsets of children exhibiting ADHD classified behavior (Sonuga-Barke, Wiersema, van der Meere & Roeyers, 2009). Recently, theorists have elaborated alternatives to fixed deficit models of ADHD. For example, delay aversion (Castellanos et al., 2006) and state regulation deficit (van de Meere, 2005) models suggest that both adaptive and maladaptive behavior in children exhibiting ADHD behaviors are context-dependent products of relations between motivational resources and task demands. Among such children, overactivation of motivational resources occurs in the context of challenging tasks, whereas underactivation occurs in less challenging tasks. In structured contexts, underactivation can produce attentional drift and motivate stimulus seeking (e.g., fidgeting, heighted motor movement) accompanied by a preference for immediate rather than delayed forms of gratification, thus evoking impulsive behavior. Overactivation of arousal under high task demands can result in failure to inhibit prepotent behavior patterns (Sonuga-Barke, Wiersema, van der Meere, & Roeyers, 2009). Support for these ideas comes from studies in which children are asked to produce or inhibit simple responses (e.g., pressing or inhibiting a button press) under conditions that vary the speed with which children are required to respond. In comparison to controls, children with ADHD classifications make more errors under both

Page 686

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Self-Sustaining Learning Trajectories in Children with Attention and Emotional-Behavioral Difficulties

high (i.e., short deadline to respond) and low (i.e., long deadline to respond) task demands but not in response to moderate demands (Benikos, Johnstone, & Roodenrys, 2013a & b; van der Meere & Stemerdink, 1999; van der Meere et al., 1995). Such findings suggest that task performance in children who exhibit ADHD behavior arises not simply as a product of general, trait-like deficits, but instead from the dynamic coupling of child and task demands. Children performed as well as controls in tasks that fall within an optimal range of difficulty; beyond that range, performance dissipated beyond that observed in controls. These studies suggest that inattentive, hyperactive and impulsive behavior are not linear results of central deficits but instead emerge through the self-organization motivational, affective and cognitive processes as they adjust to each other in response to local task demands. The Strengths and Limitations of Neurobiological Approaches Students with emotional and behavioral difficulties (EBD) are mainly known for problems involving so-called internalizing and externalizing behavior (Lane, Wehby, & Barton-Arwood, 2005). Many children with emotional and behavioral difficulties also exhibit difficulties in attention (e.g., ADHD). In recent years, however, increasing evidence also points to the negative academic outcomes typical for these students (Lane, Barton-Arwood, Nelson, & Wehby, 2008). Students with EBD continue to display significant academic delays across all placements (Reid, Gonzalez, Nordness, Trout, & Epstein, 2004) and do not seem to improve over time (Lane et al., 2008). These students thus require extra attention to support the development of academic performance. Unfortunately, the academic needs of this population continue to be unmet (Hayling, Cook, Gresham, State, & Kern, 2008). Research assessing the academic needs of EBD children has focused on the neurobiological shortcomings that mediate, among other problems, lags and delays in their academic performances. In addition, research also examines how to ameliorate these shortcomings using pharmacological (and sometimes nonpharmacological) treatments (e.g., Hoekzema et al., 2010). Overall, neurobiological research about ADHD has yielded important insights (Castellanos & Tannock, 2002; de Zeeuw, Weusten, van Dijk, van Belle, & Durson, 2012; Tomasi & Volkow, 2012). Cortese (2012) provided an overview of anatomical, functional, neurophysiological, neurochemical, and genetic processes in the development of ADHD behavior.

12:37pm

687

A primary conclusion consists of the finding that the brains of many children with ADHD differ from those of controls. Further, the processes that mediate ADHD behavior are both complex and heterogeneous, and are unlikely to be accounted for using simple pathophysiological models. As a result, there is a need to develop models that assess patterns of altered connectivity among multiple brain areas in individuals who exhibit ADHD behavior, rather than in terms of the localization of functional deficits in specific brain sites (Cortese, 2012). Despite the importance of research assessing the neurobiologial contributions to attentional and emotionalbehavioral problems, there are limitations to the neurobiological approach. First, until recently, findings from neurobiological research have produced few nonpharmaceutical interventions to address the behavioral and academic performance of ADHD and EBD classified children (Cortese, 2012). Teachers and practitioners face the task of interacting with whole children, and not simply their brains. As a result, there is a need to understand how neurobiological processes operate as but one important layer of functioning within the multinested epigenetic person–environment system. Second, within the neurobiological approach, neuro-biological deficits are often depicted as the primary causes of children’s maladaptive actions. As a result, the complex dynamics that operate between child and context are excluded from theoretical and empirical consideration. This neglect is a problem, as child and social context (i.e., teachers, cultural tools, learning and motivation systems) are inseparable as causal processes in the development of children’s attention and learning (Fabiano et al., 2013; Raver et al., 2008). An exception to this rule involves increasingly rich analysis of gene–environment interactions (G x E interactions) in the development of attentional and socioemotional problems (Nigg, Nilolas, & Burt, 2010; Pennington et al., 2009). A full accounting of the origins and development of attentional and emotional problems must build on to show how specific genes coact epigenetically with multinested environments to produce variation in attentional and behavioral problems. Setting the Stage: The Dynamics of Teaching and Learning Learning processes operate as socially situated, transactional processes (Steenbeek & Van Geert, 2013). To understand the process of teaching and learning involving children with attention and emotional difficulties, it is necessary to examine how learning is fostered or thwarted as a product of coactions that occur over time among aspects of the person–environment system. Such analyses

Page 687

Trim Size: 8.5in x 11in

688

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

require examining the (1) structure of a student’s attention and action as it arises (2) in coregulated exchanges that occur between (3) teachers and students (4) using cultural tools and practices as they operate (5) within particular sociocultural contexts. In this section, we elaborate a framework for understanding how teaching and learning operate as socially situated, transactional processes. The Nature and Importance of Coactive Scaffolding Students learn through their participation in sociocultural activities. Most often, successful learning occurs under the guidance and direction of more competent others who organize instruction, guidance, motivation, and emotion within a child’s zones of optimal learning (Mascolo, 2013; Van Geert, 2008). An important means for teachers to help students to learn is to make use of coactive scaffolding. Scaffolding consists of the process of providing a student with external support and assistance, and optimally, discarding this support after learning has taken place (Granott & Parziale, 2002; Wood, Bruner, & Ross, 1976). Scaffolding is an intrinsically dynamic notion. We use the term to refer to any process that operates outside of the direct control of the individual over a given unit of behavior that functions to raise an individual’s level of function beyond that which he or she is capable of performing him or herself (Mascolo, 2005). This principle is similar to the distinction made above between optimal and functional levels of skill that arise in contexts that provide high and low levels of contextual support. As discussed already, optimal level performance is that which a child can sustain under conditions of high contextual support; a child’s functional level is the level at which he functions in everyday contexts in the absence of such support. However, the process of scaffolding is different from the provision of high support. In contexts involving high support, a child performs a task on his own with contextual support but without instruction or assistance during the actual execution of the task. For example, a teacher may provide high support by first showing a child how to perform an addition problem that is beyond the child’s ability, and then asking the child to perform the addition problem on his own without additional instruction or support. In contrast, when a teacher or adult scaffolds a child’s performance, he or she often performs part of the task for the child (so the child can complete the remainder), or otherwise provides directives and cues to organize the child’s actions as the child is actually carrying out the task. Scaffolding occurs, for example, when a teacher provides ongoing instruction (e.g., Now carry the one) or cues (e.g.,

What’s next?) without which the child would be unable to complete the task himself. The level of functioning that children can sustain under conditions of scaffolding tends to be higher than their optimal level performance under conditions of high support. Social scaffolding functions by breaking down a child’s unmanageable task into more manageable units, thus allowing a child to coordinate novel aspects of a task in vivo—during the process of learning by doing. Scaffolding thus has the effect of holding up novel configurations of a child’s actions so that the child can coordinate those actions for himself even though he could not coordinate them by himself previously. Empirical studies have shown that the process of scaffolding (1) occurs on a widespread scale, (2) is effective in that it advances learning, and (3) has different forms (i.e., that it differs among individuals and contexts; see, e.g., an overview of scaffolding in the context of learning disabilities; Stone, 1998). In special education, the quality of scaffolding can be hampered by the ways behavioral problems interfere with the teacher’s actions (Wehby et al. 2003). This tends to occur because teachers are insufficiently trained to cope with problems in their scaffolding activities. Embodiment, Coaction, and the Initial Conditions of Learning An important feature of our approach is that we conceive of learning as an embodied process. Representational development occurs within the medium of the brain and body and is organized by context sensitive interactions between cognition and affect (Figure 15.2, Point 3). In any given context, the development of higher order representational activities can occur only under conditions of optimal affective engagement (see Yun-Dai, 2012). Affective engagement, in turn, is maintained when representational content remains rich, varied and accessible over the course of the teaching process. Secondly, in our view, the quality of teacher-student interactions cannot simply be perceived as having a mere influence on the ADHD child’s learning process but instead functions as a changing indicator of the types of dynamic influences that shape an ADHD child’s learning—through both desired and undesired pathways. Over time, iterations in teacher-student interactions consolidate into mutually responsive relationships. In this way, interaction dynamics (direct, immediate, short-term processes) then become a proximal measure of many distant factors. While research indicates that brain-related constraints on learning are present in children who exhibit ADHD behavior (Cortese, 2012), from a dynamic systems

Page 688

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Self-Sustaining Learning Trajectories in Children with Attention and Emotional-Behavioral Difficulties

perspective, such constraints should be viewed as the initial conditions for the learning process, and not necessarily fixed, stable, or deterministic factors. In principle, these starting conditions can be changed by designing interactions that focus on the neuropsychological strengths and weaknesses of particular children, and by providing cognitive and affective experience that falls within the optimal range of learning of individual children (Immordino-Yang & Damasio, 2007; Vinogradov, Fisher, & de Villers-Sidani, 2012). Over time, the consolidation of repeated short-term representational and affective experience yields long-term changes in the neural networks of the brain and corresponding changes in the child’s skill repertoire (John, Bullock, Zikopoulos, & Barbas, 2013; Winkielman, Niedenthal, & Oberman, 2009). It is only through analysis of sequences of learning over time that we can understand the circumstances under which neurological and brain-based constraints remain stable, or eventually change for the better or the worse. To trace the developmental consolidation of such processes, it is necessary to have access to densely sampled serial indicators of learning and action that occurs between EBD children and their teachers. Learning as an Interactive, Microdevelopmental Process. From our perspective, learning, which is often defined as a macrolevel variable operating over the long-term—is best studied as a microlevel process. That is, it must be studied during individual instruction sessions that occur within the learning triangle of teacher–material–learner (Figure 15.2). By tracking moment-by-moment teaching and learning over time, it is possible to identify how short-term triangular dynamics are consolidated into long-term learning trajectories. Toward this end, microgenetic methods have proved their value (Basseches & Mascolo, 2010; Siegler, 2006; Van der Aalsvoort, Van Geert, & Steenbeek, 2008) and are now broadly recognized as particularly suitable for studying learning as a transactional, coregulated process (Granott & Parziale, 2002; Mascolo, 2013; Siegler, 1995; Siegler & Stern, 1998). An important mechanism of change in this microgenetic process is iteration, that is, the fact that learning takes place in the form of iterative steps, which can be modeled as a series of successive operations, in which the output of the preceding operation provides the input of the next one (Van Geert, 1991, 1994, 2008). In an iterative sequence, each previous action of the student has an influence on the subsequent action of the teacher, and vice versa, indicating the cyclic character of iterative causality. Teacher–student relationships evolve as successive iterations and tend to

12:37pm

689

consolidate into stable structures (Bierman, 2011; Luckner & Pianta, 2011). The formation of mutually responsive teacher-student relations has a direct impact on student attitudes, which thereupon foster enhanced peer acceptance and positive peer reputations, which in turn promote increased student engagement. Especially at-risk students may have particular benefits from sensitively structured classroom climates (Bierman, 2011; Gazelle, 2006; Hughes & Chen, 2011). When studying learning-in-interaction as a dynamic, complex process, it is fundamental to capture it by means of a variable that reflects the important characteristics of the dynamics of that process. This variable is called an indicator variable, in that it captures some of the most characteristic qualitative properties of the underlying process. Indicator variables are means of representing the dynamics of a system without aiming to approximate a complete or exhaustive description, which, in the case of complex processes often hampers our understanding. Instead of examining the roles of the interaction partners separately, it is important to take the iterative action– reaction responses of both partners together as defining variable (see also Fogel, 1993, 2009). One way to assess such dynamics is by focusing on scaffolding. By way of illustration, we present two studies on scaffolding dynamics in special education, involving individual math instruction with children exhibiting ADHD and related problems. Scaffolding Dynamics During Arithmetic Lessons We examined the dynamics of short-term interaction between teachers and students during arithmetic lessons. In particular, we focused upon the responsiveness of teachers and students over the course of one-to-one math instruction. Within any single initiation-response interaction, we defined responsiveness in terms of whether the responding utterance followed meaningfully from the content of the initiating utterance. Responses that connected meaningfully to initiating utterances (responsive utterances) were classified as matches, and those that were unrelated to initiating statements (unresponsive statements) were identified as mismatches. Matching and mismatching utterances could occur in either direction between teacher and student (i.e., teacher → student, student → teacher). Thus, response matches functioned as the indicator variable in this study. We also identified teacher and student self-iterations. Teacher self-iterations refer to statements made by the teacher subsequent to a self-initiated utterance (e.g., after a short wait, answering

Page 689

Trim Size: 8.5in x 11in

690

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

a question that she had asked the student). We assumed that variations in scaffolding quality would be represented by variations in responsiveness over the course of teaching and learning. As it is difficult for even experienced teachers to provide clear instructions that maximize student self-regulation, we expected that the interaction dynamics between teachers and students would show evidence of disruption and mismatch over time. Participants included four primary school students exhibiting emotional and behavioral problems (mean age was 9.7 years) in a small town in northern Netherlands. The case focuses on one of these students (W) who exhibited behavioral problems as well as problems with attention, concentration, and impulse regulation. The data took the form of videotaped arithmetic lessons delivered biweekly over a 2-year period (32 lessons). Figure 15.9 shows the number of matches, mismatches, student self-iterations, and teacher self-iterations expressed in the scaffolding dynamics of student W and his teacher over a 2-year period. As indicated in the figure, each variable showed substantial fluctuations over time. These patterns do not substantially differ from fluctuations that appear in the data of the other three EBD students. (This finding is based on an overall variability measure and Monte Carlo permutation analysis, p = 0.66.) Focusing on changes in the percentage of matches over time, two basic phases can be distinguished. Matches occurred more frequently in the first half of the observational period than in the remaining half. A change point analysis revealed that lesson 12 marked a discontinuity in the mean number of matches observed. A Monte Carlo permutation test shows that the difference is statistically significant (p < 0.01). In addition, the standard deviation of the matches became significantly reduced after fragment 17 (p = 0.02), marking

a second transition point. Thus, the variable M showed a nonstationary pattern, with discontinuities around fragments 12 and 17. These discontinuities suggest changes in the attractor state of this student–teacher system. Figure 15.10 represents the scaffolding dynamics between W and his teacher as reflected in changes in matches, mismatches, teacher self-iterations, and student negative responses over the 2-year period. Figure 15.10 shows that the drop in matches around Lesson 12 coincides with a gradual increase of mismatches and in the level and variability of the teacher’s self-iterations. Thus, over the long term, relations among matches, mismatches, and teacher and student statements shifted from one attractor pattern to another. The shift is one from responsive, coherent, and contingent interactions prior to Lesson 12 to increasingly less responsive, coherent, and contingent interactions thereafter (see Steenbeek, Jansen, & van Geert, 2012). Figure 15.11 identifies the number of math assignments that the child was successfully completing over the course of time. The figure provides a graphical illustration of the complexity of the relationships among response matches and self-iterations on one hand and the amount of productivity on the other. The patterns demonstrate a decrease in the quality of the scaffolding between child and teacher coincided with progress in math performance, as indicated by the number of assignments completed. Our main question concerned what the short-term scaffolding dynamics would reveal about the different phases identified in the previous section. Figure 15.12 shows the sequences of the variables match, mismatch, teacher self-iteration, and student self-iteration in two prototypical instruction sessions between W and his teacher. The first session (left panel) occurred before the first change

0.80 0.70 Percentage

0.60 0.50 0.40 0.30 0.20 0.10 0.00 1

3 % match (M)

5

7 % mismatch (MM)

9 11 13 Time (analyzed fragments)

15

% teacher self iterations (TSI)

17

19

21

% student self iterations (SSI)

Figure 15.9 Changes in matches, mismatches and self-iterative utterances between a child and tutor in mathematics instruction.

Page 690

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Self-Sustaining Learning Trajectories in Children with Attention and Emotional-Behavioral Difficulties

Teacher self iterations

Response matches 0.5

0.70

0.4

0.60

Proportion

Proportion

691

0.80

0.6

0.3 0.2 0.1

0.50 0.40 0.30

0

0.20 1

8

17 16

4

21 26 32 43 56 63 70 76 18 24 30 35 48 58 68 71

1

8 4

17 16

Fragment number 0.20

21 26 32 43 56 63 70 76 18 24 30 35 48 58 68 71

Fragment number 0.05

Mismatches

0.16

Student negative responses

0.04

Proportion

Proportion

12:37pm

0.12 0.08

0.03 0.02 0.01

0.04 0.00

0.00 1

8 4

17 16

21 26 32 43 56 63 70 76 18 24 30 35 48 58 68 71

Fragment number

1

8 4

17 16

21 26 32 43 56 63 70 76 18 24 30 35 48 58 68 71

Fragment number

Figure 15.10 Independent analysis of growth for matches, mismatches, teacher self-iterations and student self-iterations.

point and represents a typical responsive session, the second session (right panel) occurred after the second change point, and represents a prototypical nonresponsive session. Figure 15.12 shows transition diagrams for both sessions. As indicated in Figure 15.12 (Session 4, upper left), in a prototypical responsive session, most of the self-iterations were produced by the teacher. This pattern was repeated in the prototypical nonresponsive session (Session 19, upper right). As indicated in Session 4 (Lower Left Corner), in the prototypical responsive session, when a match occurred (91% of the sequences), it was almost always followed by another match (93% of the sequences). In addition, the prototypical responsive session contained no mismatches. In contrast, in the proto-typical nonresponsive session (Session 19, lower right), when matches occur (41% of sequences), they were followed by a match in only 41% of the sequences, a mismatch in 41%, and by a neutral match in 15%. Thus, there were more vacillations among match, mismatch, and neutral sequences in the nonresponsive than in the responsive session (p < 0.01). Thus, although the match–mismatch ratio differs substantially across the two lessons, the ratio of teacher self-iteration and student self-iteration stayed the same. This latter observation points to the stable and inactive role of the student during the instruction sessions. These

findings support the idea that the interaction dynamics in EBD student–teacher dyads make it difficult, even for highly trained teachers, to motivate and enhance the use of self-regulatory strategies in EBD students. Arithmetic Learning Among EBD and Typically Developing Students Many studies suggest that EBD students show lags and delays in their academic performances, and that instructional practices can play an important role in reducing these delays. Several types of instructional behavior are effective for EBD students. These include providing clear and explicit instructions, enhancing opportunities to respond to questions, teaching the effective use of self-talk and self-regulation strategies (Kroesbergen & Van Luit, 2003; Sutherland, Alder, & Gunter, 2003). In addition, error correction, consistent feedback, individualized instruction, enhanced levels of instructional talk, and prompts for responding help increase student engagement and learning (Conroy, Sutherland, Snyder, & Marsch, 2008). Steenbeek, Jansen, and van Geert (2012) examined the arithmetic performance of a group of five EBD students in comparison with that of a group of five typically developing students. Student performance on arithmetic

Page 691

Trim Size: 8.5in x 11in

692

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

0.5

response matches

level

0.4 0.3 0.2

mismatches

0.1 0 1

2

3

4

5

6

7

8

9

10 11 12 13 14 15 16 17 18 19 20 21

time (analyzed fragments)

breakpoint 1

student negative responses

1

level

0.8

teacher self-iteration

0.6 0.4 0.2 0 1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20

time (analyzed fragments)

21

breakpoint 2

100

tasks/lesson

80 60 40 20 0 1

2

3

4

5

6

7

8

9

10 11 12 13 14 15 16 17 18 19 20 21

raw data

smoothed data

time (analyzed fragments)

Figure 15.11 Relation between changes in matches, mismatches and self-iterations and the growth of mathematical performance.

assignments ordered with respect to increasing complexity over a 2-year period was assessed. We expected that EBD would exhibit lower levels of performance but higher levels of inter-individual variability than typically developing students. We also expected that typically developing students would make progress at a uniform pace consistent with the class-based nature of instruction. Figure 15.13 shows changes in arithmetic performance for each group over a one-year period. As expected, typically developing students successfully completed more assignments than the EBD students. Further, as indicated in Figure 15.13, the growth curves for EBD children show greater inter-individual variability than those for the typically developing group. Although the growth curves for the typically developing group began

at the same intercept, the pace of growth, as indicated by the slope of the curves, differed from individual to individual. In contrast, the growth curves for the four EBD students differed in terms of their intercepts (i.e., the level of complexity at which students began the instruction) but were similar in terms of their slopes (i.e., learning pace). Similar slopes for EBD students occurred even though these students followed individualized learning courses in the context of individualized scaffolding situations.

Self-Sustaining Trajectories in Suboptimal Learning Dynamics These patterns of results point to the development of a self-sustaining dynamic of suboptimal educational

Page 692

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Self-Sustaining Learning Trajectories in Children with Attention and Emotional-Behavioral Difficulties Fragment 4

47%

12:37pm

693

Fragment 19

53% 36%

64%

53%

5%

47% 67%

7%

93%

95% adult

33%

child

adult

match

child

match mismatch

93%

41%

41%

91%

mismatch 41% 54%

31%

15%

31%

15% 33% 33% 7%

100%

29% neutral

9%

33% neutral

number of sums per lesson (smoothed curve)

Figure 15.12 Interaction dynamics in prototypical responsive and nonresponsive learning sessions with a boy with emotional and behavioral difficulties.

80 60 40 20 0 1

11

21

31

41

51 61 lessons

71

81

91

101

typically developing student

problematic student regular education

EBD student

student W

Figure 15.13 children.

Growth of mathematical performance in EBD

functioning in which teacher’s overinvestment and the student underinvestment support and maintain each other over time. Teacher overinvestment comes in the form of sustained patterns of scaffolding over time and through her willingness to bridge the gap between a series of assignments missed by the student. The student’s underinvestment is indicated by poor effort and a lack of interest and motivation. This pattern is supported by a highly unbalanced surplus of self-iterations produced by the teacher in comparison to the student. The teacher continuously produced long utterances and tended to answer

most of the questions she posed to the student. In contrast, the student maintained the strategy of responding with short answers (e.g., yes, no, or I don’t know). Response matches—a measure of responsivity—had their origins in the teacher’s effort to match the student, and not vice versa, and declined during the second part of the 2-year period. This self-sustaining suboptimal pattern of teacher overinvestment likely has its origins in the teacher’s desire to ensure short-term continuity in instruction, which, in the long term, proves problematic. Teachers know that students are easily distracted and tend to react to frustration in terms of negative emotion. Sensitive to the emotional reactions of their peers, negative emotions, and disorder easily spread through the class. Hence, there was a strong desire of the teacher to orient students toward their work without too much frustration in order to create room for another student to receive individual instruction. Teachers were well aware of this self-sustaining dynamic. In this way, suboptimal teaching and learning processes had their origins in the ways children’s difficulties influenced the classroom dynamics, and not by a lack of quality among the teachers. These suboptimal interactions reflect a dynamic process in which student difficulties reciprocally affect teacher activities in ways that perpetuate and amplify a downward spiral in teaching and learning (Van Geert & Steenbeek, 2005).

Page 693

Trim Size: 8.5in x 11in

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

student 1 teacher 1

1 14 27 40 53 65 79 92 105 118 131 144 157 170 183 195 209 222 235 248 261 274 287 303

learning & teaching level

time in steps (a) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

1 14 27 40 53 65 79 92 105 118 131 144 157 170 183 195 209 222 235 248 261 274 287 303

student 1 teacher 1

time in steps (b) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

student 1 teacher 1

1 14 27 40 53 65 79 92 105 118 131 144 157 170 183 195 209 222 235 248 261 274 287 303

It is possible to describe the performance of EBD students using a dynamic model of learning and teaching called the dynamic scaffolding model (Van Geert & Steenbeek, 2005). The model starts with the assumption that a student’s learning and a teacher’s teaching can be represented by means of the same ruler. For instance, a particular student’s current level in math can be represented on a measurement scale that specifies mathematical ability. The student’s teacher provides instruction and scaffolding in the form of assignments, explanations, questions and so forth. The teacher’s instruction will target a particular level of mathematical ability from the student. The level of teaching can, in principle, be represented on the same developmental ruler, (e.g., a dynamic skill ruler) as the child’s mathematical ability. While the student’s progress strongly depends on the teacher’s instruction, a teacher’s instruction can be effective only if it operates on a level that is slightly higher than the level currently achieved by the student. As the student learns, the teacher should adapt his level of instruction to the new skill level of the student. Learning and teaching are dynamically coupled processes. Thus, changes in learning depend, among other factors, on the level of teaching provided, whereas changes in teaching depend on the amount of learning accrued from preceding teaching activities. For the teacher to adapt, on the one hand, his teaching to the level of the student, and, on the other hand, to the progress made by the student, the teacher must be able to assess the student’s level and progress in real-time. Thus, if the student’s progress is observed as being low, the teacher will hardly, if at all, modify the teaching level. Students differ in terms of the distance between their actual levels of functioning and the levels of teaching under which they optimally learn. In comparison to intellectually advanced students, students with learning difficulties will exhibit smaller distances between their actual levels of functioning and the levels of teaching under which optimal learning occurs. Hence, while teaching, a teacher must not only assess the students’ levels of functioning and their degree of progress, she must also estimate the optimal distance between their levels of functioning and the level of teaching under which students can make optimal progress. The general mathematical properties of this teachinglearning model rely on the logistic change model. In so doing, they also rely on couplings between a student’s level of skill, a student’s progress in learning, and the level and progress in the teacher’s instruction (see van Geert &

Steenbeek, 2005; see also Van Dijk et al., 2013). As far as EBD students are concerned, we may assume EBD students profit more from a relatively small distance between their actual level of functioning and the level of instruction. In the studies described above, we have also seen that the teachers of EBD tend to avoid strong challenges and in general stay rather close to the students’ actual levels of functioning. As a result, student learning remains suboptimal. EBD students are “under-challenged,” as student and teacher get stuck in a mutually influencing process of underachievement and under-challenge.

learning & teaching level

Modeling School Performance in EBD Students

learning & teaching level

694

Cicchetti

time in steps (c)

Figure 15.14 Modeling growth in teaching and learning with children with attentional and emotional problems.

Page 694

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

The Development Socioemotional Adjustment in Children with Emotional and Behavioral Difficulties

The dynamic model predicts that if teachers tend to stay too close to the student’s actual levels of performance, the process will lock itself into a pattern where no real learning progress is made (Figure 15.14a). If the teacher attempts to force learning by presenting challenges above a student’s optimal level, there will be a similar self-sustaining process of suboptimal learning (Figure 15.14c). However, the model also predicts an intermediary window where the level of teaching and challenge provides an opportunity for optimal learning (see Figure 15.14b). The simulations based on the model show that the rate of student learning (which, along with other factors, is a function of the student’s effort and motivation) is also a critical factor. Lower rates of learning destroy the window of possible opportunities and create self-sustaining processes of underachievement that are difficult to change. In short, what the model deductively demonstrates is that among EBD students, learning opportunities are fragile and dependent on the teacher’s capacity for sensitive adaptation. The fragility of the teaching-learning dynamics, which the dynamic model clearly demonstrates, can explain why so many children in special education perform at lower intellectual levels than might be expected on the basis of their actual intellectual abilities.

THE DEVELOPMENT SOCIOEMOTIONAL ADJUSTMENT IN CHILDREN WITH EMOTIONAL AND BEHAVIORAL DIFFICULTIES Thus far, we have examined idiosyncratic pathways in the macro-development of dissociative structures among maltreated individuals as well as microdevelopmental trajectories of teaching and learning involving children with attention difficulties. In this section, we extend our discussion of the dynamics of development to an analysis of developmental pathways taken by different groups of individuals from childhood through adulthood. In particular, we examine the development of socioemotional functioning among children who exhibit emotional and behavioral difficulties. A common way to understand emotional and behavioral-regulation difficulties involves the internalizing–externalizing dichotomy. According to Merrell and Walker (2004): Externalizing problems include behavioral characteristics that are considered to be undercontrolled and other-directed, such as antisocial and aggressive behavior, conduct problems and delinquency, destructive and harmful behavior, and the

12:37pm

695

hyperactive impulsive manifestations of ADHD. Internalizing problems include “overcontrolled” or self directed behavioral and emotional characteristics such as depression, anxiety, social withdrawal, and somatic problems. It is important to recognize that in some cases, a child’s behavioral characteristics can be of a ‘mixed’ variety, manifesting important aspects of the both the internalizing and externalizing broad bands (p. 907).

Despite its prominence in developmental research, the internalizing–externalizing dimension raises several problems. First, the dimension is broad. As a result, there is considerable heterogeneity in the behaviors to which the terms externalizing and internalizing apply. Second, given its breadth, assessments of internalization and externalization tend to involve aggregations across a broad range of behavior collapsed across context, time, and domain of functioning. Third, the practice of representing externalization and internalization as polar opposites implies that the co-occurrence of behaviors named by these terms should be low. However, research, shows dramatic comorbidity between so-called internalizing and externalizing behavior. Most important, the internalizing–externalizing dimension seems to have limited descriptive value. While it is clear how everyday concepts like acting out or aggression can reflect forms of externalizing behavior, it is not clear why anxiety or depression reflect internalization. The externalizing–internalizing contrast has its origins in the distinction between over- and undercontrol (Block & Block, 1980). From this view, problems of aggression, attention regulation and impulsivity reflect an underdeveloped capacity for regulation (undercontrol), whereas anxiety, depression, and related states reflect a tendency toward overcontrol (Block & Block, 1980). While a large literature suggests that many (but not all) externalizing problems are linked to poor behavior regulation (Rubin, Burgess, Dwyer, & Hastings, 2003), relations between overregulation and internalization are less clear. Kagan’s (2003, 2008) work on behavioral inhibition is an exception. Kagan and his colleagues have demonstrated a degree of stability in the emotional development of children who exhibit a bias toward behavioral inhibition. They have also shown that a tendency toward behavioral inhibition is related to the development of social anxiety, alcoholism, and related conditions (Smoller et al., 2003). The concept of behavioral inhibition, however, is different from the concept of overcontrol. Behavioral inhibition reflects an emotional disposition to withdraw from novelty, whereas overcontrol implies a more controlled and self-directed process.

Page 695

Trim Size: 8.5in x 11in

696

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

The Dynamics of Adaptive and Maladaptive Socioemotional Adjustment Instead of representing maladaptive emotional processes along a single externalizing-internalizing dimension, Figure 15.15 depicts a multidimensional space-state representation of socioemotional adjustment in motiveinconsistent contexts. The horizontal dimension reflects variation in the direction of emotional action from moving toward to moving away from motive-inconsistent events. The vertical dimension reflects degree of activation, ranging from activation to immobilization. In any given motive-inconsistent situation, a person confronts a series of affective choices—alternative modes of emotional responsivity within or between situations (de Rivera, 1991). Figure 15.15 shows affective channels that avail themselves in motive-inconsistent situations. These include (1) anxiety under uncertainty; (2) moving against others in response to moral violations (e.g., anger, rage); (3) moving away from others under conditions of danger or powerlessness (e.g., fleeing in fear; hiding in shame); and (4) immobilization under conditions of hopelessness or helplessness (e.g., depression; freezing in fear). Within and between contexts, emotional functioning fluctuates dynamically throughout this state-space. Within this state-space, some affective movements may be more likely than others. For example, theorists have long postulated dynamic relations between shame and anger (Hejdenberg & Andrews, 2011; Scheff & Retzinger, 1991). Research suggests that angry aggression (moving against) is often motivated by the invocation of shame-related challenges to identity (e.g., insults, criticism; Ferguson, Eyre, & Ashbaker, 2000; Scheff, 2010; Shanahan, Jones, ACTIVATION Anxiety

uncertainty

MOVE AWAY

Shame (Hide) Fear (Flight)

inability to change object remove self

MOTIVE INCONSISTENCY

ought violation change object

inability to effect change

MOVING AGAINST

Anger, Aggression Fear (Fight)

IMMOBILIZATION/ COLLAPSE OF WILL Depression Fear (Freeze)

Figure 15.15 The dynamics of socioemotional adjustment for negative emotions.

& Thomas-Peter, 2011; Tangney, Wagner, Fletcher, & Gramzow, 1992; Wright, Gudjonsson, & Young, 2008). Conversely, children who exhibit reactive aggression (moving against others) are more likely to experience peer rejection (Morrow, Hubbard, McAuliffe, Rubin, & Dearing 2006) and shame-related feelings that result from rejection (Ji, Wei, Chen, & Zhang, 2012). Movements between anger and shame can thus foment a mutually sustaining anger ←→ shame dynamic. Several points arise from this example. First, emotional life is rarely organized around single emotional states. As a result, it is not helpful to think of socioemotional development as organized around any single class of emotions or around any single pole of a bipolar dimension. Instead, emotional life involves dynamic movements between and among different emotional states throughout the state space. Second, some emotional dynamics may be more probable than others (e.g., shame ←→ anger vs. fear ←→ shame). Third, some emotional dynamics (e.g., shame ←→ anger) may involve movement between opposite poles within the multi-dimensional state space. As a result, thinking of emotional life in terms of dynamic movements within a state space provides a framework for understanding how emotional functions self-organize into heterogeneous, fluctuating, and comorbid patterns over time.

Pathways in the Development of Adaptive and Maladaptive Patterns of Socioemotional Adjustment Patterns of socioemotional adjustment are epigenetic products of the mutual interplay among (not always obvious) nested processes that orient development in particular directions (Gottlieb, Lickliter, & Wabash, 2005; Mascolo, 2013). Given a set of initial conditions (e.g., a child’s genotype and temperament), variations in local conditions can move development along a variety of different pathways. Development, however, is an iterative process; as development occurs, developmental changes at one moment create the conditions that canalize movement in the next moment (Cox, Mills-Koonce, Propper, & Gariepy, 2010; Witherington, 2011; Van Geert, 1994). As conditions of development organize into stable, mutually strengthening patterns, developmental changes can cascade in particular directions. As development advances, deviations from particular pathways are possible, but tend to require more systemic changes in developmental conditions. Such changes can often be initiated by modifying the functioning of key developmental variables—sometimes called control parameters (Fogel & Thelen, 1987)—which then

Page 696

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

The Development Socioemotional Adjustment in Children with Emotional and Behavioral Difficulties

12:37pm

697

socioemotional outcomes identified at the top of the figure. Within the figure, statements in shaded boxes refer to the influences of socialization agents (caregivers and peers); unshaded statements refer to developmental steps along a socioemotional trajectory. There are multiple pathways to the development of adaptive, prosocial, behavioral outcomes. Three socially adaptive pathways are identified in Figure 15.16. The empathic–prosocial pathway (1) involves development toward adaptive, empathy-based prosocial action. Prosocial outcomes are most probable among children who (1) exhibit an emotional bias toward positive affect and

iterate and ripple over time, creating novel opportunities for development in alternative directions. Figure 15.16 displays a developmental web describing pathways in the development of adaptive and maladaptive patterns of socioemotional adjustment. The figure identifies patterns of converging, diverging, and overlapping trajectories in patterns of socioemotional adjustment within groups of individuals. The bottom portion identifies the temperamental starting points of development in terms of patterns of affectivity and attentional control (Evans & Rothbart, 2009). Each pathway is defined in terms of a series of movements toward or away from the varied

(9) Inhibited Victim-Aggressor

SOCIOEMOTIONAL OUTCOMES

DISINHIBITION

ADAPTIVE ACTION

Conduct Problems

Prosocial

SelfEnhancing RejectionAggression Hostility

ReactiveProactive/Callous Deviant Peer Groups

Systematic Exploitation of Weakness

Property Violations

(6a) Conduct Problems without Aggression

Indirect Aggression

Prosocial Identity

Anxiety & Depression

Depression Anxiety

Peer Acceptance

Verbal Proactive Reactive

Verbal Proactive

(6) Severe Conduct Problems

Verbal Reactive

Efficacy of Aggression

Academic Difficulty

SelfRegulation

Rejection

ATTENTIONAL CONTROL Callous/ Lack of AFFECTIVE DISPOSITION Empathy

(5) (4) Proactive Reactive Aggression Aggression

MED

LOW

Negative

(3a)

Authoritative Parenting Firm Regulation

Noncompliance

HIGH

Overcontrolling, Intrusive, Distant Parenting

Compliance Concern for Others

(2) (1) Irritable/ Prosocial/ Adaptive Adaptive

Negative/ Irritable Daring

Peer Rejection/ Victimization

Mutual Responsivity

Rejection Harsh Discipline

Frustration Provocation

(8a) Sensitive Regulation

Gentle Redirection

Sensitivity

Physical Callous/ Permissive or Harsh Discipline Aggression NonAggressive

PATHWAY

Anxiety & Depression

Peer Rejection

Academic Difficulty

SEQUENCES & PROCESSES

Anxiety Anxious Behavioral Inhibition Social Anxiety Adaptive Depression

Empathy-Based Prosocial Behavior

(7) Adolescent Onset Rebellion

Status Violations

INHIBITION

HIGH Positive/ Empathic

(3) Inhibited/ Adaptive

(8) Inhibited/ Maladaptive

LOW HIGH Inhibition/ Fear

Figure 15.16 Pathways Toward Adaptive and Maladaptive Socioemotional Adjustment. Phrases along individual trajectories indicate order of developing skills or deficits; gray boxes indicate social effects on developmental pathways; dashed lines indicate co-morbidity of anxiety and depression.

Page 697

Trim Size: 8.5in x 11in

698

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

concern for others; (2) exert high dispositional levels of effortful control over action and emotion (Eisenberg, Hofer, & Vaughan, 2007); and (3) are cared for by sensitive and authoritative parents (Kochanska & Akzan, 2006; Kochanska & Kim, 2013). Such infants and their caregivers mutually influence each other’s actions over time, and quickly self-organize into mutually responsive and emotionally attuned interactions (Kochanska & Aksan, 2004). Children who develop along the adaptive–prosocial path tend to demonstrate affective concern for others from an early age (Roth-Hanania, Davida, & Zahn-Waxler, 2011) and begin to regulate their actions in terms of parental rules during the second year of life (Kochanska & Aksan, 2006). The form of intersubjectivity established within mutually responsive relationships mediates the joint production of positive affective affect (Bornstein, 2013), the development of emotional-regulation skills (Ursache, Blair, Stifter, & Voegtline, 2013), a concern for others (Roth-Hanania, Davida, & Zahn-Waxler, 2011), and the appropriation of sociomoral standards to regulate social conduct (Kochanska & Kim, 2013). The development of prosocial action in children both fosters and is reciprocally facilitated by peer acceptance and the formation of increasingly stable peer relationships (Clark & Ladd, 2000). Figure 15.16 identifies two additional pathways toward adaptive prosocial behavior. These occur among children who exhibit emotional biases toward negative or irritable (Pathway 2, irritable–adaptive) and inhibited (Pathway 3, inhibited–adaptive) affect. These children are protected from maladaptive development by virtue of their capacity for high levels of behavioral and emotional control, coupled with parenting styles that are sensitive to children’s emotional needs. The prosocial development of children with irritable temperaments is facilitated by the use of firm but affectively positive behavioral regulation. In contrast, because less arousal is needed to motivate inhibited children, they benefit from gentle redirection in regulatory contexts (Kochanska & Aksan, 2006). Pathways in the Development of Aggression and Conduct Problems Figure 15.16 identifies four pathways in the development of conduct problems. Conduct problems encompass a broad range of rule violating behavior, including noncompliance, aggression, property violations (e.g., stealing, vandalism), status violations (e.g., running away, truancy, substance abuse), and delinquency (e.g., arrest; incarceration; Barry, Golmaryami, Rivera-Hudson, & Frick, 2013; Wakschlag, Tolan, & Leventhal, 2010). Aggression is among the more important and widely studied conduct

problems. Figure 15.16 identifies four pathways in the development of problems involving physical aggression. Researchers have differentiated between adolescent-onset and child-onset pathways in the development of aggression (Barry et al., 2013; McMahon & Frick, 2005). Figure 15.16 identifies three pathways in the development of child-onset aggression: reactive aggression (Pathway 4), proactive aggression (Pathway 5), and severe conduct problems and aggression (Pathway 6). Adolescent-onset aggression (Pathway 7) tends to be limited to adolescence and arises as the product of comparatively few risk factors. It tends to arise as an amplification of the normative process of autonomy seeking (Moffitt, 2003). Reactive aggression refers to defensive or retaliatory attempts to harm others in response to real or perceived provocation; proactive aggression involves deliberate, nonemotional, and unprovoked attempts to harm others for the purpose of instrumental gain (Kempes, Matthys, de Vries, & van Engeland, 2005). Research assessing the reactive-proactive dimension has produced considerable debate. On the one hand, research shows that reactive and proactive aggression have different origins and developmental courses (Hubbard, McAuliffe, Morrow & Romano, 2010). Reactive aggression is associated with irritable/angry infant temperament (Vitaro, Barker, Boivin, Brendgen, & Tremblay, 2006), lower dispositional capacities for effortful and attentional control (Connor, Chartier, Preen, & Kaplan, 2010), and poor impulse regulation (Connor, Chartier, Preen, & Kaplan, 2010). Children and adults who exhibit reactive aggression are more likely to develop a hostile attribution bias—a tendency to interpret ambiguous emotion in others as indicating hostile intent (Bailey & Ostrov, 2008)—and heighted sensitivity to rejection (Jacobs & Harper, 2013). Such biases may arise from a history of reciprocal provocation by parents and peers, maltreatment or other forms of threat (Dodge, Lochman, Harnish, Bates, & Pettit, 1997). These biases also tend to feed forward into later romantic relationships (Canyas, Downey, Berenson, Ayduk, & Kang, 2010) and play a role in dating violence (Brendgen, Vitaro, Tremblay, & Lavoie, 2001). In contrast, as a form of deliberate provocation, proactive anger is less likely to be accompanied by anger (Hubbard et al., 2002). It tends to be associated with lower basal physiological arousal and may be associated with high levels of dispositional sensation seeking (Wilson & Scarpa, 2011). Children who display proactive anger tend to hold beliefs about the social legitimacy of aggression and its efficacy as a means to instrumental ends (Arsenio, Adams & Gold, 2009). Teens who exhibit high levels of proactive aggression show higher levels of delinquency

Page 698

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

The Development Socioemotional Adjustment in Children with Emotional and Behavioral Difficulties

(Brendgen, Vitaro, Tremblay, & Lavoie, 2001) and lower levels of psychosocial adjustment than those who exhibit high levels of reactive aggression (Card & Little, 2006). On the other hand, researchers have noted that reactive and proactive aggression tend to co-occur (Vitaro & Brendgen, 2005) with correlations that range between .40 and .70 (Poulin & Boivin, 2000). In an important study, Barker, Tremblay, Nagin, Vitaro, and Lacourse (2006) identified three trajectories of aggressive behavior between the ages of 13 and 17. The vast majority of teens (.93) were classified as either low aggressors or moderate aggressors. A small minority (.06) exhibited high levels of aggression that peaked at age 15. However, regardless of group membership, children who aggressed exhibited similar levels of both proactive and reactive aggression. Given the massive co-occurrence of reactive and proactive aggression, some theorists have suggested abandoning the distinction (Bushman & Anderson, 2001). Recent research has produced findings that may help account for these seemingly contradictory findings. Researchers have demonstrated that early-onset aggression is often associated with the presence of callous and unemotional affective dispositions in children (Frick & Viding, 2009; McMahon & Frick, 2005). Callous and unemotional behavior includes a lack of empathy and concern for others, an absence of guilt upon wrongdoing, and the intentional manipulation of others for personal gain (Frick & Viding, 2009; Muñoz, Qualter, Padgett, 2011). Children who show early-onset aggression tend to score high on measures of callous and unemotional social engagement (Dandreaux & Frick, 2009). Frick and White (2008) reviewed research showing that children exhibiting callous emotionality are more likely to show fearless and thrill-seeking behaviors, deficits in processing emotional information, less sensitivity to punishment cues, and positive beliefs about the use of aggression. Children who exhibit callous and unemotional dispositions are more likely to exhibit conduct problems, bullying (Fanti & Kimonis, 2012) and delinquent behavior (Edens, Campbell, & Weir, 2007) that carries forward into adulthood (Frick & Viding, 2009). Aggression and conduct problems are not the results of single causes or risk factors. Risk and protective processes combine in complex ways in the probabalisitc self-organization of developmental trajectories. Fanti, Frick, and Georgio (2009) found that adolescents showing higher levels of callous and unemotional dispositions were more likely to exhibit combined proactive and reactive aggression and not pure forms of either proactive or reactive aggression (see also Thorton, Frick,

12:37pm

699

Crapazono, & Terranova, 2013). Similarly, in a large sample of Greek-Cypriot adolescents, Fanti (2013) found that teens who exhibited both high levels of conduct problems and high levels of callousness exhibited higher levels of bullying, substance abuse; inattentiveness, impulsivity, and narcissism, as well as less robust family support than teens rated as high in either conduct problems or callousness. Together, these findings support the proposition for the three child-onset pathways identified in Figure 15.16. While research suggests two distinct but overlapping pathways involving reactive (4) and proactive (5) aggression, children who exhibit these forms of aggression in the absence of callous and unemotional dispositions are less likely to be involved in severe conduct problems. Thus, severe antisocial outcomes (6) involve the co-occurrence of callous and unemotional affective disposition with a combination of both proactive and reactive aggression. As indicated in Figure 15.16, in addition primary pathways toward conduct problems, research suggests an additional path toward conduct problems without aggression (Burt, 2009; Eley, Lichtenstein, & Moffitt, 2003). Pathways in the Development of Inhibition and Social Anxiety Inhibited patterns of socioemotional adjustment are those that are organized around avoidance as a style of affective coping (Rapee, Schniering, & Hudson, 2009). Such individuals tend to approach events with a disposition toward fearful or anxious affect, show avoidance toward unfamiliar people and events, and exhibit heightened vigilance in novel contexts (Lahat, Hong, & Fox, 2011). Research suggests that children exhibiting fearful/inhibited biases early in life are more likely to develop problems related to anxiety in adolescence and adulthood (Lahat, Hong, & Fox, 2011; Muris, Brakel, Arntz, & Schouten, 2011; Schwartz, Snidman, & Kagan, 1999; Sportel, Nauta, Hullu, Jong, & Hartman, 2011). For example, Prior, Smart, Sanson, and Oberklaid (2000) found that 42% of children who were rated as shy across multiple observations in childhood exhibited anxiety problems in adolescence. In contrast, only 11% of the children not previously classified as shy exhibited anxiety problems in their large sample. Thus, while not all children classified as inhibited develop anxiety problems later in development, a sizable proportion do. Figure 15.16 identifies four pathways in the development of inhibited patterns of socioemotional adjustment. Severe social anxiety problems develop through the inhibited/maladaptive pathway (8). Children who pass through

Page 699

Trim Size: 8.5in x 11in

700

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

this pathway tend to experience fearful/inhibited temperament and a diminished capacity for attentional control (Sportel et al., 2011) coupled with parenting styles that can be classified as overprotective, overcontrolling, or intrusive or distant (Lahat et al., 2011; Lindhout, Markus, Hoogendijk, & Boer, 2009). Such parenting styles are unlikely to operate as static properties of individual parents. Instead, it is most likely such parenting styles are emergent, coregulated products of a parent’s attempt to respond to an inhibited child (Evans & Porter, 2009, 2012). While over-protective parenting fails to promote strategies regulating emotional arousal in the context of novel events, over-controlling or intrusive parenting is likely to amplify inhibited emotion and restrict a child’s capacity to master novel situations. Emotionally distant parents may foster the development of anxious attachment relationships, which increase the probability of anxiety problems in children (Warren, Huston, Egeland, & Sroufe, 1997). Anxiety problems in children who display inhibited temperaments and diminished capacities for attentional control can be mitigated by parents who introduce children to novel events with persistence and sensitivity (Pathway 8a). Conversely, although the capacity for high levels of attention control acts as a protective factor in the development of anxiety problems in children with inhibited temperament (Lahat et al., 2011; Sportel et al., 2011) insensitive or intrusive parenting can channel development away from a prosocial path and in the direction of social anxiety (Pathway 3a). While inhibited temperament biases socioemotional development toward the direction of social anxiety, it is important not to view inhibited temperament simply as a risk factor for maladaptive behavior. Shyness and behavioral inhibition is often viewed negatively, at least among Westerners (Aho, 2011). The work of Kochanksa, Aksan, and colleagues suggest that behavioral inhibition may serve important developmental functions. For example, children with inhibited temperament do not require strong arousal induction in moral socialization (Kochanska & Aksan, 2006). In an insightful study, Aksan and Kochanska (2004) showed that children who exhibited inhibited temperament and inhibition to novelty in infancy (9, 14, and 22 months) showed an increased capacity for voluntary inhibitive control at 45 months). Aksan and Kochanska (2004) suggested that the involuntary reactive inhibition (inhibited temperament) functions to decreases the speed of approach to novel situations early in life. This decrease, in turn, provides the opportunity for additional attentional focus on novel events, thus facilitating the development of voluntary effortful control. These findings

illustrate the ways nonobvious relations between developing systems promote psychological development. In the case of Aksan and Kochanska’s (2004) findings, involuntary relations between inhibition and novelty earlier in development operate iteratively to set up conditions for the development of the capacity for voluntary control. The resulting developing of voluntary control then buffers against developmental movement in the direction of social anxiety. As is the case with the development of aggression, the development of social anxiety is also accompanied by experiences of depression (Epkins & Heckler, 2011; Sportel et al., 2013). Peer rejection, victimization, social isolation, and loneliness play important and cumulative roles in the development of depression and anxiety in children, adolescents, and adults (Dempsey & Storch, 2008; Epkins & Heckler, 2011). Briggs, Nelson, and Sampino (2010) reported evidence that depression among adolescents who experience social anxiety was mediated over time by experiences of peer rejection and victimization. Silk, Davis, McMakin, Daahl, and Forbes (2012) reviewed evidence suggesting that the codevelopment of anxiety and depression in adolescence is mediated by a heightened sensitivity to self-evaluative social threat (fear of negative evaluations of the self by others) coupled with modifications in the proclivity to seek out wanted outcomes. Research indicates, for example, that anxious adolescents show a heightened attentional bias to social threat, and are more likely to interpret socially ambiguous situations in terms of threats to one’s self-evaluation (Silk et al., 2012). These same processes are also implicated in the development of depression. Silk et al. (2012) suggest that in an attempt to adapt to social threat, adolescents revise their expectations for social acceptance, ultimately resulting in increasing depression. These findings suggest that experiences of rejection, exclusion, loneliness, and related social threats to self-evaluation play cumulative roles as risk factors in the codevelopment of anxiety and depression (Epkins & Heckler, 2011). The codevelopment of anxiety and depression as a product of peer exclusion and rejection is represented in terms of the double lines associated with the inhibited pathways depicted in Figure 15.16. For some children, the anxiety-rejection link not only results in depression, but also evokes aggressive reactive and proactive attempts to lash out against rejection. The inhibited-victor aggressor pathway (Pathway 9) occurs in children with inhibited temperaments who are recipients of harsh discipline or consistent patterns of humiliation by peers. Research shows that in some circumstances (e.g., when children are humiliated or raised in violent homes) temperamentally

Page 700

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Beyond Best Practices and Evidence-Based Intervention: Translating Dynamic Systems into Prevention and Practice

inhibited children become highly aggressive in later childhood and adulthood (Watson, Fischer, Andreas, & Smith, 2004). Trajectories of Development: Making a Path by Walking Developmental pathways are the outcomes of constructive processes. Pathways should not be viewed as inherent, preformed, or predestined roads that anyone with a particular precondition must take. Unlike highways that have been planned and pre-structured by a central authority, developmental paths are more like trails that that are formed through the active process of walking (Thompson, 2007). A developmental pathway emerges slowly as if one were walking through a thicket. As each step is taken, part of the path is formed. Each step provides the conditions that orient the next step. Each movement is constrained and channelized by the obstacles and clearings that one encounters along the way. Obstacles, however, can turn into clearings with the thrash of a sickle, while clearings can invite movement away from desired destinations. Thus, pathway representations of development are like prototypical images that researchers have imposed upon an enormous variety of individual trajectories. In this way, the notion of pathway provides a particular way of summarizing large numbers of individual, self-organized trajectories that show sufficient similarity to be treated as members of common idealized class. BEYOND BEST PRACTICES AND EVIDENCE-BASED INTERVENTION: TRANSLATING DYNAMIC SYSTEMS INTO PREVENTION AND PRACTICE How can dynamic systems approach inform the goals of preventing and redirecting maladaptive behavior? Perhaps the most important—and most difficult—task is a conceptual one—namely the task of helping practitioners and policy makers think systemically about the origins and treatment of maladaptive behavior. Many current approaches to intervention and prevention continue to adopt a static, discrete state approach to the problem of maladaptive behavior. From this view, maladaptive behavior is understood as a reflection of more or less fixed pathologies located within individual persons; interventions (e.g., medication, therapy, skill development) are focused on effecting change within the individual child. With exceptions (e.g., family systems therapy; coordinated intervention teams), interventions tend to be focused on the individual child viewed as separate from the contexts

12:37pm

701

in which she acts. However, treatments (e.g., drugs, psychotherapy) are often administered in an unsystematic (or loosely connected) fashion, with different treatments administered and assessed independent of one another. Still further, therapeutic interventions—psychotherapy, social skills training and so forth—tend to occur in contexts that are isolated from everyday life (e.g., the four walls of a therapist’s office). If one thinks of psychological maladaption as a kind of pathology that resides within individuals, it makes sense to focus interventions on processes that occur within persons. However, if one understands patterns of maladaptive behavior as emergent products of individuals in social contexts, this argument begins to lose its force. In the past decades, the quest for evidence-based interventions and prevention programs based on best practices has occupied the attention of policy makers, practitioners and clinically oriented researchers. Evidence-based interventions are those that can cite scientific support for their effectiveness (Gullotta & Blau, 2008). The call for evidence-based therapies had its origins in dissatisfaction with therapeutic approaches that had long been untested, that had been based primarily on anecdotal testing, or that were judged impervious to testing. Interest in evidence-based therapies is not simply a product of scientific motives; insurance companies make evidence of the effectiveness of an intervention a requisite for the compensation of practitioners; federal agencies make funding of mental health research and programs contingent upon empirical demonstrations of therapeutic effectiveness. Evidence-based research has demonstrated the efficacy of different modes of therapy for various broad classes of maladaptive behavior. The requirement that practitioners garner empirical evidence supporting the effectiveness of psychological interventions has positive value. However, calls for evidence-based prevention and treatment tend to be based on the assumption that certain intervention and prevention strategies for particular classes of problems have universal application across contexts. That is, they tend to assume that (1) patterns of maladaptive behavior reflect more or less fixed types of problems (i.e., pathologies) that (2) lend themselves to standardized, generalized and even manualized modes of treatment. They also assume (3) that change is the result of a causal force—a property of the treatment as such, exerting an influence on every individual client similar to that causal force—plus random individual influences—that may increase or decrease the observed effect. This view reflects a continued reliance on models of science developed in the early twentieth century that seek to identify universal laws and relationships that transcend the context and content of

Page 701

Trim Size: 8.5in x 11in

702

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

target phenomena. Prevention and intervention programs that fail to acknowledge the dynamic, heterogeneous and nonlinear origins of maladaptive behavior are unlikely to have high levels of effectiveness. Further, they may mistakenly identify group-based average effects with real effects on individuals, without taking into account that some interventions might also have a causal negative effect on some clients who are compensated in the group averages by positive effects in other clients. Dynamic systems approaches seek to move beyond the limits imposed by linear ways of thinking about prevention, intervention, and assessment. Dynamic systems approaches maintain that effective prevention and intervention must be organized around a foundational appreciation of the nature of complexity (Gorman et al., 2006; Schensul, 2009). One cannot seek to understand the contextualized dynamics of maladaptive activity using principles and methodologies designed to assess relations among decontextualized variables. Systems thinking requires nothing less than a shift in mind-set: • The unit of prevention and intervention should not be the individual child separate from its social context; instead, it should focus on the multiply nested child–environment system as a whole (Antonishak & Reppucci, 2008). Maladaptive patterns of action are not fixed forms or linear expressions of inner pathologies; instead, they are emergent products of individuals (with their dispositions and histories) operating within a complex person–environment system. • A complex system is not defined as one that is composed of a simple set of multiple processes acting in isolation; instead, a complex system is one in which component processes mutually influence each other over time. If this is so, then prevention and intervention programs organized around discrete sets of uncoordinated interventions that are insensitive to local context are unlikely to be effective. Instead, prevention and intervention should focus on seeking transformations in the system as a whole. This can be accomplished by exerting control over central variables (sometimes called control variables) that, through their coactions with other variables in the system, modify the functioning of the whole. • Complex systems are not steady states. They are sensitive to the initial conditions of their component processes (Kelso, 1995; Thelen & Smith, 2006). This means that minor variations in the initial conditions of two different complex systems can, over time, produce dramatic variations in the operations of those systems. If follows that efforts at prevention and intervention that are based on generalized or standardized procedures that are insensitive to the dynamics of local systems are unlikely to be effective. Instead, interventions must

be organized around particular dynamics of local child–environment systems. Further, because small changes in local variables within a system can modify system functioning, changes introduced by prevention and intervention programs are often unforeseen and unpredictable. As a result, prevention and intervention programs must be sensitive to concrete changes in target systems as they evolve dynamically over time. Instead of seeking to produce standardized remedies intended to have broad application across diverse contexts, dynamic systems approaches maintain that the tasks of prevention and intervention become ones of seeking to exert control over the dynamics of local complex systems. In the language of dynamic systems theory, the goal of prevention and intervention would be to produce phase shifts in functioning of the system—namely, a series of cascading or rippling effects that transform relations among variables throughout the child–environment system. To illustrate, consider the simple child–environment system shown in Figure 15.17. This system describes the structure of a particular social interaction (Tseng & Seidman, 2007) between a child and a staff member in a group home for children who exhibit emotional and behavioral problems. At this school, staff members are trained in the use of time-out techniques for managing rule violations among children (8–11-year-olds) in the unit. A product of behaviorist thinking, the time-out technique was initially proposed as a way to provide a child with time out from the reinforcing contingencies assumed to be operative during the production of rule breaking behavior (Bostow & Bailey, 1969). Many practitioners, educators, and parents, however, misuse the time-out technique as a form of punishment (Kemp, 1996). Instead of operating as a break from reinforcing contingencies, the technique is used as an unpleasant consequence for a child’s misdeed. As a result, time-out procedure often has effects

(d) Economic Resources

(c) Social Resources

(b) Norms & Rules

STAFF (e) Physical, Social & Temporal Organization

2. TIME OUT

staff-child (f) relationship

3. OPPOSITION

4.

1.

Child-Child Conflict

CHILD Escalation

5.

Restrain Child

Dispositions/ History (a)

Figure 15.17 Dynamics of staff–child interaction in a group home.

Page 702

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Beyond Best Practices and Evidence-Based Intervention: Translating Dynamic Systems into Prevention and Practice

that are opposite to their intended ones. For example, in an interview study with children enrolled in a special education program, Visser, Singer, van Geert, and Kunnen (2009) found that some children experience a time-out as a moment in which they can better reflect on and prepare a negative reaction to the behavior of the other child. In the interaction depicted in Figure 15.17, two boys were engaged in a conflict while watching television together (Point 1). As a result, (2) a staff member ordered one of the boys to a time-out. The boy responded by (3) mouthing off to the staff member, who then (4) increased the amount of time that the boy was asked to spend in time-out. Over time, the boy’s oppositional behavior escalated out of control, and (5) the boy was physically restrained. In this example, the child’s unwanted behavior emerged as a product of coactive processes that operated within a complex child–environment system. The components of this simple system affect and constrain each other over time. In Figure 15.17, the shaded area indicates coregulated activity between a staff member and a child over time. The nonshaded elements indicate processes and resources that constrain child–staff interaction. The most obvious example of the mutual interplay among system components involves the positive feedback cycle that arose between the staff member and the child over the course of the encounter. In this interaction, the child’s unwanted behavioral escalation was the emergent product of coregulated processes that occurred between the child and the staff member. While this was the most proximal source of the child’s behavioral escalation, other elements of the system played an important if less obvious role. The (1) rules and norms of the group home stipulated that staff were to use the time-out technique to regulate misbehavior. Interactions between students and staff were constrained by the (2) social and (3) economic resources available to the group home—namely limits on the availability of funds and the capacity to recruit and retain qualified staff members. Further, the (4) physical (i.e., close quarters), social (i.e., the particular children involved in the conflict; the relationship between the target child and the staff member) and temporal (i.e., the placement of quiet TV time after homework and before bedtime) place constraints on the ways children can respond to each other and to the staff member. However, changes in one or more core components of this child–environment system would likely produce dramatic changes in the capacity of staff to regulate the behavior of the children who lived there. The most obvious source of transformation would involve implementing appropriate uses of the time-out procedure. This would involve training staff to use time-outs as an opportunity to break behavior sustaining contingencies and to provide an opportunity for reflection and problem-solving, and also to talk with the children about how they experience

12:37pm

703

the meaning of time-out, what it does to them emotionally, physically and rationally (Visser, Singer, van Geert, & Kunnen 2009). Success in the use of the time-out procedure—especially if coupled with sensitive emotional regulation and scaffolded behavioral redirection—would likely create a cascade of reciprocally influencing transformations. These could include the emergence of a shared ethos of improved intervention strategies among staff; increased capacity for behavioral and emotional regulation among children; increased cooperativeness between and among staff and children; and improved staff–client relationships. Over time, such subtle changes would settle into a stable and organized system of functioning. Participation in the system would have the effect of organizing children’s capacity for effective social engagement and emotional regulation within the context of the unit. It is also possible that similar results could be achieved by seeking to influence less central components of the child–environment system. For example, it is possible that misuse of time-outs in this particular organization was a product of difficulty in recruiting and retaining staff with suitable levels of experience and education. It is possible that the misuse of the time-out procedure arises from difficulties in training less qualified staff, or in retaining qualified or trained staff. If this were the case, it might be possible to transform the child–environment system by enhancing economic resources and using those resources to improve the quality of the social resources available to the organization, perhaps by recruiting and retaining more highly qualified staff who would require less intensive training in basic procedures. It is also possible that the aggressive behavior within a group of children prevails as a product of reciprocal imitation or retaliation that originates its in the aggressive activity of a minority of students. Simply changing the context often improves the behavior quite substantially (Visser, Kunnen, & Van Geert 2010). We propose a series of general principles that can guide efforts at prevention and intervention at multiple levels of functioning. Efforts at prevention and intervention can be organized around several loosely organized steps: 1. Understand the dynamics of the local nonlinear systems in question. Efforts at prevention and intervention must begin with contextualized understandings of the nature of the local systems of psychological adaptation or maladaptation under consideration (Gorman et al., 2006; Schensul, 2009). This may require access to local data related to the functioning of particular child–environment systems of interest. Efforts at prevention and intervention can be focused on systems that operate at multiple levels. These may include the level of the individual child at home or at school; of multiple children interacting with each other and

Page 703

Trim Size: 8.5in x 11in

704

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

with adults within organizations; or groups of children functioning within larger communities (Tricket & Schensul, 2009; Visser, Kunnen, & Van Geert 2010; Yoshikawa & Hsueh, 2001). Success in assessing the nature of nonlinear systems is facilitated by cooperation among interdisciplinary teams of practitioners who can coordinate their expertise relative to understanding a problem (Aiken & Hanges, 2012). For example, to assess a problem of aggression within a school system, it is important to bring together teams that can assess a broad range of variables related to the problem of aggression. These may include identifying (1) forms of observed aggression and the contexts in which they occur; (2) students at risk of being perpetrators or targets of aggression; (3) student culture and student-student relationships; (4) teacher–student interaction that arise during aggressive encounters; (5) demographics and values and of the families; (6) school culture; and (7) school–home–community relations. Varieties of methods are available to collect and analyze data for conducting such assessments, from dynamic mathematical modeling to ethnographic observation (Aiken & Hanges, 2012; Schiepek, 2003; Steenbeek & van Geert, 2013; Tricket & Schensul, 2009). 2. Identify core control variables. Having gained an understanding of the dynamics of the local system of interest, identify major control variables whose manipulation can have effects on the operation of other components in the system. For example, studies suggest that parental involvement and moral climate and children’s personal interpretive socioemotional biases are important variables in mediating the level of aggression observed in middle schools (Farrell, Henry, Mays, & Schoeny, 2011; Foà, Brugman, & Mancini, 2012). Such variables may function as important control variables that, when modified, can produce cascading changes in a multiple variables related to the production of aggression within a given school system. 3. Coordinate attempts to regulate core control variables in relation to each other. It is not enough simply to seek to modify parameters of a dynamic system in isolation of one another. Instead, research suggests that prevention and intervention programs that seek to coordinate interventions and resources in relation to each other are more able to produce systemic change than those that do not (Eron et al., 2002; Farmer, Farmer, Estell, & Hutchins, 2007). The process of coordinating possible control variables during prevention and intervention increases the likelihood of fostering synergy among the coacting variables that compose a dynamic system. 4. Monitor and respond to changes in the system continuously over time. Traditional approaches to intervention tend to assess target functioning of individuals both

prior to and subsequent to an intervention. Pre- and posttest designs yield insufficient data to allow assessment of the change dynamics in a complex system. To gain sufficient information to provide a clear understanding of the developmental course of intervention, it is necessary to assess the status of an evolving system at multiple points over the course of an intervention. A mindset of continuous assessment provides an important corrective to static pre- and posttest designs. Without multiple assessments, it is not possible to identify the types of nonlinear changes that result from coactions among system variables. Further, nonlinear systems often exhibit instabilities and fluctuations that result from minor changes in component processes. As a result, modifications in any given control variable can produce both wanted and unwanted effects (Gorman et al., 2006). Continuous monitoring of changes in system dynamics is necessary to know when and how to regulate the fluctuation of system parameters to produce optimal effects. Still further, understanding how manipulation of control parameters produces patterns of wanted and unwanted change over time is important for the development of prevention programs. The capacity to identify core control parameters provides practitioners with the ability to know where and how to allocate limited resources to produce optimal gain. CONCLUSIONS: THE SELF-ORGANIZATION OF ORDER, VARIABILITY, AND DISORDER IN STRUCTURES OF THINKING, FEELING, AND ACTING The way we conceptualize a problem affects how we study it. In the case of deviant behavior, our conceptualizations affect how we approach people who exhibit patterns of action that we classify as maladaptive. The study of deviant behavior raises difficult issues that have long captured the attention of psychologists, sociologists, philosophers, physicians, and practitioners. Although the medical model of psychopathology remains dominant, recently, it has weathered its share of controversy (Frances, 2013). From our perspective, the medical model influences theory and research in developmental psychopathology by representing psychological dysfunction as fixed structures that reflect an underlying pathology. In what follows, we discuss how dynamic systems thinking can address limitations of the psychopathology of fixed forms. From Fixed Forms to Emergent Structures When we think of psychological problems in terms of the concept of disorder, we run the risk of treating behavior and psychological problems as if they were entities or

Page 704

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

Conclusions: The Self-Organization of Order, Variability, and Disorder in Structures of Thinking, Feeling, and Acting

things. The psychopathology of fixed forms leads theorists and researchers to think of psychological problems as trait-like entities that exist within individuals. When this occurs, researchers seek to investigate the nature of the presumed trait or form: What is the nature of the trait? Where is it located? What causes it to come into being? However, from a dynamic systems perspective, behavior is not a thing that a person has; behavior is a process that occurs over time within particular social contexts. It is necessary, therefore, to study how behavior emerges over time in particular contexts and social interactions. An inhibited child is not shy all the time; a child who withdraws at school is often noisy and rambunctious at home (Kagan & Snidman, 1990). Further, when we treat psychological problems as trait-like entities, we tend to direct our attention to behavior that fits our a priori conception of the nature of the entity, and ignore behavior that does not comport with those preconceptions. For example, by focusing on the so-called trait rather than on the dynamic production of behavior in context, we fail to observe how patterns of both maladaptive and adaptive behavior arise as individuals adjust to the demands of local contexts. By studying individuals in contexts, systems approaches seek to understand the origins of order and variability, adaptive and maladaptive action, consistency and inconsistency. Similarly, if we think of patterns of maladaptive activity as discrete and trait-like forms, we are likely to be surprised when we find evidence that presumably fixed categories of disorder in fact fade into one another and assume a variety of diverse and overlapping forms. In contrast, if we seek to understand how patterns of psychological adaptation and maladaptation develop over time in individual children and contexts, it would not be surprising to find adaptive and maladaptive variation in behavior as individuals with different psychological resources and histories adjust to the contextual demands. Conceptualizing Dysfunction: From Pathology to Adaptive Conflict In medicine, pathology refers to a discernible deviation in the structure and functioning of the body. Drawing on this concept, the term psychopathology is typically employed in two ways. The first uses the concept of pathology as a metaphor for deviant behavior. From this view, psychopathology is to healthy behavior as a disease is to a healthy organ. Problems arise when one treats this metaphorical construction as a literal statement. Unlike the body, psychological functioning (or mind) is not an entity or thing. As a result, it is not the kind of thing that can be diseased. Treating psychopathology as a disease obscures the social, scientific, and pragmatic processes

12:37pm

705

through which psychological processes are identified as deviant or maladaptive. The standards we use to identify psychological dysfunction are different from those used to determine physical pathology (Banner, 2013). Judgments that any given set of psychological processes is dysfunctional or disordered can be made only against the backdrop of some framework for identifying what is functional or ordered. Such judgments are ultimately normative and evaluative in nature. While many seek to define psychological dysfunction in neutral or nonevaluative terms, this is not possible. Judgments of human action necessarily operate against the backdrop of inescapable evaluative frameworks (Boysen, 2007; Taylor, 1989). The second use of the term psychopathology is arguably less metaphorical, but nonetheless troublesome. From this view, psychological problems are regarded as symptoms of an underlying pathology located outside of the sphere of behavior itself (e.g., in the brain or body). This view is less metaphorical in the sense that it holds that psychological problems are actually products of bodily (brain) pathology. Several problems arise from this view (Banner, 2013). First, this approach suggests that there should be some correspondence between an underlying biological pathology and psychological symptoms. However, as indicated throughout this chapter, research shows patterns of psychological dysfunction are highly variegated and multiply determined. Patterns of psychological dysfunction emerge over time and function as dynamic systems. However, the proposition that psychological problems are caused by brain disorders begs the question of how to discriminate psychological dysfunction from psychological health. Whatever the role of biological disorder in psychological dysfunction, judgments of psychological dysfunction can only made by invoking standards for defining psychological functioning as optimal or nonoptimal. The identification of biological dysfunction does not in and of imply that any concomitant psychological processes are dysfunctional. From a dynamic systems perspective, persons—like all organisms—are complex adaptive systems (Lickliter & Honeycutt, 2013; Masten, 2007; van Geert, 1994). A person’s psychological processes operate as modes of adapting to their environs. Adaptation reflects the capacity of an organism to adjust its behavior to environmental demands. From this view, it might be better to think of modes of psychological dysfunction as systemic difficulties in adaptation, rather than as distinct forms of pathology, disorder, disease, or illness. The concepts of adaptation and maladaptation are relational concepts; they make reference to the goodness of fit between the resources of the organism and the demands of its environment (Gottlieb & Halpern, 2002; Lerner, Hess, & Nitz, 1990). Thus, maladaptive behavior is behavior that causes problems for an

Page 705

Trim Size: 8.5in x 11in

706

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

individual, dyad, or community. The concept of dysfunction as adaptive conflict reflects a more person-centered rather than pathology-centered understanding psychological dysfunction (von Eye, 2009). Instead of seeking to understand the essence of a given category of psychological disorder, emphasis is placed upon the processes by which individual develop adaptive and maladaptive ways of relating to their worlds over time. From Aggregates to Individual Trajectories From a dynamic systems perspective, human behavior self-organizes through the dynamic coupling of person-in-context. Thus to understand the process of psychological development, it becomes necessary to examine how novel psychological structures arise through the moment-by-moment coactions among aspects of the person–environment system. This requires moving beyond methods based primarily on aggregate judgments collapsed across time and context and toward direct observation of integrative patterns of behavior during the process of their developmental formation. Traditionally, much research in developmental psychopathology has been variable centered rather than person centered (von Eye & Bergman, 2003). Research directed toward identifying relations among abstract variables using aggregated data tends to be nonergodic and thus fails to capture the nature of processes that occur at the individual level of functioning (Molenaar & Ram, 2010). Further, much research is conducted using participant judgments on rating scales and checklists without actually observing the activity of individual persons. Such methods fail to examine the data that is most essential to understand the process by which order and variability develop over time: the dynamic coupling of individual and social context over time. The use of direct observation of individual-context interaction allows researchers to track how complex patterns of adaptive and maladaptive functioning codevelop in single individuals (Granic & Hollenstein, 2006; Kagan, Snidman, McManis, Woodward, & Hardway, 2002). The Challenges Ahead Over the past 20 years, dynamic systems approaches to maladaptive development have themselves undergone considerable development. Nonetheless, dynamic systems models of development are only beginning to show their promise. There remain important challenges to be overcome. The Importance of Psychological Theory in Dynamic Systems Analyses of Development Dynamic systems theory has its origins in the fields of physics, mathematics, biology, and chemistry. Since then, it has proliferated in virtually all areas of scientific inquiry.

The principles of dynamic systems theory are broadly applicable to the study of any system that demonstrates complexity. Over the past 20 years, there has been an astonishing proliferation of work in many areas of psychology and developmental science. While the principles of dynamic systems theory are broadly applicable, it is perhaps misleading to speak of dynamic systems thinking as a theory. Dynamic systems theory operates more as a meta-theory that provides principles and methods for understanding and modeling dynamic phenomena (Witherington, 2007). If this is so, meaningful dynamic systems modeling—mathematical or otherwise—requires a solid foundation of psychological theorizing (Cowan, 2003; Spencer, Austin, & Schutte, 2012; Van Geert, 1994). This is not to say that dynamic systems approaches should necessarily work from a hypothetico-deductive framework. Dynamic systems models of development are still in their infancy. Much work remains in developing dynamic system methods that are capable of representing the overwhelming complexity of human psychological functioning as it unfolds over time. As this work ensues, researchers are producing extraordinarily rich analyses of the dynamic coupling of local psychological processes as they unfold and develop in a wide variety of psychological domains (e.g., Ferrer & Helm, 2013; McAssey, Helm, Hsieh, Sbarra, & Ferrer, 2013). Such studies, even when they do not proceed from a specific theoretical model, are providing an extremely rich matrix of findings that can inform the development of psychological theory. Thus, while dynamic systems theory provides meta-theoretical and meta-methodological tools for developmental science, to exploit the promise of dynamic systems thinking, it is important to wed systems meta-theory more closely with genuinely psychological theorizing. Elaborating Quantitative and Qualitative Methods for Assessing Complexity in Normative and Nonnormative Development The task of studying dynamic complexity in development raises obvious methodological questions. Traditional methods that focus on parsing variance at the level of populations typically fail to shed light on the dynamic processes that operate at the level of the development individual action (Molenaar & Campbell, 2009; Ong & Zautra, 2009). Dynamic systems researchers have developed a multiplicity of methods that track dynamic change in ways that are sensitive to the individual development. These methods include mathematical modeling (Ferrer & Steele, 1982; Kunnen, 2012); state space grids (Hollenstein, 2013); microdevelopment of individual and dyadic activity (Mascolo & Fischer, 2015; Thomas, Hopwood, Woody, Ethier, & Sadler, 2013); analyses of alternative developmental pathways (Mascolo & Fischer, 2015); self-report

Page 706

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

References

and diary methods (Navarro, Arrieta, & Ballén, 2007); thick description (Fogel, Garvey, Hsu, & West-Stroming, 2006); and qualitative analyses (Haggis, 2008; Sommerfeld & Hollenstein, 2011). Traditional methods—including experimental design (Gottlieb, 2001) and factor analysis (Ram, Brose, & Molenaar, 2013)—are also applicable to the dynamic analysis of complex systems. These approaches are sensitive to issues relating to developmental timing (Granic, 2005); synchronicity of unfolding psychological processes (McAssey, Helm, Hsieh, Sbarra, & Ferrer, 2013); multiple levels of analysis (Van Geert & Lichtwarck-Aschoff, 2005); sequence (Molenaar & Goode, 2013); and structural transformation (Mascolo & Fisher, 2015). While there has been enormous progress made in the elaboration of dynamic systems methodology, a series of challenges remain. First, most of this work has been performed with normative populations. There is a need to apply and develop these methods further for the study of the dynamic development of maladaptive behavior patterns. Second, despite the impressive progress in the development of dynamic systems methods, the issue of representing developmental complexity remains. For example, in pursuing ergodic methods that are sensitive to individual development, researchers are increasingly performing longitudinal analyses of multiple single individuals. This raises the question of the extent to which it is desirable and possible to use idiographic methods to make nomothetic generalizations (Fisher, Newman, & Molenaar, 2011). Further, given the enormous complexity of normative and nonnormative development, the question of how to differentiate core control variables from other variables relevant to the development of any given system or systems of activity remains a central challenge. Third, as research comes to address increasingly complex units of behavior, there will be a need to find ways to assess developmental complexity at multiple levels of analysis in ways that are accessible, meaningful and intelligible but not overly simplistic. Toward this end, there is a need to explore ways to coordinate quantitative and qualitative research methods. Similarly, it is essential to keep in mind that the study of developmental complexity does not always require complex designs. We should never lose sight of the principle that methods are driven by problems (Bergman & Vargha, 2013). Simple (but not simplistic) methods can often reveal vast complexity. Examples of these include but are not limited to the use of experimental designs to study the effect of variations in rearing conditions on behavioral development (Gottlieb, 1991, 2002; Jablonka & Lamb, 2006); thick description of historical-relational trajectories of development (Fogel, Garvey, Hsu, & West-Stroming, 2006; Piaget, 1952); and the microdevelopmental tracking of developmental complexity in real time (Mascolo & Fischer, 2015).

12:37pm

707

Accounting for Structural Change in Dynamic Assessments of Development While dynamic systems approaches have produced powerful methods for the quantitative analysis of nonlinear trajectories of development, it is equally important to understand the structural aspects of developmental change. Many attempts to identify trajectories of maladaptive development focus on quantitative changes in the frequency of or rate of maladaptive activity over time. While useful, such analyses fail to account for the patterns of differentiation and integration that define a sequence as a genuinely developmental one. The concept of development is not simply a synonym for change over time or age-related change. Development is a directional and transformational concept (Bibace & Kharlamov, 2013; Raeff, 2011; Van Geert, 1986). To say that a psychological structure is developing implies that it is undergoing successive structural changes defined with reference to some actual, hypothetical, or idealized developmental outcome (see Basseches & Mascolo, 2010; Sroufe, 2007; Van Geert, 1986, 1987). For example, the structure of indirect aggression in adolescence is different from the reactive aggression of a preschooler (Arnocky, Sunderani, Miller, & Vaillancourt, 2012; Card & Little, 2006); the dynamics of shame, worthlessness and rage experienced among adult victims of early sexual abuse differ from the raw unarticulated experiences of victimization in childhood (Isely, Isely, Freiburger, & McMackin, 2008; Rahm, Renck, & Ringsberg, 2013). To measure structural change, there is a need to invoke some type of developmental yardstick to measure the transformations in the structure of normative and nonnormative patterns of thinking, feeling and acting over time. Dynamic skill theory (Fischer, 1980; Mascolo & Fischer, 2015) provides one such yardstick. Fischer’s analysis of developmental transformations in dissociative splitting (described earlier in this chapter) provides but one example of how researchers can account for structural transformations in nonnormative patterns of thinking, feeling, and acting over the course of development. REFERENCES Aber, J. L., Allen, J. P., Carlson, V., & Cicchetti, D. (1989). The effects of maltreatment on development during early childhood: Recent studies and their theoretical, clinical, and policy implications. In D. Cicchetti & V. Carlson (Ed.), Child maltreatment: Theory and research on the causes and consequences of child abuse and neglect (pp. 579–619). Cambridge, UK: Cambridge University Press. doi: 10.1017/ CBO9780511665707.019 Aho, K. (2010). The psychopathology of American shyness: A hermeneutic reading. Journal for the Theory of Social Behaviour, 40, 190–2. Aiken, J. R., & Hanges, P. J. (2012). Research methodology for studying dynamic multiteam systems: Application of complexity science. In S. J. Zaccaro, M. A. Marks, & L. A. DeChurch (Eds.), Multiteam systems: An organization form for dynamic and complex environments (pp. 431–458). New York, NY: Routledge/Taylor & Francis Group.

Page 707

Trim Size: 8.5in x 11in

708

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

Ainsworth, M. D. S., Blehar, M., Waters, E., & Wall, S. (1978). Patterns of attachment: A psychological study of the strange situation. Hillsdale, NJ: Erlbaum. Aksan, N., & Kochanska, G. (2004). Links between systems of inhibition from infancy to preschool years. Child Development, 75, 1477–1490. doi: 10.1111/j.1467-8624.2004.00752.x Antonishak, J., & Reppucci, N. (2008). Ecological and community level influences on child development. In T. P. Gullotta & G. M. Blau (Eds.), Handbook of childhood behavioral issues: Evidence-based approaches to prevention and treatment (pp. 69–86). New York, NY: Routledge/ Taylor & Francis Group. Arnocky, S., Sunderani, S., Miller, J. L., & Vaillancourt, T. (2012). Jealousy mediates the relationship between attractiveness comparison and females’ indirect aggression. Personal Relationships, 19(2), 290–303. doi: 10.1111/j.1475-6811.2011.01362.x Arsenio, W. F., Adams, E., & Gold, J. (2009). Social information processing, moral reasoning, and emotion attributions: Relations with adolescents reactive and proactive aggression. Child Development, 80, 1739–1755. doi: 10.1111/j.1467-8624.2009.01365.x Ayoub, C. C., Fischer, K. W., & O’Connor, E. E. (2003). Analyzing development of working models for disrupted attachments: The case of family violence. Attachment and Human Development, 5, 97–119. doi: 10.1080/1461673031000108478 Ayoub, C. C., Rogosh, F., Toth, S. L., O’Connor, E., Cicchetti, D., Rappolt-Schlichtmann, G., & Fischer, K. W. (2006). Cognitive and emotional differences in young maltreated children: A translational application of dynamic skill theory. Development and Psychopathology 18, 670–706. doi: 10.1017/S0954579406060342 Ayoub, C., Willett, J. B., & Robinson, D. (1993b). Families at risk of child maltreatment: Entry-level characteristics and growth in family functioning. Child Abuse and Neglect, 16, 495–511. Badcock, P. B. (2012). Evolutionary systems theory: A unifying meta-theory of psychological science. Review of General Psychology, 16(1), 10–23. doi: 10.1037/a0026381 Bailey, C., & Ostrov, J. M. (2008). Differentiating Forms and Functions of Aggression in Emerging Adults: Associations with Hostile Attribution Biases and Normative Beliefs. Journal of Youth and Adolescence, 37(6), 713–722. doi: 10.1007/s10964-007-9211-5 Banner, N. F. (2013). Mental disorders are not brain disorders. Journal of Evaluation in Clinical Practice, 19, 509–513. Barker, E. D., Tremblay, R. E., Nagin, D. S., Vitaro, F., & Lacourse, E. (2006). Development of male proactive and reactive physical aggression during adolescence. Journal of Child Psychology and Psychiatry, 47, 783–790. doi: 10.1111/j.1469-7610.2005.01585.x Barry, C. T., Golmaryami, F. N., Rivera-Hudson, N., & Frick, P. J. (2013). Evidence-based assessment of conduct disorder: Current considerations and preparation for DSM-5. Professional Psychology: Research and Practice, 44, 56–63. doi: 10.1037/a0029202 Basseches, M., & Mascolo, M. F. (2010). Psychotherapy as a developmental process. New York, NY: Routledge/Taylor & Francis Group. Baumeister, R. F., Smart, L., & Boden, J. M. (1996). Relation of threatened egotism to violence and aggression: The dark side of high self-esteem. Psychological Review, 103, 5–33. doi: 10.1037/0033295X.103.1.5 Bavelier, D., Green, C. S., Schrater, P., & Pouget, A. (2012). Brain plasticity through the life span: Learning to learn and action video games. Annual Reviews of Neuroscience, 35, 391- 416. Benikos, N., Johnstone, S. J., & Roodenrys, S. J. (2013a). Short-term training in the Go/Nogo task: Behavioural and neural changes depend on task demands. International Journal of Psychophysiology, 87(3), 301–312. doi: 10.1016/j.ijpsycho.2012.12.001 Benikos, N., Johnstone, S. J., & Roodenrys, S. J. (2013b). Varying task difficulty in the Go/Nogo task: The effects of inhibitory control, arousal, and perceived effort on ERP components. International Journal of Psychophysiology, 87, 262–272. doi: 10.1016/j.ijpsycho.2012.08.005

Bergman, L. R., & Vargha, A. (2013). Matching method to problem: A developmental science perspective. European Journal of Developmental Psychology, 10(1), 9–28. doi: 10.1080/17405629.2012.732920 Bibace, R., & Kharlamov, N. A. (2013). The spiral: The concept of development after Werner and Kaplan. Culture & Psychology, 19(4), 453–462. doi: 10.1177/1354067X13500324 Bierman, K. L. (2011). The promise and potential of studying the “invisible hand” of teacher influence on peer relations and student outcomes: A commentary. Journal Of Applied Developmental Psychology, 32(5), 297–303. doi: 10.1016/j.appdev.2011.04.004 Blackburn, R., Logan, C., Donnelly, J. P., & Renwick, S. D. (2008). Identifying psychopathic subtypes: Combining an empirical personality classification of offenders with The Psychopathy Checklist-Revised. Journal of Personality Disorders, 22(6), 604–622. doi: 10.1521/pedi. 2008.22.6.604 Block, J. H., & Block, J. (1980). The role of ego-control and ego-resiliency in the organization of behavior. In W. A. Collins (Ed.), Minnesota symposium on child psychology (Vol. 13, pp. 39–101). Hillsdale, NJ: Erlbaum. doi: 10.1521/pedi.2008.22.6.604 Boylan, K., Vaillancourt, T., Boyle, M., Szatmari, P. (2007). Comorbidity of internalizing disorders in children with oppositional defiant disorder. European Child and Adolescent Psychiatry 16(8): 484–494. Bongers, I. L., Koot, H. M., van der Ende, J., & Verhulst, F. C. (2003). The normative development of child and adolescent problem behavior. Journal of Abnormal Psychology, 112, 179–192. doi: 10.1037/0021-843X.112.2.179 Borsboom, D., Cramer, A. O. J., Schmittmann, V. D., Epskamp, S., & Waldorp, L. J. (2011). The small world of psychopathology. PloS one, 6, e27407. doi: 10.1371/journal.pone.0027407 Bornstein, M. H. (2013). Mother-infant attunement: A multilevel approach via body, brain, and behavior. In M. Legerstee, D. W. Haley, & M. H. Bornstein (Eds.), The infant mind: Origins of the social brain (pp. 266–298). New York, NY: Guilford Press. Bostow, D. E., & Bailey, J. B. (1969). Modification of severe disruptive and aggressive behavior using brief timeout and reinforcement procedures. Journal of Applied Behavior Analysis, 2(1), 31–37. doi: 10.1901/jaba.1969.2-31 Boysen, G. A. (2007). An evaluation of the DSM concept of mental disorder. Journal of Mind and Behavior, 28, 157–173. Brendgen, M., Vitaro, F., Tremblay, R. E., & Lavoie, F. (2001). Reactive and proactive aggression: Predictions to physical violence in different contexts and moderating effects of parental monitoring and caregiving behavior. Journal of Abnormal Child Psychology, 29, 293–304. doi: 10.1023/A:1010305828208 Bretherton, I., & Waters, E. (1985). Growing points of attachment theory and research. Monographs of the Society for Research in Child Development, 50(1-2, Serial No. 209). Breuer, J., & Freud, S. (1955). Studies on hysteria (A. Strachey & J. Strachey, Trans.). In J. Strachey (Ed.), Standard edition of the complete psychological works of Sigmund Freud (Vol. 2). London, UK: Hogarth Press. (Original publication 1895.) Brown, A. L., & Reeve, R. (1987). Bandwidths of competence: The role of supportive contexts in learning and development. In L. S. Liben (Ed.), Development and learning: Conflict or congruence? Hillsdale, NJ: Erlbaum. Brugman, D., & Mancini, T. (2012). School moral atmosphere and normative orientation to explain aggressive and transgressive behaviours at secondary school. Journal of Moral Education, 41(1), 1–22, doi: 10.1080/03057240.2011.652520 Bushman, B. J., & Anderson, C. A. (2001). Is it time to pull the plug on hostile versus instrumental aggression dichotomy? Psychological Review, 108, 273–279. doi: 10.1037/0033-295X.108.1.273 Cairns, R. B. (1983). The emergence of developmental psychology. In P. H. Musen & W. Kessen, Handbook of Child Psychology: Vol. 1: History, Theory, and Methods (4th ed., pp. 41–102). New York: Wiley.

Page 708

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

References Calverley, R., Fischer, K. W., & Ayoub, C. (1994). Complex splitting of self-representations in sexually abused adolescent girls. Development and Psychopathology, 6, 195–213. Canadian Broadcasting Corporation. (1990). To a safer place. Video program. Ottowa, Ontario. Carlson, V., Cicchetti, D., Barnett, D., & Braunwald, K. G. (1989). Finding order in disorganization: Lessons from research on maltreated infants’ attachments to their caregivers. In D. Cicchetti & V. Carlson (Eds.), Child maltreatment: Theory and research on the causes and consequences of child abuse and neglect (pp. 494–528). Cambridge, UK: Cambridge University Press. Card, N. A., & Little, T. D. (2006). Proactive and reactive aggression in childhood and adolescence: A meta-analysis of differential relations with psychosocial adjustment. International Journal of Behavioral Development, 30, 466–480. doi: 10.1177/0165025406071904 Case, R., & Edelstein, W. (Eds.) (1993). The new structuralism in cognitive development: Theory and research on individual pathways. In Contributions to human development, 23. Basel, Switzerland: Karger. Castellanos, X., & Tannock, R. (2002). Neuroscience of attentiondeficit/hyperactivity disorder: the search for endophenotypes. Nature Reviews Neuroscience 3, 617–628. Cepeda, N., Blackwell, K., & Munakata, Y. (2013). Speed isn’t everything: Complex processing speed measures mask individual differences and developmental changes in executive control. Developmental Science, 16, 269–286. doi: 10.1111/desc.12024 Chen, C., McComas, J. J., Hartman, E., & Symons, F. J. (2011). A prospective sequential analysis of the relation between physical aggression and peer rejection acts in a high-risk preschool sample. Early Education and Development, 22, 574–592. doi: 10.1080/10409289.2010.481706 Cicchetti, D. (1984). The emergence of developmental psychopathology. Child Development, 55(1), 1–7. doi: 10.2307/1129830 Cicchetti, D. (1989). How research on child maltreatment has informed the study of child development: Perspectives from developmental psychopathology. In D. Cicchetti & V. Carlson (Eds.), Child maltreatment: Theory and research on the causes and consequences of child abuse and neglect (pp. 377–431). Cambridge, UK: Cambridge University Press. doi: 10.1017/CBO9780511665707.014 Cicchetti, D. (2013). Annual research review: Resilient functioning in maltreated children—Past, present, and future perspectives. Journal of Child Psychology and Psychiatry, 54(4), 402–422. Cicchetti, D. (2010). Developmental psychopathology. In M. E. Lamb, A. M. Freund, & R. M. Lerner (Eds.), The handbook of life-span development, Vol 2: Social and emotional development (pp. 511–589). Hoboken, NJ: John Wiley & Sons Inc. Cicchetti, D., & Beeghly, M. (1987). Symbolic development in maltreated youngsters: An organizational perspective. In D. Cicchetti & M. Beeghly (Eds.), Atypical symbolic development. New directions for child development (Vol. 36, pp. 5–29). San Francisco, CA: Jossey-Bass. doi: 10.1002/cd.23219873605 Cicchetti, D., & Toth, S. L. (2005). Child maltreatment. Annual Review of Clinical Psychology, 1, 409–438. Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50(1–2), 16–25. doi: 10.1111/j.1469-7610.2008.01979.x Cherkes-Julkowski, M., & Mitlina, N. (1999). Self-organization of mother–child instructional dyads and later attention disorder. Journal of Learning Disabilities, 32, 6–21. doi: 10.1146/annurev.clinpsy. 1.102803.144029 Clark, K. E., & Ladd, G. W. (2000). Connectedness and autonomy support in parent–child relationships: Links to children’s socioemotional orientation and peer relationships. Developmental Psychology, 36, 485–498. doi: 10.1037/0012-1649.36.4.485 Cole, P. M., & Putnam, F. W. (1992). Effect of incest on self and social functioning: A developmental psychopathology perspective.

12:37pm

709

Journal of Consulting and Clinical Psychology, 60,174–184. doi: 10.1037/0022-006X.60.2.174 Connor, D. F., Chartier, K. G., Preen, E. C., & Kaplan, R. F. (2010). Impulsive aggression in attention-deficit-hyperactivity disorder: Symptom severity, comorbidity, and attention-deficit/hyperactivity disorder subtype. Journal of Child and Adolescent Psychopharmacology, 20, 119–126. doi: 10.1089/cap.2009.0076 Conroy, M. A., Sutherland, K. S., Snyder, A. L., & Marsh, S. (2008). Classwide interventions effective instruction makes a difference. Teaching Exceptional Children, 40(6), 24–30. Cortese, S., (2012). The neurobiology and genetics of AttentionDeficit/Hyperactivity Disorder (ADHD): What every clinician should know. European Journal of Paediatric Neurology, 16(5), 422–433. doi: http://dx.doi.org/10.1016/j.ejpn.2012.01.009 Cowan, N. (2003). Comparisons of developmental modeling frameworks and levels of analysis in cognition: connectionist and dynamic systems theories deserve attention, but don’t yet explain attention. Developmental Science, 6(4), 440–447. doi: 10.1111/1467-7687.00299 Cox, M. J., Mills-Koonce, R., Propper, C., & Gariépy, J. (2010). Systems theory and cascades in developmental psychopathology. Development and Psychopathology, 22, 497–506. doi: 10.1017/S0954579410000234 Cramer, A. O. J., Waldorp, L. J., der Maas van, & Borsboom, D. (2010a). Comorbidity: A network perspective. Behavioral and Brain Sciences, 33, 137–150. Cramer, A. O. J., Waldorp, L. J., der Maas van, & Borsboom, D. (2010b). Complex realities require complex theories: Refining and extending the network approach to mental disorders. Behavioral and Brain Sciences, 33, 178–193 Crittenden, P. M. (1988). Relationships at risk. In J. Belsky & T. Nezworski (Eds.), Clinical implications of attachment theory (pp. 136–174). Hillsdale, NJ: Erlbaum. Danchin, É., Charmantier, A., Champagne, F. A., Mesoudi, A., Pujol, B., & Blanchet, S. (2011). Beyond DNA: integrating inclusive inheritance into an extended theory of evolution. Nature Reviews Genetics, 12(7), 475–486. Dandreaux, D. M., & Frick, P. J. (2009). Developmental pathways to conduct problems: A further test of the childhood and adolescent-onset distinction. Journal of Abnormal Child Psychology, 37, 375–385. De Zeeuw, P.; Weusten, J., van Dijk, S., van Belle, J., & Durston, S. (2012) Deficits in cognitive control, timing and reward sensitivity appear to be dissociable in ADHD. PLoS ONE, 7(12): e51416. doi: 10.1371/journal.pone.0051416 Dempsey, A. G., & Storch, E. A. (2008). Relational victimization: The association between recalled adolescent social experiences and emotional adjustment in early adulthood. Psychology in the Schools, 45, 310–322. doi: 10.1002/pits.20298 de Rivera, J. (2006). Conceptual encounter: The experience of anger. In C. T. Fischer (Ed.), Qualitative research methods for psychologist: Introduction through empirical studies (pp. 213–245). San Diego, CA: Elsevier Academic Press. doi: 10.1016/B978-012088470-4/50011-9 de Rivera, J. (1991). The structure and dynamics of emotion. In D. J. Ozer, J. R. Healy, & A. J. Stewart (Eds.) , Perspectives in personality, Vol. 3: Part A: Self and emotion; Part B: Approaches to understanding lives (pp. 191–212). London, UK: Jessica Kingsley Publishers. Dirks, M. A., De Los Reyes, A., Briggs-Gowan, M., Cella, D., & Wakschlag, L. S. (2012). Annual research review: Embracing not erasing contextual variability in children’s behavior—theory and utility in the selection and use of methods and informants in developmental psychopathology. Journal of Child Psychology and Psychiatry, 53(5), 558–574. Dishion, T. J., Nelson, S. E., Winter, C. E., & Bullock, B. M. (2004). Adolescent friendship as a dynamic system: Entropy and deviance in the etiology and course of male antisocial behavior. Journal of Abnormal Child Psychology, 32, 651–663. doi: 10.1023/B:JACP. 0000047213.31812.21

Page 709

Trim Size: 8.5in x 11in

710

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

Dishion, T. J., Eddy, J., Haas, E., Li, F., & Spracklen, K. (1997). Friendships and violent behavior during adolescence. Social Development, 6, 207–223. doi: 10.1111/1467-9507.ep11631575 Dodge, K. A., Lochman, J. E., Harnish, J. D., Bates, J. E., & Pettit, G. S. (1997). Reactive and proactive aggression in school children and psychiatrically impaired chronically assaultive youth. Journal of Abnormal Psychology, 106, 37–51. doi: 10.1037/0021-843X.106.1.37 Dodge, K. A., & Pettit, G. S. (2003). A biopsychosocial model of the development of chronic conduct problems in adolescence. Developmental Psychology, 39, 349–371. doi: 10.1037/0012-1649.39.2.349 Edens, J. F., Campbell, J. S., & Weir, J. M. (2007). Youth psychopathy and criminal recidivism: A meta-analysis of the psychopathy checklist measures. Law and Human Behavior, 31, 53–75. doi: 10.1007/ s10979-006-9019-y Eisenberg, N., Hofer, C., & Vaughan, J. (2007). Effortful Control and Its Socioemotional Consequences. In J. J. Gross (Ed.), Handbook of emotion regulation (pp. 287–306). New York, NY: Guilford Press. Eley, T. C., Lichtenstein, P., & Moffitt, T. E. (2003). A longitudinal behavioral genetic analysis of the etiology of aggressive and nonaggressive antisocial behavior. Development and Psychopathology, 15, 383–402. doi: 10.1017/S095457940300021X Epkins, C. C., & Heckler, D. R. (2011). Integrating etiological models of social anxiety and depression in youth: Evidence for a cumulative interpersonal risk model. Clinical Child and Family Psychology Review, 14, 329–376. doi: 10.1007/s10567-011-0101-8 Erdelyi, M. H. (1985). Psychoanalysis: Freud’s cognitive psychology. New York: W.H. Freeman. Eron, L., Huesmann, R., Spindler, A., Guerra, N., Henry, D., & Tolan, P. (2002). A cognitive-ecological approach to preventing aggression in urban settings: Initial outcomes for high-risk children. Journal of Consulting and Clinical Psychology, 70(1), 179–194. doi: 10.1037/0022-006X.70.1.179 Estes, W. K., (1956). The problem of inference from curves based on group data. Psychological Buletin, 53, 134–140. Evans, D. E., & Rothbart, M. K. (2009). A two-factor model of temperament. Personality and Individual Differences, 47, 565–570. doi: 10.1016/j.paid.2009.05.010 Evans, C. A., Nelson, L. J., & Porter, C. L. (2012). Making sense of their world: Sensory reactivity and novelty awareness as aspects of temperament and correlates of social behaviours in early childhood. Infant and Child Development, 21(5), 503–520. doi: 10.1002/icd.1754 Evans, C. A., & Porter, C. L. (2009). The emergence of mother-infant co-regulation during the first year: Links to infants’ developmental status and attachment. Infant Behavior & Development, 32, 147–158. doi: 10.1016/j.infbeh.2008.12.005 Famularo, R., Kinscherff, R., Fenton, T., & Ayoub, C. (1994). Maternal and child posttraumatic stress disorder in cases of child maltreatment. Child Abuse and Neglect, 18, 27–36. Fanti, K. A. (2013). Individual, social, and behavioral factors associated with co-occurring conduct problems and callous-unemotional traits. Journal of Abnormal Child Psychology, 41, 811–824. doi: 10.1007/ s10802-013-9726-z Fanti, K. A., & Kimonis, E. R. (2012). Bullying and victimization: The role of conduct problems and psychopathic traits. Journal of Research on Adolescence, 22, 617–631. doi: 10.1111/j.1532-7795.2012. 00809.x Fanti, K. A., Frick, P. J., & Georgiou, S. (2009). Linking callousunemotional traits to instrumental and non-instrumental forms of aggression. Journal of Psychopathology and Behavioral Assessment, 31, 285–298. Farmer, T. W., Farmer, E. M. Z., Estell, D. B., & Hutchins, B. C. (2007). The developmental dynamics of aggression and the prevention of school violence. Journal of Emotional and Behavioral Disorders, 15(4), 197–208.

Farrell, A. D., Henry, D. B., Mays, S. A., & Schoeny, M. E. (2011). Parents as moderators of the impact of school norms and peer influences on aggression in middle school students. Child Development, 82(1), 146–161. doi: 10.1111/j.1467-8624.2010.01546.x Feffer, M. (1982). The structure of Freudian thought. NewYork, NY: International Universities Press. Feiring, C., Cleland, C. M., & Simon, V. A. (2010). Abuse-specific self-schemas and self-functioning: A prospective study of sexually abused youth. Journal of Clinical Child and Adolescent Psychology, 39, 35–50. doi: 10.1080/15374410903401112 Ferguson, T. J., Eyre, H. L., & Ashbaker, M. (2000). Unwanted identities: A key variable in shame–anger links and gender differences in shame. Sex Roles, 42, 133–157. Ferguson-Smith, A. C. (2011). Genomic imprinting: the emergence of an epigenetic paradigm. Nature Reviews Genetics, 12(9), 663. Ferrer, E., & Helm, J. L. (2013). Dynamical systems modeling of physiological coregulation in dyadic interactions. International Journal of Psychophysiology, 88(3), 296–308. doi: 10.1016/j.ijpsycho.2012 .10.013 Ferrer, E., & Steele, J. (2012). Dynamic systems analysis of affective processes in dyadic interactions using differential equations. In J. R. Harring & G. R. Hancock (Eds.), Advances in longitudinal methods in the social and behavioral sciences (pp. 111–134). Charlotte, NC: IAP Information Age Publishing. Fischer, K. W. (1980). A theory of cognitive development: The control and construction of hierarchies of skills. Psychological Review, 87, 477–531. doi: 10.1037/0033-295X.87.6.477 Fischer, K. W. (2008). Dynamic cycles of cognitive and brain development: Measuring growth in mind, brain, and education. In A. M. Battro, K. W. Fischer, & P. J. Léna (Eds.), The educated brain: Essays in neuroeducation (pp. 127–150). New York, NY: Cambridge University Press. doi: 10.1017/CBO9780511489907.010 Fischer, K. W., Bullock, D., Rotenberg, E. J., & Raya, P. (1993). The dynamics of competence: How context contributes directly to skill. In R. H. Wozniak & K. W. Fischer (Eds.), Development in context: Acting and thinking in specific environments (pp. 93–117). Hillsdale, NJ: Erlbaum. Fischer, K. W., & Pipp, S. L. (1984). Development of the structures of unconscious thought. In K. Bowers & D. Meichenbaum (Ed.), The unconscious reconsidered (pp. 88–148). New York, NY: Wiley. Fischer, K. W., & Rose, S. P. (1994). Dynamic development of coordination of components in brain and behavior: A framework for theory and research. In G. Dawson & K. W. Fischer (Eds.), Human behavior and the developing brain (pp. 3–66). New York, NY: Guilford Press. Fischer, K. W., Shaver, P., & Carnochan, P. G. (1990). How emotions develop and how they organize development. Cognition and Emotion, 4, 81–127. doi: 10.1080/02699939008407142 Fischer, K. W., & Watson, M. W. (1981). Explaining the Oedipus conflict. In K. W. Fischer (Ed.), Cognitive development. New directions for child development (Vol. 12, pp. 79–92). San Francisco, CA: Jossey-Bass. Fischer, K. W., Yan, Z., & Stewart, J. (2003). Adult cognitive development: Dynamics in the developmental web. In J. Valsiner & K. Connolly (Eds.), Handbook of developmental psychology (pp. 491–516). Thousand Oaks, CA: Sage. Fisher, A. J., Newman, M. G., & Molenaar, P. M. (2011). A quantitative method for the analysis of nomothetic relationships between idiographic structures: Dynamic patterns create attractor states for sustained posttreatment change. Journal of Consulting and Clinical Psychology, 79(4), 552–563. doi: 10.1037/a0024069 Flavell, J. (1982). On cognitive development. Child Development, 53, 1–10. Foà, C., Brugman, D., & Mancini, T. (2012) School moral atmosphere and normative orientation to explain aggressive and transgressive behaviours at secondary school. Journal of Moral Education, 41:1, 1–22, doi: 10.1080/03057240.2011.652520

Page 710

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

References Fogel, A. (1993). Developing through relationships: Origins of communication, self, and culture. Chicago, IL: University of Chicago Press. Fogel, A., Garvey, A., Hsu, H., & West-Stroming, D. (2006). Change processes in relationships: A relational-historical research approach. New York, NY: Cambridge University Press. doi: 10.1017/ CBO9780511489686. Fogel, A., & Thelen, E. (1987). Development of early expressive and communicative action: Reinterpreting the evidence from a dynamic systems perspective. Developmental Psychology, 23, 747–761. doi: 10.1037/0012-1649.23.6.747 Foulkes, D. (1982). Children’s dreams: Longitudinal studies. New York, NY: Wiley. Frances, A. (2013). Saving normal: An insider’s revolt against out-of-control psychiatric diagnosis, DSM-5, Big Pharma, and the medicalization of ordinary life. New York, NY: William Morrow & Co. Freud, A. (1966). The ego and the mechanisms of defense (C. Baines, Trans.). New York, NY: International Universities Press. (Original publication 1936.) Freud, S. (1961/1963). Introductory lectures on psychoanalysis. In J. Strachey (Ed.), Standard edition of the psychological works of Sigmund Freud (Vol. 15 & 16). London, UK: Hogarth Press. (Original publication 1917.) Freud, S. (1961). The ego and the id (J. Riviere & J. Strachey, Trans.). In J. Strachey (Ed.), Standard edition of the complete psychological works of Sigmund Freud (Vol. 19). London, UK: Hogarth Press. (Original publication 1923.) Frick, P. J., & Viding, E. (2009). Antisocial behavior from a developmental psychopathology perspective. Development and Psychopathology, 21, 1111–1131. doi: 10.1017/S0954579409990071 Frick, P. J., & White, S. F. (2008). Research review: The importance of callous-unemotional traits for developmental models of aggressive and antisocial behavior. Journal of Child Psychology and Psychiatry, 49, 359–375. doi: 10.1111/j.1469-7610.2007.01862.x Frijda, N. H. (1986). The emotions. Cambridge, UK: Cambridge University Press. Frijda, N. H. (2009). Emotion experience and its varieties. Emotion Review, 1, 264–271. doi: 10.1177/1754073909103595 Gullotta, T. P., & Blau, G. M. (2008). (Eds.), Handbook of childhood behavioral issues: Evidence-based approaches to prevention and treatment. New York, NY: Routledge/Taylor & Francis Group Gogtay, N., & Thompson, P. M. (2010). Mapping Gray Matter Development: Implications for typical development and vulnerability to psychopathology. Brain and Cognition, 72(1), 6. Goldin-Meadow, S., Alibali, M., & Church, R. B. (1993). Transitions in concept acquisition: Using the hand to read the mind. Psychological Review, 100, 279–297. doi: 10.1037/0033-295X.100.2.279 Goodman, S. H., & Brand, S. R. (2009). Maternal depression and infant mental health. Handbook of infant mental health. 3rd Ed. New York, NY: Guilford. Gorenstein, E. E., & Newman, J. P. (1980). Disinhibitory psychopathology: A new perspective and a model for research. Psychological Review, 87(3), 301–315. doi: 10.1037/0033-295X.87.3.301 Granic, I. (2005). Timing is everything: Developmental psychopathology from a dynamic systems perspective. Developmental Review, 25(3–4), 386–407. doi: 10.1016/j.dr.2005.10.005 Granic, I., & Hollenstein, T. (2006). A survey of dynamic systems methods for developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology, Vol 1: Theory and method (2nd ed., pp. 889–930). Hoboken, NJ: John Wiley & Sons Inc. Granic, I., & Patterson, G. R. (2006). Toward a comprehensive model of antisocial development: A dynamic systems approach. Psychological Review, 113, 101–131. doi: 10.1037/0033-295X.113.1.101 Granott, N. (2002). How microdevelopment creates macrodevelopment: Reiterated sequences, backward transitions, and the zone of

12:37pm

711

current development. In N. Granott & J. Parziale (Eds.), Microdevelopment: Transition processes in development and learning (pp. 213–242). Cambridge, UK: Cambridge University Press. doi: 10.1017/ CBO9780511489709.009 Gottlieb, G., & Halpern, C. (2002). A relational view of causality in normal and abnormal development. Development and Psychopathology, 14, 421–435. doi: 10.1017/S0954579402003024 Gottlieb, G. (1991). Experiential canalization of behavioral development: Results. Developmental Psychology, 27(1), 35–39. doi: 10.1037/ 0012-1649.27.2.35 Haan, N. (1977). Coping and defending. New York, NY: Academic Press. Haggis, T. (2008). ‘Knowledge must be contextual’: Some possible implications of complexity and dynamic systems theories for educational research. Educational Philosophy and Theory, 40(1), 158–176. Hastings, P. D., Fortier, I., Utendale, W. T., Simard, L. R., & Robaey, P. (2009). Adrenocortical functioning in boys with attentiondeficit/hyperactivity disorder: Examining subtypes of ADHD and associated comorbid conditions. Journal of Abnormal Child Psychology, 37, 565–578. doi: 10.1007/s10802-008-9292-y Harter, S., & Buddin, B. (1987). Children’s understanding of the simultaneity of two emotions: A five-stage developmental sequence. Developmental Psychology, 23, 388–399. doi: 10.1037/0012-1649. 23.3.388 Hawes, D. J., Dadds, M. R., Brennan, J., Rhodes, T., & Cauchi, A. (2013). Revisiting the treatment of conduct problems in children with callous-unemotional traits. Australian and New Zealand Journal of Psychiatry, 47, 646–653. Hayling, C. C., Cook, C., & Gresham, F. M. State, T., & Kern, L. (2008). An analysis of the status and stability of the behaviors of students with emotional and behavioral difficulties. Journal of Behavioral Education, 17(1), 24–42. Hechtman, L. (2009). Comorbidity and neuroimaging in attention-deficit hyperactivity disorder. Canadian Journal of Psychiatry, 54, 649–650. doi: 10.1177/0004867413484092 Heider, K. (1991). Landscapes of emotion: Mapping three cultures of emotion in Indonesia. New York, NY: Cambridge University Press. Hejdenberg, J., & Andrews, B. (2011). The relationship between shame and different types of anger: A theory-based investigation. Personality and Individual Differences, 50, 1278–1282. doi: 10.1016/j.paid. 2011.02.024 Hencke, R. W. (1996). Self stories: Effects of children’s emotional styles on their appropriation of self-schemata. Unpublished doctoral dissertation, Harvard University, Cambridge, MA. Herman, J. (1992). Trauma and recovery. New York, NY: Basic Books. Herman, K. C., Ostrander, R., Walkup, J. T., Silva, S. G., & March, J. S. (2007). Empirically derived subtypes of adolescent depression: Latent profile analysis of co-occurring symptoms in the Treatment for Adolescents with Depression Study (TADS). Journal of Consulting and Clinical Psychology, 75(5), 716–728. doi: 10.1037/ 0022-006X.75.5.716 Hilgard, E. R. (1977). Divided consciousness. New York, NY: Wiley. Hoekzema, E., Carmona, S., Tremols, V., Gispert, J. D., Guitart, M., Fauquet, J., & Vilarroya, O. (2010). Enhanced neural activity in frontal and cerebellar circuits after cognitive training in children with attention-deficit/hyperactivity disorder. Human Brain Mapping, 31(12), 1942–50. doi: 10.1002/hbm.20988 Hollenstein, T. (2013). State space grids: Depicting dynamics across development. New York, NY: Springer Science + Business Media. doi: 10.1007/978-1-4614-5007-8 Howe, M. L., & Lewis, M. D. (2005). The importance of dynamic systems approaches for understanding development. Developmental Review, 25(1), 247–251. Hubbard, J. A., McAuliffe, M. D., Morrow, M. T., & Romano, L. J. (2010). Reactive and proactive aggression in childhood and

Page 711

Trim Size: 8.5in x 11in

Cicchetti

712

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

adolescence: Precursors, outcomes, processes, experiences, and measurement. Journal of Personality, 78, 95–118. doi: 10.1111/j.1467-6494. 2009.00610.x Hughes, J., & Chen (2011). Reciprocal effects and prediction of academic self-efficacy. Journal of Applied Developmental Psychology, 32, 247–256. Hulette, A. C., Kaehler, L. A., & Freyd, J. J. (2011). Intergenerational associations between trauma and dissociation. Journal of Family Violence, 26, 217–225. doi: 10.1007/s10896-011-9357-5 Huxley, J. (1942). Evolution: The Modern Synthesis. Immordino-Yang, M. H., & Damasio, A. R. (2007). We feel, therefore we learn: The relevance of affective and social neuroscience to education. Mind, Brain and Education, 1, 3–10.

the future of psychiatric canon and practice. Philosophy, Ethics, and Humanities in Medicine, 7(2). 5341–5347. Kempes, M., Matthys, W., de Vries, H., & van Engeland, H. (2005). Reactive and proactive aggression in children: A review of theory, findings and the relevance for child and adolescent psychiatry. European Child & Adolescent Psychiatry, 14, 11–19. doi: 10.1007/ s00787-005-0432-4 Kelso, J. (1995). Dynamic patterns: The self-organization of brain and behavior. Cambridge, MA: MIT Press. Kemp, F. (1996). The ideology of aversive treatment as applied to clients and colleagues. Child & Family Behavior Therapy, 18(1), 9–27. doi: 10.1300/J019v18n01_02

Isely, P. J., Isely, P., Freiburger, J., & McMackin, R. (2008). In their own voices: A qualitative study of men abused as children by Catholic clergy. Journal of Child Sexual Abuse, 17(3–4), 201–215.

Kenardy, J., Smith, A., Spence, S. H., Lilley, P.-R., Newcombe, P., Dob, R., & Robinson, S. (2007). Dissociation in children’s trauma narratives: An exploratory investigation. Journal of Anxiety Disorders, 21, 456–466. doi: 10.1016/j.janxdis.2006.05.007

Ivanova, M. Y., & Israel, A. C. (2006). Family stability as a protective factor against psychopathology for inner-city children receiving psychological services. Journal of Clinical Child and Adolescent Psychology, 35, 564–570.

Kim, J., & Cicchetti, D. (2006). Longitudinal trajectories of self-system processes and depressive symptoms among maltreated and nonmaltreated children. Child Development, 77, 624–639. doi: 10.1111/ j.1467-8624.2006.00894.x

Izard, C. E., Ackerman, B. P., Schoff, K. M., & Fine, S. E. (2000). Self-organization of discrete emotions, emotion patterns, and emotion–cognition relations. In M. D. Lewis & I. Granic (Eds.), Emotion, development, and self-organization: Dynamic systems approaches to emotional development (pp. 15–36). New York, NY: Cambridge University Press. doi: 10.1017/CBO9780511527883.003

Kochanska, G., & Aksan, N. (2006). Children’s conscience and self-regulation. Journal of Personality, 74, 1587–1617.

Jablonka, E., & Lamb, M. J. (2006). The evolution of information in the major transitions. Journal of Theoretical Biology, 239(2), 236–246.

Kochanska, G., & Aksan, N. (2004). Development of mutual responsiveness between parents and their young children. Child Development, 75, 1657–1676. doi: 10.1111/j.1467-8624.2004.00808.x

Jacobs, N., & Harper, B. (2013). The effects of rejection sensitivity on reactive and proactive aggression. Aggressive behavior, 39(1), 3–12. Jaffee, S. R., Caspi, A., Moffitt, T. E., Dodge, K. A., Rutter, M., Taylor, A., & Tully, L. A. (2005). Nature × nurture: Genetic vulnerabilities interact with physical maltreatment to promote conduct problems. Development and Psychopathology, 17(1), 67–84. doi: 10.1017/S0954579405050042 Ji, L., Wei, X., Chen, L., & Zhang, W. (2012). Peer relationship adversities and children’s aggression during late childhood: The mediating roles of self-conception and peer beliefs. Acta Psychologica Sinica, 44, 1479–1489. doi: 10.3724/SP.J.1041.2012.01479 John, Y. J., Bullock, D., Zikopoulos, B., & Barbas, H. (2013). Anatomy and computational modeling of networks underlying cognitive-emotional interaction. Frontiers in Human Neuroscience, 7, 1–26. doi: 101 .10.3389/fnhum.2013.00101 Kagan, J. (2007). A trio of concerns. Perspectives on Psychological Science, 2, 361–376. doi: 10.1111/j.1745-6916.2007.00049.x Kagan, J. (1997). Conceptualizing psychopathology: The importance of development profiles. Development and Psychopathology, 9, 321–334. doi: 10.1017/S0954579497002071 Kagan, J., & Snidman, N. (1990). Temperamental contributions to human development: The biological characteristics of infants influence their initial behavior to unfamiliar contexts. Research & Clinical Center for Child Development, 12, 59–70. Kagan, J., Snidman, N., McManis, M., Woodward, S., & Hardway, C. (2002). One measure, one meaning: Multiple measures, clearer meaning. Development and Psychopathology, 14(3), 463–475. doi: 10.1017/S0954579402003048 Kahn, R. E., Byrd, A. L., & Pardini, D. A. (2013). Callous-unemotional traits robustly predict future criminal offending in young men. Law and Human Behavior, 37, 87–97. doi: 10.1037/b0000003 Kawa, S., & Giordano, J. (2012). A brief historicity of the Diagnostic and Statistical Manual of Mental Disorders: Issues and implications for

Kochanska, G., Aksan, N., & Joy, M. E. (2007). Children’s fearfulness as a moderator of parenting in early socialization: Two longitudinal studies. Developmental Psychology, 43, 222–237. doi: 10.1111/ j.1467-6494.2006.00421.x

Kroesbergen, E. H., & Van Luit, J. E. (2003). Mathematics interventions for children with special educational needs a meta-analysis. Remedial and special education, 24(2), 97–114. Kunnen, S. (2012a). A dynamic systems approach to adolescent development. New York, NY: Psychology Press. Kunnen, S. (2012b). The art of building dynamic systems models. In S. Kunnen (Ed.), A dynamic systems approach to adolescent development (pp. 99–116). New York, NY: Psychology Press. Lahat, A., Hong, M., & Fox, N. A. (2011). Behavioural inhibition: Is it a risk factor for anxiety?. International Review of Psychiatry, 23, 248–257. doi: 10.3109/09540261.2011.590468 Lane, K. L., Barton-Arwood, S., Nelson, J. R., & Wehby, J. H. (2008). Academic performance of students with emotional and behavioral disorders served in a self-contained setting. Journal of Behavioral Education, 17, 43–62. Lane, K. L., Wehby, J. H. & Barton-Arwood, S. (2005). Students with and at-risk for emotional and behavioral disorders: Meeting their social and academic needs. Preventing School Failure, 49, pp. 6–9. Lewandowski, G. R., & Strohmetz, D. B. (2009). Actions can speak as loud as words: Measuring behavior in psychological science. Social and Personality Psychology Compass, 3, 992–1002. doi: 10.1111/ j.1751-9004.2009.00229.x Lamborn, S. D., Fischer, K. W., & Pipp, S. (1994). Constructive criticism and social lies: A developmental sequence for understanding honesty and kindness in social interactions. Developmental Psychology, 30, 495–508. doi: 10.1037/0012-1649.30.4.495 Lazarus, R. S. (1991). Emotion and adaptation. New York. NY: Oxford University Press. Lerner, R. M., Hess, L. E., & Nitz, K. (1990). A developmental perspective on psychopathology. In M. Hersen & C. G. Last (Eds.), Handbook of child and adult psychopathology: A longitudinal perspective (pp. 9–32). Elmsford, NY: Pergamon Press.

Page 712

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

References Lewis, M. D. (1996). Self-organising cognitive appraisals. Cognition and Emotion, 10, 1–25. doi: 10.1080/026999396380367 Li, J., Wang, L., & Fischer, K. W. (2004). The organization of Chinese shame concepts. Cognition and Emotion, 18, 767–797. Lindhout, I. E., Markus, M. H., Hoogendijk, T. G., & Boer, F. (2009). Temperament and parental child-rearing style: Unique contributions to clinical anxiety disorders in childhood. European Child & Adolescent Psychiatry, 18, 439–446. doi: 10.1007/s00787-009-0753-9 Loeber, R., Stouthamer-Loeber, M., & Green, S. M. (1991). Age at onset of problem behaviour in boys, and later disruptive and delinquent behaviours. Criminal Behaviour and Mental Health, 1, 229–246. Lickliter, R., & Honeycutt, H. (2013). A developmental evolutionary framework for psychology. Review of General Psychology, 17, 184–189. doi: 10.1037/a0032932 Luckner, & Pianta, R. (2011). Teacher–student interactions in fifth grade classrooms: Relations with children’s peer behavior. Journal of Applied Developmental Psychology, 32, 266–275. doi: 10.1016/j.appdev. 2011.02.010 Maddux, J. E., Gosselin, J. T., & Winstead, B. A. (2005). Conceptions of psychopathology: A social constructionist perspective. In J. E. Maddux & B. A. Winstead (Eds.), Psychopathology: Foundations for a contemporary understanding (pp. 3–18). Mahwah, NJ: Lawrence Erlbaum Associates Publishers. Main, M., & Hesse, E. (1990). Parents’ unresolved traumatic experiences are related to infant disorganized attachment status: Is frightened and/or frightening parental behavior the linking mechanism? In M. T. Greenberg, D. Cicchetti, & E. M. Cummings (Eds.), Attachment in the preschool years (pp. 161–182). Chicago, IL: University of Chicago Press. Mascolo, M. F. (2005). Change processes in development: The concept of coactive scaffolding. New Ideas in Psychology, 23, 185–196. doi: 10.1016/j.newideapsych.2006.05.002 Mascolo, M. F., & Fischer, K. W. (2010). The dynamic development of thinking, feeling, and acting over the lifespan. In R. M. Lerner & W. F. Overton (Eds.), Handbook of life-span development. Vol. 1: Biology, cognition, and methods across the lifespan (pp. 149–194). Hoboken NJ: Wiley. doi: 10.1002/9780470880166.hlsd001006 Mascolo, M. F., Fischer, K. W., & Li, J. (2003). Dynamic development of component systems of emotions: Pride, shame and guilt in China and the United States. In R. J. Davidson, K. R. Scherer, & H. Goldsmith (Eds.), Handbook of affective sciences (pp. 375–408). New York, NY: Oxford University Press. Mascolo, M. F., Harkins, D., & Harakal, T. (2000). The dynamic construction of emotion: Varieties in anger. In M. D. Lewis & I. Granic (Eds.), Emotion, development, and self-organization: Dynamic systems approaches to emotional development (pp. 125–152). New York, NY: Cambridge University Press. doi: 10.1017/ CBO9780511527883.007 Martel, M. M., Gremillion, M., Roberts, B., von Eye, A., & Nigg, J. T. (2010). The structure of childhood disruptive behaviors. Psychological Assessment, 22(4), 816–826. doi: 10.1037/a0020975 Martel, M. M., Goth-Owens, T., Martinez-Torteya, C., & Nigg, J. T. (2010). A person-centered personality approach to heterogeneity in attention-deficit/hyperactivity disorder (ADHD). Journal of Abnormal Psychology, 119, 186–196. doi: 10.1037/a0017511 Masten, A. S. (2007). Resilience in developing systems: Progress and promise as the fourth wave rises. Development and Psychopathology, 19, 921–930. McAssey, M., Helm, J., Hsieh, F., Sbarra, D. A., & Ferrer, E. (2013). Methodological advances for detecting physiological synchrony during dyadic interactions. Methodology, 9(2), 41–53. doi: 10.1027/ 1614-2241/a000053

12:37pm

713

McAuliffe, M. D., Hubbard, J. A., Rubin, R. M., Morrow, M. T., & Dearing, K. F. (2006). Reactive and proactive aggression: Stability of constructs and relations to correlates. The Journal of Genetic Psychology: Research and Theory on Human Development, 167, 365–382. doi: 10.1017/S0954579407000442 McMahon, R. J., & Frick, P. J. (2005). Evidence-based assessment of conduct problems in children and adolescents. Journal of Clinical Child and Adolescent Psychology, 34, 477–505. doi: 10.1207/ s15374424jccp3403_6 Merrell, K. W., & Walker, H. M. (2004). Deconstructing a definition: Social maladjustment versus emotional disturbance and moving the EBD field forward. Psychology in the Schools, 41(8), 899–910. doi: 10.1002/pits.20046 Mischel, W., & Mischel, F. (1958). Psychological aspects of spirit possession: A reinforcement analysis. American Anthropologist, 60, 249–260. Molenaar, P. M. (2004). A manifesto on psychology as idiographic science: Bringing the person back into scientific psychology, this time forever. Measurement: Interdisciplinary Research and Perspectives, 2, 201–218. doi: 10.1207/s15366359mea0204_1 Molenaar, P. C., & Campbell, C. G. (2009). The new person-specific paradigm in psychology. Current Directions in Psychological Science, 18(2), 112–117. doi: 10.1111/j.1467-8721.2009.01619.x Molenaar, P. M., & Goode, T. (2013). New methods for sequential behavior analysis. In G. J. Madden, W. V. Dube, T. D. Hackenberg, G. P. Hanley, & K. A. Lattal (Eds.), APA handbook of behavior analysis, Vol. 1: Methods and principles (pp. 267–280). Washington, DC: American Psychological Association. doi: 10.1037/13937-012. Molenaar, P. M., & Ram, N. (2010). Dynamic modeling and optimal control of intraindividual variation: A computational paradigm for nonergodic psychological processes. In S. Chow, E. Ferrer, & F. Hsieh (Eds.), Statistical methods for modeling human dynamics: An interdisciplinary dialogue (pp. 13–37). New York, NY: Routledge/Taylor & Francis Group. Muñoz, L. C., Qualter, P., & Padgett, G. (2011). Empathy and bullying: Exploring the influence of callous-unemotional traits. Child Psychiatry and Human Development, 42, 183–196. doi: 10.1007 /s10578-010-0206-1 Muris, P., van Brakel, A. L., Arntz, A., & Schouten, E. (2011). Behavioral Inhibition as a Risk Factor for the Development of Childhood Anxiety Disorders: A Longitudinal Study. Journal of Child and Family Studies, 20(2), 157–170. doi: 10.1007/s10826-010-9365-8 Myin-Germeys, I., Marcelis, M., Krabbendam, L., Delespaul, P., & van Os, J. (2005). Subtle fluctuations in psychotic phenomena as functional states of abnormal dopamine reactivity in individuals at risk. Biological Psychiatry. 58, 105–110. Nandi, A., Beard, J. R., Galea, S. (2009). Epidemiologic heterogeneity of common mood and anxiety disorders over the lifecourse in the general population: A systematic review. BMC Psychiatry, 9, 31. Navarro, J., Arrieta, C., & Ballén, C. (2007). An approach to the study of the dynamics of work motivation using the diary method. Nonlinear Dynamics, Psychology, and Life Sciences, 11(4), 473–498. Nigg, J. T. (2005). Neuropsychologic theory and findings in attention-deficit/hyperactivity disorder: The state of the field and salient challenges for the coming decade. Biological Psychiatry, 57(11), 1424–1435. doi: 10.1016/j.biopsych.2004.11.011 Orobio de Castro, B., Brendgen, M., Van Boxtel, H., Vitaro, F., & Schaepers, L. (2007). ’Accept me, or else’: Disputed overestimation of social competence predicts increases in proactive aggression. Journal of Abnormal Child Psychology, 35, 165–178. doi: 10.1007/ s10802-006-9063-6 Osgood, C. E., Suci, G. J., & Tannenbaum, P. (1957). The measurement of meaning. Urbana, IL: University of Illinois Press.

Page 713

Trim Size: 8.5in x 11in

714

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

Pearce, M. J., Jones, S. M., Schwab-stone, M. E., & Ruchkin, V. (2003). The protective effects of religiousness and parent involvement on the development of conduct problems among youth exposed to violence. Child development, 74(6), 1682–1696.

Reijntjes, A., Kamphuis, J. H., Prinzie, P., Boelen, P. A., van der Schoot, M., & Telch, M. J. (2011). Prospective linkages between peer victimization and externalizing problems in children: A meta-analysis. Aggressive Behavior, 37, 215–222. doi: 10.1002/ab.20374

Peer, J., Strachan, E., & Spaulding, W. (2008). Heterogeneity in behavioral treatment response in severe mental illness. Journal of Nervous and Mental Disease, 196, 198–206.

Reijntjes, A., Thomaes, S., Bushman, B. J., Boelen, P. A., de Castro, B., & Telch, M. J. (2010). The outcast-lash-out effect in youth: Alienation increases aggression following peer rejection. Psychological Science, 21, 1394–1398. doi: 10.1177/0956797610381509

Pepler, D. J., Craig, W. M., Connolly, J. A., Yuile, A., McMaster, L., & Jiang, D. (2006). A developmental perspective on bullying. Aggressive Behavior, 32, 376–384. doi: 10.1002/ab.20136 Piaget, J. (1951). Play, dreams, and imitation in childhood (C. Gattegno & F.M. Hodgson, Trans.). New York, NY: Harcourt Brace. (Original publication 1946.) Piaget, J. (1957). Logique et équilibre dans les comportements du sujet. Études d’Épistémologie Génétique, 2, 27–118. Piaget, J., & Szeminska, A. (1952). The child’s conception of number (C. Gattegno & F.M. Hodgson, Trans.). London, UK: Routledge & Kegan Paul. (Original publication 1941.)

Romero-Canyas, R., Downey, G., Berenson, K., Ayduk, O., & Kang, N. (2010). Rejection sensitivity and the rejection-hostility link in romantic relationships. Journal of Personality, 7, 119–148. doi: 10.1111/j.1467-6494.2009.00611.x Rose, L. T., Rouhani, P., & Fischer, K. W. (2013). The Science of the Individual. Mind, Brain, & Education, 7, 3, 152–158. Roth-Hanania, R., Davidov, M., & Zahn-Waxler, C. (2011). Empathy development from 8 to 16 months: Early signs of concern for others. Infant Behavior & Development, 34, 447–458. doi: 10.1016/j.infbeh.2011.04.007

Platt, S. A., & Sanislow, C. A. (1988). Norm-of-reaction: Definition and misinterpretation of animal research. Journal of Comparative Psychology, 102, 254–261. doi: 10.1037/0735-7036.102.3.254

Rubin, K. H., Burgess, K. B., Dwyer, K. M., & Hastings, P. D. (2003). Predicting preschoolers’ externalizing behaviors from toddler temperament, conflict, and maternal negativity. Developmental Psychology, 39, 164–176. doi: 10.1037/0012-1649.39.1.164

Poulin, F., & Boivin, M. (2000). Reactive and proactive aggression: Evidence of a two-factor model. Psychological Assessment, 12, 115–122. doi: 10.1037/1040-3590.12.2.115

Rutter, M. (2005). Multiple meanings of a developmental perspective on psychopathology. European Journal of Developmental Psychology, 2, 221–252.

Prior, M., Smart, D., Sanson, A., & Oberklaid, F. (2000). Does shy-inhibited temperament in childhood lead to anxiety problems in adolescence? Journal of the American Academy of Child & Adolescent Psychiatry, 39, 461–468. doi: 10.1097/00004583200004000-00015 Putnam, F. W. (1989). Diagnosis and treatment of multiple personality disorder. New York, NY: Guilford. Putnam, F. W. (1994). Dissociation and disturbances of self. In D. Cicchetti & S. L. Toth (Eds.), Disorders and dysfunctions of the self (Vol. 5, pp. 251–266). Rochester, NY: University of Rochester. Putnam, F. W., Helmers, K., & Trickett, P. K. (1993). Development, reliability, and validity of a child dissociation scale. Child Abuse & Neglect, 17, 731–741. Raeff, C. (2011). Distinguishing between development and change: Reviving organismic-developmental theory. Human Development, 54(1), 4–33. doi: 10.1159/000324334 Rahm, G., Renck, B., & Ringsberg, K. C. (2013). Psychological distress among women who were sexually abused as children. International Journal of Social Welfare, 22(3), 269–278. doi: 10.1111/j.1468-2397.2012.00898.x Ram, N., Brose, A., & Molenaar, P. M. (2013). Dynamic factor analysis: Modeling person-specific process. In T. D. Little (Ed.), The Oxford handbook of quantitative methods: Statistical analysis (Vol. 2, pp. 441–457). New York, NY: Oxford University Press.

Saarni, C., & Lewis, M. (Eds.). (1993). Lying and deception in everyday life. New York, NY: Guilford. Sapienza, J. K., & Masten, A. S. (2011). Understanding and promoting resilience in children and youth. Current Opinion in Psychiatry, 24, 267–273. doi: 10.1097/YCO.0b013e32834776a8 Schensul, J. (2009). Community, culture and sustainability in multilevel dynamic systems intervention science. American Journal of Community Psychology, 43(3–4), 241–256. doi: 10.1007/s10464-009-9228-x Schiepek, G. (2003). A dynamic systems approach to clinical case formulation. European Journal of Psychological Assessment, 19(3), 175–184. doi: 10.1027/1015-5759.19.3.175 Schmittmann, V. D., Cramer, A. O. J., Waldorp, L. J., Epskamp, S., Kievit, R. A., & Borsboom, D. (2013). Deconstructing the construct: A network perspective on psychological phenomena. New Ideas in Psychology, 31, 43–53. doi: 10.1016/j.newideapsych.2011.02.007 Shanahan, S., Jones, J., Thomas-Peter, B. (2011). Are you looking at me, or am I? Anger, aggression, shame and self-worth in violent individuals. Journal of Rational Emotive Cognitive Behavioral Therapy, 29(1), 77–91. doi: 10.1007/s10942-0090105-1 Shaver, P. R., Schwartz, J., Kirson, D., & O’Connor, C. (1987). Emotion knowledge: Further exploration of a prototype approach. Journal of Personality and Social Psychology, 52, 1061–1086. doi: 10.1037/0022-3514.52.6.1061 Shen, L., & Bigsby, E. (2010). Behavioral activation/inhibition systems and emotions: A test of valence vs. action tendency hypotheses. Communication Monographs, 77, 1–26. . doi: 10.1080/03637750903514268

Rapee, R. M., Schniering, C. A., & Hudson, J. L. (2009). Anxiety disorders during childhood and adolescence: Origins and treatment. Annual Review of Clinical Psychology, 5311-341. doi: 10.1146/annurev.clinpsy.032408.153628

Siegler, R. S., & Jenkins, E. (1989). How children discover new strategies. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.

Raya, P. (1996). Pretense in pair-play therapy: Examining the understanding of emotions in young at-risk children. Unpublished doctoral dissertation, Harvard University, Cambridge, MA.

Siegler, R. S., & Stern, E. (1998). Conscious and unconscious strategy discoveries: A micorgenetic analysis. Journal of Experimental Psychology: General, 127, 337–397.

Reid, R., Gonzalez, J. E., Nordness, P. D., Trout, A., & Epstein, M. H. (2004). A meta-analysis of the academic status of students with emotional/behavioral disturbance. The Journal of Special Education, 38, 130–144.

Silk, J. S., Davis, S. S., McMakin, D. L., Dahl, R. E., & Forbes, E. E. (2012). Why do anxious children become depressed teenagers? The role of social evaluative threat and reward processing. Psychological Medicine, 42, 2095–2107. doi: 10.1017/S0033291712000207

Page 714

Trim Size: 8.5in x 11in

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

References Smith, L. B. (2005). Cognition as a dynamic system: Principles from embodiment. Developmental Review, 25, 278–298. doi: 10.1016/ j.dr.2005.11.001 Sneed, J. R., Hamagami, F., McArdle, J. J., Cohen, P., & Chen, H. (2007). The dynamic interdependence of developmental domains across emerging adulthood. Journal of Youth and Adolescence, 36, 351–362. doi: 10.1007/s10964-006-9081-2 Sommerfeld, P., & Hollenstein, L. (2011). Searching for appropriate ways to face the challenges of complexity and dynamics. British Journal of Social Work, 41(4), 668–688. doi: 10.1093/bjsw/bcr057 Sonuga-Barke, E., Bitsakou, P., & Thompson, M. (2010). Beyond the dual pathway model: Evidence for the dissociation of timing, inhibitory, and delay-related impairments in attention-deficit/hyperactivity disorder. Journal of the American Academy of Child & Adolescent Psychiatry, 49, 345–355. doi: 10.1097/00004583-201004000-00009 Sonuga-Barke, E. S., Wiersema, J. R., van der Meere, J. J., & Roeyers, H. (2010). Context-dependent dynamic processes in attention deficit/hyperactivity disorder: Differentiating common and unique effects of state regulation deficits and delay aversion. Neuropsychology Review, 20, 86–102. doi: 10.1007/s11065-009-9115-0 Sonuga-Barke, E. S., Sergeant, J. A., Nigg, J., & Willcutt, E. (2008). Executive dysfunction and delay aversion in attention deficit hyperactivity disorder: Nosologic and diagnostic implications. Child and Adolescent Psychiatric Clinics of North America, 17(2), 367–384. doi: 10.1016/j.chc.2007.11.008 Spencer, J. P., Austin, A., & Schutte, A. R. (2012). Contributions of dynamic systems theory to cognitive development. Cognitive Development, 27(4), 410–418. doi: http://dx.doi.org/10.1016/j.cogdev .2012.07.006 Spencer, T. J., Biederman, J., & Mick, E. (2007). Attentiondeficit/hyperactivity disorder: Diagnosis, lifespan, comorbidities, and neurobiology. Journal of Pediatric Psychology, 32, 631–642. doi: 10.1093/jpepsy/jsm005 Sportel, B., Nauta, M. H., de Hullu, E., & de Jong, P. J. (2013). Predicting internalizing symptoms over a two year period by BIS, FFFS and attentional control. Personality and Individual Differences, 54, 236–240. doi: 10.1016/j.paid.2012.08.043 Sroufe, L. (2007). The place of development in developmental psychopathology. In A. S. Masten (Ed.), Multilevel dynamics in developmental psychopathology: Pathways to the future (pp. 285–299). New York, NY: Taylor & Francis Group/Lawrence Erlbaum Associates. Steenbeek, H., Jansen, L., & van Geert, P. (2012). Scaffolding dynamics and the emergence of problematic learning trajectories. Learning and Individual Differences, 22, 64–75. doi: 10.1016/j.lindif.2011.11.014 Steenbeek, H., & van Geert, P. (2013). The emergence of learning-teaching trajectories in education: a complex dynamic systems approach. Nonlinear Dynamics, Psychology, and Life Sciences, 17(2), 233–267. Steenbeek, H., & Van Geert, P. (2008). An empirical validation of a dynamic systems model of interaction: do children of different sociometric statuses differ in their dyadic play? Developmental Science, 11, 253–281. doi: 10.1111/j.1467-7687.2007.00655.x Steenbeek, H. W., & van Geert, P. L. C. (2007). A theory and dynamic model of dyadic interaction: Concerns, appraisals, and contagiousness in a developmental context. Developmental Review, 27(1), 1–40. doi: 10.1016/j.dr.2006.06.002 Stein, Z., Dawson, T., & Fischer, K. W. (2010). Redesigning testing: Operationalizing the new science of learning. In M. S. Khine & I. M. Saleh (Eds.), New science of learning: Cognition, computers, and collaboration in education (pp. 207–224). New York, NY: Springer. Stone, C. (1998). The metaphor of scaffolding: Its utility for the field of learning disabilities. Journal of Learning Disabilities, 31, 344–364. doi: 10.1177/002221949803100404

12:37pm

715

Sullivan, H. S. (1953). The interpersonal theory of psychiatry. New York, NY: Norton. Sutherland, K. S., Alder, N., & Gunter, P. L. (2003). The effect of increased rates of opportunities to respond on the classroom behavior of students with emotional/behavioral disorders. Journal of Emotional and Behavioral Disorders, 11(4), 239–248. Tangney, J. P., Wagner, P., Fletcher, C., & Gramzow, R. (1992). Shamed into anger? The relation of shame and guilt to anger and self-reported aggression. Journal of Personality and Social Psychology, 62(4), 669–675. doi: 10.1037/0022-3514.62.4.669 Tavory, I., Jablonka, E., & Ginsburg, S. (2012). Culture and Epigenetics. Terr, L. C. (1991). Childhood traumas: An outline and overview. American Journal of Psychiatry, 148, 10–20. Thelen, E., & Smith, L. B. (2006). Dynamic systems theories. In R. M. Lerner & W. Damon (Eds.), Handbook of child psychology: Theoretical models of human development (6th ed., Vol. 1, pp. 258–312). Hoboken, NJ: John Wiley & Sons Inc. Thomas, K. M., Hopwood, C. J., Woody, E., Ethier, N., & Sadler, P. (2013). Momentary assessment of interpersonal process in psychotherapy. Journal of Counseling Psychology, 1-14, 61(1). doi: 10.1037/a0034277 Thornton, L. C., Frick, P. J., Crapanzano, A., & Terranova, A. M. (2013). The incremental utility of callous-unemotional traits and conduct problems in predicting aggression and bullying in a community sample of boys and girls. Psychological Assessment, 25, 366–378. doi: 10.1037/a0031153 Tomasi, D. , & Volkow, N. D. (2012). Abnormal functional connectivity in children with attention-deficit/hyperactivity disorder. Biological Psychiatry, 71, 443–450. doi: 10.1016/j.biopsych.2011.11.003 Trickett, E., & Schensul, J. (2009). Summary comments: Multi-level community based culturally situated interventions. American Journal of Community Psychology, 43(3–4), 377–381. doi: 10.1007/s10464009-9237-9 Tseng, V., & Seidman, E. (2007). A systems framework for understanding social settings. American Journal of Community Psychology, 39(3–4), 217–228. doi: 10.1007/s10464-007-9101-8 Ursache, A., Blair, C., Stifter, C., & Voegtline, K. (2013). Emotional reactivity and regulation in infancy interact to predict executive functioning in early childhood. Developmental Psychology, 49, 127–137. doi: 10.1037/a0027728 Vaillant, G. E. (1977). Adaptation to life. Boston, MA: Little, Brown. Valentino, K., Cicchetti, D., Rogosch, F. A., & Toth, S. L. (2008). True and false recall and dissociation among maltreated children: The role of self-schema. Development and Psychopathology, 20, 213–232. doi: 10.1017/S0954579408000102 van Derbur Atler, M. (1991). Address presented at the Kempe National Center for Prevention and Treatment of Child Abuse and Neglect, Denver, CO. van der Aalsvoort, G., van Geert, P., & Steenbeek, H. (2008). Possibilities and drawbacks of microgenetic methodology related to research on learning and instruction. In: Kumpulainen, K. (Ed.). Investigating Classroom interaction: Methodologies in action. pp. 245–278. Sense publishers. Rotterdam. van der Meere, J., & Stemerdink, N. (1999). The development of state regulation in normal children: An indirect comparison with children with ADHD. Developmental Neuropsychology, 16, 213–225. van Geert, P. (2009). Nonlinear complex dynamical systems in developmental psychology. In S. J. Guastello, M. Koopmans, & D. Pincus (Eds.), Chaos and complexity in psychology: The theory of nonlinear dynamical systems (pp. 242–281). New York, NY: Cambridge University Press.

Page 715

Trim Size: 8.5in x 11in

716

Cicchetti

c15.tex V3 - Volume I - 12/10/2015

12:37pm

What Can Dynamic Systems Models of Development Offer to the Study of Developmental Psychopathology?

van Geert, P. (2004). A dynamic systems approach to diagnostic measurement of SLI. In L. Verhoeven & H. van Balkom (Eds.), Classification of developmental language disorders: Theoretical issues and clinical implications (pp. 327–348). Mahwah, NJ: Lawrence Erlbaum Associates Publishers. Van Geert, P. (1994). Dynamic systems of development: Change between complexity and chaos. New York, NY: Harvester Wheatsheaf. Van Geert, P. (1986). The concept of development. Advances in Psychology, 36, 3–50. van Geert, P. P., & Steenbeek, H. H. (2005). The dynamics of scaffolding. New Ideas in Psychology, 23(3), 115–128. doi: 10.1016 /j.newideapsych.2006.05.003 van Rijn, P. W., & Molenaar, P. M. (2005). Logistic models for single-subject time series. In L. van der Ark, M. A. Croon, & K. Sijtsma (Eds.), New developments in categorial data analysis for the social and behavioral sciences (pp. 125–145). Mahwah, NJ: Lawrence Erlbaum Associates Publishers. doi: 10.1080/08856250802596717 Vinogradov, S., Fisher, M., & de Villers-Sidani, E. (2012). Cognitive training for impaired neural systems in neuropsychiatric illness. Neuropsychopharmacology, 37, 43–76. doi: 10.1038/npp.2011.251 Visser, M., Kunnen, S. E., & Van Geert, P. L. (2010). The impact of context on the development of aggressive behavior in special elementary school children. Mind, Brain, and Education, 4(1), 34–43 Visser, M., Singer, E., van Geert, P. L., & Kunnen, S. E. (2009). What makes children behave aggressively? The inner logic of Dutch children in special education. European Journal of Special Needs Education, 24(1), 1–20. doi: 10.1080/08856250802596717 Vitaro, F., & Brendgen, M. (2005). Proactive and reactive aggression: A Developmental perspective. In R. E. Tremblay, W. W. Hartup, & J. Archer (Eds.), Developmental origins of aggression (pp. 178–201). New York, NY: Guilford Press. von Eye, A., & Bergman, L. R. (2003). Research strategies in developmental psychopathology: Dimensional identity and the person-oriented approach. Development and Psychopathology, 15, 553–580. Vondra, J. I., Barnett, D., & Cicchetti, D. (1989). Perceived competence and actual competence among maltreated and comparison school children. Development and Psychopathology, 1, 237–255. doi: 10.1017/S0954579400000432 Wakschlag, L. S., Tolan, P. H., & Leventhal, B. L. (2010). Research review: ’Ain’t misbehavin’: Towards a developmentally-specified nosology for preschool disruptive behavior. Journal of Child Psychology and Psychiatry, 51, 3–22. doi: 10.1111/j.1469-7610.2009.02184.x Wallbott, H. G., & Scherer, K. R. (1993). Cultural differences in experiencing shame and guilt. In J. P. Tangney & K. W. Fischer (Ed.), The self-conscious emotions: Pride, shame, and guilt (pp. 465–487). New York, NY: Guilford. Warren, S. L., Huston, L., Egeland, B., & Sroufe, L. (1997). Child and adolescent anxiety disorders and early attachment. Journal of the American Academy of Child & Adolescent Psychiatry, 36, 637–644. doi: 10.1097/00004583-199705000-00014 Waschbusch, D. A., Porter, S., Carrey, N., Kazmi, S., Roach, K. A., & D’Amico, D. A. (2004). Investigation of the heterogeneity of disruptive behaviour in elementary-age children. Canadian Journal of Behavioural Science, 36, 97–11. doi: 10.1037/h0087221 Watson, M. W., Fischer, K. W., Andreas, J., & Smith, K. W. (2004). Pathways to aggression in children and adolescents. Harvard Educational Review, 74, 404–430. Watson, M. W., & Getz, K. (1991). The relationship between oedipal behaviors and children’s family role concepts. Merrill-Palmer Quarterly, 36, 487–505. Weaver, I. C., Champagne, F. A., Brown, S. E., Dymov, S., Sharma, S., Meaney, M. J., Szyf, M. (2005). Reversal of maternal programming

of stress responses in adult offspring through methyl supplementation: altering epigenetic marking later in life. Journal of Neuroscience, 25(47), 11045–11054. Wehby, J. H., Lane, K. L., & Falk, K. B. (2003). Academic instruction for students with emotional and behavioral disorders. Journal of Emotional and Behavioral Disorders, 11, 194–197. Westen, D., & Cohen, R. (1992). The self in borderline personality disorder: A psychodynamic perspective. In Z. Segal & S. Blatt (Eds.), Self-representation and emotional disorder: Cognitive and psychodynamic perspectives (pp. 334–368). New York, NY: Guilford Press. Westen, D., Malone, J. C., & DeFife, J. A. (2012). An empirically derived approach to theclassification and diagnosis of mood disorders. World Psychiatry, 11, 172–180. White, B. A., & Kistner, J. A. (2011). Biased self-perceptions, peer rejection, and aggression in children. Journal of Abnormal Child Psychology, 39, 645–656. doi: 10.1007/s10802-011-9506-6 Wichstrøm, L., Berg-Nielsen, T., Angold, A., Egger, H., Solheim, E., & Sveen, T. (2012). Prevalence of psychiatric disorders in preschoolers. Journal of Child Psychology and Psychiatry, 53, 695–705. doi: 10.1111/j.1469-7610.2011.02514.x Wilson, L. C., & Scarpa, A. (2011). The link between sensation seeking and aggression: A meta-analytic review. Aggressive Behavior, 37, 81–90. doi: 10.1002/ab.20369 Winkielman, P., Niedenthal, P. M., & Oberman, L. M. (2009). Embodied perspective on emotion-cognition interactions. In J. A. Pineda (Ed.), Mirror neuron systems: The role of mirroring processes in social cognition (pp. 235–257). Totowa, NJ: Humana Press. Winsler, A. (1998). Parent–child interaction and private speech in boys with ADHD. Applied Developmental Science, 2, 17–39. doi: 10.1207/s1532480xads0201_2 Witherington, D. C. (2011). Taking emergence seriously: The centrality of circular causality for dynamic systems approaches to development. Human Development, 54, 66–92. doi: 10.1159/000326814 Witherington, D. C. (2007). The dynamic systems approach as metatheory for developmental psychology. Human Development, 50, 127–153. doi: 10.1159/000100943 Witherington, D. C., & Crichton, J. A. (2007). Emotion, 7, 628–637. doi: 10.1037/1528-3542.7.3.628 Witherington, D. C., & Margett, T. E. (2009). Systems and dynamic systems: The search for inclusive merger. Human Development, 52, 251–256, doi: 10.1159/000215075 Wolf, C., & Linden, D. E. J. (2012). Biological pathways to adaptability – interactions between genome, epigenome, nervous system and environment for adaptive behavior. Genes, Brain and Behavior, 11: 3–28. doi: 10.1111/j.1601-183X.2011.00752.x Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. Wright, K., Gudjonsson, G. H., & Young, S. (2008). An investigation of the relationship between anger and offence-related shame and guilt. Psychology, Crime & Law, 14, 415–423. doi: 10.1080 /10683160701770369 Wundt, W. (1907). Outlines of psychology (Judd, C.H., Trans.). New York, NY: Stechert. (Original publication 1905). Yoshikawa, H., & Hsueh, J. (2001). Child development and public policy: Toward a dynamic systems perspective. Child Development, 72(6), 1887–1903. doi: 10.1111/1467-8624.00384 Zachar, P., & Kendler, K. S. (2010). Philosophical issues in the classification of psychopathology. In T. Millon, R. F. Krueger, & E. Simonsen (Eds.), Contemporary directions in psychopathology: Scientific foundations of the DSM-V and ICD-11 (pp. 127–148). New York, NY: Guilford Press.

Page 716

CHAPTER 16

A Survey of Dynamic Systems Methods for Developmental Psychopathology ISABELA GRANIC, TOM HOLLENSTEIN, and ANNA LICHTWARCK-ASCHOFF

INTRODUCTION 717 THE DEVELOPMENTALIST’S DILEMMA 718 PRINCIPLES OF DYNAMIC SYSTEMS 719 State Space, Attractors, and Dynamic Stability 720 Interrelations Between Time Scales 720 Hierarchically Embedded Levels of Organization 721 Perturbations, Phase Transitions, and Nonlinear Change 721 RESEARCH DESIGN STRATEGIES INFORMED BY DS PRINCIPLES 723 SUITABLE DATA FOR DS ANALYSES 725 DS METHODS 727 Real-Time Measures 727

Event History Analysis 731 Developmental-Time Measures 738 Latent Class Analysis 740 STATE SPACE GRID ANALYSIS: A GRAPHICAL AND STATISTICAL MIDDLE ROAD 746 State Space Grids 746 State Space Grid Analysis: Within Grids 748 State Space Grid Analysis: Between-Grid Analysis 750 FUTURE DIRECTIONS: IMPLICATIONS FOR CLINICAL RESEARCH 750 CONCLUSION 752 REFERENCES 752

INTRODUCTION

paradigm. Many of them have developed heuristically rich, complex models based on open systems concepts, but little headway has been made in finding alternative analytic tools appropriate for testing these models—a predicament Richters (1997) dubbed the developmentalist’s dilemma. In this chapter, we suggest that methods derived from dynamic systems (DS) theory may be useful for addressing this dilemma. There are three main objectives: (1) to outline key dynamic systems principles and highlight their commensurability with developmental psychopathologists’ core conceptual concerns; (2) to provide a survey of research designs, methodological techniques, and measurement strategies currently being used and refined by developmental DS researchers; and (3) to elaborate on one specific DS method, state space grid analysis and provide several empirical examples using this method. The state space grid approach (Lewis, Lamey, & Douglas, 1999) was developed as a middle road between DS methods that are mathematically demanding on the one hand (and, thus, often inaccessible or inappropriate for developmental psychopathology) and purely descriptive on the other. We have included most DS methods currently available for developmentalists. The methodological techniques and specific examples we have selected were chosen because

Decades ago, Lewin (1931) criticized psychology for its overreliance on methodologies that were originally designed for studying closed, physical systems. These methods, he warned, were inappropriate for the study of complex, developmental processes. Over 70 years later, the same criticisms continue to be raised with increasing urgency. Specifically, leading theorists (Cicchetti & Cohen, 1995b; Ford & Lerner, 1992; Hinde, 1992; Kagan, 1992; Keating, 1990; Overton & Horowitz, 1991; Richters, 1997) suggest that there is a fundamental incompatibility between developmentalists’ organismic, open systems models and the mechanistic research methods with which these models are tested. Developmental psychopathologists have been particularly concerned with their inherited mechanistic An earlier, less detailed version of this chapter appeared in Development and Psychopathology, 15 (2003). We gratefully acknowledge the financial support of Grant 1 R21 MH 67357 from the National Institute of Mental Health (to D. Pepler and I. Granic) and Grant 410-2003-1335 from the Social Sciences and Humanities Research Council of Canada (to the first author). Color versions of Figures 16.7 and 16.14 are available at http://onlinelibrary.wiley.com/book/10.1002/9781118963418 717

718

DS Methods

they seemed most appropriate for addressing the types of research questions developmental psychopathologists tend to pursue. In the end, however, our main purpose is to provide a clear enough picture of various DS methods to inspire developmental psychopathologists to expand their analytic repertoire and, thus, test their systems-based models more directly. THE DEVELOPMENTALIST’S DILEMMA Before proceeding to discuss DS theory and its methodological bag of tricks, we will review briefly the strong rationale for developing new methods for the field of developmental psychopathology (also see Cicchetti & Cohen, 1995b). Developmental psychopathologists have adopted an organismic, holistic, transactional framework for conceptualizing individual differences in normal and atypical development (e.g., Cicchetti, 1993; Cicchetti & Cohen, 1995a, 1995b; Cummings, Davies, & Campbell, 2000; Garmezy & Rutter, 1983; Sameroff, 1983, 1995; Sroufe & Rutter, 1984). These scholars often frame their models in terms of organizational principles and systems language. The systems theories that inform these models include General Systems Theory (Sameroff, 1983, 1995; von Bertalanffy, 1968); Developmental Systems Theory (Ford & Lerner, 1992); the ecological framework (Bronfenbrenner, 1979); contextualism (Dixon & Lerner, 1988); the transactional perspective (Dumas, LaFreniere, & Serketich, 1995); the organizational approach (Cicchetti & Schneider-Rosen, 1986; Erickson, Egeland, & Pianta, 1989; Garmezy, 1974; Sroufe, 1979, 1986; Sroufe & Rutter, 1984); the holistic-interactionistic view (Bergman & Magnusson, 1997); and the epigenetic view (Gottlieb, 1991, 1992). As a class of models, these approaches focus on process-level accounts of human behavior and on the context dependence and heterogeneity of developmental phenomena. They are concerned with the equi- and multifinality of development, the hierarchically embedded nature of intrapersonal (e.g., neurochemical activity, cognitive and emotional processes), interpersonal (e.g., parent–child relationships; peer networks), and higher order social systems (e.g., communities, cultures). They are also fundamentally concerned with the mechanisms that underlie change and novelty (as well as stability) in normal and clinically significant trajectories. As a result of inadequate measurement techniques, however, the complex developmental models informed by the language of systems thinking remain largely untested (Richters, 1997). For example, in the field of childhood aggression, a number of leading scholars have become concerned with highlighting the heterogeneity of aggressive

youth and advocating the development of causal models that recognize the equifinality of aggression (e.g., Cicchetti & Richters, 1993; Hinshaw & Zupan, 1997; Moffitt, 1993). But it remains difficult to test these models because most of our current research methods and analytic techniques (e.g., regression analysis, t-tests, path analysis) rely on strategies that aggregate overtly similar subjects into one group or another (e.g., aggressive and nonaggressive children) to conduct group-level statistical analyses. Thus, although we may know that aggressive children show the same behavioral pattern for very different reasons (e.g., abuse, permissive parenting, marital conflict, birth of a new sibling), this variability cannot be systematically addressed because multivariate analytic strategies carry an a priori assumption of within-group homogeneity. This is not just a niggling statistical detail. Several leading methodologists have argued that in most cases, these assumptions are completely unfounded and have likely led to serious misinterpretations of data (e.g., Hinshaw, 1999; Lykken, 1991; Meehl, 1978; Richters, 1997). How did this gap between methods and models come about? Numerous past critics (e.g., Hinde, 1992; Meehl, 1978; Lykken, 1991; Overton & Horowitz, 1991; Richters, 1997) have pointed to psychology’s original sin for an explanation: in an effort to gain credibility and align itself with the hard sciences, psychology appropriated the methods and analytic techniques of mechanistic, nineteenth-century physical sciences. This paradigm is inappropriate for the study of self-organizing, active, reactive, interactive, and adaptive organisms (i.e., the stuff of psychology). The irony is that psychology’s embrace of this mechanistic paradigm came at the same time as the physical and biological sciences were advancing a radically new one—one based on open systems concepts (Overton & Horowitz, 1991; Richters, 1997). For some domains of psychology, adopting techniques from statistical mechanics may not be as paralyzing as it has become for developmental psychology (Thelen & Smith, 1994, 1998; Thelen & Ulrich, 1991; van Geert, 1998a, 1998c). At the heart of developmental questions, however, is how things change. By what process do novel structures (e.g., formal operational thought) or skills (e.g., emotion regulation, walking, language) emerge? The pioneers of developmental science (i.e., Piaget, Vygotsky, Werner) concerned themselves with the pursuit of abstract laws or properties that govern development: the structural explanation of how development unfolds (Cicchetti, 1990; van Geert, 1998a, 1998c). As van Geert (1998a) argued, however, change and the emergence of novelty may no longer be the focus of contemporary developmental psychology. This state of affairs can be traced to the

Principles of Dynamic Systems

“adoption of a statistics that was designed for different purposes, namely distinguishing populations characterized by some special feature . . . and estimating the linear association between the variance of some independent variable on the one hand and a dependent variable on the other” (p. 146). The adoption of such statistics seems to have missed the point of the original questions laid out by the founding scholars of developmental science (Thelen & Smith, 1994; van Geert, 1998a). Developmental psychopathology, having grown in part from this tradition, inherited the same schism. But developmental psychopathology also grew from several other disciplines that were less hampered by this statistical baggage. Embryology (e.g., Kuo, 1967; Spemann, 1938) was one of these disciplines (for a review, see Cicchetti, 1990). In the pursuit of understanding normal embryological functioning, “early embryologists derived the principles of a dynamically active organism and of a hierarchically integrated system that were later used in investigations of the processes of abnormal development within the neurosciences, embryology, and experimental psychopathology” (Cicchetti, 1990, p. 4). These insights formed the basis for the organizational approach in developmental psychopathology. Another avenue of influence comes from the discipline of ethology, which views the organism as a comprehensive whole and advocates a “multidomain, multidiscipline approach to psychopathology,” (Cicchetti, 1990, p. 19), an approach that is at the heart of developmental psychopathology (see also Hinde, 1983). Interestingly, developmental psychopathology also shares its roots with the founders of child psychoanalysis (e.g., A. Freud, Klein, Winnicott, Bowlby, Erikson) who, like Piaget and Vygotsky, formalized theories based on detailed observations of children in their natural environments (Cicchetti, 1990; Cicchetti & Cohen, 1995b). These original, individual-based, ethnographic methods, however, are rarely applied in contemporary research (but see Cicchetti & Aber, 1998, for an exception). Yet it seems clear to us that these organismic, systems-oriented roots lend themselves well to DS principles and the methods they suggest. As it stands now, the field of developmental psychopathology seems to be at an impasse. On one hand, some scholars have suggested that systems approaches to studying development have provided an interesting metaphor, but offer little more (Cox & Paley, 1997; Reis, Collins, & Bersheid, 2000; Vetere & Gale, 1987). Thus, one option is to give up the grail, abandon this well-intentioned enterprise, and build simpler, more linear models that can be tested with established statistical rigor. Another option is offered by Richters (1997):

719

“Resolving the developmentalist’s dilemma will require more than a recognition of the inadequacies of the existing paradigm. It will require intensive efforts to develop indigenous research strategies, methods, and standards with fidelity to the complexity of developmental phenomena” (p. 226). In the closing of his essay, Richters offered some general instructions for how this new generation of studies should proceed: (1) There should be intense focus placed on understanding individuals and the causal structures that underpin specific individuals’ development, with particular attention paid to well-characterized exemplars (i.e., nonextreme) cases; (2) no single method should be held as superior or inferior (e.g., case based, variable based, cross sectional, longitudinal, historical, ethnographic), but instead methodological pluralism should be encouraged and may vary in degree depending on the phenomena investigated; (3) ritualized hypothesis testing should be generally abandoned for more exploratory, creative approaches that emphasize the discovery process; and (4) a narrow focus on explained variance and prediction should take a secondary role to explanatory power. These directions provide the springboard from which to consider the potential contributions dynamic systems methods can make to addressing systems-inspired models.

PRINCIPLES OF DYNAMIC SYSTEMS Developmental psychopathologists will be familiar with most of the concepts in DS theory because of their long-standing familiarity with systems concepts in general. Nevertheless, for the sake of clarity and precision, we believe it is important to delineate this framework from the systems approaches mentioned previously (Lewis & Granic, 2000). Formally, a dynamical system is a set of mathematical equations that specify how a system changes over time. The various patterns and processes that emerge from this set of equations rely on a technical language originally developed in the fields of mathematics and physics. The concepts derived from this mathematical framework comprise the principles of dynamic systems approaches. The terms that are most commonly associated with this framework include attractors, repellors, state space, perturbations, bifurcations, catastrophe, chaos, hysteresis, complexity, nonlinearity, far from equilibrium, and so on. Thus, what we refer to as dynamic systems theory or dynamic systems principles is a meta-theoretical framework (Witherington, 2007) that encompasses a set of abstract principles that have been applied in different disciplines (e.g., physics, chemistry, biology, psychology)

720

DS Methods

and to various phenomena (e.g., lasers, ant colonies, brain dynamics) at vastly different scales of analysis (from cells to economic trends and from milliseconds to millennia). Consistent with developmental DS theorists (e.g., Fogel, 1993; Lewis, 2000; Thelen & Smith, 1994, 1998), we use the term dynamic systems to refer to the systems themselves (not the equations) that change over time. DS principles provide a framework for describing how novel forms emerge and stabilize through a system’s own internal feedback processes (Prigogine & Stengers, 1984). This process is known as self-organization and refers to the spontaneously generated (i.e., emergent) order in complex, adaptive systems. In fields as various as physics (e.g., Haken, 1977), chemistry (e.g., Prigogine & Stengers, 1984), biology (e.g., Kauffman, 1993), and neuroscience (Freeman, 1995), DS principles have proven essential for providing process-level accounts of the structure and organization of system behavior, and changes in that structure over time (Lewis & Granic, 2000). DS principles resonate with most systems concepts in general. Dynamic systems approaches to development emphasize the multiple reciprocal interactions among system elements that are hierarchically nested and mutually influential. The context or ecology in which the system is embedded is critical for understanding a dynamic system’s behavior (Hollenstein, 2012). Also consistent with various systems perspectives, development is conceptualized as movement toward greater levels of complexity through the interplay between positive and negative feedback cycles (Lewis & Granic, 1999b). As will become clear from our selection of methods, we are most strongly influenced by the pioneering work of Esther Thelen, Linda Smith, Alan Fogel, Marc Lewis, and Paul van Geert—developmental psychologists who brought DS principles to the attention of the field at large. Because our focus is on methods specifically, a thorough theoretical discussion of DS concepts and their relevance to developmental science is precluded; thus, the reader is strongly encouraged to supplement the current theoretical introduction with reviews by Thelen & Smith (1994, 1998), Fogel (1993), and Lewis (2000; 2011). In the following discussion, we highlight some key principles and then move on to their methodological implications.

range of these possibilities. Stable patterns emerge through feedback among many lower order (more basic) system elements; these emergent patterns are referred to as attractors in DS terminology. In real time, attractors may be understood as absorbing states that “pull” the system from other potential states. Behavior moves in a trajectory across the state space toward these attractors. Over developmental time, attractors represent recurrent patterns that have stabilized and are increasingly predictable. As noted by Thelen and Smith (1994), all developmental acquisitions can be described as attractor patterns that emerge over weeks, months, or years. As recurring stable forms, attractors are often represented topographically as valleys on a dynamic landscape (Figure 16.1). The deeper and wider the attractor, the more likely it is that behavior falls into it and remains there, and the more resistant it is to small changes in the environment. Repellors, or states that the system tends to avoid or be repelled from, are represented as hills on this landscape—states in which the system cannot rest. As the system develops, a unique state space, defined as a model of all possible states a system can attain, is configured by several attractors and repellors. Critically, living systems are characterized by multistability (Kelso, 1995); that is, their state space (i.e., behavioral repertoire) includes several coexisting attractors. Contextual constraints probabilistically guide behavior toward a particular attractor at any given moment in time. As we will see later, the concepts of state space, attractors and multistability have informed several research designs and methodologies in recent years. The operationalization of these principles, either graphically, mathematically, or heuristically, have helped DS researchers uncover previously undetected behavioral variability, as well as the processes by which this variability stabilizes into unpredicted, but nevertheless stable, attractor patterns. Interrelations Between Time Scales DS researchers are always fundamentally concerned with the interplay between different time scales. From a DS

State Space, Attractors, and Dynamic Stability Dynamic, self-organizing systems share several key properties, some of which have already been mentioned. One key feature of open systems is that, although theoretically they have the potential to exhibit an enormous number of behavioral patterns, they tend to stabilize in a limited

Figure 16.1 A state space with three attractors (the wells) and one repellor (the hill). Source: Hollenstein T. (2007). State space grids: Analyzing dynamics across development. International Journal of Behavioral Development, 31, 384–396.

Principles of Dynamic Systems

perspective, the same principles of change and stability can be applied at the moment-to-moment scale (real time) as they can to developmental time (weeks, months, years). The interplay between nested time scales is constant and reciprocal (Thelen & Smith, 1994, 1998). Self-organization at the real-time scale constrains self-organization at the developmental scale, which in turn constrains real-time behavior (van Gelder & Port, 1995). Thelen and Smith (1998) elaborated: . . . each behavioral act occurs over time . . . but every act changes the overall system and builds a history of acts over time . . . Habituation, memory, learning, adaptation, and development form one seamless web built on processes over time—activities in the real world. (p. 593)

Research designs based on DS principles almost always measure behavior on at least two time scales (Hollenstein, 2012; Hollenstein, Lichtwarck-Aschoff, & Potworowski, 2013). The manner and extent to which the two or more levels of analysis are related to each other is subsequently examined. Thus, DS-informed studies often involve collecting real-time, observational data over repeated sessions in a longitudinal design such that moment-tomoment behavioral patterns, and changes in those patterns, can be traced along a longitudinal trajectory (Lee & Karmiloff-Smith, 2002). Hierarchically Embedded Levels of Organization Dynamic systems are not only nested processes in time, they are also coupled hierarchically. Mechanisms in development may be defined at different levels of organization (e.g., physiological, emotional, behavioral, social) and these levels are thought to be embedded, linked by feedback processes, and mutually constraining. Importantly, DS theorists insist that no level of organization is any more “basic” or “primary” in terms of causality. Thus, for example, from a DS perspective, neural processes are accorded no higher causal privilege than psychological or social processes. Instead, “descriptions of change of many components are needed so that multilevel processes and their mutual interactions can be fully integrated . . . Moreover, explanations at every level must be consistent and ultimately reconcilable (Thelen & Smith, 1998, pp. 596–597). This view of causality contrasts with most approaches in psychology (but see Cicchetti & Dawson, 2002a). Rather than simple linear relationships, DS researchers often understand relations among multiple levels in terms of circular causality (Hakens, 1977). Circular causality suggests that interactions among lower order elements provide the means by which higher order patterns emerge

721

and these emergent attractors exert top-down influences to maintain the entrainment of lower level components. Developmental psychopathologists, particularly those in the neurosciences, have recently become interested in understanding how different levels of analysis can shed light on typical and atypical development (Calkins & Fox, 2002; Cicchetti & Dawson, 2002b; Nelson et al., 2002). The DS meta-theoretical framework, with concepts such as circular causality, may help researchers organize their findings and parsimoniously integrate the various theories about different levels into one coherent model. Indeed, it is likely that common dynamic principles of self-organization govern different levels of organization (e.g., neural, cardiac, emotional, behavioral, societal); if this is so, then synergistic DS methods that can be applied similarly at different levels of analysis will provide a powerful means by which integrative models can be developed and empirically supported. The extent to which real- and developmental-time scales are interrelated and hierarchical levels of organization are embedded is further clarified when considering perturbations and their relation to phase transitions. Perturbations, Phase Transitions, and Nonlinear Change Through the amplification properties of positive feedback, nonlinear changes in the organizational structure of a dynamic system can be observed. As shown in Figure 16.2, these abrupt changes are referred to as phase transitions and they occur at points of bifurcation, or junctures in the

Figure 16.2 Alternative developmental trajectories. Phase transitions occur at regular junctures in development. Increased variability at phase transitions is shown in magnified segment. Source: From Lewis, M. D. & Douglas, L. (1998) A dynamic systems approach to cognition-emotion interactions in development (pp. 159–188), in What Develops in Emotional Development? M. F. Mascolo & S. Griffins (Eds.). New York: Plenum Press.

722

DS Methods

system’s development. At these thresholds, small fluctuations have the potential to disproportionately affect the interactions of other elements leading to the emergence of new forms. Novelty does not have to originate from outside the system; it can emerge spontaneously through feedback within the system. During a phase transition, systems are extremely sensitive to perturbations. Between these points, however, self-organizing systems tend toward coherence and stability. Phase transitions are characterized by interrelated changes in real and developmental time. In developmental time, periods of stability and relative predictability are followed by a period of disequilibrium in which established patterns are destabilized. After this period of flux, developmental systems restabilize and settle into new habits of interactions. Corresponding to this developmental profile, real-time behavior during a phase transition is more variable, flexible, and more sensitive to perturbations; behavior may change from one state to another frequently, and is less likely to settle in any one state for very long (Thelen & Ulrich, 1991). However, before and after the phase transition, real-time behavior is far less variable; only a small number of behavioral states are available to the system, and once the system settles into one of these stable patterns, it tends to remain there for an extended time period (e.g., Thelen & Smith, 1994; van Geert, 1998b; van der Maas & Molenaar, 1992). DS researchers have used the concept of perturbations on a real-time scale as an empirical design innovation to test the relative stability of observed behavioral patterns. Perturbations have the potential to abruptly push the system from one stable pattern to another (Fogel, 1993; Thelen & Smith, 1994). But this is only a potential—whether and how a system becomes reorganized is determined by its underlying structure. A perturbation can be considered a focal aspect of environmental context, and sensitivity to perturbations (which typifies phase transitions) is a special case of the more general sensitivity of dynamic systems to their contexts—a sensitivity that changes with system development gradually or more suddenly. Thus, context-sensitivity for DS researchers is not just “a form of jargon for anything environmental, as if invoking the term suggests compliance with current scientific and conceptual canons” (Boyce et al., 1998, p. 145). DS researchers systematically observe changes in behavior, as it varies with contextual forces, to infer the underlying structure of the system (e.g., Fogel, 1993; Granic & Lamey, 2002; Lewis & Granic, 1999b, 2000; Thelen & Smith, 1994; Thelen & Ulrich, 1991). On a developmental scale, principles of nonlinear change, phase transitions and perturbations have been most often used for the explicit purpose of studying

the structural profile of developmental transitions (e.g., Fogel & Thelen, 1987; Granic, Hollenstein, Dishion, & Patterson, 2003; Lewis, 2000; Lewis & Granic, 2000; Thelen & Ulrich, 1991; van Geert, 1991, 1994). NeoPiagetian scholars such as van der Maas and Molenaar (1992) have used a particular type of dynamic model, the cusp-catastrophe model, to represent the nonlinear nature of stage transitions. Borrowing from Gilmore (1981), they suggest a number of criteria or flags that can be used to operationalize a transition. Among the transition flags are: a sudden jump from one parameter value to another, evidence of hysteresis (i.e., when the same conditions elicit different behaviors, depending on the immediate prior history of the system), anomalous variance, and an increased sensitivity to perturbations. Transitions in motor, cognitive, linguistic, and socioemotional development have been successfully modeled by the application of variants of these flags (Case et al., 1996; Lewis et al., 1999, 2004; van der Maas & Molenaar, 1992; van Geert, 1991, 1994). Structural changes in parent–adolescent interactions at the early adolescent stage transition have also been shown to exhibit the properties of a phase transition (Granic, Dishion, & Hollenstein, 2003; Granic et al., 2003). Many of the DS concepts described here are clearly resonant with other systems views. But as a point of distinction, we suggest that four principles central to the DS framework are either neglected or less emphasized in other approaches and hold great promise for new empirical directions in developmental psychopathology. First, DS theory is primarily concerned with the emergence of novelty through the process of self-organization, whereas, with some notable exceptions (Ford & Lerner, 1992), most of the emphasis in more general systems views is on mechanisms of stability (i.e., negative feedback processes, for examples see cybernetic models; Granic, 2000; Lewis & Granic, 1999a). Second, although systems views may acknowledge the nonlinear nature of change in developmental systems, this is the hallmark principle of DS approaches to development and it has led to the exploration of radical new strategies such as catastrophe theory (e.g., van der Maas & Molenaar, 1992), developmental growth curve modeling (van Geert, 1991, 1994), and the study of phase transitions (e.g., Thelen & Ulrich, 1991; Thelen & Smith, 1994). Third, variability represents critical information in DS research. Conventionally, variability in developmental data has been seen as the result of measurement error and, thus, a source of noise that should be minimized. DS theorists make a radical departure from this approach: variability is considered a rich source of information, indexing impending change and “the essential ground for exploration and selection” (van Geert & van Dijk, 2002, p. 342). Methods that

Research Design Strategies Informed by DS Principles

tap changes in variability are a mainstay of DS researchers (e.g., Thelen & Ulrich, 1991; van Geert & van Dijk, 2002). Finally, DS theorists are fundamentally concerned with the interrelations between time scales of development and put a great deal of emphasis on understanding the unfolding patterns of real-time behavior (Thelen & Smith, 1998). This final principle is critical in terms of its methodological implications, and it is most often ignored in other systems frameworks. In the following sections, we describe a number of dynamic systems approaches to research designs and measurement strategies. To limit the scope of our review, we will not be discussing the exciting work emerging in the neurosciences (a field that has long embraced the principles of self-organization). This work is clearly relevant to developmental psychopathologists and the reader is referred to the neuroimaging methods detailed in the most recent volume of Developmental Psychopathology, 3rd edition. We also spend less time proportionally on the mathematical modeling techniques than the graphical, descriptive and statistical ones because we believe that the latter group of methodologies are generally more accessible and may ultimately prove more appropriate for the types of research questions put forward by developmental psychopathologists. We first review some general DS-informed research design concepts and strategies put forth by Thelen and her colleagues (e.g., Thelen & Smith, 1994; Thelen & Ulrich, 1991). Next, we discuss the types of data most suitable for DS analyses. We follow by providing a list of graphical techniques and quantitative strategies appropriate for the analysis of real-time and developmental data; these descriptions are supplemented with actual or hypothetical examples relevant to developmental psychopathologists. We also highlight the limitations inherent in some of the techniques and argue that a newly developed DS methodology, state space grid analysis, may help address a number of these weaknesses. The last part of this paper provides a detailed description of the state space grid technique, which combines graphical methods with statistical analysis in a way that maintains fidelity to DS concepts. We provide several examples of programs of research and individual studies that have used variations of this approach.

RESEARCH DESIGN STRATEGIES INFORMED BY DS PRINCIPLES Thelen and her colleagues have explicitly laid out a methodological strategy for developmental psychologists interested in dynamic analyses (Thelen & Smith, 1994,

723

1998; Thelen & Ulrich, 1991). Their strategy includes six steps: 1. Identify the collective variable of interest. A collective variable must be an observable phenomenon (not a construct or latent variable) that captures the coordination of the elements of a multidimensional system. Because Thelen and Ulrich (1991) were interested in motor development, they chose the phasing of alternating steps as the collective variable that condensed the many aspects of interlimb coordination. Changes in this collective variable can then be tracked over developmental time. This is the first challenging step—unlike physical systems, it is difficult to identify a collective variable in psychological systems. Extensive developmental observations and experiments are recommended as a first step toward this goal. An example relevant to developmental psychopathologists might be the observed intensity of a child’s oppositional behavior—a collective variable that may capture the coordination of mood states, arousal level, appraisal processes, and so on (these processes themselves would need to be assessed in multiple contexts). 2. Describe the attractors for that system. This involves mapping the real-time trajectory of the collective variable in various contexts across different developmental periods and identifying its relative stability. Thus, the contexts in which a child’s oppositional behavior is most stable can be identified (e.g., with a parent), as well as the contexts in which such behavior is less stable (i.e., more easily changed), is never observed, or is rarely observed. High stability indicates an attractor state. Attractor states may be tested by examining the variability of the collective variable given particular contexts (e.g., how often does the child become oppositional in response to a request to clean up at home versus at school). Additional features can help identify attractors as well. First, an attractor may be present if behavior takes a relatively short time to return back to the state once it has left it (latency to return to the attractor is short). Second, behavior tends to stay in an attractor for longer durations than other states. Third, it should take a much larger perturbation to move behavior out of a strong attractor state relative to weaker attractors or other regions of the state space. This last point is also a reminder that living systems exhibit multistability. The implication is that researchers should attempt to identify all or many of the attractors that are available to a particular system. Then the critical question becomes: what are the conditions under which behavior will gravitate toward one attractor versus another? Returning to

724

DS Methods

the oppositional child, most conventional approaches focus on describing and predicting the child’s negative, destructive behavior. But a DS approach would map out not only the negative behavior attractors but also the interested/engaged states, the joyful and the neutral states, and, importantly, the pathways to and from these attractors. Although these latter states may be weaker attractors, the goal would be to track the contextual conditions under which the child moves from more positive to more negative attractors and vice versa to get a better picture of the underlying dynamics of this child’s behavioral landscape. 3. Map the individual developmental trajectories of the collective variable. This step requires collecting observations at many time points in a longitudinal design (also see Fogel, 1993). The density of time samples depends on the developmental period in question (i.e., in infancy, weekly observations may be needed, whereas in late childhood data collected monthly might suffice; Fogel, 1993). Then, developmental profiles can be graphed on a case-by-case basis and the similarities and differences among profiles can be described. Stable (i.e., fixed or cycling) segments of the time series denote an attractor. At this stage, the multifinality and equifinality of developmental trajectories can be discovered. Some developmental profiles may start out looking similar and then, from very small differences that become amplified, trajectories may diverge. Other developmental profiles may show the opposite pattern of different initial conditions being “pulled” toward a particular attractor. The key at this stage is to create individual profiles rather than aggregate across subjects; otherwise, the variability inherent in developmental processes will be obscured. As described in more detail later, an alternative to averaging developmental profiles is to cluster individual trajectories into groups that share profile characteristics (Hollenstein, Granic, Stoolmiller, & Snyder, 2004). In developmental psychopathology, a similar point has been emphasized by researchers doing case-based, or person-oriented, analysis (e.g., Bergman & Magnusson, 1997). 4. Identify phase transitions in development. As described earlier, transitions in development are characterized by increased variability, a breakdown of stable patterns, and the emergence of new forms. The various catastrophe flags described earlier can help researchers identify points of transition. Transition periods are critical to mark because they allow researchers to access and manipulate mechanisms underlying change. This point

is particularly relevant for developmental psychopathologists interested in clinical interventions. For instance, there may be normative stage transitions in children’s development during which, as a result of normal maturational processes, the coordination among system elements begins to break down, previous attractors are destabilized, and there is an increased potential for new patterns to emerge (e.g., Granic, Hollenstein, Dishion, & Patterson, 2003; Lewis & Granic, 1999b; Lewis, Zimmerman, et al., 2004; Lewis et al., 1999). Clinical interventions may have their greatest impact if they are targeted at these sensitive periods. The idea of fluctuations at phase transitions is not new to developmental psychopathology, although it has rarely been pursued empirically. One fascinating example comes from a study conducted by Inhelder (1976), who found that some mentally retarded children’s skill performance oscillated between various stage levels prior to reaching a new stage. Interpreting these findings, Cicchetti (1990) suggested two possibilities: the processes accounting for these children’s development may simply be different from those of normal children, or, more intriguingly, “their oscillations are universal phenomena that become visible only in retarded children because of the slower nature of their development” (p. 16; see also Cicchetti & Sroufe, 1976). Bertenthal (1999) further emphasizes the importance of variability at phase transitions. He suggests that variability is not just an index of change, but actually helps drive change. According to Bertenthal, “variability offers flexibility, which drives development following Darwinian principles. Principles of variation and selection cause successful behaviors to be stored and repeated more frequently than the less successful” (van Geert & van Dijk, 2002, p. 343; Thelen & Smith, 1994). The theoretical implications for understanding intervention effects are compelling. Successful interventions may induce a phase transition during which behavioral, cognitive and affective variability increases, providing the fertile ground from which more positive, or less distressing patterns can be selected, repeated, and potentially stabilized. This possibility can be empirically verified by using simple descriptive statistics (e.g., looking for an increase in standard deviations and variance, a breakdown of correlations, and changes in entropy; additional measures of variance are described later) or more formal techniques (described in the section on state space grids). 5. Identify control parameters. In DS language, control parameters are the “agents of change.” The purpose

Suitable Data for DS Analyses

of tracking the collective variable across different contexts and developmental transition points is to ultimately identify the mechanisms underlying processes of change. Control parameters are not simply independent variables (although they can be considered a special type). Usually, independent variables are static measures that are assumed to have a linear effect on outcomes. Control parameters are better understood as mediators with special properties; continuous and small changes in the values of these parameters can result in abrupt threshold effects on a collective variable. Moreover, these nonlinear changes occur at different values depending on whether the control parameter is increasing or decreasing. For example, through fine-grained longitudinal observations, Thelen and Ulrich (1991) were able to identify overall changes in muscle mass as the control parameter that was related to improvements in infants’ treadmill stepping. In many areas of developmental psychopathology, however, this step is the most difficult, because psychological systems are incredibly complex and the problem of identifying and manipulating one or very few causal mechanisms is often insoluble. Moreover, a control parameter may not always be something that can be manipulated (e.g., temperament, parental depression). Nevertheless, DS researchers urge us to at least keep the concept in mind. 6. Manipulate control parameters to experimentally generate phase transitions. Despite its difficulty, this suggestion is a familiar one to many developmental psychopathologists. Simply put, once a causal factor has been inferred from careful descriptive analysis, it should be experimentally manipulated to examine whether it does indeed trigger the expected shift in behavior. In this respect, intervention studies are an exceptional avenue for testing the role of specific control parameters in developmental psychopathology (c.f. Dishion, Bullock, & Granic, 2002; Dishion & Patterson, 1999; Eddy & Chamberlain, 2000; Forgatch & DeGarmo, 1999). One of the best examples of following this proposed strategy comes from the work on the etiology and treatment of aggressive behavior. For instance, based on decades of microsocial observational studies with families, coercion has been identified as a mediating causal mechanism underlying the etiology of childhood aggression (e.g., Patterson, 1982; Patterson, Reid, & Dishion, 1992). To confirm this supposition experimentally, Forgatch and DeGarmo (1999) investigated the impact of a randomized control intervention that aimed to decrease the rate at which parents engaged in coercive

725

interchanges. Results confirmed that, indeed, changes in coercion resulted in decreases in children’s aggressive behavior. From a DS perspective, parental discipline strategies, for example, could have been the control parameter that was adjusted through the intervention (i.e., coercion decreased). Despite the advantages of Thelen and colleagues’ approach, many researchers are likely to experience some problems with implementing their general strategy. First, it requires collecting continuous time-series data (e.g., physiological data, behavioral observations coded in real time); this type of data is time-consuming and expensive to collect. More importantly, time-series data may not capture the type of information pertinent to many developmental psychopathologists. Second, and related, unlike motor or cognitive development in which some skill or task performance increases or decreases quantitatively over time, psychopathology may not involve such graded changes (Lewis & Granic, 1999b; Lewis et al., 2004). Instead, the development of psychopathology may better be characterized as emergent patterns of interconnected changes in different domains (e.g., biological, cognitive, emotional) that are nonlinearly related to one another and change qualitatively, as well as quantitatively (i.e., it is often categorical or ordinal). We will address this issue at greater length when we discuss state space grid analysis and the limitations it addresses in this regard. SUITABLE DATA FOR DS ANALYSES Developmental psychopathologists collect several types of data that are appropriate for DS analyses. Because a DS perspective is fundamentally about changes in time, the most important data characteristic is that there are multiple measurements over time. Thus, questionnaire data collected at one time point would be inappropriate for tapping dynamic processes (e.g., Cummings et al., 2000). Below, we list the four types of data that lend themselves most easily to DS analyses and name some examples of analytic techniques that can be conducted. These techniques are then explained in more detail in the following section. Table 16.1 lists the techniques, the appropriate data types for each, and provides some (non-exhaustive) examples of empirical papers that have applied the various DS methods. 1. Observational data—continuous. Includes data obtained in time units less than one second; these data are often composed of physiological measures. The density of

726

DS Methods

TABLE 16.1 Summary of DS Techniques and DS Concepts and Examples of Studies That Have Applied These Techniques Techniques Real-time

DS concept

Examples

Case studies Time series

Self-organization Attractors

Phase plots

Phase space attractors

Event history analysis

Attractors

Fourier analysis

Attractors

Karnaugh maps

Phase transition State space Feedback Attractors

Fogel (1990, 1993) Bakeman & Gottman (1997) Granic & Dishion (2003) Sabelli, Carlson-Sabelli, Patel, Levy, & Diez-Martin (1995) Totterdell, Briner, Parkinson, & Reynolds (1996) Snyder, Stoolmiller, Wilson, & Yamamoto (2003) Lougheed, Hollenstein, Lichtwarck-Aschoff, & Granic (2015) Newtson (1994,1998) Schroeck (1994) Dumas, Lemay, & Dauwalder (2001)

Coupled equations

Nonlinear dynamics

Attractors Phase space feedback Entropy Chaos & determinism

SSG

State space attractors Phase transitions Perturbation variability

Developmental Descriptive developmental profile Latent class analysis

Dynamic growth modeling

Connectionist modeling

Phase space attractors Self-organization Phase transition Attractors Self-organization Phase transition Phase space attractors Self-organization Phase transition feedback Attractors Bifurcations Phase transition Chaos

Catastrophe modeling

Phase transition Hysteresis

SSG

Phase transition State space

data points allows for the more sophisticated techniques adopted from the natural sciences, where this type of data is most common. These techniques are often based on time-series analysis and include the domain of elaborate mathematical models based on coupled equations and dynamical catastrophe models. Quantitative measures of chaos—Lyupanov exponents,

Gottman, Guralnick, Wilson, Swanson, & Murray (1997) Nowak & Vallacher (1998) Ryan, Gottman, Murray, Carrere, & Swanson (2000) Gottman, Murray, Swanson, Tyson, & Swanson (2002) Guastello (2001) Heath (2000) Newell & Molenaar (1998) Dishion, Nelson, Bullock, & Winter (2004) Pincus (2001) Lewis, Lamey, & Douglas (1999) Hollenstein, Granic, Stoolmiller, & Snyder (2004) Granic & Lamey (2002) Granic, O’Hara, Pepler, & Lewis (2007) Hollenstein (2007, 2013) Hollenstein & Lewis (2006) Hollenstein, Allen, & Sheeber (in press) Thelen & Ulrich (1991) Thelen & Smith (1994) Van Geert & van Dijk (2002) van der Maas (1998)

van Geert (1994, 1998) Ruhland & van Geert (1998) Thelen, Schoner, Scheier, & Smith (2001) Steenbeek & van Geert (2005, 2007, 2008) McClelland (1979) Rumelhart & McClelland (1986) Johnson & Morton (1991) Smith (1993, 1995) Smith & Thelen (2003) Thelen et al. (2001) Hartleman, van der Maas, & Molenaar (1998) van der Maas & Molenaar (1992) Granic, Dishion, & Hollenstein (2003) Granic, Hollenstein, Dishion, & Patterson (2003) Lewis, Zimmerman, Hollenstein, & Lamey (2004) Lewis, Lamey, & Douglas (1999) van der Giessen, Hollenstein, Hale, Koot, Meeus, & Branje (2015)

entropy, correlational dimensions, and so on—can also be derived from such continuous data. Furthermore, all of the methods available to the other data types are available with this kind of data, either as time series or summaries of time series. 2. Observational data—discrete. Includes live and taped observational data that are converted to codes along

DS Methods

time units as small as one second. These codes can represent the sequence of behavior for one or more subjects and are typically inappropriate for time-series techniques unless they have a sufficiently large number of data points. By applying DS graphical techniques such as state space grids, Karnaugh maps, recurrence plots, and phase plots, the temporal patterns embedded in the temporal sequence can be uncovered. These data can be used to identify attractors, perturbations, phase transitions, and other DS patterns. 3. Longitudinal data—short. This type of data can include, for example, hourly/daily/weekly self-report measures (e.g., diary or beeper studies), repeated phone interviews, or repeated observational sessions (e.g., therapy sessions). The time points may be frequent enough to allow the use of some of the real-time techniques available to the first type of data or may be analyzed using techniques applicable to developmental-time data (type 4). 4. Longitudinal data—long. Includes any combination of the above data types collected at different time points that may span weeks, months, or years. The main distinction from data type 3 is the time span between first and last measurements. Data collected in three or more waves is often used to depict change, growth, or intervention effects and a variety of DS methods including state space grid analysis, dynamic growth modeling, developmental profile analysis, and catastrophe modeling can be applied. Developmental phase transitions are detectable through the application of these techniques.

DS METHODS Real-Time Measures As most teachers of research methods and statistics in general will insist, eyeballing your data is an important part of the analytic procedure. For DS researchers, graphical techniques provide the core of their analytic armament. Perhaps because dynamic systems theory is a descriptive framework and because it aims to describe phenomena in geometric terms (recall our discussion of behavioral trajectories on a state space), plotting data is the mainstay of DS researchers (Abraham, Abraham, & Shaw, 1992; Norton, 1995). A number of these real- and developmental-time graphical methods can be complemented with various quantitative tools. We have grouped the following methods under the rubric of real-time methods because that is how they have tended to be applied. But it is important to note

727

that because the same principles of self-organization apply at different time scales, many of these techniques can be adapted easily to study phenomena in developmental time. Case Studies Perhaps the one common recommendation DS researchers offer is to start with fine-grained, real-time observations of the phenomenon of interest and follow this behavior across a significant developmental period. One of most basic first steps toward this end is the careful description of case studies. Fogel’s (1990, 1993) research on mother– infant relationship processes is exemplary in this respect. His microgenetic research relies heavily on detailed descriptions of videotaped interactions, as they proceed in real time. He uses these case histories as a “means to seek patterns in sequences of action in a context, in both real and developmental time scales” (Fogel, 1990, p. 343). Metaphors based on DS principles serve as guides for identifying dynamically stable dyadic patterns (frames) and changes in those patterns across development. Although these narrative descriptions are rich in detail and provide ample fodder for generating hypotheses, they are intentionally not quantified at this stage. As such, this method does not address developmental psychopathologists’ search for quantitative techniques to test their conceptual models. To quantify impressions from case studies and statistically test hypotheses, Fogel and his colleagues used a number of techniques (for a review, see Lavelli, Pantoja, Hsu, Messinger, & Fogel, 2005) including simple descriptive statistics, individual growth curve modeling (Fogel, 1995; Hsu & Fogel, 2001), conditional probabilities (Hsu & Fogel, 2003a), and event history analysis (Hsu & Fogel, 2003b) to measure the emergence, strength, and dissipation of parent–infant attractors. These methods are discussed in later sections of this chapter. Time-Series Analysis The group of methods that fall under time-series analysis offer some useful approaches to characterizing the real-time behavior of dynamic systems. The methods include visual representations as well as simple statistics that can be used to examine the qualities and predictive power of hypothesized attractor patterns. To illustrate the utility of this approach, we summarize the rationale, time-series procedures, and results from a study by Granic and Dishion (2003) using data from antisocial and normal friendship interactions. Past research suggested that antisocial adolescents can be distinguished from their prosocial counterparts by the

DS Methods

extent to which they engage in reciprocal deviant talk (e.g., talk about stealing, lying; Dishion et al, 1995, 1996, 1997). Observational studies showed that antisocial peers had a higher mean duration of deviant talk (or “rule-break” talk) than prosocial peers. Central tendency measures, however, do not speak directly to the processes underlying these interactions. Moreover, they obscure potentially critical temporal patterns. To come closer to a process-level explanation, we began by conceptualizing deviant talk as an attractor for antisocial, but not prosocial peers (Granic & Dishion, 2003). Our interest was not in examining whether one group showed more deviant talk than another, but whether, over the course of an interaction, antisocial adolescents became stuck in an antisocial talk attractor. That is, did they become increasingly engrossed in topics organized around deviancy? One way to explore this hypothesis was to examine whether, as the interaction unfolded and antisocial dyads repeatedly returned to talking about deviant topics, they also spent increasingly more time in that deviant pattern. Time-series procedures were well suited for our purposes. Deviant (or rule-break talk) and normative talk was coded continuously from videotaped interactions between best friends. Time-series plots were derived for each dyad, with the duration of each deviant talk bout on the y axis and each successive bout represented along the x axis (Figures 16.3 and 16.4). The slope of that time-series (i.e., the standardized beta) was then calculated using simple regression analysis. Slope is one parameter that can

Duration in RB Bout (sec)

50

40

30

20

10

10 1

9 5

18 13

26 22

36 30

44 40

52 48

60 56

71 65

76

Successive RB Bouts Across Time

Figure 16.3 Example of a time-series for an antisocial youth with a positive rule-break bout slope. Source: From Granic, I., & Dishion, T. J. (2003). Deviant talk in adolescent friendships: A step toward measuring a pathogenic attractor process. Social Development, 12, 314–334.

20

Duration in RB Bout (sec)

728

15

10

5

0 0

3

4

7 9 11 13 15 17 19 21 23 25 Successive RB Bouts Across Time

Figure 16.4 Example of a time-series for a normal adolescent with a negative rule-break bout slope. Source: From Granic, I., & Dishion, T. J. (2003). Deviant talk in adolescent friendships: A step toward measuring a pathogenic attractor process. Social Development, 12, 314–334.

be identified from a time series. Others include cyclicity (number of cycles in a series, length of cycles) and autocorrelation. We used the slope measure in a somewhat unique way to highlight a key DS principle. If deviant talk indeed functioned as an attractor for antisocial youth, then we expected to see a time-series that showed a positive slope, as exemplified in Figure 16.3. If it was not an attractor for a dyad, then we expected to see a time-series with either a flat or negative slope, as shown in Figure 16.4. Thus, each participant was assigned a deviant-talk slope value and these values were then used in regression analyses to predict antisocial outcomes 3 years later. The results are summarized in Table 16.2. As hypothesized, the attractor index (the slope of deviant talk bouts) predicted serious authority conflict (e.g., number of arrests, school expulsion) and drug abuse three years later. These results were particularly strong because they remained statistically significant even after controlling for arguably the three most predictive risk factors in childhood: prior deviant behavior, family coercion, and deviant peer associations in childhood. We also showed that the mean duration of deviant talk bouts was not sufficient to predict these outcomes. Thus, these findings showed that reconceptualizing deviant talk as an attractor is not simply an exercise in relabeling; there are specific hypotheses that are engendered by applying this approach. From a DS perspective, although the frequency and amount of time spent talking about deviant topics are important, even more critical is the process by which dyadic behavior continues

DS Methods TABLE 16.2 Regression Results Predicting Adolescent Authority Conflict and Substance Abuse in Midadolescence from the Strength of the Deviant Talk Attractor in Early Adolescence

Steps in regression

Antisocial behavior

Substance abuse

R2 change Total R2

R2 change Total R2

1. Child deviancy

.283

.162

2. Child family coercion

.02

.30

.02

.18

3. Child deviant peer affiliation

.072

.36

.061

.24

4. Attractor strength (slope of deviant talk)

.052

.41

.061

.30

< .05. < .01. 3 p < .001. 1p 2p

to be drawn toward, and held, in this deviant pattern. Both antisocial and prosocial peers discussed deviant acts and breaking rules. It was the dynamic (time-based) characteristics of these conversations—as revealed by a simple time-series analysis–that differentiated antisocial and normal friendships. Bakeman and Gottman (1997) used a similar analytic approach to analyze data from marital couples’ interactions using this time-series technique, except they plotted the interevent interval between successive displays of negative affect (i.e., the time in between one negative affect and the next, across the interaction). They showed that, for distressed dyads, the time between each negative affect display became shorter and shorter over the course of the conversation (i.e., return time to the attractor became shorter

729

over the interaction). Thus, negative affect was an absorbing state for distressed, but not happily married, couples. Recurrence Quantification Analysis Recurrence quantification analysis (RQA; Marwan, Carmen Romano, Thiel, & Kurths, 2007; Webber & Zbilut, 2005) is a nonlinear time series analysis technique that has found many applications across scientific disciplines (Webber, Marwan, Facchini, & Giuliani, 2009). We here discuss an adaptation of the RQA for categorical data because these are likely the kind of data that clinical researchers and developmental psychopathologists will use. In studies of behavior observation the data are usually time series that represent the occurrence of a particular type of behavior at a particular point in time. In that respect the data may be characterized as a temporal sequence of unordered behavioral responses, that is, as categorical time-serial data. RQA quantifies the temporal order in time series at each and all possible timescales or delays during the observation. RQA for categorical data is based on the dichotomous distance matrix, in which it is noted for all behaviors when each behavior that occurs at some instance in the time series also occurs at some other instance in that same time series, either earlier or later. Such a reoccurring value is called a recurrent point, which is the basic measure of RQA. A two-dimensional graphical representation of this matrix is called the recurrence plot. In a recurrence plot all the points are plotted that represent recurring values (Figure 16.5). When recurrent points occur in subsequent time steps in the time series, line or

RP of session 1

RP of session 2

RP of session 3

RP of session 4

RP of session 5

RP of session 6

HDLL: 2.750

HDLL: 4.261

HDLL: 4.261

HDLL: 2.198

HDLL: 2.777

HDLL: 2.966

RP of session 1

RP of session 2

RP of session 3

RP of session 4

RP of session 5

RP of session 6

HDLL: 2.232

HDLL: 2.735

HDLL: 2.154

HDLL: 2.236

HDLL: 1.910

HDLL: 2.719

Figure 16.5 Recurrence plots of two exemplary dyads across observation sessions. Gray diagonal line represents the line of identity. Dyad in the upper row is showing the peak in entropy and dyad in the lower row does not show a peak in entropy. Source: With permission to reprint figure from Dr. David Pincus, Nonlinear Dynamics, Psychology, and Life Sciences (NDPLS) Permissions Editor. From Lichtwarck-Aschoff, A., Hasselman, F., Cox, R., Pepler, D., & Granic, I. (2012). A characteristic destabilization profile in parent–child interactions associated with treatment efficacy for aggressive children. Nonlinear Dynamics, Psychology, and Life Sciences, 16, 353–379.

730

DS Methods

block structures emerge in the recurrence plot. These structures reflect that certain sequences of values are repeatedly occurring (i.e., recurring) over time. One could say the dyad returns to a previously displayed pattern of behaviors. The (average) length of such structures indicates how stable (or prominent) this particular pattern of behavior is in the interaction. Note that various structures that are located at different regions in the recurrence plot may represent different behavioral patterns, not only in duration but also in terms of their particular constituent behaviors (i.e., content). As can be seen in Figure 16.5, the recurrence plots are symmetrical around the line of identity (see grey diagonal line). This line represents the time series as it was observed. Several different measures can be derived from those recurrence plots (e.g., determinism, predictability, entropy, trapping time, and divergence; for an elaborate discussion see Webber & Zbilut, 2005). In a recent study Lichtwarck-Aschoff and colleagues (2012) applied RQA to examine changes in entropy to investigate whether successful treatment outcomes were associated with having gone through a phase transition. Repeated mother–child interactions of children being treated for aggression were analyzed over the course of a 12-week treatment period (treatment was a combined Cognitive Behavioral Therapy and Parent

Management Training). Interventions can be seen as causing perturbations to a (mother–child) system. To be effective in eliciting a qualitative shift to a more healthy attractor state, treatment should trigger increasing levels of variability and disorganization up until the point that the system can no longer resist and is forced to give up the old maladaptive attractor state (see also Granic & Patterson, 2006). The temporary change in the natural variability of the system, the increased within-system variability, is functional in the sense that it allows the system to explore new patterns and configurations (De Weerth & Van Geert, 2002; Thelen & Smith, 1994; Van Geert & Van Dijk, 2002). After the destabilization phase, the system settles down again, organizing into a new (dynamically) stable regime of the system (Van Orden, Kloos, & Wallort, 2009). Thus, it was expected to find a change pattern, signaling an impeding phase transition that was characterized by a temporary increase in random fluctuations or breaking down of stable recurring patterns of dyadic behaviors following the start of the intervention. The RQA measure that quantifies this particular aspect of the dynamics is the Shannon entropy of the distribution of diagonal line structures in the recurrence plot (see also Stephen et al., 2009; Stephen, Dixon, & Isenhower, 2009). A subsequent latent class growth curve analysis

Results of latent growth curve cluster analysis 4.5 Entropy of the length of diagonal line structures in the rp (CI.95)

LCGA: No peak in Entropy LCGA: Peak in Entropy 4

3.5

3

2.5

2

1.5

1

2

3

4

5

6

Session

Figure 16.6 Result from the LCGA analysis of values for entropy of the diagonal line structures in the recurrence plots of each dyad across the six observation sessions (95% confidence intervals are based on 15,000 bootstrap replications). Source: With permission to reprint figure from Dr. David Pincus, Nonlinear Dynamics, Psychology, and Life Sciences (NDPLS) Permissions Editor. From Lichtwarck-Aschoff, A., Hasselman, F., Cox, R., Pepler, D., & Granic, I. (2012). A characteristic destabilization profile in parent–child interactions associated with treatment efficacy for aggressive children. Nonlinear Dynamics, Psychology, and Life Sciences, 16, 353–379.

DS Methods

including the frequency, duration, total time, and the time since the last occurrence of the behavior can be included in these models to determine the absorbing quality of a hypothesized attractor. Furthermore, time-invariant (e.g., gender, child’s problem behavior, maternal depression) and time-varying (e.g., changes in context, affective behavior of the partner, time) covariates can be added to these models as predictors of the transition rate. Figure 16.7 shows a schematic diagram of an interactive sequence. The bottom plot line represents the time line of events being modeled. In this case, there are only two states possible: in the proposed attractor or out of that state. This example illustrates a dichotomous response set, but it is not necessary for event streams to be binary for event history analyses; multiple states can be modeled as competing risks. The three episodes in this example correspond to the start and end times of the behavior of interest. The first episode lasts 14 seconds, the second episode lasts 18 seconds, and the third is 14 seconds long. With just this information it is possible to model the probability of transitions into the attractor as one instantaneous rate, and the duration since the last episode can be added into the model as a predictor. Presumably, whatever behavior is being researched does not occur in a vacuum. There is the nature of the task at hand, the influence of other people and even factors unique to each individual that must be considered. Thus, as shown by the time-varying covariates (upper 2 lines) in Figure 16.7, more elaborate models can tease apart the relative contributions of various covariates to the recurrence of a particular behavior. In a series of studies Snyder,

on the sequences of entropy values revealed two distinct classes of dyads, with one class showing a clear peak in entropy over the six measurement points (Figure 16.6). The latent class membership variables showed a significant systematic relationship with observed dyad improvement (as rated by clinicians). The class with the peak in entropy over the sessions consisted largely of treatment improvers (Figure 16.6). Further analysis revealed that improvers and nonimprovers could not be distinguished based on content-specific changes (e.g., more positivity or less negativity during the interaction). This study thus supports the DS view that treatment functions through creating the conditions that facilitate dynamic reorganization (Schiepek, 2009) rather than a linear input–output model in which a series of equally powerful weekly interventions force the system to a particular reaction or end state. Event History Analysis As can be seen with the previous examples, and others to follow, it is not always necessary to invent new methods to explore DS-inspired hypotheses. Another method that can help researchers explore the different characteristics of attractor processes is event history analysis. A variant of survival analysis and Cox regression models, hazard analysis calculates transition rates (also called hazard rate or latency to recurrence) of state changes in categorical time series (Allison, 1984; Blossfeld & Rohwer, 2002; Stoolmiller & Snyder, 2006). The rate of transition in and out of a behavioral state (e.g., tantrum, laugh) can be interpreted as an index of attractor strength. Several other measures

First Episode

731

Third Episode

Second Episode

Covariate 2 Covariate 1

Event of Interest 0

0

14 0

5

10

15

18

Episode Time

20

25

30

35

0

40

45

14

50

55

Session Time

Figure 16.7 Schematic diagram of a behavioral interaction with some of the components that can be analyzed through event history analysis. See footnote 1. Source: Adapted from: Snyder, J., Stoolmiller, M., Wilson, M., & Yamamoto, M. (2003). Child anger regulation, parental responses to children’s anger displays, and early child antisocial behavior. Social Development, 12, 335–360.

732

DS Methods

Stoolmiller, and colleagues have used hazard analyses to investigate a wide range of parameters of the state space for emotion response states in parent–child interactions (with clinical and control children), and the relationship of those parameters to time fixed and time-varying covariates. In the first study a repeated-events, competing-risks hazard analysis was used to estimate the return time for child anger displays using both time fixed and time-varying predictors (Snyder, Stoolmiller, Wilson, & Yamamoto, 2003). In state space terms, return time may be interpreted as an indication of the strength of the anger attractor state. Results showed that the longer it took a child to down regulate an anger display, the more quickly that anger display would reoccur. Further, the authors found that the speed of child’s return time to an anger display (at time t + 1) was amplified by the cumulative frequency of parent negative responses (a cumulative time-varying covariate). Using a time-fixed covariate, poor parental discipline was also related to a faster return rate to anger. Individual differences in the hazard for return to anger displays were also associated with children’s conduct problems (as rated by teachers and parents). In a second study a complementary hazard analysis was applied to investigate the relationship of the child’s hazard to move to an angry start state (given mom was aversive) to a neutral state to child overt and covert antisocial behavior (Dagne & Snyder, 2009). It was found that children high on overt antisocial behavior relative to the average child were slower to move to a neutral emotional state. Children high on covert antisocial behavior relative to the average child, in contrast, were faster to move to a neutral affective state. Additionally, maternal self-reported trait hostility (a time-fixed covariate) was associated with a reduced hazard for mothers to move from an aversive to a nonaversive state in the context of child anger displays, which indicates that hostile mothers are slower to cool down in negative encounters with their children. In a third study Stoolmiller and Snyder (2006) examined the hazard of children’s transition from a neutral state to positive, sad/fearful, or angry emotion states when their mothers initiated an aversive behavior. It appeared that on average, the hazard for children to move to a negative emotion state (either angry or sad/fearful) increased as mothers initiated an aversive behavior. Children high on antisocial behavior were significantly slower to move from neutral to sadness/fear, but there was no relationship between antisocial behavior and children’s hazard to move from neutral to anger. Finally, in a recent study Snyder, Bockman, and Stoolmiller (2012) investigated the hazard of down-regulating

anger and sadness/fear relative to neutral in relation to children’s antisocial behavior. The results showed that overt and covert antisocial behaviors were associated with an underutilization and rapid down regulation of sadness/fear relative to neutral. Further, children in overt antisocial behavior also showed an overutilization and slow down regulation of anger relative to neutral affect. These studies demonstrate that hazard analysis is a sophisticated and flexible mathematical approach to model a system’s behavior by capturing temporal and well as sequential shifts among various emotion states. These models can be run on individuals or pooled across an entire sample and it is possible to model the state space of emotion displays from a dyadic as well as an individual perspective (Stoolmiller & Snyder, 2006). It is not only the potential of identifying predictors of attractor strength that is the appeal for DS researchers but the fact that time itself can be used as a predictor. Probably the most difficult process of event history analysis is restructuring data into the appropriate format. Despite this difficulty, we expect that event history analysis will be used with increasing frequency as a method to model attractor dynamics. Fourier analysis or spectral decomposition can be used for finding periodicities (i.e., cyclic patterns) in time-series data such as those used in phase plots. In general, this procedure treats a time-series as a conglomerate wave form, breaks it down into a collection of pure waves (each of uniform frequency), and identifies the most prominent waves (Schroeck, 1994). Newtson (1994, 1998) used this method to analyze the coupled dynamics of dyadic interactions. The relative amplitudes and temporal synchrony of these behavior waves were associated with the degree of mutuality or competition in interpersonal relationships. In other applications, different types of attractors including oscillating and periodic attractors can be identified. For behavioral scientists, however, this method has its limitations. Like all time-series procedures, it requires the researcher to collapse meaningful categorically coded observational data into one or very few continuous dimensions (e.g., Bakeman & Gottman, 1997). For example, most observational coding schemes used in developmental psychopathology (e.g., SPAFF—Gottman, McCoy, Coan, & Collier, 1996; FPC—Dishion, Gardner, Patterson, Reid, & Thibodeaux, 1983; MAX—Izard, 1979) code discrete behaviors like contempt, argue, belligerence, and whining. To conduct Fourier analysis, these codes would have to fall along a single dimension (i.e., intensity of negativity). This type of collapsing is often either conceptually unfeasible or it is unappealing because of its oversimplification.

DS Methods

Karnaugh Maps Inspired by synergetics, a type of dynamics developed by Haken (1977), Dumas, Lemay, and Dauwalder (2001) adapted this technique from Boolean algebra to study parent–child interactions. Karnaugh maps depict all possible combinations of up to four binary variables in one table or grid. A simple two-variable Karnaugh map is basically a four-cell cross-tabulation of event frequencies with one dichotomous variable to each axis. Threeand four-variable maps are somewhat more complicated. Figure 16.8 shows a schematic four-variable Karnaugh map. Each variable (A, B, C, and D) can have a value of 0 or 1, and each is displayed on one side of the square table (bottom, left, top, and right). Each cell in the map can be represented as a unique combination of the binary values of the four variables. Thus, a Karnaugh map is a state space of all possible states of the system and represents the relative frequencies in each state. Dumas and colleagues extended the application of these maps by plotting the transitions between states to depict temporal patterns across the space. In their study, Dumas et al. (2001) plotted every minute of six-hour home observations according to four parent and child dichotomous variables that included control, compliance, aversive and positive behavior. Each behavior was plotted according to the four-variable configuration

Variable B = 1

Variable C = 1

Variable C = 0

0000

0010

1010

1000

0100

0110

1110

1100

0101

0111

1111

1101

0001

0011

1011

1001

Variable A = 0

Variable D = 1

Variable B = 0

and successive behaviors were connected by a trajectory. They were primarily interested in comparing the maps of clinically referred dyads with randomly selected controls. They analyzed the recurring transactions (i.e., attractors) that were found most frequently in each group and aggregated these findings graphically on one summary Karnaugh map. Their results showed differences in transactional patterns between the groups, mostly involving positive interactions and cycles of maternal control and child compliance. In addition to the graphical depiction of interaction sequences, the authors computed a complexity index that was designed to quantify each map on a continuum from completely deterministic to completely random. They found that all maps, regardless of group assignment, were neither random nor deterministic but somewhere in between. While this approach was unique and interesting (particularly from a methodological standpoint), from a DS perspective there is no reason to think that social behavior, especially in dyads with a rich history, is ever random or ever completely determined. Nevertheless, the complexity measure has a great deal of potential. For example, it might be used to determine whether stable coercive parent–child interaction patterns become less determined (i.e., the old attractor patterns break down) during developmental transitions such as puberty.

Variable D = 0

Variable B = 0

Variable C = 0

733

Variable A = 1

Figure 16.8 Schematic of a four-variable Karnaugh map. Each cell is a unique combination of four binary variables. The arrow shows a sample transition from one state, where only variable D is present (i.e., only D equals 1), to the next state in time, where all 4 variables are present (i.e., all equal 1).

734

DS Methods

A very similar approach has been applied in a study by Lichtwarck-Aschoff and colleagues (2009) to investigate intraindividual variability on a different time scale: diary entries of mother–daughter conflict episodes. Indices of the episode-to-episode variability in emotional states (i.e., composition of emotions) were used to assess whether girls in conflictual mother–daughter relationships would show signs of being stuck in a certain (emotional) conflict pattern. In line with research on real-time variability (Granic et al., 2007; Hollenstein & Lewis, 2006; Hollenstein et al., 2004) it was hypothesized that conflictual relationships would be characterized by a lack of variability across episodes. However, in contrast to real-time measures, which reflect changes from one emotion to another moment by moment, diary entries describe emotional states of entire conflict episodes that can be composed by several different emotions. Thus, changes in a multidimensional (emotional) space have to be assessed. (Note that this emotional space is therefore different from the state space grids developed by Lewis et al., 1999, described later in this chapter.) To do so, measures based on symbol dynamics (Daw, Finney, & Tracy, 2003), which actually extends the Karnaugh maps in terms of the dimensions that can be included, were developed. The basic rationale behind this approach is the transformation of the system’s trajectory or geographical movement in time into a sequence of specific symbols, corresponding to partitions within the state space (Dale & Spivey, 2005). First, the emotional state space for each girl was computed, which contained all her reported emotions across all her conflict episodes. Second, within each girl’s state space all the emotional states that represented how a girl felt during a particular conflict were described. An emotional state could take the form of a single emotion or a combination of (many) different emotions. The mathematical representation of every state consisted of a particular string of ones (emotions being present) and zeros (emotions being absent). One way to assess the emotional variability was to count the number of different emotional states across all conflict episodes. Further, the Hamming distance (Hamming, 1950; Te¸sileanu & Meyer-Ortmanns, 2006) was used to compute additional variability measures. The Hamming distance calculates the distance (absolute difference) between two symbolic strings (here, two emotional states). The distance between the emotional states then indicates how much the emotional states differ from each other and in this sense represents the variation between the states (compare this with the intergrid distance used by Lewis et al., 1999). In this study a sequential and an overall Hamming distance was computed.

As expected the results revealed an inverse U-shaped relation between girls’ emotional variability and the number of conflicts. Thus, the emotional variability decreased when relationships became highly conflictual. Moreover, girls who showed limited variability in emotional states across conflict episodes tended to attach the same emotional state to divergent conflict topics. For these girls, many different conflict contexts still induced the same emotional state. In other words, these girls felt the same, no matter what. Hence, the conflictual mother–daughter systems seemed to be stuck or frozen into a small number of emotional states and had lost their situational sensitivity (i.e., sensitivity and flexibility in response to different conflict topics). Karnaugh maps and symbol dynamics hold considerable promise for several reasons. First, because these methods allow for the representation of behavior in multiple dimensions (including time), a great deal of the complexity in (interactional) behavior can be nicely captured. In case of the Karnaugh maps the temporal quality of dyadic behavior is also maintained and can be tracked easily through visual inspection. Moreover, the systemic properties of dyadic interactions are kept intact, in contrast to conventional methods that often require analyzing each dyad member separately. The approach described in the Lichtwarck-Aschoff et al. (2009) study offers the possibility to study changes and variability of multidimensional behavior which is especially relevant for studies on larger time scales where measures often reflect some aggregated unit of analysis (retrospective reports of conflict episodes rather than real-time presentations of emotions). However, some limitations need to be addressed before these techniques can be applied to a broad range of behavioral data. One problem is that data have to be collected or converted into dichotomous variables that occur concurrently. Developmental psychopathologists are often interested in data that cannot be meaningfully transformed into dichotomous values; continuous or categorical data cannot be adequately captured with these techniques. Also, these methods do not provide direct quantifiable tests of the strength of (dyadic) patterns, nor opportunities for comparing patterns. In this respect, the new developments in state space grid analysis may provide some promising avenues. Some of the quantitative measures that may be derived from state space grids may also be applicable to Karnaugh maps. Coupled Equations Most generally, this method refers to the use of paired equations derived from two synchronized time-series

DS Methods

that produce parameters that describe the underlying dynamics of a system. The use of coupled equations may involve mathematically demanding procedures and often require fine-grained time-series data or simulated data. One particularly successful application in the field has been the work of John Gottman and colleagues, who used coupled differential equations to model the dynamics of marital couples and to predict from those dynamics couples who will remain married or end up divorcing (Gottman, Coan, Carrere, & Swanson, 1998; Ryan, Gottman, Murray, Carrere, & Swanson, 2000). They also used this method to study how peer interactions influence the behavior of developmentally delayed versus normal children (Gottman, Guralnick, Wilson, Swanson, & Murray, 1997) and parent–infant interaction (Gottman, Murray, Swanson, Tyson, & Swanson, 2002). Gottman and colleagues’ modeling procedures, based on mathematical biology, are an attempt to mathematically formalize theoretical assertions made by general systems theorists such as von Bertalanffy (1968). Their model itself represents a theory about the mechanisms and dynamics of marriage. Their approach is to model maritial interactions using the mathematics of difference and differential equations. These equations express hypothesized mechanisms of change over time. Gottman and colleagues offered their model of marital interactions as a first attempt to create an explanatory theory of marriage; they recognize the value of cycling from theory development to modeling to experiments and back to theory development. As the authors comment, “The theory may be dead wrong, but it is precise” (Gottman et al., 2002, p. 67). The modeling process is guided by posing two questions: (1) What are the steady states; and (2) which of these states are stable (i.e., the attractors) and which are unstable (i.e., repellors)? The first step is to set up the phase space such that each axis represents a variable corresponding to each participant. (This procedure is not the same as the one we described earlier in lag-1 phase plots.) The next step was to identify the steady states by graphing the null clines, curves in phase space for which variables stay constant. These curves help to determine the steady states by partitioning the phase space into regions where changes in each variable are increasing or decreasing. Gottman et al.’s (2002)’s technique uses a time-series of coded observational data to create an equation for each person in the interaction. For this method, Gottman et al. uses variables that are codes or categories of codes that fall along a single continuous dimension from very negative to very positive with neutral in the middle. Equations are created to reflect the theoretical presumption that one

735

person’s behavior at time t is a function of his or her own behavior at time t – 1 as well as the other participant’s behavior at time t – 1. Thus, there is an uninfluenced and an influenced factor in each equation that determines the behavior at each time t. The uninfluenced component is what each person brings to the interaction independent of the other. The uninfluenced set point can be thought of as a temperamental factor that is assumed to be the attractor a person, left on his or her own, will eventually approach. Mathematically, this is denoted by a parameter that represents the rate of change toward the uninfluenced set point, also called emotional inertia, plus a constant that is specific to the person. The influenced component is a mathematical function representing the other person’s previous behavior. These functions can be used to plot curves that describe the degree of influence based on one assumption: positive behavior influences positively and negative behavior influences negatively. Thus, the equations for a husband, H, and wife, W, are Wt+1 = r1 Wt + a + I(Ht ) Ht+1 = r2 Ht + b + I(Wt ) where r is the emotional inertia, a and b are constants specific to the individual, and I is the influence function. These equations are then manipulated to estimate the parameters and to determine the uninfluenced and influenced set points. Plotting both of the null clines (the curves for which the values of the variables remain constant) in phase space allows for the identification of stable steady states—the points where the two curves intersect. Consider the example in Figure 16.9a, which shows a bilinear null clines for a husband and wife. There is one positive and one negative stable state, or attractor, for this couple (the circles at the intersection of the two null clines). W

W

H

A

H

B

Figure 16.9 Null clines and the stability of steady states. Steady states are circled and represented by the point at which the null clines intersect. Source: Adapted from Gottman, J. M. (2002). The mathematics of marriage: Dynamic nonlinear models. London, UK: MIT Press.

736

DS Methods

Over the course of the interaction their behavior is drawn to one or the other of these attractors. One of the interesting uses of this modeling technique is its depiction of how a system exhibits sensitivity to initial conditions and how marriages can change over time. For example, situations at work may lead the husband to feel a lot of stress and he may begin to express this stress in his interactions with his wife. This would be reflected in a change in the parameter b in the model that corresponds to the husband’s uninfluenced state, or what he brings to the interaction. A change in this constant shifts his influence function to the left (Figure 16.9b) to the point where the positive attractor no longer exists. Thus, this type of modeling can distinguish between superficial change (i.e., a move to different stable state) and deeper change (i.e., a fundamental change in the influence function that eliminates or adds a stable state to the marital couple’s behavioral repertoire). Gottman and colleagues (2002) have used this basic model and more elaborated versions to identify mechanisms related to specific types of marriages. Some couples only have one stable steady state, if it is positive the marriage is happy, if it is negative the marriage is unhappy. Typically, the situation is more complex. Happy couples also tend to have both positive and negative attractors. In fact, high influence and high inertia couples tend to have multiple steady states and thus will most likely have at least one negative attractor. It may be the relative strengths of these attractors that best distinguish distressed and nondistressed couples. Gottman and colleagues’ approach to mathematical modeling seems to hold exciting potential for developmental psychopathologists who are interested in understanding social processes of various kinds. For example, their model can be easily adjusted to simulate interactions that have different types of power imbalances: parent–child interactions, therapist–client sessions, sibling interactions, bully– victim relations, and so on. Developmental psychopathologists have constructed detailed theoretical models that hypothesize specific influence functions among social partners (e.g., coercion processes in families with aggressive children, authoritative parenting, and anxiety-disordered children). Gottman et al.’s approach challenges scholars in the field to translate these theories into mathematically specified mechanisms. This approach may be particularly rewarding if developmental principles can be formalized and included in these endeavors. Nonlinear Dynamics These methods are derived from mathematical procedures in physics and other sciences that aim to measure and

model nonlinear phenomena (Abraham, Abraham, & Shaw, 1990; Heath, 2000; Norton, 1995). The simplest of these is nonlinear regression wherein the parameters are exponential functions or the predictive combinations are not simply additive. Other applications from this area are those related to chaos theory (Newell & Molenaar, 1998). One can use nonlinear dynamic techniques to find the embedding dimension, entropy, determinism, recurrence, or fractal dimension of any time series of sufficient length and sufficient precision. This class of techniques is typically applied to continuous time-series data (type 1), often physiological. For those developmental psychopathologists who are increasingly collecting this sort of data (e.g., heart rate, skin conductance, neural activity), this class of methods holds a great deal of promise, particularly if there are reasons to hypothesize nonlinearities. Moreover, there are many software packages available for free or at minimal cost that can calculate these measures for any appropriate time series. The challenge, therefore, lies not in the calculation but in the psychological interpretation of these measures. There are a number of diverse applications of this class of techniques. One of the main reasons for using these methods is to determine the degree of orderliness, complexity, or predictability in the behavior of a given system. One example in developmental psychopathology comes from the work of Dishion, Nelson, Bullock, and Winter (2004). They applied a technique that calculated the degree of complexity of peer interactions, in the hopes of characterizing and differentiating prosocial versus antisocial peer interactions. The authors drew upon information theory and applied Shannon’s entropy measure (Shannon & Weaver, 1949) as an index of the degree of disorganization, uncertainty, or information in a dyadic interaction. In information theory, uncertainty and information are related because “information about an unknown event is more useful as one knows less about the event” (Wickens, 1989, p. 231). For example, if 100% of a set of communications is of one type (e.g., deviant talk), there is total certainty and no information is conveyed from one communication to another. This is a case of low entropy. Conversely, if there are a variety of types of communications and all are equally probable, then there is total uncertainty and a maximal amount of information is gained by finding out which event occurred. This is an example of maximum entropy. Shannon’s entropy is typically used with contingency tables where the values in each cell represent the probability of a different two-event sequence. A simple way to think of entropy is the average number of dichotomous

DS Methods

decisions (yes–no) that need to be made to classify an event in that table (Wickens, 1989). Thus, in the previous example, the situation where only one event is observed would have a minimal entropy value of 1 because only one dichotomous decision is necessary to classify any event. Shannon’s entropy is calculated as follows: Entropy = H = Σpij (log2 (1∕pij )) where pij is the conditional probability in the cell that is in row i and column j of a contingency table. Dishion et al. (2004) used turn-taking observational data of boys interacting with their closest friend to calculate entropy. Their conversational turns were coded into five categories: negative engagement, directive, converse, positive engagement, and other. Thus, there was a 5 × 5 matrix of two-event sequences for each peer dyad. The frequencies in each cell of the matrix were used to calculate the conditional probabilities necessary for the entropy calculation. The findings from this study showed that all of the friendship dyads’ interactions (regardless of whether or not they were antisocial) became more ordered (less entropic) over three time points, ages 14, 16, and 18 years old. This finding in itself is compelling. It supports the DS view that development is characterized by increasing orderliness—a crystallization of interaction patterns. As relationships of all kinds develop, they stabilize and, thus, become more predictable over time. When the authors compared early-onset antisocial boys with prosocial boys, they found that antisocial boys’ interactions corresponded to higher entropy scores, suggesting a more disorganized pattern, and were characterized by more deviant talk than the typically developing boys. These results may suggest that antisocial boys are less skilled than prosocial boys at carrying on a fluid, appropriately contingent conversation. Some other potentially useful applications of this approach are found in clinical therapy research (e.g., Badalamenti & Langs, 1992; Burlingame, Fuhriman, & Barnum, 1995; Butz, Chamberlain, & McCown, 1997; Lichtenberg & Heck, 1986; Pincus, 2001; Tschacher, Scheier, & Grawe, 1998). Lichtenberg and Heck (1986) examined the “Gloria” psychotherapy tapes and compared the results of Shannon’s entropy measure, Markov chain analysis, and sequential analysis. Each of these approaches showed some merit and each had associated limitations. In terms of therapeutic process and outcome, two therapy studies have applied Shannon’s entropy measure with somewhat contradictory results: one found that, over the course of intervention, verbal utterances became increasingly complex (Badalamenti & Langs, 1992) while the

737

other found that client and therapist measures of therapy sessions became increasingly coherent (i.e., less complex, according to the entropy measure) and this tendency was related to treatment success (Tschacher et al., 1998). Finally, a study conducted by Pincus (2001) set out to demonstrate that when treatment works, therapy sessions become both more complex and more coherent. This study may be particularly relevant to developmental psychopathologists because it was based on assumptions from family systems theory (Minuchin, 1974). Pincus was interested in empirically testing one of the basic assumptions of family systems therapy: that families maintain a state of equilibrium through negative feedback processes which dampen deviations. These deviations (e.g., depressive withdrawal, angry outbursts) manifest as symptoms for a particular family member, but also serve to maintain homeostasis in the family (e.g., Minuchin, 1974; Minuchin & Fishman, 1974). For example, a child may become oppositional to (consciously or unconsciously) distract the family from the real problem: intense marital discord. When the child acts out, the parents join forces and attend to the disruptive behavior, which precludes escalation of the marital conflict (dampening process); thus, the family maintains a state of equilibrium, however dysfunctional it may be. The goal of family therapy is twofold: “to disrupt or block these negative feedback processes, promoting family disequilibrium and reorganization” and to “capitalize on natural deviation-amplifying processes within family systems in order to bring about systemic change” (positive feedback; Pincus, 2001, p. 145). In short, the goal of therapy is to induce a phase transition in the family system. As a first step toward validating these assumptions, Pincus conducted a pilot study in which he analyzed the conversations of one family undergoing therapy. He used an analytic technique called orbital decomposition (OD; Guastello, Hyde, & Odak, 1998), which combined Shannon’s entropy and measures of Lyapunov dimensionality, as well as others, to identify hierarchical patterns, the extent of randomness in a time series, and the optimal level of analysis for any particular series (i.e., the number of units in a string that best captures the information for a time series). The results confirmed that, indeed, the family exhibited evidence of both coherence and complexity across treatment sessions. Because this was a pilot study, we are hesitant to make too much of these findings. But OD may hold particular promise for developmental psychopathologists because it was developed for analyzing categorical time series data, instead of the usual continuous series. As such, it addresses a major limitation that has

738

DS Methods

been repeatedly raised with the other time series techniques reviewed here. The methods for analyzing nonlinear dynamics described in this section have been applied to single time series. These are appropriate for some biological measures with high temporal resolution (e.g., heart rate) but another form of continuous data, electrical activity in the brain, presents a more complex challenge. There are increasingly more developmental psychopathologists who are integrating brain activity measures in their research programs (e.g., Calkins & Fox, 2002; Cicchetti & Dawson, 2002a; Nelson et al., 2002). Although there is clearly not enough space to discuss the various brain imaging methods in the current review, a word on these techniques is warranted. Ongoing electrical activity in the brain, measured by electroencephalogram (EEG), can involve as many as 256 separate time series, one per electrode. Methods that analyze these kinds of complex data are designed to find relations and patterns among and across these various time series. For example, developmental neuroscientists interested in the self-organization of brain dynamics have found real-time synchronization across distal cortical areas to be correlated with attentional states or focused perception (Lewis, 2005). Transient, organized activity in different regions oscillating in phase (phase synchrony) and cortical coherence (a similar measure that incorporates both amplitude and phase) are detected through various temporal correlation methods and have been found to correspond with attentional (e.g., Engels, Fries, & Singer, 2001) and emotional (Paré & Collins, 2000) events. We close this section with a cautionary note that applies to any dynamical mathematical technique—because these techniques are derived from other fields, oftentimes it may be easier to input numbers into the equations than to sensibly interpret the results of the computations. A number of DS researchers have been criticized for interpreting their mathematical functions too literally, either the data inserted into the equations do not conform to the mathematical assumptions or, perhaps more seriously, they are potentially arbitrary in terms of their psychological meaning. These methods hold a great deal of promise, but a great deal of theoretical and empirical work remains to be done before the psychological meaning of nonlinear dynamic techniques is clearly understood. Developmental-Time Measures Descriptive Developmental Profile Analysis The empirical work by Thelen and colleagues (e.g., Thelen & Smith, 1994; Thelen & Ulrich, 1991) has been characterized as descriptive in that it is nonparametric, uses

descriptive statistics, and often relies heavily on displaying individual developmental profiles graphically. These researchers most often collect continuous time-series data (e.g., number of alternating steps, degree of displacement of foot, proportion of stepping cycle) over repeated occasions across a significant developmental period. They create developmental profiles on a case-by-case basis and describe the similarities and differences among these profiles. A core concern in this type of analysis is to identify periods of transition during which variability dramatically increases and old behavioral habits dissolve, giving rise to new ones (i.e., from crawling to walking). The most straightforward way to graph developmental growth profiles is to simply plot scores for the collective variable (e.g., alternate stepping) on the y axis across several developmental time points on the x axis, and to do so for each case individually. Van Geert and van Dijk (2002) offered a number of additional methods for visualizing and describing intraindividual time-series data over repeated observations. These methods were specifically created for measuring changes in variability across developmental time and they are well suited for one-dimensional continuous longitudinal data. In addition, with any individual developmental profile plot, several trajectories can be grouped into clusters that show similar profiles by using variations of a cluster analytic procedure (e.g., Hollenstein, Granic, Stoolmiller, & Snyder, 2004). One simple but elegant plotting technique is referred to as the moving min-max graph (van Geert & van Dijk, 2002). This graph represents the developmental trend line but also the variability around this trend line in a bandwidth of observed scores. A moving window is used—that is, a time frame selected by the researcher (e.g., 2 weeks, 3 months) is moved over one position all the way across the time series such that each window partly overlaps with the previous window. There is no absolute window length that is required, but van Geert and van Dijk (2002) suggest that the window size should be about one-tenth the size of the entire data set and no less than five data points. The maximum and minimum score at each time window is plotted and then the bandwidth of these scores can be examined in terms of its stability and change. In addition, a moving average that corresponds to the moving min-max window can be computed and displayed along the middle of the bandwidth. Figure 16.10 shows a graph taken from van Geert and van Dijk (2002) in which they showed the developmental trend for one child’s acquisition of spatial prepositions. They used a time window of 18 days across a 10-month period. The graph shows that there is a moderate sized bandwidth in the beginning of the trajectory and then, after a couple of oscillations, a

DS Methods 3

MLU-w

2.5 2 1.5

1/18/1999

12/18/1998

11/18/1998

10/18/1998

9/18/1998

8/18/1998

7/18/1998

6/18/1998

5/18/1998

4/18/1998

3/18/1998

1

Observation date Min

Average

Max

Figure 16.10 Moving min-max graph representing one child’s acquisition of spatial prepositions (time window of 18 days, last window, 15 days) From van Geert, P., & van Dijk, M. 2002). Focus on variability: New tools to study intra-individual variability in developmental data. Infant Behavior and Development, 25, 340–370.

large increase in the bandwidth is maintained until the end of the observation period. The authors interpreted this plot as suggesting a developmental transition (one that showed the characteristics of a phase transition) in the use of spatial prepositions. For a more detailed examination of the values within the bandwidth, the authors also propose using an “altitude line graph,” which simply depicts intermediate values as well (e.g., the second highest and second lowest values, percentiles) with a line connecting these values. The result is a plot that resembles the contour lines in a geographical relief map. A number of other variations of these variability charts can be found in van Geert and van Dijk (2002). To explore developmental trends in individual or group data, developmentalists commonly plot idealized or processed estimations of a data series (e.g., regression models, smoothing techniques) that obscure information on variability (this is what they are designed to do: smooth out fluctuations). In general, developmentalists constrain themselves to a very small repertoire of smoothing techniques and, thus, lose potentially critical information about the characteristics of transitions and the idiosyncratic nature of development (van Geert & van Dijk, 2002). Van Geert and van Dijk (2002) argued for the application of more sophisticated smoothing models “which follow the actual rise and fall of the data as faithfully as one wishes” (p. 351). They provide alternatives, many of which are options available in commonly used statistical packages including SPSS. Some of these techniques include spline models and local polynomial regression models, or loess

739

smoothers, which allow researchers to detect fluctuations and nonlinear trends in their data sets (Simonoff, 1996; van Geert & van Dijk, 2002). To complement the impressions gained from the graphical developmental profiles, DS researchers may use simple statistics such as within-subject measures of variance (e.g., standard deviations) and correlations to track increases in variability across developmental transition points. The well-known standard deviation (SD) measure is sometimes problematic, however, because of its sensitivity to the mean (i.e., if distributions are censored at a low score, the SD tends to increase when the mean increases), making SDs difficult to compare across individual subjects or samples. To address this limitation, another less common measure of variability can be used: the coefficient of variation (CV; van Geert & van Dijk, 2002). The CV is simply the SD divided by the mean and, although it too has limitations, it can be used as an alternative to the SD. Skewness is another easily calculated measure that provides a great deal of information about the pattern of variability in a given trajectory (van Geert & van Dijk, 2002). A skewed distribution is one potential indicator of bimodality, one of the catastrophe flags previously mentioned. In DS terms, bimodality can provide information about the extent to which a particular skill has been acquired or a pattern has been consolidated. For example, in the context of acquiring theory of mind (ToM), a 3.5-year-old child who is just beginning to understand that others may hold false beliefs about the child will likely show very few correct responses to ToM tests over several weeks of repeated testing; these will show up as positive outliers. In general, the child will still be responding “incorrectly,” as if she has no false belief understanding; thus, the distribution will be positively skewed. At the age of 4, however, this child will likely show a negatively skewed distribution of responses—her ToM skills will have consolidated and now only very rarely will she respond as if she did not understand that others’ can hold false beliefs. A similar example might be used for thinking about children in treatment. For instance, at first, a child undergoing cognitive-behavioral therapy for problems with anxiety may respond to frightening or challenging situations with predominantly avoidant strategies and very few proactive, assertive behaviors (i.e., the distribution will be positively skewed, with most behaviors falling toward the avoidant end of the continuum and a few outliers falling toward the assertive end). But if treatment is successful, this child may end therapy showing a pattern of responses that is negatively skewed. The point at which the distribution shifts from being positively to negatively skewed may provide key information about the timing and profile of change

740

DS Methods

across treatment. From a DS perspective, it is likely that this child’s distribution of avoidant-assertive scores from pre- to posttreatment will shift from positively skewed to bimodal to negatively skewed—that is, at no point will he exhibit responses that show a normal distribution. This pattern would suggest an abrupt shift (a phase transition) in response styles may be related to treatment success. The main point here is that instead of trying to normalize distributions at any stage, information about skewness and variability can provide insights about the nature of transitions and developmental change. Standard statistical packages (e.g., SPSS, SAS) provide tests for assessing skewness. A Kolmogorow-Smirnov test provides a statistic for sample sizes (or number of observations) greater than 50 and the Shapiro-Wild test for sample sizes less than 50 (van Geert & van Dijk, 2002). In addition, van Geert and van Dijk (2002) described bootstrapping methods that provide researchers a random sample with which to compare statistically significant differences in skewness values across a given developmental trajectory. Finally, individual growth curve modeling procedures can be applied to examine different profiles of change on both the individual and group level. This method is familiar to many developmentalists; it has been applied to examine different trajectories of antisocial behavior (e.g., Nagin & Tremblay, 1999), infant vocal development (Hsu & Fogel, 2001), infant-mother interactions (vanden Boom & Hoeksma, 1994), as well as many other domains. Several easily accessible computer programs are available to run these modeling procedures including multilevel analysis and hierarchical linear modeling (HLM; Willett, 1997). Latent Class Analysis Descriptive developmental profile analyses, as we have described them, are generally restricted to continuous one-dimensional data. Categorical and nominal data are difficult to use with these approaches (without somehow transforming them into continuous data). In this regard, combining different types of developmental profile analyses with methods that can capture content-specific changes in categorical or ordinal variables may be important. Another elegant option for investigating the bimodal properties of a distribution is offered by van der Maas (1998): latent class analysis (LCA). This technique is appropriate for analyzing categorical data. A number of developmental psychopathologists familiar with modeling procedures such as structural equation modeling (SEM) and HLM may also be aware of this method. LCA is conceptually similar to cluster analytic procedures in that

it identifies subgroups or types. But it has additional power because it also provides a maximum likelihood estimation and is adaptable to different data structures with no a priori assumptions about the distributions. As a cognitive developmentalist, van der Maas (1998) was particularly interested in this method for modeling the bimodal response distributions identified on various cognitive tasks at developmental transiton points or across various contexts (Quinlan, van der Maas, Jansen, Booij, & Rendell, 2007; van Rijn, Someren, & van der Maas, 2003). He argued that cognitive strategies can be best understood as attractors that are more or less available at different ages and in different contexts. In his approach, response patterns to multiple tasks are analyzed in terms of a cusp catastrophe model and multistable states (bimodality) are identified through LCA procedures. By presenting several examples of the application of rule and strategy detection in multiple domains, van der Maas (1998) showed that LCA may be one of the more promising psychometric models for identifying attractors and multistable states with categorical data. Dynamic Growth Modeling A group of scholars from the Netherlands can be credited for having pioneered the use of dynamic growth models in the study of cognitive developmental transitions (e.g., van der Maas, 1998; van der Maas & Molenaar, 1992; van Geert, 1994, 1995ab, 1997). The class of techniques advocated by these researchers is often grouped under the heading nonlinear dynamics, in reference to the modeling techniques and equations that are associated with this branch of mathematics. Although the techniques have often been applied to real-time data, these scholars explore various applications to developmental modeling instead. Dynamic growth modeling was developed to simulate change over time (or growth) using logistic difference equations. Van Geert (1994) has used this procedure to model the processes underlying stage-like transitions in the growth of syntactic forms. The basic DS premise of van Geert’s technique is that development of cognitive capacities is much like the self-organized proliferation of multiple species over the course of evolution. Van Geert models cognitive growers that constitute a complex system of intraindividual and environmental relations. Like the real-time coupled equations described earlier in reference to Gottman’s work, the modeling procedure is realized with iterative equations in which behavioral states are updated by the amplifying and dampening forces inherent in the system’s (i.e., the child’s) experiential history combined with current external (contextual) resources and limitations. The

DS Methods

growth models depict the kinds of nonlinear developmental profiles predicted by stage theorists (e.g., Piaget and Vygotsky). They also emphasize the use of graphical procedures to plot empirical data from longitudinal studies of children’s cognitive skill acquisition to match these empirical profiles to the data from simulations derived from the equations. The mathematical formulations used in these models are common in other disciplines; what makes these models unique is the application of both psychological and DS concepts to identify the mechanisms of growth that model observed developmental profiles. Another critical strength of this approach is that the process of developing simulations forces the researcher to specify the null hypothesis explicitly. The conventional a priori comparison between the hypothesized result and a lack of any effect is replaced by a specific null hypothesis of an alternative pattern of results. More recently van Geert and his colleagues have broadened their cognitive developmental focus and have become interested in modeling children’s peer play interactions (Steenbeek & van Geert, 2005, 2007, 2008). This new modeling endeavor may be more relevant to developmental psychopathologists because of its content (i.e., peer interactions), but more importantly, it may prove more generalizable to a large body of social interaction problems that developmental psychopathologists are interested in tackling. Similar to previous endeavors, the procedure begins with a theoretical model. In this case, Steenbeek and van

Geert (2005, 2007) developed a theoretical model that stipulated that differences in peer interactions with rejected versus popular children can be understood as a function of the following reciprocally interacting factors: each child’s goals (autonomy and involvement), the drive that results from the extent to which these goals are met, the emotional appraisal at each given moment, and the resulting emotional expression and behavior. These theoretical factors are then formalized into a mathematical model that stipulates the mechanism differentiating rejected and popular children’s dyadic interactions. Because this is a DS-based model, the mathematical formulation necessarily includes reciprocal, iterative processes and feedback cycles within and across occasions, as well as sensitivity to random or chance variations, and the potential for nonlinear results. Next, the simulation is run repeatedly and the results of the simulation are compared to theoretical expectations. Finally, the results of the simulation are compared to empirical data from real dyadic interactions. The model includes a number of input parameters that can be set at the outset. These values can be start values for a particular variable or the strength of the connection between two variables. Rejected and popular children, for example, are assigned different preference values according to the original theoretical model. Figure 16.11 is a representation of the simulation model with all of the reciprocal and iterative processes depicted by arrows. Table 16.3 shows the different parameters and what values can be manipulated at the outset. The table also shows the parameter

time t+1

time

child 1

preferred level

preferred level

realized level

realized level

741

child 2

drive

behavior

behavior

drive

emotional appraisal

emotional expression

emotional expression

emotional appraisal

drive

behavior

behavior

drive

emotional appraisal

emotional expression

emotional expression

emotional appraisal

preferred level

realized level

preferred level

realized level

Figure 16.11 Steenbeek and van Geert’s (2005) peer interaction simulation model. The first row of boxes represents the first moment in time, t, and the second row represents the second moment, t + 1. Thin arrows represent the iterative feedback components (the output of one iteration is the input for the next). Source: Steenbeek, H., and van Geert, P., 2005. A Dynamic Systems Model of Dyadic Interaction during Play of Two Children. European Journal of Developmental Psychology, 2(2), 105–145.

742

I little bit stronger than A

I much stronger than A I = average A = high Positive = moderate Negative = difficult Positive = big Negative = big Continuity = average Symmetry = high

1 Concerns 2 Influence of behaviour on realization of concerns 3 Relation between emotional appraisal and emotional expression 4 Influence of emotional expression on preference of concerns 5 Nonintentional principles of behaviour

Continuity = average Symmetry = average

Positive = average Negative = average

Positive = moderate Negative = moderate

I = average A = average

I stronger than A

Average

Continuity = average Symmetry = average

Positive = average Negative = average

Positive = moderate Negative = moderate

I = average A = average

I stronger than A

Average

Average dyad

Continuity = average Symmetry = low

Positive = average Negative = average

Positive = moderate Negative = moderate

I = high A = average

I little bit stronger than A

Popular

Continuity = average Symmetry = high

Positive = big Negative = big

Positive = moderate Negative = difficult

I = average A = high

I much stronger than A

Average

Popular dyad

Notes: a The table can be read as follows: a child with a rejected status that plays with a play partner with an average status (rejected dyad; upper left of the table) has as setting for the first input parameter group concerns that his/her concern involvement is much stronger than his/her concern autonomy, which is expressed as “I much stronger than A”. b I = concern involvement, A = concern autonomy. The average dyad has standard settings (average, moderate), Everything that differs from these standard settings is printed in italics (high, low, difficult, big). c Information about the corresponding ranges of numerical values of these settings of input-parameter groups for children of different statuses in the context of playing with a play partner with a different status can be found in Appendix 2. Source: Steenbeek, H., and van Geert, P. (2005). A dynamic systems model of dyadic interaction during play of two children. European Journal of Developmental Psychology, 2(2), 105–145.

Continuity = average Symmetry = low

Positive = average Negative = average

Positive = moderate Negative = moderate

I = high A = average

Average

Rejected

Rejected dyad

Status of child Status of play partner

Type of dyad

TABLE 16.3 Parameter profiles for various types of dyads

DS Methods

profiles for each type of dyad (the various combinations with rejected, popular, and average children). The model is run, for example, 1,000 times with the particular inputs specified in Table 16.3 for children with different sociometric status. One of the unexpected results that emerged from running the simulations was that, contrary to what previous literature suggested, the data representing rejected children playing with normal-status children showed more positive emotional expressions than that for popular children with normal-status children (Steenbeek & van Geert, 2005, 2008). When these simulation results were compared to actual empirical data, the results matched (statistical tests described by the authors in detail were run to compare the simulated and empirical results). The authors concluded that, because rejected children seldom had a chance to play with other children in general, the new experience of playing with someone else led to the expression of more happy emotions than expected. The popular children, on the other hand, were not as pleased with the experience because they were not playing with their regular friends and had plenty of opportunities to play with other children. The authors suggest that the modeling results are consistent with contextualist, functionalist, and dynamic systems perspectives. Indeed, their models remain true to fundamental DS assumptions. Broadly, the simulation technique demonstrated that emotions and behavior: . . . Are not a function of] personality characteristics that are relatively independent of the contexts and of the local and temporal realization of the concerns . . . [instead, the findings] emphasize the importance of iterative and interacting processes that take place in real time. It overthrows the classical distinction between independent and dependent variables in favor of a process of mutually interacting and simultaneous processes. (Steenbeek & van Geert, 2005)

Van Geert (1998) emphasized the critical role mathematical modeling plays in developmental research: “In order to find out the implications of theories, they have to be transformed into mathematical models that capture the major dynamic principles of such models and that can be used to explore the range of developmental trajectories under all possible or likely parameter conditions” (p. 155; also see Newtson, 1994; van Geert, 1994). But despite van Geert’s concerted efforts to make his approach accessible, its impact on developmental psychology and developmental psychopathology may be limited because of its use of mathematical procedures that are daunting to most psychologists. More importantly, a great deal of the developmental phenomena that researchers are interested

743

in examining are not conducive to this sort of modeling because they are not easily quantified as continuous variables. Also, regardless of the domain in which this method is applied, and as with all simulation techniques, the correspondence of the parameters in the model with genuine psychological mechanisms is often difficult to evaluate and runs the risk of seeming arbitrary. Nevertheless, this modeling technique and others may prove useful for developmental psychopathologists who have very precise hypotheses that can be faithfully translated into mathematical functions that specify mechanisms of change, influence, and interaction. Connectionist Modeling Connectionist and dynamic systems theories have separate histories, but share a great deal of commonalities (Spencer & Thelen, 2003). Although it is not quite accurate to characterize connectionism as a dynamic systems methodology, we would be missing an important contribution if we were not to include this modeling approach. Many (but not all) connectionist models are indeed dynamic systems and researchers working with these models often turn to concepts such as attractors, bifurcations, and chaos to understand their properties (Smith & Samuelson, 2003). Connectionism is often characterized as a framework for simulating idealized brain-like, or neural network, processing. This is because connectionist modeling starts with the propagation of activation across a network of abstract units that can be hypothetically likened to networks of neurons that fire or do not, depending on the context. Knowledge, representation, or meaning is understood as being stored not in one or two units, but in connection weights between processing units. Learning and development is evidenced by changes in connections between these units. These changes result from statistical regularities in the input to the network and these regularities are governed by the Hebbian learning principle: units that fire together, wire together. Thus, “the connection weights between layers [of units]—the response of the network to a specific input—depend on the statistical regularities in the network’s history of experiences” (Thelen & Smith, 1998, p. 580). Through their own real-time activity, without any additional intervention, connectionist networks change their own connections and, thus, teach themselves. One of the best-known examples of a connectionist model is Rumelhart and McClelland’s (1986) demonstration of a network that acquired irregular and regular forms of past-tense, and overgeneralized (like real children were thought to do) the irregular forms. Their breakthrough was to show that a network could behave as if it was

744

DS Methods

governed by higher order rules with no such rules having been prespecified. In most connectionist models, activation among units is a continuous temporal process. Differential equations are used to simulate this continuous activation process such that weighted connections link the rate of change in one variable to the inputs received from the other units (e.g., Munakata & McClelland, 2003). In many recent models, the activation process is assumed to be nonlinear. A simple example is provided by a nonlinear adjustment to McClelland’s (1979) cascade model: dnetai = k(Σaj wij –ai ) dt

j

ai = f (neti ) where a represents the activation of receiving unit i, aj is the activation of sending unit j, wij represents the connection weight from unit j to unit i, and k is a time constant that sets the rate of change. netai refers to the net input to unit i, and f is a nonlinear function (Munakata & McClelland, 2003). In terms of development, connectionist simulations based on variations of the nonlinear differential equation have been developed to understand category and word learning (Plunkett, 1993; Smith, 1993, 1995), children’s knowledge of dimension words and their selective attention to each dimension (Gasser & Smith, 1996; Smith, 1993), the development of face recognition (Johnson & Morton, 1991), and the A-not-B error (Munakata, 1998), among others. In terms of the A-not-B error, Thelen and colleagues (Thelen, Schoener, Scheier, & Smith, 2001; Smith & Thelen, 2003) extended past connectionist models by applying the dynamic systems principles of multicausality and nested time scales and developed a dynamic field model. They incorporated the idea that instabilities in network activities (i.e., variability) index qualitative changes in systemic behavior. Spencer and Schoner (2003) summarize this dynamic connectionist model: “Instabilities disrupt the one-to-one input–output mapping and make dynamic fields nonstandard neural networks” (p. 406). Connectionist modeling, like other dynamical modeling techniques listed, requires either a great deal of technical expertise or collaboration among researchers with varying talents (e.g., computer programming, mathematics, developmental theory). We are aware of very few efforts in developmental psychopathology to apply this type of simulation procedure, but there may be a number of phenomena in the field that could benefit from such

an approach. For example, different types of friendship formations on the playground may be modeled through connectionist procedures. Children seem to self-organize into coalitions or cliques very quickly at the beginning of a school year. These emergent groupings may be modeled as networks of units that become strengthened over time. Different types of coalitions (e.g., bullies and victims versus supportive friendship relationships) may exhibit unique properties (e.g., bullies/victims may be less stable, more perturbable, and they may be less likely to include new members than other types of friendships). These unique properties may suggest adjustments to existing theoretical models of friendship formation and even have the potential to suggest change mechanisms that may be targeted through intervention. Readers interested in applying connectionist modeling are directed to Rethinking Innateness: A Connectionist Perspective on Development (Elman et al., 1996) and the companion book Exercises in Rethinking Innateness: A Handbook for Connectionist Simulations (Plunkett & Elman, 1997). These volumes provide a thorough primer in connectionist modeling and give novices an opportunity for hands-on practice. The authors make important distinctions between mechanisms of change and development and the content of those changes. Of special concern is the attempt to understand emergence and nested interactions at multiple time scales. The exercise book is a practical step-by-step manual that includes computer software and several example simulations that model, for example, stage-wise development and how children learn the past tense. Catastrophe Modeling This is a mathematical procedure wherein several control parameters can be used to account for discontinuous change in one of seven topological forms (catastrophe models; Thom, 1975; Guastello, 1995). As such, it has some appeal for developmental psychopathologists interested in incorporating insights from developmental stage theories or modeling nonlinear changes in behavior due to treatment or trauma. These models show how each of the possible combinations of control parameters result in different values of a dependent variable. The potential values of the dependent variable are represented on a plane, much like a state space. Nonlinear shifts can be depicted as a fold or curl in the plane showing where behavior changes suddenly rather than continuously. The simplest and most commonly used catastrophe model is a cusp catastrophe, as depicted in Figure 16.12. The top surface of this three-dimensional figure is the

DS Methods

behavior plane A

B

E F C

D G control plane

collective variable

splitting factor control parameter

Figure 16.12 Basic cusp catastrophe model.

behavior plane, which is made up of all the possible values of the response variable (i.e., collective variable). The location of a point in the behavior plane is predicted by the values of the control parameter and the splitting factor (the axes on the bottom plane called the control plane), which are equivalent to independent variables in these models. The top left position on the behavior plane (point A), for instance, is a combination of a low value of the control parameter and a low value of the splitting factor. Keeping the splitting factor constant for the moment, follow the movement from point A to point B. Here we see that there is a linear response in behavior as the control parameter increases continuously. However, when the values of the splitting factor are high, there are some values of the control parameter that result in two possible locations on the behavior plane (points D and E or F and G). This cusp fold in the behavior plane is the region that describes the nonlinear response dynamics of the system. As the control parameter increases from the lowest value at point C, the behavior increases through points G and D. However, at point D the collective variable value suddenly jumps to point E on the plane. When the control parameter decreases from its maximum value, behavior passes through point E to point F in a linear response. Now at point F, behavior suddenly jumps down to point G, a much lower value for the collective variable. This difference in response depending on whether the control parameter is increasing or decreasing is one of the unique features represented by cusp models and exemplifies the process of hysteresis. For example, consider the relationship between frustration and anger. The first step in catastrophe modeling is to identify that a catastrophe exists. The catastrophe flags mentioned earlier are used for this purpose. Presence of hysteresis, divergence from linear response, sudden jumps, bimodality, anomalous variance, and critical slowing down

745

can all serve as flags. For the current example, suppose that through experimentation a researcher discovers a bimodality in the distribution of anger responses to frustration. In cases where frustration is moderate, anger is either high or low but never in between. This would be an indication that a cusp catastrophe could be an appropriate model. Moreover, suppose this bimodality only occurs in high reward contexts, and that frustration and anger have a positive linear relationship in low reward contexts. Thus, for this example, anger is the collective variable of interest, frustration is the control parameter, and amount of reward in the context is the splitting factor (Figure 16.13). In a low-reward task (e.g., no motivational incentive other than following instructions), as frustration increases, the anger response would increase gradually (path A to B in Figure 16.13). In a high-reward task (e.g., in a gambling context), as frustration increases, the anger response would break from linearity at point D. Anger would suddenly jump to a higher level at point E. If frustration continues to increase, there would again be a linear increase in anger. However, frustration would have to subside well below the levels at which the previous jump occurred (point F) before anger decreased nonlinearly to its former level. The next step of the modeling procedure is to fit the model to empirical data, much like linear regressions are fit to data. Catastrophe models have been used to model diverse developmental transitions such as conservation of liquid (van der Maas & Molenaar, 1992); acquisition of syntax (Ruhland, & van Geert, 1998); infant reaching with grasping (Wimmers, Savelsbergh, van der Kamp, & Hartelman, 1998); dating (Tesser & Achee, 1994); blame responses of victims of assault (Lanza, 1999); adolescent smoking (Byrne, Mazanov, & Gregson, 2001); alcohol use (Clair, 1998); and anxiety in preuniversity students (Haslett, Smyrnios, & Osbourne, 1998). These and other

behavior plane A

B

E

F C

D G

Anger (collective variable)

control plane Reward (splitting factor)

Frustration (control parameter)

Figure 16.13 Cusp catastrophe model representing the relation between anger and frustration in different contexts.

746

DS Methods

examples indicate that, despite the difficulty in specifying or measuring control parameters, as discussed earlier, important insights about the profiles and mechanisms of change can be gained. Catastrophe theory and models will likely become more applicable as more researchers address the nonlinearities in their data and work to measure hypothesized control parameters. STATE SPACE GRID ANALYSIS: A GRAPHICAL AND STATISTICAL MIDDLE ROAD The various DS techniques introduced thus far have considerable potential for addressing some of the analytic challenges faced by developmental psychopathologists. However, we have also pointed out some obstacles for implementing these techniques. In general, most techniques require continuous data, whereas ordinal and categorical variables are more common in developmental psychopathology, especially in observational studies. In addition, many of the techniques are either solely descriptive, preventing researchers from testing the strength and reliability of their findings, or require complex mathematical procedures that may be inaccessible or irrelevant to most developmental psychopathologists. Recently, hybrid strategies have been developed, providing a middle ground by combining graphical techniques that capture the descriptive richness of DS concepts with simple statistical procedures that stay true to systems assumptions (Hollenstein, 2007, 2012, 2013; Lamey et al., 2004; Lewis et al., 1999). The state space grid method is a graphical and statistical strategy that links the analysis of real- and developmental time patterns and allows for the identification of individual and group differences. Thus, the flexibility of this methodology has proven valuable for developmental psychopathologists. We begin by describing this technique and the various measures that can be derived for statistical analysis. State Space Grids Recall that DS theorists use the concept of a state space to represent the range of behavioral habits, or attractors, for a given system. In real time, behavior is conceptualized as moving along a trajectory on this hypothetical landscape, being pulled toward certain attractors and freed from others (Figure 16.1). Based on these abstract formalizations, Lewis, Lamey, and Douglas (1999) developed a graphical approach that utilizes categorical time series structured as two ordinal variables that define the state space for any particular system. Though these state variables can represent two measures of one individual’s

behavior (e.g., Lewis et al., 1999, Lewis et al., 2004) so far the most frequent use of state space grids has been the analysis of dyadic interactions (Branje, 2008; Cerezo, Trenado, & Pons-Salvador, 2012; Connell, Hughes-Scalise, Klostermann, & Azem, 2011; Dishion et al., 2004; Gardner & Wampler, 2008; Granic, Dishion, & Hollenstein, 2003; Granic & Lamey, 2002; Granic et al., 2003, 2007; Granic et al., 2012; Hollenstein, 2007; Hollenstein & Lewis, 2006; Hollenstein et al., 2004; Lunkenheimer, Albrecht, & Kemp, 2012; Lunkenheimer et al., 2001, 2012; Mainhard, Pennings, Wubbels, & Brekelmans, 2012; Moore et al., 2012; Ribeiro, Bento, Salgado, Stiles, & Gonçalves, 2011; Sravish, Tronick, Hollenstein, & Beeghly, 2013; van Dijk, Hunnius, & van Geert, 2012; van der Giessen, Branje, Frijns, & Meeus, 2013; van der Giessen et al., 2015). With this application, a dyad’s trajectory (i.e., the sequence of behavioral states) is plotted as it proceeds in real time on a grid representing all possible behavioral combinations. Categories of one dyad member’s (e.g., parent) behavior forms the columns on the x axis and the other member’s (e.g., child) behavior forms the rows on the y axis. Each cell of the grid is the joint x–y state (e.g., parent–child) and plot points in these cells represent the occurrences of these joint states, and their durations if applicable. Moreover, a trajectory is drawn through the successive dyadic points in the temporal sequence they were observed. Any time the behavior changes a line is drawn connecting the new point and the previous point. For example, a hypothetical trajectory representing 15 seconds of coded behavior is presented in Figure 16.14. With the duration of each joint state represented by the size of the plot point and the direction of each transition indicated by an arrow, this sequence begins with the parent and child in conflict (shared negativity/hostility, bottom left cells) and ends in repair (shared positivity, top right cell). The SSG methodology can tap both the content of interactions (e.g., mutual hostility, permissive parenting) and the structure of these interactions. When we refer to structure, we mean the patterning of behavior, regardless of content, such as the relative flexibility versus rigidity of interaction patterns (Hollenstein, Lichtwarck-Aschoff, & Potworowski, in press). With SSGs, we are able to examine whether behavior clusters in very few or many states (i.e., cells), or regions (i.e., a subset of cells) of the state space. We can also track how long the trajectory remains in some cells but not others, and how quickly it returns to particular cells. If a dyadic trajectory remains in a small number of cells, and makes very few transitions between cells, this system may be thought of as rigid, inflexible, or stuck. In contrast, a trajectory that moves around to many cells in the state space grid and makes frequent changes between

State Space Grid Analysis: A Graphical and Statistical Middle Road

Neutral Negative Hostile

Child affect

Positive

Parent Child Simple Example.trj

Hostile

Negative

Neutral

Positive

Parent affect

Figure 16.14 Example of a state space grid with a hypothetical trajectory representing 10 seconds of coded behavior, one arrow head per second. Plotting begins in the lower left part of the cell and moves in a diagonal as each second is plotted, ending in the upper right. See footnote 1.

these cells may indicate a highly flexible, or variable, system. We can identify attractors as those cells to which behavior is drawn repeatedly, in which it rests over extended periods, and/or to which it returns quickly. Moreover, as discussed in the following sections, a range of variables that capture the relative stability of particular attractors may be derived from SSGs and these values can be tested statistically for changes in real and developmental time.

747

A major advantage of SSGs is that they provide an intuitively appealing way to view complex, interactional behavior; thus they are first and foremost a useful tool for exploratory analysis. An examination of the heterogeneity of family interactions with aggressive children may help illustrate this point (Granic & Lamey, 2002). SSGs were used to explore differences in the parent–child interactions of “pure” externalizing children (EXT) and children comorbid (MIXED) for externalizing and internalizing problems. This study is useful not only for demonstrating how the grids work, but also to demonstrate design innovations based on DS principles that are useful with or without SSGs, in this case, a systematic perturbation. Parents and clinically referred children discussed a problem for four minutes and then tried to “wrap up and end on a good note” in response to a signal (the perturbation) within the next two minutes. The perturbation was intended to increase the emotional pressure on the dyad, triggering a reorganization of their behavioral system. We hypothesized that, as a function of differences in the underlying structure of their relationships, EXT and MIXED dyads would be differentially sensitive to the perturbation and would reorganize to different parts of the state space. Prior to the perturbation, however, we expected dyads’ interactions to look relatively similar. Separate grids were constructed for the pre- and postperturbation interaction sessions. For this study, the lines (trajectories) are less important to notice than the points that show clustering in particular cells. Figure 16.15 shows an example of an

Preperturbation

Postperturbation

Positive

Neutral

Neutral

Negative

Negative

Hostile

Hostile

CHILD

Positive

Hostile

Negative

Neutral

Positive

Hostile

Negative

Neutral

Positive

PARENT

Figure 16.15 Pre- and postperturbation state space grids for an EXT dyad. Source: From Granic, I., & Lamey, A. V. (2002). Combining dynamic-systems and multivariate analyses to compare the mother–child interactions of externalizing subtypes. Journal of Abnormal Child Psychology, 30, 265–283.

748

DS Methods

interaction between a pure externalizing child and his parent, pre- and postperturbation. As exemplified in these grids, EXT dyads tended to go to the permissive region (child hostile–parent neutral/positive) of the state space grid, as well as other regions (i.e., mutual neutrality and negativity), before the perturbation. After the perturbation, EXT dyads tended to remain and stabilize in the permissive region. Figure 16.16 represents the interaction of a MIXED dyad. Similar to EXT dyads, the MIXED dyads occupied the permissive region, as well as other areas, before the perturbation. But in contrast with the EXT group, MIXED dyads tended to move toward the mutual hostility, or mutual negativity, region of the state space grid after the perturbation. Granic and Lamey (2002) concluded that the perturbation was a critical design innovation that provided the means by which clinical subtypes could be differentiated. Another contribution of this study was a more general one—the use of SSGs, with their rich case-by-case temporal narratives, provided a technique to further parse interaction processes that have been previously assumed to represent one coherent pattern. In this case, the coercive process (Patterson, 1982; Patterson, Reid, & Dishion, 1992; Snyder et al., 1994) was shown to constitute two separate microsocial patterns—two separate attractors on a state space (Granic & Patterson, 2006). Moreover, the conditions under which dyads would be drawn toward one region or the other were found to differ for subtypes. The use of SSGs to uncover heterogeneous processes may

be relevant to a variety of phenomena in developmental psychopathology including variability in the real-time unfolding of attachment patterns (Coleman & Watson, 2000), in bullying interactions on the playground (Pepler, Craig, & O’Connell, 1999) and in parent–adolescent interactions during puberty (Granic, Dishion, et al., 2003; Granic, Hollenstein, et al., 2003). State Space Grid Analysis: Within Grids In general, there are two ways to use SSGs for analysis: deriving measures from one specific state space arrangement (within-grid analysis) and analyzing similarities and differences between grids, such as with longitudinal designs (between-grid analysis). Within a given SSG, patterns can be quantified and used as variables for statistical analyses. Measures that capture the temporal and spatial patterning of behavior have been developed for time-based (e.g., second-by-second coding, daily diary entries) as well as event-based (e.g., conversational turns) data (Hollenstein, 2013). These variables can all be obtained via the SSG software program, GridWare (Lamey et al., 2004), available for free on the Internet (www.statespacegrids.org) and may be used to map intra- or interpersonal behavioral trajectories. Here, we highlight two within-grid analysis approaches relevant to the development of psychopathology (for a comprehensive review, see Hollenstein, 2013): attractor (or content) analysis and whole-grid (or structural) analysis. In general, long durations or frequent recurrences of

CHILD

Preperturbation

Postperturbation

Positive

Positive

Neutral

Neutral

Negative

Negative

Hostile

Hostile Hostile Negative

Neutral

Positive

Hostile

Negative

Neutral

Positive

PARENT

Figure 16.16 Pre- and postperturbation state space grids for a MIXED dyad. Source: From Granic, I., & Lamey, A. V. (2002). Combining dynamic-systems and multivariate analyses to compare the mother–child interactions of externalizing subtypes. Journal of Abnormal Child Psychology, 30, 265–283.

State Space Grid Analysis: A Graphical and Statistical Middle Road

turn), but a high return time for grid A, cell 1/0. Thus, multiple attractor strength indices can be used for subsequent statistical analyses. Whole-grid analysis can be used to derive measures of the overall variability or structure of trajectories on the SSG. There are two classes of overall variability measures: dispersion-based and transition-based. Measures of dispersion capture the overall spread of the trajectory based on the number of unique cells visited. With duration-based data, this dispersion can also be calculated controlling for the relative durations in each cell—cells with longer duration count more in the index than cells with low durations. Thus, high levels of dispersion indicate a more even distribution of behavior across the entire state space. Transitions are somewhat independent of dispersion. A trajectory can have a high number of transitions (i.e., the number of lines in the trajectory on an SSG plot) among many cells or only a few. Thus, a simple count of the number of transitions (or transitions-per-minute for studies with trajectories of unequal length) can indicate the dynamic flexibility of the system, as complementary information to dispersion. As discussed earlier, another transition-based measure is entropy, which is based on the lag 1 event sequences—higher entropy indicates less predictability. Both dispersion and transition measures of dyadic variability have been shown to correlate with children’s mood and behavior problems. Low variability, interpreted as low flexibility or high rigidity, in parent–child interactions has been associated with externalizing and internalizing problems in early childhood (Hollenstein et al., 2004), lack of response to treatment for externalizing problems (Granic et al., 2007), and maternal depression (Lunkenheimer et al., 2012). Returning to Figure 16.17, grids A and C show high dispersion, compared with grid B, and grid A shows a low stability 1 value, compared with grid B.

behavior in a particular cell or region suggest an attractor on the state space, and these hypothetical attractors can be compared and tested within individuals across development as well as between individuals. Attractor analysis can be theoretically driven (expected attractor states are compared across subjects; e.g., Granic & Lamey, 2002) or empirically derived for each case (e.g., Lewis et al., 1999). Lewis et al. (1999) is still the best example of empirically derived attractors for analysis. First, attractor cells or regions were identified via a winnowing technique that used the individual cell durations. Through iterations removing one cell at a time, the number of possible attractor states was winnowed down to only those cells in which durations exceeded chance. Each trajectory was found to have at least one cell—and sometimes two cells—that qualified as attractors. Once identified, analytical options for theoretically and empirically derived attractors are the same. For within-grid analyses, this primarily consists of determining the strength or relative pull of the attractor on the system’s trajectory. On top of relatively straightforward attractor indices such as cumulative, proportional, mean, or maximum durations in cells or groups of cells (i.e., regions) of interest, attractor strength can be measured via return time. Return time is the latency to return to a cell or region following an event in that cell or region. This can be measured in units of time, number of events, or number of unique cells visited en route. For example, Figure 16.17 shows three grids with attractor regions shaded in gray. Grid B shows a duration in cells 2/2 (cell labels follow the x/–y convention) and 2/3 and a very low duration in cell 1/3. In grid C, cell 3/1 shows a high mean duration (each time behavior goes to that cell, it tends to stay there for some time) whereas cell 1/1 in that same grid shows a low perseverance value. Finally, grid B shows a very low return time for cell 2/2 (every time behavior leaves that cell, it returns in approximately one

4

4

3

3 E

2

0

S 5

4

3

(a)

2

1

E

4 E

3 S

2

1

2

1

1

0

0 5

4

749

3

(b)

2

1

S 5

4

3

2

1

(c)

Figure 16.17 Examples of three state space grids from Lewis, Lamey, and Douglas (1999). Source: From Lewis, M. D., Lamey, A. V., & Douglas, L. (1999). A new dynamic systems method for the analysis of early socioemotional development. Developmental Science, 2, 457–475.

750

DS Methods

State Space Grid Analysis: Between-Grid Analysis After computing the within-grid parameters that are most relevant for a particular research question, longitudinal statistical techniques (most of which are quite familiar to developmental psychopathologists) may be applied. We recommend using these statistical procedures in such a way that maintains the integrity of the individual (or dyadic) case (e.g., growth clusters). However, multivariate analyses including one-way analyses of variance (ANOVAs), regressions, and SEM can just as easily be run on the grid variables. For example, we have shown over the course of five longitudinal observations that parent–boy variability (dispersion and transitions) peaked when the son was 13–14, consistent with the hypothesis that adolescence is a developmental phase transition of temporarily heightened variability in interpersonal interactions. The use of within-grid measures of repeated observations in this way is statistically relatively straightforward, thus we will not elaborate further here However, two between-grid techniques enable the kind of person-centered analyses of interest to system-minded researchers. Lewis et al. (2004) developed two techniques to depict month-to-month stability and change in toddler’s socioemotional habits. Observed once a month for 11 months (ages 14–24 months), each subject had 11 SSGs of the behavioral trajectories. The study was designed to test the hypothesis that the 18–20-month transition was a developmental phase transition—a temporary period of instability and fluctuation as the children developed from one stable behavioral pattern to another. Thus, the authors derived two indices from SSGs to see if the lowest month-to-month stability (i.e., greater change) occurred at 18–20 months. The first technique they used was to calculate the intergrid distance score (IDS). Grid-to-grid Euclidian distance scores yield a global metric of the difference in behavioral landscapes from month to month, based on the sum of squared differences across all cell pairs. For each grid cell pair, the difference in duration values over two consecutive months is calculated, then squared, and then these values are summed for all cells. Next, the square root of this sum is taken as the distance score between the two grids. The second technique was to put all SSG cell variables for all participants at every month into one cluster analysis. Each SSG was then identified as a member of a homogenous cluster. Then, returning back to a case-by-case computation, a cluster change score was calculated: a value of zero was assigned if the cluster membership of one month was the same as the subsequent month and a value of 1 was assigned if the cluster membership changed from one

month to the next. Thus, from 11 SSGs, 10 cluster change scores were calculated for each case. Both the IDS and cluster change score were highest during the 18–20-month period as hypothesized. Thus, SSG-to-SSG change can reveal patterns of development that go beyond using within-grid measures in standard statistical analysis. Although the SSG method is clearly still in its early stages of development, we are encouraged by its potential. One of the important advantages to this technique is its inherent flexibility. At the very least, it is a visual, exploratory tool to develop and refine hypotheses. At best, this methodology provides a source from which novel, temporal-based, process-level predictors may be tapped and used to strengthen current models of normal and atypical development. Researchers are not limited to using continuous time-series, as is the case with many other DS methods. Categorical and ordinal data are also appropriate for this type of analysis. Also, the grids are malleable in that they can represent systemic behavior on the individual as well as dyadic level. In addition to the aforementioned examples, changes in peer, romantic couples, and sibling interactions, for example, can easily be tracked using SSGs. In fact, apart from the difficulties in visually representing the data, the variables derived from the grids can be extended past the two dimensions on which we have focused. For instance, SSGs have been used to analyze triadic (Lavictoire et al., 2012) and group (Martin et al., 2005) interactions. Another benefit of this approach is the extent to which it remains user-friendly and does not require expertise in mathematical modeling. In addition, we have developed a SSG software program that will output grids, compute all the grid parameters mentioned previously, and export these measures for statistical analyses on these parameters: GridWare (Lamey et al., 2004), available for free on the Internet (www.statespacegrids.org). FUTURE DIRECTIONS: IMPLICATIONS FOR CLINICAL RESEARCH There are no limits to the diverse topics and varied problems that can potentially be addressed with the application of DS methods; clearly listing all of them here would be impossible. In this final section, we have chosen to limit our discussion to the implications for clinical research, an area in which many developmental psychopathologists share an interest. Because DS methods are specifically designed to capture change processes, and because the study of psychopathology often breaks down into the study of individual patterning, one of the most exciting potential

Future Directions: Implications for Clinical Research

applications of DS methods may be in treatment research. Although randomized controlled trials have helped to identify the most effective interventions for a variety of childhood mental health problems, there remains variability in outcomes (e.g., Eyberg, Nelson, & Boggs, 2008; In-Albon & Schneider, 2007) and almost no understanding about the underlying processes and mechanisms of change (Hinshaw 2002; Kazdin, 2000, 2001, 2002, 2009). This problem highlights our lack of understanding of the change process itself. Information about the mechanisms responsible for the success of interventions is critical for guiding clinicians in making informed decisions about how to tailor interventions to different contexts and for unique individuals and families. Also, identifying mechanisms of change is a crucial step toward more effective program dissemination in community settings (Kazdin, 2000). DS principles and methods should be able to provide a microsocial, process-level account of how family and peer relationships change over the course of treatment and why some may fail to do so (Gardner & Wampler, 2008; Granic et al., 2007; Lichtwarck-Aschoff et al., 2012). In the following, we present three hypotheses that can be generalized to any evidence-based intervention. First, phase transitions in normative development may be critical to mark because they allow clinicians and researchers to more efficiently, and perhaps more successfully, access and manipulate mechanisms of change. There may be normative stage transitions in children’s development (e.g., early childhood 3–5 years, early adolescence 11–14 years), as well as idiosyncratic transition periods (e.g., divorce or birth of a sibling) during which, the coordination among system elements begins to break down, previous attractors are destabilized, and new patterns have the potential to emerge (e.g., Lewis et al., 1999). During phase transitions, the system is much more open to environmental shifts, and seemingly small changes have the potential to radically alter the trajectory of relationships and individuals. As a result, prevention and intervention efforts that target youth mental health problems may have their maximal effect during these periods. Thus, based on DS ideas we would expect that clinical interventions and prevention efforts will be most effective if they are targeted at these sensitive periods. The same intervention targeted before or after a phase transition is hypothesized to be less successful. To test this hypothesis rigorously, it would be important to design treatment studies that examine the differential impact of the same evidence based intervention before, during, and after a recognized transition period. Second, psychotherapy researchers suggest that, for improvements to be made, treatment must trigger a

751

reorganization of affective, cognitive, and behavioral systems (e.g., Caspar, Rothenfluh, & Segal, 1992; Greenberg, Rice, & Elliott, 1996; Mahoney, 1991). To induce a major reorganization, “old patterns must be shaken loose or destabilized to allow for new configurations to emerge or to be discovered” (Hayes & Strauss, 1998, p. 940). Thus, “destabilization is viewed as a necessary and natural process that allows for growth and change” (p. 940). Although this destabilization period has been theoretically proposed, very few empirical studies have investigated this profile of change in therapeutic contexts partly because, until recently, we lacked the appropriate methodological tools for doing so (Cicchetti & Cohen, 1995). The study by Lichtwarck-Aschoff et al. (2012) discussed earlier explicitly investigated the destabilization profile in aggression treatment and found that children who improved through treatment showed characteristics of a reorganization in the parent–child dyad. In line with that study we expect that treatment gains—across a wide range of psychopathologies and interventions—will be evident only after a phase transition, operationalized as a significant increase in the variability of behavioral patterns. Without evidence of a destabilization period, treatment is expected to be less successful. It is important to keep in mind that system destabilization, or a lack thereof, is part of a process description, revealing potential forms of treatment progressions. They are never the cause of success or failure of the treatment itself. Therefore, it is important that future efforts are targeted on unraveling the underlying causes of the reorganization that are triggered by the treatment perturbation. One way to do this would involve a model building and simulation approach (discussed in this chapter; e.g., Gottman, Murray, Swanson, Tyson, & Swanson, 2002; Van Geert, 1994; Van Geert & Lichtwarck-Aschoff, 2005), emphasizing the client’s internal dynamics, together with the interaction with external influences such as treatment. A second way of understanding the causal mechanisms in terms of change processes is the investigation of (changes in) control parameters (Van Orden et al., 2009). A prime candidate for a control parameter within the context of treatment is motivation to change (Guastello, 1995, 2002; Hayes & Strauss, 1998; Kowalik, Schiepek, Kumpf, Roberts, & Elbert, 1997; Mahoney, 1991; Schiepek, 2009). Regarding organizational change, Guastello (1995, 2002) described pressure for change and resistance to change as two orthogonal control parameters. When resistance is low, even minor pressures to change (i.e., intervention efforts) may lead to gradual and small changes in behavior. When resistance is high, old patterns of behavior will not

752

DS Methods

change if pressure to change is low. However, when both resistance and pressure to change are high, radical, nonlinear change is likely to occur. This conception is in line with the struggle-and-work-through hypothesis (Chamberlain et al., 1984; Stoolmiller et al., 1993) where it has been demonstrated that families that show low resistance from baseline to termination are unlikely to show significant change and benefit from treatment (Stoolmiller at al., 1993). Hence, the time-serial investigation of changes in control parameters (e.g., motivation to change) in relation to structural changes in real-time behavior is needed to enhance our understanding of causal mechanisms producing successful treatment outcomes. Finally, an implicit, if not explicit, assumption made by most clinicians and clinical researchers is that, when treatment works, family interactions change from being emotionally negative or angry to positive or supportive. But according to emotion theorists (e.g., Izard 1977; Magai & McFadden 1995; Tomkins 1963), the expression of anger or any other negative emotion is not pathogenic; all emotions are adaptive and important to express in appropriate contexts. It is the effective—flexible—regulation of these emotions that is critical for healthy development (e.g., Bonanno, Papa, Lalande, Westphal, & Coifman, 2004). Being emotionally flexible is the hallmark of healthy functioning at any age (Kashdan & Rottenberg, 2010). Being locked into any one emotional state is often the definition of certain psychopathologies: perseveration of fearful vigilance is the mark of anxiety, unyielding sadness is diagnostic of depression, and long-lasting elation is mania. Developmentalists agree that, in general, regulation skills are learned and practiced in the context of parent–child interactions. We go on to propose that flexibility in these interactions (i.e., the ability to shift from one emotional state to another, according to contextual demands) may be more important for healthy development than the complete avoidance of negative affect (see also Gottman & Notarius, 2000, for a similar suggestion for marital relationships). Indeed the study by Granic and colleagues (2007) discussed earlier showed that children that improved through treatment could be distinguished from the ones that did not based on the flexibility of their parent–child interactions. The improvers had changed their rigid interactions into emotionally flexible and reparative ones not simply into positive interactions. But flexibility can also be assessed on different levels. As an illustration, we are currently investigating whether flexibility is associated with positive treatment outcomes in an anxiety treatment for children on the basis of changes in the coordination of body movements in parent–child interactions (see for

a similar approach Ramseyer & Tschacher, 2011). Here we expect that improver dyads show a more balanced body synchronization in contrast to a parent dominated coordination. In sum, parent–child and peer interactions are predicted to become more flexible as a function of successful treatment. Rigidity in parent–child and peer interactions on the other hand gives rise to a variety of cascading constraints that contribute to the emergence and maintenance of problem behavior. These three hypotheses provide an illustration of how DS ideas and methodologies can promote the study of change processes in clinical research. Providing a finer degree of resolution than pre-post designs and thereby revealing predictors, moderators, mediators, and mechanisms of change (Hayes et al., 2007). The evidence-based treatment movement has set out to identify what works for whom by implementing randomized clinical trials. The ideas and methodologies described in this chapter offer the possibility of enriching this work to better understand the when, how, and why of change.

CONCLUSION From the beginning of its establishment as a discipline, one of the core priorities in developmental psychopathology has been methodological diversity (e.g., Cicchetti & Cohen, 1995b; Cummings et al., 2000; Richters, 1997). One reason for encouraging this analytic pluralism is the recognition of the disparity between systems-based models of developmental psychopathology and the inadequate methodological tools that are available to test them (Richters, 1997). We have argued that DS approaches to development offer research methods that show greater fidelity to the complex, heterogeneous, temporal nature of developmental phenomena. Clearly no set of analytic methods can address the mismatch of methods and models entirely; thus, we are not arguing for the complete abandonment of well-established techniques. Instead, our purpose in providing a survey of DS methods is to encourage developmental psychopathologists to begin examining empirically questions that may have previously seemed out of analytic reach. REFERENCES Abraham, F. D., Abraham, R. H., & Shaw, C. D. (1990). A visual introduction to dynamical systems theory for psychology. Santa Cruz, CA: Aerial Press. Allison, P. D. (1984). Event history analysis: Regression for longitudinal event data. Beverly Hills, CA: Sage.

References Badalamenti, A. F., & Langs, R. J. (1992). The thermodynamics of psychotherapeutic communication. Behavioral Science, 37, 152–180. doi: 10.1002/bs.3830370302 Bakeman, R., & Gottman, J. M. (1997). Observing interaction: An introduction to sequential analysis (2nd ed.). Cambridge, UK: Cambridge University Press. Bergman, L. R., & Magnusson, D. (1997). A person oriented approach in research on developmental psychopathology. Development and Psychopathology, 9, 291–319. doi: 10.1017/S095457949700206X Bertenthal, B. (1999). Variation and selection in the development of perception and action. In G. Savelsbergh, H. van der Maas, & P. van Geert (Eds.), Non-linear developmental processes (Vol. 175, pp. 105–121). Amsterdam, The Netherlands: Royal Netherlands Academy of Arts and Sciences. Blossfeld, H., & Rohwer, G. (2002). Techniques of event history modeling: New approaches to causal analysis (2nd ed.). Mahwah, NJ: Erlbaum. Bonanno, G. A, Papa, A., Lalande, K., Westphal, M., & Coifman, K. (2004). The importance of being flexible: The ability to both enhance and suppress emotional expression predicts long-term adjustment. Psychological Science, 15, 482–487. doi: 10.1111/j.0956-7976 .2004.00705.x Boyce, W. T., Frank, E., Jensen, P. S., Kessler, R. C., Nelson, C. A., & Steinberg, L. (1998). Social context in developmental psychopathology: Recommendations for future research from the MacArthur Network on Psychopathology and Development. Development & Psychopathology, 10, 143–164. doi: 10.1017/S0954579498001552 Branje, S. J. (2008). Conflict management in mother–daughter interactions in early adolescence. Behaviour, 145, 1627–1651. doi: 10.1163/ 156853908786131315 Bronfenbrenner, U. (1979). The ecology of human development: Experiments by nature and design. Cambridge, MA: Harvard University Press. Burlingame, G. M., Fuhriman, A., & Barnum, K. R. (1995). Group therapy as a nonlinear dynamical system: Analysis of therapeutic communication for chaotic patterns. In F. D. Abraham & A. R. Gilgen (Eds.), Chaos theory in psychology (pp. 87–105). Westport, CT: Greenwood Press. Butz, M. R., Chamberlain, L. L., & McCown, W. G. (1997). Strange attractors: Chaos, complexity, and the art of family therapy. New York, NY: John Wiley & Sons Inc. Byrne, D. G., Mazanov, J., Gregson, R. A. M. (2001). A cusp catastrophe analysis of changes to adolescent smoking behavior in response to smoking prevention programs. Nonlinear Dynamics, Psychology, & Life Sciences, 5, 115–137. doi: 10.1023/A:1026467401041 Calkins, S. D., & Fox, N. A. (2002). Self-regulatory processes in early personality development: A multilevel approach to the study of childhood social withdrawal and aggression. Development and Psychopathology, 14, 477–498. doi: 10.1017/S095457940200305X Capaldi, D. M., & Patterson, G. R. (1987). An approach to the problem of recruitment and retention rates for longitudinal research. Behavioral Assessment, 9, 169–177. Case, R., Okamoto, Y., Griffin, S., McKeough, A., Bleiker, C., Henderson, B., & Stephenson, K. M. (1996). The role of central conceptual structures in the development of children’s thought. Monographs of the Society for Research in Child Development, 246, 1-295. doi: 10.2307/1166077 Caspar, F., Rothenfluh, T., & Segal, Z. (1992). The appeal of connectionism for clinical psychology. Clinical Psychology Review, 12, 719–762. doi: 10.1016/0272-7358(92)90022-Z Cerezo, M. A., Trenado, R. M., & Pons-Salvador, G. (2012). Mother–infant interaction and quality of child’s attachment: A

753

nonlinear dynamical systems approach. Nonlinear Dynamics in Psychology and the Life Sciences, 16, 243–267. Chamberlain, P., Patterson, G. R., Rid, J. B., Kavanagh, K., & Forgatch, M. S. (1984) Observation of client resistance. Behavior Therapy, 15, 144–155. doi: 10.1016/S0005-7894(84)80016-7 Cicchetti, D. (1990). A historical perspective on the discipline of developmental psychopathology. In J. E. Rolf, A. S. Masten, D. Cicchetti, & K. H. Nuechterlein (Eds.), Risk and protective factors in the development of psychopathology (pp. 2–28). New York, NY: Cambridge University Press. Cicchetti, D. (1993). Developmental psychopathology: Reactions, reflections, projections. Developmental Review. Special Issue: Setting a path for the coming decade: Some goals and challenges, 13, 471–502. doi: 10.1006/drev.1993.1021 Cicchetti, D., & Cohen, D. J. (1995a). Developmental psychopathology, Vol. 1: Theory and methods. Wiley series on personality processes. New York, NY: Wiley. Cicchetti, D., & Cohen, D. J. (1995b). Perspectives on developmental psychopathology. In D. Cicchetti & Cohen, D. J. (Eds.), Developmental psychopathology, Vol. 1: Theory and methods. Wiley series on personality processes (pp. 3–20). New York, NY: John Wiley & Sons. Cicchetti, D., & Dawson, G. (2002a). Special issue: Multiple levels of analysis, development and psychopathology, Vol. 14. New York, NY: Cambridge University Press. Cicchetti, D., & Dawson, G. (2002b). Editorial: Multiple levels of analysis. Development and Psychopathology, 14, 417–420. doi: 10.1017/ S0954579402003012 Cicchetti, D., & Richters, J. E. (1993). Developmental considerations in the investigation of conduct disorder. Development and Psychopathology, 5, 331–334. doi: 10.1017/S0954579400004429 Cicchetti, D., & Schneider-Rosen, K. (1986). Theoretical and empirical considerations in the investigation of the relationship between affect and cognition in atypical populations of infants. In C. E. Izard, J. Kagan, & R. B. Zajonc (Eds.), Emotions, cognition, and behavior. (pp. 366–406). New York, NY: Cambridge University Press. Cicchetti, D., & Sroufe, L. A. (1976). The relationship between affective and cognitive development in Down’s syndrome infants. Child Development, 47, 920–929. doi: 10.2307/1128427 Clair, S. (1998). A cusp catastrophe model for adolescent alcohol use: An empirical test. Nonlinear Dynamics, Psychology, & Life Sciences, 2, 217–241. doi: 10.1023/A:1022376002167 Connell, A. M., Hughes-Scalise, A., Klostermann, S., & Azem, T. (2011). Maternal depression and the heart of parenting: Respiratory sinus arrhythmia and affective dynamics during parent–adolescent interactions. Journal of Family Psychology, 25, 653–662. doi: 10.1037/ a0025225 Coleman, P., & Watson, A. (2000). Infant attachment as a dynamic system. Human Development, 43, 295–313. doi: 10.1159/000022691 Cox, M. J., & Paley, B. (1997). Families as systems. Annual Review of Psychology, 48, 243–267. doi: 10.1146/annurev.psych.48.1.243 Cummings, E. M., Davies, P. T., & Campbell, S. B. (2000). Developmental psychopathology and family process: Theory, research, and clinical implications. New York, NY: Guilford Press. Dagne, G., & Snyder J. (2011). Relationship of maternal negative moods to child emotion regulation during family interaction. Development and Psychopathology, 23, 211–223. doi: 10.1017/ S095457941000074X Dale, R., & Spivey, M. J. (2005). From apples and oranges to symbolic dynamics: A framework for conciliating notions of cognitive representation. Journal of Experimental & Theoretical Artificial Intelligence, 17, 317–342. doi: 10.1080/09528130500283766

754

DS Methods

Daw, C. S., Finney, C. E. A., & Tracy, E. R. (2003). A review of symbolic analysis of experimental data. Review of Scientific Instruments, 74, 915–930. doi: 10.1063/1.1531823 De Weerth, C., & Van Geert, P. (2002). Changing patterns of infant behavior and mother-infant interaction: Intra- and interindividual variability. Infant Behavior & Development, 24, 340–374. doi: 10.1016/ S0163-6383(02)00083-8 Dishion, T. J., Bullock, B. M., & Granic, I. (2002). Pragmatism in modeling peer influence: Dynamics, outcomes, and change processes. How prevention intervention studies in the field of developmental psychopathology can inform developmental theories and models. [Special Issue]. Development and Psychopathology, 14, 969–981. doi: 10.1017/ S0954579402004169 Dishion, T. J., Capaldi, D. M., Spracklen, K. M., & Li, F. (1995). Peer ecology of male adolescent drug use. Development and Psychopathology, 7, 803–824. doi: 10.1017/S0954579400006854 Dishion, T. J., Eddy, J. M., Haas, E., Li, F., & Spracklen, K. (1997). Friendships and violent behavior during adolescence. Social Development, 6, 207–223. doi: 10.1111/1467-9507.00034 Dishion, T. J., French, D. C., & Patterson, G. R. (1995). The development and ecology of antisocial behavior. In D. Cicchetti & D. Cohen (Eds.), Manual of developmental psychopathology (pp. 421–471). New York, NY: Wiley. Dishion, T. J., Gardner, K., Patterson, G. R., Reid, J. B., & Thibodeaux, S. (1983). The family process code: A multidimensional system for observing family interactions. Unpublished technical manual. Eugene, OR: Oregon Social Learning Center. Dishion, T. D., Nelson, S. E., Bullock, B. M., & Winter, C. E. (2004). Adolescent friendship as a dynamic system: Entropy and deviance in the etiology and course of male antisocial behavior. Journal of Abnormal Child Psychology, 32, 651–663. doi: 10.1023/ B:JACP.0000047213.31812.21 Dishion, T. J., & Patterson, G. R. (1999). Model building in developmental psychopathology: A pragmatic approach to understanding and intervention. Journal of Clinical Child Psychology, 28, 502–512. doi: 10.1207/S15374424JCCP2804_10 Dishion, T. J., Spracklen, K. M., Andrews, D. W., & Patterson, G. R. (1996). Deviancy training in male adolescent friendships. Behavior Therapy, 27, 373–390. doi: 10.1016/S0005-7894(96)80023-2 Dixon, R. A., & Lerner, R. M. (1988). A history of systems in developmental psychology. In M. H. Bornstein & M. E. Lamb (Eds.), Developmental psychology: An advanced textbook (2nd ed., pp. 3–50). Hillsdale, NJ: Erlbaum. Dumas, J. E., LaFrenier, P. J., & Serketich, W. J. (1995). “Balance of power”: A transactional analysis of control in mother-child dyads involving socially competent, aggressive, and anxious children. Journal of Abnormal Psychology, 104, 104–113. doi: 10.1037/ 0021-843X.104.1.104 Dumas, J. E., Lemay, P., & Dauwalder, J. P. (2001). Dynamic analyses of mother-child interactions in functional and dysfunctional dyads: A synergetic approach. Journal of Abnormal Child Psychology, 29, 317–329. doi: 10.1023/A:1010309929116 Eddy, J. M., & Chamberlain, P. (2000). Family management and deviant peer association as mediators of the impact of treatment condition on youth antisocial behavior. Journal of Consulting & Clinical Psychology, 68, 857–863. doi: 10.1037/0022-006X.68.5.857 Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D., & Plunkett, K. (1996). Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press. Erickson, M. F., Egeland, B., & Pianta, R. (1989). The effects of maltreatment on the development of young children. In D. Cicchetti & V. Carlson (Eds.), Child maltreatment: Theory and research on the causes

and consequences of child abuse and neglect (pp. 647–684). New York, NY: Cambridge University Press. Eyberg, S. M., Nelson, M. M., & Boggs, S. R. (2008). Evidence-based psychosocial treatments for children and adolescents with disruptive behavior. Journal of Clinical Child and Adolescent Psychology, 37, 215–237. doi: 10.1080/15374410701820117 Fogel, A. (1990). The process of developmental change in infant communicative action: Using dynamic systems theory to study individual ontogenies. In J. Colombo & J. Fagen (Eds.), Individual differences in infancy: Reliability, stability, prediction (pp. 341–358). Hillsdale, NJ: Erlbaum. Fogel, A. (1993). Developing through relationships: Origins of communication, self and culture. Chicago, IL: University of Chicago. Fogel, A. (1995). Development and relationships: A dynamic model of communication. Advances in the Study of Behavior, 24, 259–290. doi: 10.1016/S0065-3454(08)60396-4 Fogel, A., & Thelen, E. (1987). The development of early expressive and communicative action: Re-interpreting the evidence from a dynamic systems perspective. Developmental Psychology, 23, 747–761. doi: 10.1037/0012-1649.23.6.747 Ford, D. H., & Lerner, R. M. (1992). Developmental systems theory: An integrative approach. Newbury Park, CA: Sage. Forgatch, M. S., & DeGarmo, D. S. (1999). Parenting through change: An effective prevention program for single mothers. Journal of Consulting & Clinical Psychology, 67, 711–724. doi: 10.1037/ 0022-006X.67.5.711 Freeman, W. (1995). Societies of brains: A study in the neuroscience of love and hate. Hillsdale, NJ: Erlbaum. Gardner, B. C., & Wampler, K. S. (2008). Uncovering dynamical properties in the emotional system of married couples. Contemporary Family Therapy, 30, 111–126. doi: 10.1007/s10591-008-9056-4 Garmezy, N. (1974). The study of competence in children at risk for severe psychopathology. In E. J. Anthony & C. Koupernik (Eds.), The child in his family: Children at psychiatric risk. Oxford, UK: John Wiley & Sons. Garmezy, N., & Rutter, M. (Eds.). (1983). Stress, coping, and development in children. Baltimore, MD: Johns Hopkins University Press. Gasser, M., & Smith, L. B. (1996). Learning nouns and adjective meanings: A connectionist account. Language and Cognitive Processes, 13, 269–306. doi: 10.1080/016909698386537 Gilmore, R. (1981). Catastrophe theory for scientists and engineers. New York, NY: Wiley. Gottlieb, G. (1991). Epigenetic systems view of human development. Developmental Psychology, 27, 33–34. doi: 10.1037/0012-1649. 27.1.33 Gottlieb, G. (1992). Individual development and evolution: The genesis of novel behavior. London, UK: Oxford University Press. Gottman, J. M., Coan, J., Carrere, S., & Swanson, C. (1998). Predicting marital happiness and stability from newlywed interactions. Journal of Marriage and the Family, 60, 5–22. doi: 10.2307/353438 Gottman, J. M., Guralnick, M. J., Wilson, B., Swanson, C. C., & Murray, J. D. (1997). What should be the focus of emotion regulation in children? A nonlinear dynamic mathematical model of children’s peer interaction in groups. Development and Psychopathology, 9, 421–452. doi: 10.1017/S0954579497002113 Gottman, J. M., McCoy, K., Coan, J., & Collier, H. (1996). The specific affect coding system (SPAFF) for observing emotional communication in marital and family interaction. Mahwah, NJ: Erlbaum. Gottman, J. M., Murray, J. D., Swanson, C. C., Tyson, R., & Swanson, K. R. (2002). The mathematics of marriage: Dynamic nonlinear models. London, UK: MIT Press.

References Gottman, J. M., & Notarius, C. I. (2000). Decade review: Observing marital interaction. Journal of Marriage and the Family, 62, 927–947. doi: 10.1111/j.1741-3737.2000.00927.x Granic, I. (2000). The self-organization of parent–child relations: Beyond bi-directional models. In M. D. Lewis & I. Granic (Eds.), Emotion, development and self-organization: Dynamic systems approaches to emotional development (pp. 267–297). New York, NY: Cambridge University Press. Granic, I. (2003, April). Dynamic systems methods tap previously undetected socialization processes associated with antisocial outcomes. Paper presented at the Biennial Meeting of the Society for Research in Child Development, Tampa, FL. Granic, I., & Dishion, T. (2003). Deviant talk in adolescent friendships: A step toward measuring a pathogenic attractor process. Social Development, 12, 314–334. doi: 10.1111/1467-9507.00236 Granic, I., Dishion, T. J., & Hollenstein, T. (2003). The family ecology of adolescence: A dynamic systems perspective on normative development. In G. Adams & M. Berzonsky (Eds.), Blackwell Handbook of Adolescence (pp. 60–91). New York: Blackwell. Granic, I., Hollenstein, T., Dishion, T. J., & Patterson, G. R. (2003). Longitudinal analysis of flexibility and reorganization in early adolescence: A dynamic systems study of family interactions. Developmental Psychology, 39, 606–617. doi: 10.1037/0012-1649.39.3.606

755

Hartelman, P., van der Maas, H., & Molenaar, P. (1998). Detecting and modeling developmental transitions. British Journal of Developmental Psychology. Special Issue: Developmental Transitions, 16, 97–122. doi: 10.1111/j.2044-835X.1998.tb00751.x Haslett, T., Smyrnios, K. X, Osbourne, C. (1998). A cusp catastrophe analysis of anxiety levels in pre-university students. Journal of Psychology, 132, 5–24. doi: 10.1080/00223989809599260 Hayes, A. M., & Strauss, J. L. (1998). Dynamic systems theory as a paradigm for the study of change in psychotherapy: An application to cognitive therapy for depression. Journal of Consulting and Clinical Psychology, 66, 939–947. doi: 10.1037/0022-006X.66.6.939 Hayes, A. M., Laurenceau, J-P., Feldman, G., Strauss, J. L & Cardaciotto, L-A. (2007). Change is not always linear: The study of nonlinear and discontinuous patterns of change in psychotherapy. Clinical Psychology Review, 27, 715–723. doi: 10.1016/j.cpr.2007.01.008 Heath, R. A. (2000). Nonlinear dynamics: Techniques and applications in psychology. Mahwah, NJ: Lawrence Erlbaum Associates. Hinde, R. A. (1983). Ethology and child development. In M. Haith & J. Campos (Eds.), Carmichael’s manual of child psychology, Vol. 2 (pp. 27–94). New York, NY: Wiley. Hinde, R. A. (1992). Some complexities in aggressive behavior. In A. Fraczek & H. Zumkley (Eds.), Socialization and aggression (pp. 3–10). New York, NY: Springer-Verlag.

Granic, I., Meusel, L. A., Lamm, C., Woltering, S., & Lewis, M. D. (2012). Emotion regulation in children with behavior problems: Linking behavioral and brain processes. Development and psychopathology, 24, 1019–1029. doi: 10.1017/S095457941200051X

Hinshaw, S. P., & Zupan, B. A. (1997). Assessment of antisocial behavior in children and adolescents. In D. M. Stoff, J. Breiling, & J. D. Maser (Eds.), Handbook of antisocial behavior (pp. 36–50). New York, NY: Wiley.

Granic, I., & Patterson, G. R. (2001, July). Innovations in the study of coercive microsocial processes in middle childhood: New methods and models based on dynamic systems principles. Paper presented at the meeting of the International Society for Research in Child and Adolescent Psychopathology, Vancouver, British Columbia.

Hinshaw, S. P. (1999, June). Risk-factor research in developmental psychopathology. Symposium conducted at the biennial meeting of the International Society for Research in Child and Adolescent Psychopathology, Barcelona.

Granic, I., & Patterson, G. (2006). Towards a comprehensive model of antisocial development: A dynamic systems approach. Psychological Review, 113, 101–131. doi: 10.1037/0033-295X.113.1.101 Granic, I., & Lamey, A. V. (2002). Combining dynamic-systems and multivariate analyses to compare the mother-child interactions of externalizing subtypes. Journal of Abnormal Child Psychology, 30, 265–283. doi: 10.1023/A:1015106913866 Granic, I., O’Hara, A., Pepler, D., & Lewis, M. (2007). A dynamic systems analysis of parent–child changes associated with successful real-world interventions for aggressive children. Journal of Abnormal Child Psychology, 35, 845–857. doi: 10.1007/s10802-007-9133-4 Greenberg, L. S., Rice, L. N., & Elliott, R. K. (1996). Facilitating emotional change: The moment-by-moment process. New York, NY: Guilford Press. Guastello, S. J. (1995). Chaos, catastrophe, and human affairs: Applications of nonlinear dynamics to work, organizations, and social evolution. Mahwah, NJ: Lawrence Erlbaum Associates. Guastello, S. J. (2002). Managing emergent phenomena: Nonlinear dynamics in work organizations. Mahwah, NJ: Lawrence Erlbaum Associates. Guastello, S. J., Hyde, T., & Odak, M. (1998). Symbolic dynamic patterns of verbal exchange in a creative problem solving group. Nonlinear Dynamics, Psychology, and Life Sciences, 2, 35–58. doi: 10.1023/ A:1022324210882 Haken, H. (1977). Synergetics An introduction: Nonequilibrium phase transitions and self-organization in physics, chemistry and biology. Berlin, Germany: Springer-Verlag. Hamming, R. W. (1950). Error detecting and error correcting codes. Bell Systems Technical Journal, 25, 147–160. doi: 10.1002/j.15387305.1950.tb00463.x

Hollenstein, T. (2007). State space grids: Analyzing dynamics across development. International Journal of Behavioral Development, 31, 384–396. doi: 10.1177/0165025407077765 Hollenstein, T. (2012). The dynamic systems perspective: What is the system? In L. C. Mayes & M. Lewis (Eds.), The Cambridge handbook of environment in human development. (pp. 138–151). New York, NY: Cambridge University Press. Hollenstein, T. (2013). State space grids: Depicting dynamics across development. New York, NY: Springer. Hollenstein, T., Allen, N. B., & Sheeber, L. (in press). Affective patterns in triadic family interactions: Associations with adolescent depression. Development and Psychopathology. Hollenstein, T., Granic, I., Stoolmiller, M., & Snyder, J. (2004). Rigidity in parent–child interactions and the development of externalizing and internalizing behavior in early childhood. Journal of Abnormal Child Psychology, 32, 595–607. doi: 10.1023/B:JACP.0000047209.37650.41 Hollenstein, T., & Lewis, M. D. (2006). A state space analysis of emotion and flexibility in parent– child interactions. Emotion, 6, 656–662. doi: 10.1037/1528-3542.6.4.656 Hollenstein, T., Lichtwarck-Aschoff, A., & Potworowski, G. (2013). A Model of socioemotional flexibility at three time scales. Emotion Review, 5, 397–405. doi: 10.1177/1754073913484181 Hsu, H., & Fogel, A. (2001). Infant vocal development in a dynamic mother-infant communication system. Infancy, 2, 87–109. doi: 10.1207/S15327078IN0201_6 Hsu, H., & Fogel, A. (2003a). Social regulatory effects of infant non-distress vocalization on maternal behaviors. Developmental Psychology, 39, 976–991. doi: 10.1037/0012-1649.39.6.976 Hsu, H., & Fogel, A. (2003b). Stability and transitions of mother-infant face-to-face communication during the first six months: A

756

DS Methods

micro-historical approach. Developmental Psychology, 39, 1061–1082. doi: 10.1037/0012-1649.39.6.1061 Inhelder, B. (1976). Some pathologic phenomena analyzed in the perspective of developmental psychology. In B. Inhelder & H. Chipman (Eds.), Piaget and his school (pp. 221–227). New York, NY: Springer. In-Albon, T., & Schneider, S. (2007). Psychotherapy of childhood anxiety disorders: a Meta-analysis. Psychotherapy and Psychosomatics, 76, 15–24. doi: 10.1159/000096361 Izard, C. E. (1977). Human emotions. New York, NY: Plenum. Izard, C. E. (1979). The maximally discriminative facial movement coding system (MAX). Newark, NJ: Instructional resources center, University of Delaware. Johnson, M. H., & Morton, J. (1991). Biology and cognitive development: The case of face recognition. Oxford, UK: Blackwell. Kagan, J. (1992). Yesterday’s premises, tomorrow’s promises. Developmental Psychology, 28, 990–997. doi: 10.1037/0012-1649.28.6.990 Kashdan, T. B., & Rottenberg, J. (2010). Psychological flexibility as a fundamental aspect of health. Clinical Psychology Review, 7, 865–878. doi: 10.1016/j.cpr.2010.03.001 Kauffman, S. A. (1993). The origins of order: Self-organization and selection in evolution. New York, NY: Oxford University Press. Kazdin, A. E. (2000). Psychotherapy for children and adolescents: Directions for research and practice. New York, NY: Oxford University Press. Kazdin, A. E. (2001). Progression of therapy research and clinical application of treatment require better understanding of the change process. Clinical Psychology, 8, 143–151. doi: 10.1093/clipsy.8.2.143 Kazdin, A. E. (2002). Psychosocial treatments for conduct disorder in children and adolescents. In P. E. Nathan & J. M. Gorman (Eds.), A guide to treatments that work (2nd ed., pp. 57–85). London, UK: Oxford University Press. Kazdin, A. E. (2009). Understanding how and why psychotherapy leads to change. Psychotherapy Research, 19, 418–428. doi: 10.1080/10503300802448899 Keating, D. P. (1990). Adolescent thinking. In G. R. Elliott & S. S. Feldman (Eds.), At the threshold: the developing adolescent (pp. 54–89). Cambridge, MA: Harvard University Press. Kelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior. Cambridge, MA: Bradford/MIT Press. Kowalik, Z. J., Schiepek, G., Kumpf, K., Roberts, L. E., & Elbert, T. (1997). Psycho-therapy as a chaotic process 2: The application of nonlinear analysis methods on quasi time series of the client-therapist interaction: A nonstationary approach. Psychotherapy Research, 7, 197–218. doi: 10.1080/10503309712331331973 Kuo, Z. Y. (1967). The dynamics of behavior development. New York, NY: Random House. Lamey, A., Hollenstein, T., Lewis, M. D., & Granic, I. ( 2004). GridWare (Version 1.1). [Computer software]. http://statespacegrids.org Lanza, M. L. (1999). Catastrophe theory: Application of nonlinear dynamics to assault victim responses. Journal of the American Psychiatric Nurses Association, 5, 117–121. doi: 10.1177/1078390 39900500403 Lavelli, M., Pantoja, A. P. F., Hsu, H., Messinger, D., & Fogel, A. (2005). Using microgenetic designs to study change processes. In D. G. Teti (Ed.), Handbook of research methods in developmental psychology. London, UK: Blackwell. Lee, K., Karmiloff-Smith, A. (2002). Macro- and microdevelopmental research: Assumptions, research strategies, constraints, and utilities. In N. Granott & J. Parziale (Eds.), Microdevelopment: Transition processes in development and learning. Cambridge studies in cognitive

perceptual development (pp. 243–292). New York, NY: Cambridge University Press. Lewin, K. (1931). The conflict between Aristotelian and Galilean models of thought in contemporary psychology. Journal of General Psychology, 5, 147–177. doi: 10.1080/00221309.1931.9918387 Lewis, M. D. (2000). The promise of dynamic systems approaches for an integrated account of human development. Child Development, 71, 36–43. doi: 10.1111/1467-8624.00116 Lewis, M. D. (2011). Dynamic systems approaches: Hot enough? Cool enough? Child Development Perspectives, 5, 279–285. doi: 10.1111/ j.1750-8606.2011.00190.x Lewis, M. D., & Douglas, L. (1998). A dynamic systems approach to cognition-emotion interactions in development (pp. 159–188), in What Develops in Emotional Development? M. F. Mascolo & S. Griffins (Eds.) New York: Plenum Press. Lewis, M. D., & Granic, I. (1999a). Self-organization of cognitionemotion interactions. In T. Dalgleish & M. Power (Eds.), The handbook of cognition and emotion (pp. 683–701). Chichester, UK: Wiley. Lewis, M. D., & Granic, I. (1999b). Who put the self in self-organization? A clarification of terms and concepts for developmental psychopathology. Development and Psychopathology, 11, 365–374. doi: 10.1017/S0954579499002102 Lewis, M. D., & Granic, I. (Eds.) (2000). Emotion, development and self-organization: Dynamic systems approaches to emotional development. New York, NY: Cambridge University Press. Lewis, M. D., Lamey, A. V., & Douglas, L. (1999). A new dynamic systems method for the analysis of early socioemotional development. Developmental Science, 2, 457–475. doi: 10.1111/1467-7687.00090 Lewis, M. D., Zimmerman, S., Hollenstein, T., & Lamey, A. V. (2004). Reorganization of coping behavior at 1 1/2 years: Dynamic systems and normative change. Developmental Science, 7(1), 56–73. doi: 10.1111/j.1467-7687.2004.00323.x Lichtenberg, J. W., & Heck, E. J. (1986). Analysis of sequence and pattern in process research. Journal of Counseling Psychology, 33, 170–180. doi: 10.1037/0022-0167.33.2.170 Lichtwarck-Aschoff, A., Hasselman, F., Cox, R., Peppler, D., & Granic, I. (2012). A characteristic destabilization profile in parent–child interactions associated with treatment efficacy for aggressive children. Nonlinear Dynamics, Psychology, and Life Sciences, 16, 353–379. Lichtwarck-Aschoff, A., Kunnen, S., &, van Geert, P. (2009). Here we go again: A dynamic systems perspective on emotional rigidity across parent–adolescent conflicts. Developmental Psychology, 45, 1364–1375. doi: 10.1037/a0016713 Lougheed, J., Hollenstein, T., Lichtwarck-Aschoff, A., & Granic, I. (2015). Maternal regulation of child affect in externalizing and typically-developing children. Journal of Family Psychology, 29, 10–19. Lunkenheimer, E. S., Albrecht, E. C., & Kemp, C. J. (2013). Dyadic flexibility in early parent–child interactions: Relations with maternal depressive symptoms and child negativity and behaviour problems. Infant and Child Development, 22, 250–269. doi: 10.1002/icd.1783 Lunkenheimer, E. S., Hollenstein, T., Wang, J., & Shields, A. M. (2012). Flexibility and attractors in context: Family emotion socialization patterns and children’s emotion regulation in late childhood. Nonlinear Dynamics in Psychology and the Life Sciences, 16, 269–291. Lunkenheimer, E. S., Olson, S. L., Hollenstein, T., Sameroff, A., & Winter, C. (2011). Dyadic flexibility and positive affect in parent–child coregulation and the development of children’s behavior problems. Development and Psychopathology, 23, 577–591. doi: 10.1017/ S095457941100006X Lykken, D. T. (1991). What’s wrong with psychology anyway? In D. Cicchetti & W. M. Grove (Eds.), Thinking clearly about psychology (Vol. 1, pp. 3–39). Minneapolis, MN: University of Minnesota Press.

References Mahoney, M. J. (1991). Human change processes: The scientific foundations of psychotherapy. New York, NY: Basic Books, Inc. Magai, C., & McFadden, S. H. (1995). The role of emotions in social and personality development: History, theory, and research. New York, NY: Plenum Press. Mainhard, M. T., Pennings, H. J., Wubbels, T., & Brekelmans, M. (2012). Mapping control and affiliation in teacher–student interaction with state space grids. Teaching and Teacher Education, 28, 1027–1037. doi: 10.1016/j.tate.2012.04.008 Marwan, N., Carmen Romano, M., Thiel, M., & Kurths, J. (2007). Recurrence plots for the analysis of complex systems. Physics Reports, 438, 237–329. doi: 10.1016/j.physrep.2006.11.001 McClelland, J., L. (1979). On the time relations of mental processes: An examination of systems of processes in cascade. Psychological Review, 86, 287–330. doi: 10.1037/0033-295X.86.4.287 Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, Sir Ronald, and the slow progress of soft psychology. Journal of Consulting and Clinical Psychology, 46, 1–42. doi: 10.1037/0022-006X.46.4.806 Minuchin, S. (1974). Families & family therapy. Cambridge, MA: Harvard University Press. Minuchin, S., & Fishman, C. H. (1974). Family therapy techniques. Cambridge MA: Harvard University Press. Moffitt, T. E. (1993). “Adolescence-limited” and “life-course persistent” antisocial behaviour: A developmental taxonomy. Psychological Review, 100, 674–701. doi: 10.1037/0033-295X.100.4.674 Moore, G. A., Powers, C. J., Bass, A. J., Cohn, J. F., Propper, C. B., Allen, N. B., & Lewinsohn, P. M. (2013). Dyadic interaction: Greater than the sum of its parts? Infancy, 18, 490–515. doi: 10.1111/j.15327078.2012.00136.x Munakata, Y. (1998). Infant perseveration and implications for object permanence theories: A PDP model of the AB task. Developmental Science, 1, 161–184. doi: 10.1111/1467-7687.00021 Munakata, Y., & McClelland, J. L. (2003). Connectionist models of development. Connectionism and dynamic systems theory: Are these really different approaches to development? [Special issue]. Developmental Science, 6, 413–429. doi: 10.1111/1467-7687.00296 Nelson, C. A., Bloom, F. E., Cameron, J. L., Amaral, D., Dahl, R. E., & Pine, D. (2002). An integrative, multidisciplinary approach to the study of brain-behavior relations in the context of typical and atypical development. Development and Psychopathology, 14, 499–520. doi: 10.1017/S0954579402003061 Newell, K. M., & Molenaar, P. (1998). Applications of nonlinear dynamics to developmental process modeling. Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Newtson, D. (1994). The perception and coupling of behavior waves. In R. R. Vallacher & A. Nowak (Eds.), Dynamical systems in social psychology (pp. 139–166). San Diego, CA: Academic Press. Newtson, D. (1998). Dynamical systems and the structure of behavior. In K. M. Newell & P. C. Molenaar (Eds.), Applications of nonlinear dynamics to developmental process modeling (pp. 129–160). Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Norton, A. (1995). Dynamics: An introduction. In R. F. Port & T. van Gelder (Eds.), Mind as motion: Explorations in the dynamics of cognition (pp. 45–68). Cambridge, MA: MIT Press. Nowak, A., & Vallacher, R. R. (1998). Dynamical social psychology. New York, NY: Guilford Press. Overton, W. F., & Horowitz, H. A. (1991). Developmental psychopathology: Integrations and differentiations. In D. Cicchetti & S. L. Toth (Eds.), Rochester symposium on developmental psychopathology, Vol. 3: Models and integrations, (pp. 1–42). Rochester, NY: University of Rochester Press.

757

Quinlan, P. T., Van Der Maas, H. L. J., Jansen, B. R. J., Booij, O., & Rendell, M. (2007). Re-thinking stages of cognitive development: An appraisal of connectionist models of the balance scale task. Cognition, 103, 413–459. doi: 10.1016/j.cognition.2006.02.004 Patterson, G. R. (1982). Coercive family processes. Eugene, OR: Castalia. Patterson, G. R., Reid, J. B., & Dishion, T. (1992). Antisocial boys. Eugene, OR: Castilia. Pepler, D., Craig, W., & O’Connell, P. (1999). Understanding bullying from a dynamic systems perspective. In A. Slater & D. Muir (Eds.), The Blackwell reader in developmental psychology (pp. 440–451). Malden, MA: Blackwell Publishers. Pincus, D. (2001). A framework and methodology for the study of nonlinear, self-organizing family dynamics. Nonlinear Dynamics, Psychology, and Life Sciences, 5, 139–173. doi: 10.1023/A:1026419517879 Plunkett, K. (1993). Lexical segmentation and vocabulary growth in early language acquisition. Journal of Child Language, 20, 43–60. doi: 10.1017/S0305000900009119 Plunkett, K., & Elman, J. L. (1997). Exercises in rethinking innateness: A handbook for connectionist simulations. Cambridge, MA: MIT Press. Port, R. F., & van Gelder, T. (1995). Mind as motion: Explorations in the dynamics of cognition. Cambridge, MA: MIT Press. Prigogine, I., & Stengers, I. (1984). Order out of chaos. New York, NY: Bantam. Ramseyer, F., & Tschacher, W. (2011). Nonverbal synchrony in psychotherapy: Coordinated body-movement reflects relationship quality and outcome. Journal of Consulting and Clinical Psychology, 79, 284–295. doi: 10.1037/a0023419 Reis, H. T., Collins, W. A., & Bersheid, E. (2000). The relationship context of human behavior and development. Psychological Bulletin. Special Issue: Psychology in the 21st Century, 126, 844–872. doi: 10.1037/0033-2909.126.6.844 Ribeiro, A. P., Bento, T., Salgado, J., Stiles, W. B., & Gonçalves, M. M. (2011). A dynamic look at narrative change in psychotherapy: A case study tracking innovative moments and protonarratives using state space grids. Psychotherapy Research, 21, 54–69. doi: 10.1080/ 10503307.2010.504241 Richters, J. E. (1997). The Hubble hypothesis and the developmentalist’s dilemma. Development and Psychopathology, 9, 193–229. doi: 10.1017/S0954579497002022 Ruhland, R., & van Geert, P. (1998). Jumping into syntax: Transitions in the development of closed class words. British Journal of Developmental Psychology, 16, 65–95. doi: 10.1111/j.2044-835X.1998.tb00750.x Rumelhart, D. E., & McClelland, J. L. (Eds.). (1986). Parallel distributed processing: Explorations in the microstructure of cognition: Vol. 1. Foundations. Cambridge, MA: Bradford Books/MIT Press. Ryan, K. D., Gottman, J. M., Murray, J. D., Carrere, S., & Swanson, C. (2000). Theoretical and mathematical modeling of marriage. In M. D. Lewis & I. Granic (Eds.), Emotion, development and self-organization: Dynamic systems approaches to emotional development (pp. 349–372) New York, NY: Cambridge University Press. Sabelli, H. C., Carlson-Sabelli, L., Patel, M., Levy, A., & Diez-Martin, J. (1995). Anger, fear, depression, and crime: Physiological and psychological studies using the process method. In R. Robertson & A. Combs (Eds.), Chaos theory in psychology and the life sciences (pp. 65–88). Mahwah, NJ: Lawrence Erlbaum Associates. Sameroff, A. J. (1983). Developmental systems: Contexts and evolution. In W. Kessen (Ed.), History, theory, and methods (4th ed., Vol. 1, pp. 237–294). New York, NY: Wiley. Sameroff, A. J. (1995). General systems theories and developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental

758

DS Methods

psychopathology, Vol. 1: Theory and methods. Wiley series on personality processes (pp. 659–695). New York, NY: Wiley & Sons. Schiepek, G. (2009). Complexity and nonlinear dynamics in psychotherapy. European Review, 17, 331–356. doi: 10.1017/S1062798709000763 Schroeck, F. E. (1994). New mathematical techniques for pattern recognition. In R. R. Vallacher & A. Nowak (Eds.), Dynamical systems in social psychology (pp. 139–166). San Diego, CA: Academic Press. Shannon, C. E., & Weaver, W. (1949). The mathematical theory of communication. Urbana, IL: University of Illinois Press. Simonoff, J. S. (1996). Smoothing methods in statistics. New York, NY: Springer-Verlag. Smith, L. B. (1993). The concept of same. Advances in Child Development and Behavior, 24, 216–253. Smith, L. B. (1995). Self-organizing processes in learning to learn words: Development is not induction. In C. A. Nelson (Ed.), The Minnesota symposia on child psychology: Basic and applied perspectives on learning, cognition, and development (Vol. 28, pp. 1–32). Mahwah, NJ: Erlbaum. Smith, L. B., & Samuelson, L. K. (2003). Different is good: Connectionism and dynamic systems theory are complementary emergentist approaches to development. Connectionism and dynamic systems theory: Are these really different approaches to development? [Special issue]. Developmental Science, 6, 434–439. doi: 10.1111/1467-7687. 00298 Smith, L. B., & Thelen, E. (2003). Development as a dynamic system. Trends in Cognitive Science, 7, 343–348. doi: 10.1016/S1364-6613 (03)00156-6 Snyder, J., Edwards, P., McGraw, K., Kilgore, K., & Holton, A. (1994). Escalation and reinforcement in mother-child conflict: Social processes associated with the development of physical aggression. Development and Psychopathology, 6, 305–321. doi: 10.1017/ S0954579400004600 Snyder, J., Stoolmiller, M., Wilson, M., & Yamamoto, M. (2003). Child anger regulation, parental responses to children’s anger displays, and early child antisocial behavior. Social Development. 12, 335–360. doi: 10.1111/1467-9507.00237 Spemann, H. (1938). Embryonic development and induction. New Haven, CT: Yale University Press. Spencer, J. P., & Schoner. (2003). Bridging the representational gap in the dynamic systems approach to development. Connectionism and dynamic systems theory: Are these really different approaches to development? [Special issue]. Developmental Science 6, 392–412. doi: 10.1111/1467-7687.00295 Spencer, J. P., & Thelen, E. (2003). Connectionism and dynamic systems theory: Are these really different approaches to development? [Special issue]. Developmental Science, 6, 375–377. doi: 10.1111/14677687.00293 Sravish, A. V., Tronick, E., Hollenstein, T., & Beeghly, M. (2013). Dyadic Flexibility during the Face-to-Face Still-Face Paradigm: A dynamic systems analysis of its temporal organization. Infant Behavior and Development, 36(3), 432–437. doi: 10.1016/j.infbeh.2013.03.013 Sroufe, L. A. (1979). The coherence of individual development: Early care, attachment, and subsequent developmental issues. American Psychologist, 34, 834–841. doi: 10.1037/0003-066X.34.10.834 Sroufe, L. A. (1986). Appraisal: Bowlby’s contribution to psychoanalytic theory and developmental psychology—attachment, separation, loss. Journal of Child Psychology & Psychiatry & Allied Disciplines. Special Issue: 30th anniversary of the Association for Child Psychology and Psychiatry and Allied Disciplines, 27, 841–849 Sroufe, L. A., & Rutter, M. (1984). The domain of developmental psychopathology. Child Development, 55, 17–29. doi: 10.2307/1129832

Steenbeek, H., & van Geert, P. (2005). A dynamic systems model of dyadic child-peer play interactions. European Journal of Developmental Psychology, 2, 105–145. doi: 10.1080/17405620544000020 Steenbeek, H., & van Geert, P. (2007). A dynamic systems approach to dyadic interaction in children: Emotional expression, action, dyadic play, and sociometric status. Developmental Review, 27, 1–40. doi: 10.1016/j.dr.2006.06.002 Steenbeek, H., & van Geert, P. (2008). The empirical validation of a dynamic systems model of interaction; Do children of different sociometric statuses differ in their dyadic play interaction. Developmental Science, 11, 253–281. doi: 10.1111/j.1467-7687.2007.00655.x Stephen, D. G., Boncoddo, R. A., Magnuson, J. S., & Dixon, J. A. (2009). The dynamics of insight: Mathematical discovery as a phase transition. Memory & Cognition, 37, 1132–1149. doi: 10.3758/MC.37.8.1132 Stephen, D. G., Dixon, J. A., & Isenhower, R. W. (2009). Dynamics of representational change: Entropy, action and cognition. Journal of Experimental Psychology: Human Perception and Performance, 35, 1811–1832. doi: 10.1037/a0014510 Stoolmiller, M., & Snyder, J. (2006). Modeling heterogeneity in social interaction processes using multi-level survival analysis. Psychological Methods, 26, 164–177. doi: 10.1037/1082-989X.11.2.164 Te¸sileanu, T., & Meyer-Ortmanns, H. (2006). Competition of languages and their hamming distance. International Journal of Modern Physics C, 17(2), 259–278. doi: 10.1142/S0129183106008765 Tesser, A., & Achee, J. (1994). Aggression, love, conformity, and other social psychological catastrophes. In R. R. Vallacher & A. Nowak (Eds.), Dynamical systems in social psychology (pp. 95–109). San Diego, CA: Academic Press. Thelen, E., & Smith, L. B. (1994). A dynamic systems approach to the development of cognition and action. Cambridge, MA: Bradford/MIT Press. Thelen, E., & Smith, L. B. (1998). Dynamic systems theories. In W. Damon (Ed.), Handbook of Child Psychology: Theoretical models of human development (5th ed., Vol. 1, pp. 563–634). New York, NY: Wiley. Thelen, E., & Ulrich, B. D. (1991). Hidden skills: A dynamic systems analysis of treadmill stepping during the first year. Monographs of the Society for Research in Child Development, 56(1), 1–98. doi: 10.2307/1166099 Thelen, E., Schoner, G., Scheier, C., & Smith, L. B. (2001). The dynamics of embodiment: A field theory of infant perseverative reaching. Behavioral and Brain Sciences, 24, 1–34. doi: 10.1017/S0140525X01003910 Thom, R. (1975). Structural stability and morphogenesis. New York, NY: Benjamin-Addison-Wesley. Tomkins, S. S. (1963). Affect, imagery, consciousness: Vol. II. The negative affects. New York, NY: Springer. Tschacher, W., Scheier, C., & Grawe, K. (1998). Order and pattern formation in psychotherapy. Nonlinear Dynamics, Psychology, and Life Sciences, 2, 195–216. doi: 10.1023/A:1022324018097 van der Maas, H. (1995). Beyond the metaphor? Cognitive Development, 10, 621–642. doi: 10.1016/0885-2014(95)90029-2 van der Maas, H. (1998). The dynamical and statistical properties of cognitive strategies: Relations between strategies, attractors, and latent classes. In K. M. Newell & P. C. M. Molenaar (Eds.), Applications of nonlinear dynamics to developmental process modeling (pp. 161–176). Mahwah, NJ: Erlbaum. van der Maas, H., & Molenaar, P. (1992). Stagewise cognitive development: An application of catastrophe theory. Psychological Review, 99, 395–417. doi: 10.1037/0033-295X.99.3.395 van der Giessen, D., Branje, S. J., Frijns, T., & Meeus, W. H. (2013). Dyadic variability in mother-adolescent interactions: Developmental trajectories and associations with psychosocial functioning. Journal of Youth and Adolescence, 42, 96–108. doi: 10.1007/s10964-012-9790-7

References van der Giessen, D., Hollenstein, T., Hale, W. W., Koot, H. M., Meeus, W., & Branje, S. (2015). Emotional variability in mother-adolescent conflict interactions and internalizing problems of mothers and adolescents: Dyadic and individual processes. Journal of Abnormal Child Psychology, 43, 339–353. doi 10.1007/s10802-014-9910-9 van Dijk, M., Hunnius, S., & van Geert, P. (2012). The dynamics of feeding during the introduction to solid food. Infant Behavior and Development, 35, 226–239. doi: 10.1016/j.infbeh.2012.01.001

759

van Geert, P., & Lichtwarck-Aschoff, A. (2005). A dynamic systems approach to family assessment. European Journal of Psychological Assessment, 21, 240–248. doi: 10.1027/1015-5759.21.4.240 van Geert, P., & van Dijk, M. (2002). Focus on variability: New tools to study intra-individual variability in developmental data. Infant Behavior and Development, 25, 340–370. doi: 10.1016/S01636383(02)00140-6

van Geert, P. (1991). A dynamic systems model of cognitive and language growth. Psychological Review, 98, 3–53. doi: 10.1037/0033295X.98.1.3

van Orden, G. C., Kloos, H., & Wallort, S. (2009). Living in the pink: Intentionality, wellbeing and complexity. In C. Hooker (Ed.), Handbook of the philosophy of science (pp. 639–682). Amsterdam, The Netherlands: Elesvier.

van Geert, P. (1994). Dynamic systems of development: Change between complexity and chaos. New York, NY: Prentice Hall/Harvester Wheatsheaf. van Geert, P. (1995). Dimensions of change: A semantic and mathematical analysis of learning and development. Human Development, 38, 322–331. doi: 10.1159/000278338 van Geert, P. (1997a). Nonlinear dynamics and the explanation of mental and behavioral development. Journal of Mind & Behavior, 18, 269–290.

van Rijn, H., Van Someren, M., & Van der Maas, H. (2003). Modeling developmental transitions on the balance scale task. Cognitive Science, 27, 227–257. doi: 10.1016/S0364-0213(02)00116-7 Vetere, A., & Gale, A. (1987). The family: A failure in psychological theory and research. In A. Vetere, A. Gale, S. Lewis, C. Jolly, & S. Reynolds (Eds.), Ecological studies of family life (pp. 1–17). Chichester, UK: Wiley. von Bertalanffy, L. (1968). General systems theory. New York, NY: Braziller.

van Geert, P. (1997b). Que sera sera: Determinism and nonlinear dynamic model building in development. In A. Fogel, M. C. D. P. Lyra, & J. Valsiner (Eds.), Dynamics and indeterminism in developmental and social processes (pp. 13–38). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.

Webber Jr., C. L., & Zbilut, J. P. (2005). Recurrence quantification analysis of nonlinear dynamical systems. In M. A. Riley & G. C. Van Orden (Eds.), Tutorials in contemporary nonlinear methods for the behavioral sciences (pp. 26–94). Washington, DC: National Science Foundation. Webber Jr., C. L., Marwan, N., Facchini, A., & Giuliani, A. (2009). Simpler methods do it better: Success of recurrence quantification analysis as a general purpose data analysis tool. Physics Letters A, 373, 3753–3756. doi: 10.1016/j.physleta.2009.08.052 Wickens, T. D. (1989). Multiway contingency table analysis for the social sciences. Hillsdale, NJ: Lawrence Erlbaum Associates. Wimmers, R. H., Savelsbergh, G. J. P., van der Kamp, J., & Hartelman, P. (1998). A developmental transition in apprehension modeled as a cusp catastrophe. Developmental Psychobiology, 32, 23–35. doi: 10.1002/(SICI)1098-2302(199801)32:13.0.CO;2-V

van Geert, P. (1998a). A dynamic systems model of basic developmental mechanisms: Piaget, Vygotsky, and beyond. Psychological Review, 105, 634–677. doi: 10.1037/0033-295X.105.4.634-677 van Geert, P. (1998b). Dynamic modeling of cognitive and language development: From growth processes to sudden jumps and multimodality. In K. M. Newell & P. C. Molenaar (Eds.), Applications of nonlinear dynamics to developmental process modeling (pp. 129–160). Mahwah, NJ: Lawrence Erlbaum Associates, Inc. van Geert, P. (1998c). We almost had a great future behind us: The contribution of non-linear dynamics to developmental-science-in-themaking. Developmental Science, 1, 143–159. doi: 10.1111/1467-7687 .00020

Witherington, D. C. (2007). The dynamic systems approach as metatheory for developmental psychology. Human Development, 50, 127–153. doi: 10.1159/000100943

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

CHAPTER 17

Missing Data TODD D. LITTLE, KYLE M. LANG, WEI WU, and MIJKE RHEMTULLA

STATISTICAL ISSUES: WHAT HAPPENS WHEN DATA GO MISSING? 761 Bias 761 Type I Error 761 Type II Error/Power 761 MECHANISMS OF MISSINGNESS 763 Missing Completely at Random 764 Missing at Random 766 Missing Not at Random 766 MODERN MISSING DATA METHODS 767 FULL INFORMATION MAXIMUM LIKELIHOOD 767 Multiple Imputation 768 PRACTICAL CONSIDERATIONS 774 Assumptions 775

Auxiliary Variables 776 Fraction of Missing Information 777 Assessing Model Fit with Missing Data 777 Mediation Analysis 780 When to Use FIML Versus MI 782 REVIEW OF MISSING DATA PRACTICES IN PSYCHOLOGICAL RESEARCH 782 WHY CHANGE? 784 PLANNED MISSING DATA DESIGNS 784 Multiform Planned Missing Protocols 785 Two-Method Planned Missing Design 787 Wave-Missing Designs 789 CONCLUSIONS 790 REFERENCES 791

Missing data are ubiquitous in most human science data, and many prominent social scientists and statisticians have written on the subject of addressing nonresponse in social science studies (e.g., Enders, 2013; Graham, Cumsille, & Shevock, 2012; Graham, Hofer, Donaldson, MacKinnon, & Schafer, 1997; Graham, Hofer, & Piccinin, 1994; Hofer & Hoffman, 2007; Little & Schenker, 1995; Nakagawa & Freckleton, 2010; Sinharay, Stern, & Russell, 2001). The treatments for missing data vary considerably across the disciplines covered under the human sciences umbrella. Unfortunately, the traditional procedures that have been in common use through the last century are prone to bias except in very circumscribed conditions. Fortunately,

recent advances in the capabilities of computer hardware and software have made a new breed of contemporary approaches both accessible and easy to implement. These contemporary treatments are versatile and powerful in the ways that unplanned missing data can be handled. And, as we will describe in detail, the power of the modern missing data approaches can be leveraged to allow remarkable uses in the form of planned missing data designs. The goal of this chapter is to supplement, in an accessible manner, the burgeoning literature on modern missing data treatments, with a particular focus on planned missing data designs. Later in this chapter, we will describe in some detail the two contemporary approaches: full information maximum likelihood (FIML) estimation and multiple imputation (MI). For book-length treatments of these contemporary approaches to addressing missing data, see Allison (2002), Enders (2010), Graham (2012), Little and Rubin (2002), and van Buuren (2012). Our discussions around the topic of missing data will focus on examples derived from clinical research in the area of developmental psychopathology (see also Little, Jorgensen, Lang, & Moore, 2014). Our goal is to fuel the movement toward a data-collection paradigm shift

Partial support for this project was provided by grant NSF 1053160 to the first and third authors, a Banting postdoctoral fellowship from the Social Sciences and Humanities Research Council of Canada to the fourth author, and the Center for Research Methods and Data Analysis at the University of Kansas (when it was under the direction of Todd D. Little). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding agencies. We would like to thank Richard Kinai for assisting with creating Table 17.2.

760

Page 760

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Statistical Issues: What Happens When Data Go Missing?

that embraces planned missing data designs and improves study protocols to better address the inevitable unplanned missing data that will arise. In this regard, missing data are not a bane in the research process, but rather appropriate application of modern methods provide solace for the unplanned missing values and dramatic opportunities in the context of planned missing data designs.

STATISTICAL ISSUES: WHAT HAPPENS WHEN DATA GO MISSING? Regardless of the reasons for missing data, the problem cannot simply be ignored (Allison, Goodson, & Neilands, 2008; Raghunathan, 2004), and the presence of nonresponse necessarily complicates the statistical analysis (Brown, 1994; Cheung, 2007; Finkbeiner, 1979; Gleason & Staelin, 1975; Haitovsky, 1968; Harel, Hofer, Hoffman, Pedersen, & Johansson, 2007; Kim & Bentler, 2002; Lee, 1986; Little, Howard, McConnell, & Stump, 2011; Little, Schnabel, & Baumert, 2000; Longford, 2005; McArdle, 1994; McArdle & Hamagami, 1992; McArdle & Hamagami, 2001; Rovine, 1994; Song & Lee, 2002; Wothke, 2000). More specifically, three statistical issues emerge when data have gone astray: bias, type I error, and type II error (i.e., power). Bias The first key issue is bias. Is there anything systematic about why the data are missing? If so, then the remaining data for the selected sample are no longer representative of the population from which the sample was drawn. Any parameter estimates derived from the remaining data will no longer generalize to the population and any conclusions will be, to some degree, invalid (Little & Rubin, 2002; Schafer, 1997). As we will illustrate in more detail below, the classical, or ad hoc, treatments for missing data (e.g., listwise or pairwise deletion) do nothing to correct for potential bias, whereas the modern, or principled, treatments attempt to correct for any biasing influences (Little & Rubin, 2002; Rubin, 1987). Modern missing data treatments address the bias induced by a systematic influence associated with the missing data by adjusting the parameter estimates to reflect, as closely as possible, what they would have been had there been no missing data (Rubin, 1996). The extent to which this is possible depends on the causes of missing data, and whether those causes are represented in the data (i.e., the missing data mechanism; more on this later). In other

12:38pm

761

words, when data go missing we lose information both in the statistical sense of lower power and in the substantive sense that we know less than what we should know. Type I Error The second issue to contend with in the face of missing data is type I error. Many traditional methods for dealing with missing data (e.g., any single-imputation technique) result in standard errors that are smaller than they should be, given that the data are not complete. The problem here is similar to what would happen if one collected data from 20 people but computed a standard error using N = 200. The standard error is based on an assumption that the data contain more statistical precision than they really do, so the estimated standard error is too small. Too-small standard errors translate to a higher-than-nominal probability of finding a significant result when there is no effect in the population. Type II Error/Power The third issue when data go missing is power loss (i.e., higher type II error rates). As implied in the previous paragraph, missing data means less statistical precision, so it will always result in less powerful analyses than a complete data set (unless the missing values were completely redundant with the nonmissing values; Davey & Savla, 2010; Graham, Olchowski, & Gilreath, 2007; Savalei & Rhemtulla, 2012). Listwise deletion leads to the greatest loss in power of all missing data methods (Enders, 2010; Little & Rubin, 2002; Schafer & Graham, 2002). Single imputation methods, such as mean substitution or regression imputation, will lead to greater power; however, these methods rarely lead to unbiased estimates and are associated with increased type I error rates (Enders, 2010; Rubin, 1987, 1996; Schafer & Graham, 2002). Modern treatments to missing data are efficient, which means they result in the most powerful analysis possible, while also eliminating much of the bias due to selective missingness (Graham et al., 2007). Table 17.1 presents a list of missing data treatments along with descriptions of how each treatment deals with (or fails to deal with) bias and type I and II error. As is plain from this table, the conditions under which the traditional ad hoc methods will be valid are quite limited and often unrealistic in contemporary studies across the human sciences. To fully understand the reasons for the poor performance of classical methods, we need to define the three mechanisms for missing data.

Page 761

Trim Size: 8.5in x 11in

762

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

TABLE 17.1 Methods of Treating Missing Data Method Deletion

Bias/power/error

Description

Listwise deletion/complete case analysis

Bias under MAR & MNAR

Pairwise deletion/available case analysis

Bias under MAR & MNAR

All cases with any missing values are deleted. Analyses are done on the remaining cases. Under MCAR, this technique produces unbiased parameter estimates. Standard errors (and thus confidence intervals and p-values) are larger, reflecting the loss of data. As the proportion of missing data increases, power can decrease substantially. Under MAR/MNAR, the sample after deletion is not representative of the complete data set, resulting in biased parameter estimates (e.g., means, correlations, regression coefficients). Bias increases as the complete case sample grows more different from the complete sample, and as the proportion of missing data increases. Each statistic is computed using cases that have complete data available for that statistic. A case with missing data on V3 is not used to find the correlation between V2 and V3, but is used to find the correlation between V1 and V2. Power loss is less than with listwise deletion but still a problem. Estimates are biased under MAR and MNAR. Every statistic is based on a different sample size, which may result in an impossible correlation matrix, which can cause serious estimation problems. As sample size is a key element in calculating standard errors, lack of a constant N results in some software packages using the average sample size per variable for these computations. This has the potential of underestimating/overestimating the standard error for some variables. Lack of a single value for a sample size affects the accuracy of the standard error and the testing of model fit. Single-imputation techniques address prediction error but not error of estimation, leading to deflated standard error estimates. This leads to inflated type I error rates. When variance is underestimated, correlations suffer and significance testing tends to result in more type II errors. Each missing value is replaced with that variable’s mean. The imputed values have zero variance because they all take on the same value. As such, estimates of the variable’s variance and relations with other variables (e.g., covariances, correlations, regression coefficients) are underestimated. This problem exists regardless of the missingness mechanism and gets worse as the proportion of missing data increases. Each missing value is replaced with its predicted value based on a regression equation using all other variables as predictors. Information from the complete variables is used to fill in the incomplete variables. The imputed values are perfectly predicted by the other variables in the data set, with no error added. Variances will be underestimated (because error variance is lacking), and variable relations (correlations, covariances, regressions) will be overestimated. This problem exists regardless of the missingness mechanism, and gets worse as the proportion of missing data increases. As with regression imputation, each missing value is replaced with its predicted value, but it has a normally distributed residual error term added to the imputed predicted score. The added residual variance adds variability to the imputed variable and adds noise to the estimated relations between variables. Thus, under MCAR and MAR, parameter estimates are unbiased. Under MNAR, the required information to achieve accurate estimates is missing, so bias persists under MNAR. As with every single-imputation technique, the imputed data are treated like complete data, so standard errors will resemble what they would have looked like with complete data. In other words, the information lost due to missing data is not accounted for, so standard errors will be too small. Inflated type I error rates (under every mechanism) result. The EM algorithm produces maximum likelihood (ML) estimates of the covariance matrix 𝚺 and mean vector 𝝁 from incomplete data. The algorithm iterates through two steps. In the expectation (E) step, the missing values are filled in with their expected values after conditioning on the observed data (similar to stochastic regression imputation). The subsequent maximization (M) step applies complete data ML estimation to the imputed data from the E-step to derive updated estimates of the covariances and the means. These updated estimates are used by the next E-step to build the regression equations used to fill in the missing data. The algorithm repeatedly iterates between the E- and M-steps until the difference in the maximized likelihood values from adjacent M-steps falls below a specified criterion. The imputed values from the last E-step may be saved as a single imputed data set. EM is an important building block of more complex missing data methods (e.g., Enders, 2003; 2004; Enders & Peugh, 2004; Lauritzen, 1995; Savalei & Bentler, 2009). The estimated 𝚺 and 𝝁 are the same as those produced by FIML. However, a single EM imputation has the same problem as any other single imputation, which is that uncertainty cannot be accounted for in a single-imputed data set. Type I error rates will be inflated.

Loss of power under MCAR, MAR, & MNAR

Loss of power under MCAR, MAR, & MNAR

Single imputation

Mean substitution

Regression imputation

Bias under MCAR, MAR, & MNAR Error inflation (type I error) under MCAR, MAR, & MNAR Bias under MCAR, MAR, & MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

Stochastic regression imputation

Bias under MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

Expectationmaximization (EM) imputation

Bias under MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

(continued)

Page 762

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Mechanisms of Missingness

12:38pm

763

TABLE 17.1 (Continued) Method Deletion

Bias/power/error

Description

Hot deck imputation

Bias under MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

Cold deck imputation

Bias under MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

Averaging available items (person-mean imputation / prorated scale score)

Bias under MCAR, MAR, & MNAR Error inflation under MCAR, MAR, & MNAR

Both hot and cold deck imputation techniques are common in survey research. Hot deck imputation works by finding participants who match the case with missing data on other variables. For example, missing scores on job performance are replaced by randomly sampling job performance scores from cases with identical age and gender. Depending on how the imputation is actually performed, this method might or might not preserve the correlations among variables. This method is nonparametric and less sensitive to model misspecification. It is not good for small sample sizes. In cold deck imputation, information from external sources is used to determine the matching variables. An external source could be another available data set from which to impute missing values. This technique results in small standard errors (as missing information is not taken into account) and adjusted R2 , denoting decreased variability in the data used for these analyses. A technique used on multiple item scale data with item-level missing data. A scale score is computed by averaging the available items for each case. This technique is equivalent to replacing the missing values for each case with the mean of that case’s values. This technique only works for MCAR assuming parallel items: all items have same mean, same true-score variance, and the same error variance—this is rarely the case. When this assumption is violated, estimates will be biased even when data are MCAR. More specifically, coefficient alpha estimates will be biased, as will measures of association and variance.

Last observation carried forward

Bias under MCAR, MAR, & MNAR Error inflation (type I error) under MCAR, MAR, & MNAR

Missing values in longitudinal data are replaced with the latest previous observation. A key assumption here is that the most recent observation is the best guess for subsequent missing values. The implication is that scores remain constant after the last observed measurement, an assumption that is not reasonable in longitudinal data. Parameter estimates will be biased in unpredictable directions, depending on the particular features of the data. The magnitude of bias is not predictable. Treatment group differences can be attenuated. Type I error rates are inflated through failure to account for the effect of missing information.

Bias under MNAR only Power is maximized and error rates are correct for MCAR and MAR

In MI, multiple values for each missing data-point are imputed. The information lost due to missing data in estimating model parameters is quantified using the variability in parameter estimates between imputed data sets. As such, standard error estimates are unbiased, resulting in accurate type I error rates. Imputation uses all the information in the data set, resulting in unbiased parameter estimates under MCAR and MAR. MI commonly assumes multivariate normal data, but imputation algorithms for other distributions are possible. With the correct assumptions on data distributions and missing data mechanism and a large number of imputations, MI is asymptotically unbiased and efficient. FIML estimates model parameters and standard errors directly using all the available data. FIML aims at identifying the population parameters that are most likely to have generated the observed data (i.e., maximize the likelihood of observing sample data). FIML is a model-based missing data solution, which means that its accuracy depends on the accuracy of the model being estimated. Typically multivariate normality is assumed, though it is possible (but computationally intensive) to use FIML with binary and other distributions. When the assumptions are met, FIML leads to asymptotically unbiased and efficient estimates.

Modern methods Multiple imputation (MI)

Full information maximum likelihood (FIML)

Bias under MNAR only Power is maximized and error rates are correct for MCAR and MAR

MECHANISMS OF MISSINGNESS Missing data can occur because a specific datum for given person is not present (e.g., a participant skipped an item due to an error or not wanting to answer) or because data on all variables at a specific measurement occasion for a specific person is not present (e.g., a participant dropped out of a longitudinal study after the second wave of data collection). The first type of missing data is often

referred to as item missing. The second type of missing data is often called wave missing, or if the wave missingness occurs because of participants permanently dropping out of a study, attrition. Terms like attrition, wave missing, and item missing are general descriptors of the types of missing data that occur because a desired response was not obtained (Enders, 2010; Rubin, 1987; Shafer & Graham, 2002). This type of missing data should be distinguished from structural missingness. Structural missingness refers

Page 763

Trim Size: 8.5in x 11in

764

Cicchetti

12:38pm

Missing Data

to elements of a data matrix that are not possible to collect (e.g., items that ask about pregnancy for participants who are not or cannot become pregnant, empty cells following a negative response to a “gatekeeper” item, or data for participants after they are deceased). The various descriptors of the types of nonresponse that can occur should not be confused with the mechanisms of missing data. A mechanism of missing data is the causal process associated with whether a data point is missing or not but not necessarily what the value of the missing information would have been.1 Missingness mechanisms distinguish the causes of missingness. The pattern of missing data can be described using a missingness matrix, R, which has the same dimensions as the data matrix but contains 0 whenever a datum is observed and 1 whenever it is missing. The key question in establishing the missingness mechanism is: what predicts R? There are three potential sources of prediction, which are the three missingness mechanisms (Rubin, 1976). These causes together account for all missing data. In describing these mechanisms, we will refer to Table 17.2, which provides their definitions in terms of the formal probability distributions of the missing data patterns as originally described by Rubin (1976). Figure 17.1 provides a graphic schematic of the mechanisms of missing data. By way of a running example, we will describe a hypothetical study of the relationship between parental involvement (a scale administered as part of a larger parent 1

c17.tex printer - Volume I - 12/10/2015

Strictly speaking, the missingness mechanism can’t provide any information about what the missing value may have been. It is only a way to describe the process that gave rise to the nonresponse. From the perspective of Rubin (1976), every cell in a data set is associated with two values: an actual data value Yij (which may or may not be observed); and a missingness indicator Rij (which takes a value of 1 when the corresponding Yij is missing and 0 if it is observed). The missingness mechanisms, as they are traditionally conceptualized, are concerned only with describing the process by which Rij takes a value of 1 or 0 (i.e., whether or not Yij is missing or observed), but they do not address what the correct value for any missing Yij might be. This is an important distinction. Knowing what the missing values would have been implies knowledge of the sampling distribution of the missing values. Knowing the missingness mechanism, on the other hand, is a weaker form of information that implies only knowledge of the more general process by which a value is or is not observed (but not anything about what that value may look like). Rubin (1987, p. 17) overtly distinguishes between these two types of information when he says: “In general, it is impossible to estimate a population quantity such as the mean without making assumptions about either the distribution of values for nonrespondents or the process that creates nonresponse.”

protocol) and child school adjustment (a scale administered as part of a larger teacher protocol). As a purely hypothetical point of discussion, we will pick an arbitrary true correlation of .5 between these two constructs. This correlation of .5 exists in the population and in the hypothetical sample that is drawn from the population; that is, sampling error is held constant so we can focus on the influence of the missing data mechanisms. Missing Completely at Random The first mechanism of missing data is termed missing completely at random (MCAR). When the data are MCAR it means that the missing data patterns have zero associations with any variables included in the data set or with any unmeasured variables that are associated with the missing values themselves.2 This lack of association does not mean that a datum is not missing for a reason—under MCAR, the reasons are essentially random. In this regard, the label for this mechanism clearly communicates, even to naïve readers, the reason for the missingness. In this regard, MCAR can be a truly random process or a process that happens to have a zero association with all variables in an investigation. Unfortunately, MCAR data are not as common as we would hope. As indicated in Table 17.1, 2

Every incident of nonresponse, as with any natural occurrence, has some cause. If MCAR missing had no relationship to any variable at all, then it would not exist. In other words, if there is no reason for a value to go missing, it will be observed. This claim can lead to confusion if we describe MNAR as the probability of missingness being predictable by some set of unmeasured variables. This misunderstanding can be largely remedied if we recognize that this definition of MNAR is what Enders (2010) called indirect MNAR. The issue can be illuminated by expanding the description of indirect MNAR slightly. Indirect MNAR is only problematic because the unobserved variables that predict R are also related to the Ymis . This mutual association induces a spurious correlation between Ymis and R. This is one reason that the original Rubin (1976) definitions for the missingness mechanisms are given only in terms of the statistical patterns in the data set. In practice, whether the MNAR is direct or indirect, the observed pattern of relationships in the data will be the same. Namely, there will be an apparent relationship between Ymis and R, even after controlling for the relationship between Yobs and R. As discussed above, the probability of MCAR missing is necessarily predictable by some set of unmeasured variables (otherwise the data wouldn’t be missing). However, unlike the indirect MNAR case, those unmeasured variables are not associated with the missing values. Therefore, in the MCAR case, there is no spurious correlation induced between Ymis and R, and there is no apparent connection between Ymis and R in the data set.

Page 764

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Mechanisms of Missingness

12:38pm

765

TABLE 17.2 Missing Data Mechanisms Mechanism

Mathematical definition1

Verbal definition

Missing completely at random (MCAR)

P(R|YObs , YMis ) = P(R)

Missingness on Y is unrelated to any of the variables of interest. The missing data are a pure random sample of the complete data.

Missing at random (MAR)

P(R|YObs , YMis ) = P(R|YObs )

Missingness on Y is related to the observed data and not Y itself. The missing data are a random sample of the complete data, after conditioning on the observed data.

Missing not at random (MNAR)

P(R|YObs , YMis ) = P(R|YObs , YMis )

The missingness on Y is dependent on the missing values themselves. The distribution of the missing values is unrecoverable because the predictors of nonresponse are missing.

1R

represents the missing data pattern (i.e., 1 if the value is missing, 0 if the value is observed). YMis is the notation for the values of the missing data; YObs represents the values of the observed scores.

Missing Data Values (Attrition and nonresponse)

β=0 β≠0 Ignorable

β≠0

Nonignorable

MCAR

MAR

MNAR

A truly random process

A measured/predictable process

An unmeasured/unpredictable process

r=0

r≠0 r=0

Figure 17.1 The mechanisms of missing data and their potential associations with the missing data and the missing values. (Note: The MCAR variables are necessarily uncorrelated with MAR variables and MNAR variables. MAR and MNAR variables can have varying overlap and the higher the overlap the less the influence of the MNAR process can be. The regression betas reflect the implied strength of the multiple linear prediction of the missing values from the set of possible variables that are classified as either in MCAR, MAR, or MNAR set.) Source: Little (2013). Used with permission of Todd D. Little.

many classic methods for treating missing data assume that all missing data are due to the MCAR mechanism (Little & Rubin, 2002). When the data are MCAR it means that every missing data treatment, except for single imputation techniques, will produce unbiased parameter estimates. When the pairwise deletion method is used with MCAR missing data, for example, estimates will be perfectly accurate. Here, the problem is that the missing data have undermined the power of the statistical model and type II errors are now likely to arise (i.e., failing to detect a true effect as being significant; Schafer & Graham, 2002). Figure 17.1 shows the lack of association between the MCAR mechanism and the missing data as a predictive relationship that is necessarily zero and the correlation of the MCAR mechanism with the other two mechanisms is also necessarily zero. Table 17.2 shows that the conditional probability of the missingness, R, is not associated with the values of the missing variables (YMis ) or with the values of

the observed variables (YObs ), and thus conditional independence holds and the missingness can be ignored. Using the hypothetical example above, some children are missing the teacher-rated adjustment score because one teacher had recently moved, one teacher had the flu, one teacher refused because the child had just arrived in his classroom, one teacher lost the protocol, one teacher was distracted by a health issue, and so on. In each of these cases, a unique reason for each missing data point is to blame—the missing data are due to the whims of truly chance events, none of which are related to any of the variables under examination. Similarly, if a random subset of the completed adjustment protocols were accidentally shredded, this would also constitute an MCAR mechanism. In these scenarios, a principled missing data tools will, on average, reproduce the .5 correlation that exists in the population. In other words, the truly chance events have not introduced bias in the estimated correlation. The

Page 765

Trim Size: 8.5in x 11in

766

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

only problem in this scenario is the loss in power that the lost data have caused. In this scenario the cause of missingness is “ignorable” because it does not need to be explicitly accounted for in the imputation or analysis models (as opposed to the missing not at random mechanism; Rubin, 1976). In other words, when using principled missing data tools, the MCAR mechanism will not have an impact on the estimated parameters of an analysis model (Figure 17.1).

The MAR assumption is often tested by predicting the missing data pattern from variables that are on the data set (e.g., by using the Little, 1988, test for MCAR or logistic regression). If the results of these tests are nonsignificant, then one of two reasons is possible: either the data are MCAR, or they are MNAR. Unfortunately, the ability to differentiate between these two mechanisms is challenging. In many ways, finding evidence of predictive links would be a preferred outcome of a missing data screening because it would indicate evidence of an ignorable MAR process.

Missing at Random The second mechanism of missing data is labeled, confusingly, missing at random (MAR). When data are MAR it means that the causes of the missing data are in the observed data. Here, the values of the missing data can be regarded as a random effect that can be predicted by other variables in the data set. The MAR mechanism, therefore, is also an ignorable mechanism in so far as values of the missing data can be accurately recovered using a modern missing data model. In this regard, therefore, both MCAR and MAR are considered ignorable because the consequences for conclusions are negligible; that is, when employing modern, principled missing data tools, the parameter estimates of any statistical model fit to the data will be unbiased when either or both of these mechanisms are involved. As seen in Table 17.2, under the MAR mechanism, the only predictors of the missing data pattern (R) are contained in the observed data (Yobs ). So the missing values (Ymis ) can be considered a purely random sample of the complete data (Y), after accounting for the values of the observed data (Yobs ). Using the hypothetical example above, some children are missing the parental involvement scores because a large percentage of primary care givers failed to fill out portions of the parental protocol. The reason these primary care providers did not complete the protocol is because they are consumed with the responsibilities and demands of care providing. The questionnaire protocol contains variables that record important characteristics of these care providers that were collected during the intake assessment by a trained interviewer. The variables include the person’s role as a primary care provider or not, the number of hours per day spent providing care, employment status, the number of hours spent working per week, and the like. These variables are strong predictors of the reason for the missing responses. Using these variables in a modern treatment procedure would recover the lost information and would provide unbiased parameter estimates from a given statistical model.

Missing Not at Random The third mechanism of missing data is the nefarious one. Here, the data are missing for a systematic reason but this systematic reason for the missingness is not accessible to the researcher. In this situation, data are MNAR. That is, there is not a predictor of the missingness that a researcher can use. In our example, parents who have low involvement in their children’s education may be the least likely to turn in the survey for all the same reasons that they have low involvement to begin with (e.g., being overly busy). This selective responding results in a situation where the lowest values on parental involvement are missing, and the variable that predicts missingness is parental involvement itself. In situations like these, the variable needed to predict the missingness is, itself, missing. Another way to describe this mechanism is that the data are missing for a reason, but the variable representing the reason is not on the observed data set (i.e., because it is part of YMis in Table 17.2). Under the MNAR mechanism, the equation for predicting the missingness (R) does not simplify—unlike MCAR and MAR, the missing values themselves (YMis ) predict unique information in R. This means that the information that is available in the data set is not sufficient to recover the relations among variables. The MNAR circumstance is a non-ignorable situation because it will have consequences on the validity of any conclusions. The information needed to correct the data for the reasons is simply not available because it is missing. Because the observed data cannot predict the missing values, if these missing values were imputed, the imputations would be constructed as though the missingness mechanism were MCAR. However, the results would be biased because the reason for the missingness is unrecoverable. Models that are designed to recover missing information when the data are MNAR are beyond the scope of our discussion because (1) the methods for preventing bias are complex models that require making strong untestable assumptions (see Enders, 2010, 2011; Muthen, Kaplan, &

Page 766

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Full Information Maximum Likelihood

Hollis, 1987); and (2) with planning, an MNAR mechanism can be converted to a MAR mechanism (Collins, Schafer, & Kam, 2001; Schafer & Graham, 2002; Yoo, 2009; Yuan, 2009a). For example, a classic example of an MNAR mechanism applies to income reported in net or gross annual salary. Very rich participants are more likely to leave this information missing than are the less wealthy participants. In this case the cause of the missingness is itself missing. There is no variable on the data set that would predict it and, therefore, it would not be recoverable. On the other hand, the research protocol could include variables that are known correlates of income. For example, job title, years of education, square-footage of the primary home, make and model of the primary automobile, and political affiliation. Variables such as these will, in a multivariate combination, provide information that now correlates with the pattern of missingness on income. Variables such as these are proxies of the missing information. A well-planned protocol would strive to have such proxies of important information that might be missing as an MNAR process. In the developmental psychopathology literature, sensitive variables such as physical, emotional, and sexual abuse might be missing because the person has been abused in some way. If the protocol contains variables that are expected to predict abuse the process would fall under the MAR mechanism because these variables predict missingness.

MODERN MISSING DATA METHODS Table 17.1 reviews a range of classic techniques for dealing with missing data, from the simplest (e.g., listwise deletion, mean imputation) to more complex (e.g., stochastic regression imputation, hot deck imputation). As described there, none of these methods produce ideal results. At best, some of these methods can produce unbiased parameter estimates when the missingness is MCAR (e.g., listwise or pairwise deletion) or when it is MAR (e.g., stochastic regression imputation). At worst, some of the methods produce biased results even under MCAR (e.g., mean imputation, regression imputation). In addition, all of these methods either lack power (i.e., any deletion method) or result in inflated type I error rates (i.e., any single imputation method). Though it can be justifiable to use one of these methods when there is only a very small amount of missingness, using any of these techniques to deal with a substantial amount of missing data risks biased estimates, low power, or inflated type I error. In contrast to all of these methods, multiple imputation and full information maximum likelihood produce unbiased and fully efficient

12:38pm

767

(i.e., powerful) parameter estimates when their assumptions are met. In the next sections, we describe these two methods in detail.

FULL INFORMATION MAXIMUM LIKELIHOOD Full information maximum likelihood (FIML) is a state-of-the-art missing data tool that represents the culmination of an extensive history of research into maximum likelihood solutions to the missing data problem (e.g., Beale & Little, 1975; M. B. Brown, 1974; C. H. Brown, 1983; Carter & Meyers, 1973; Chen & Fienberg, 1976; Dempster, Laird, & Rubin, 1977; Enders, 2001b; Enders & Bandalos, 2001; Gold & Bentler, 2000; Hartley, 1958; Hartley & Hocking, 1971; Healy & Westmacott, 1956; Jamshidian & Bentler, 1999; Kenward, Lesaffre, & Molenberghs, 1994; Kenward & Molenberghs, 1998; Little & Rubin, 1989; Orchard & Woodbury, 1972; Rubin, 1991). The framework for FIML was formalized by Anderson (1957) as a way to derive ML estimates directly from incomplete data. Through clever manipulation of the likelihood function, FIML can hide any missingness from the estimator and allow one to fit a statistical model directly in the presence of incomplete data, without any need to estimate values for the missing data (Arbuckle, 1996). This approach has a distinct advantage over EM and imputation-based missing data tools. Because there is no estimation of values for the missing data, there is also no artificial certainty introduced into the treated data (as opposed to EM and single imputation techniques) nor is there any need to quantify the uncertainty in the accuracy of the imputed values (as must be done with multiple imputation). To get a conceptual idea of how FIML operates, consider your computer’s monitor. The image on this monitor is made up of many rows of pixels, just as a data set is made up of many rows of responses. If a pixel dies in your monitor, you are still able to understand the image on the screen because you can use the information from surrounding pixels to infer the whole image, passing over the damaged one. This principle will hold true even if a relatively large proportion of the pixels in your monitor were to fail, so long as there are enough pixels remaining for you to connect the dots and extrapolate the implied complete image from the partial image on your screen. Similarly, the FIML log likelihood function (equation 17.2) achieves an analogous effect when used to fit a statistical model to incomplete data. By using only what is known from the observed data, FIML can estimate around the missing data and infer what the whole model should look

Page 767

Trim Size: 8.5in x 11in

768

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

like without needing to know what the missing responses would truly be. In this way, just as you can look at a damaged computer monitor and still understand how the complete image would appear, FIML can be applied to an incomplete data set to produce estimates that correctly describe the entire sample. While the underlying mathematics can get complicated, the overarching principles guiding FIML estimation are relatively simple. First we must consider how a maximum likelihood estimator transforms the information contained in the observed data into a set of parameter estimates. Since FIML is based on normal-theory ML estimation, the multivariate normal loglikelihood function is a natural starting point for this discussion. Equation 17.1 gives the multivariate normal loglikelihood function: 𝓁 (𝝁, 𝚺|Y) ] N [ ∑ k 1 1 T −1 = − ln (2𝜋) − ln |Σ| − (Yi − 𝝁) Σ (Yi − 𝝁) , 2 2 2 i=1 (17.1) where Y is a N × k matrix of incomplete data, 𝝁 is the k-variate vector of unknown, population-level means of the variables in Y, and 𝚺 is the k × k, unknown, population-level covariance matrix of the variables in Y. The derivation of this equation falls beyond the scope of this chapter. However, for the sake of the current discussion, the reader need only recognize that equation 17.1 represents a reformulation of the multivariate normal distribution that provides a more computationally efficient way to compute ML estimates of 𝝁 and 𝚺. The beauty of FIML estimation is that it accomplishes its objective with only a small modification to equation 17.1. In the presence of missing data, equation 17.1 becomes the following: 𝓁FIML (𝝁, 𝚺|Y) ] N [ ∑ k 1 | | 1 T −1 = − ln (2𝜋) − ln |Σi | − (Yi − 𝝁i ) Σi (Yi - 𝝁i ) . 2 2 2 i=1 (17.2) The only difference between equations 17.1 and 17.2 is that equation 17.2 includes subscripted versions of 𝝁 and 𝚺. This subtle distinction allows each row in Y to have unique versions of 𝝁 and 𝚺, which allows the dimensions of 𝝁 and 𝚺 to differ between observations. When the dimensions of 𝝁 and 𝚺 can differ across observations, it is possible for the final aggregated estimates of those parameters to be based on only the observed data, thereby hiding the nonresponse from the estimator and removing the need to delete any

data or fill in any values for the missing data. This small change to the likelihood function is the mathematical procedure that allows FIML to estimate around the missing data and infer a complete set of parameter estimates from incomplete input. As long as there is sufficient overlap in the questions with observed responses (i.e., the coverage is adequate), FIML estimation can borrow enough information across observations to effectively remove any need to consider what the true values of the missing data would have been. In practice, FIML estimation is quite convenient because it works behind the scenes to produce accurate parameter estimates without requiring the user to give much thought to the missing data (Enders, 2006, 2010; Muthen & Muthen, 2012). Consider the case of many participants filling out the same survey (i.e., a standard cross-sectional design). Even if some of those subjects do not complete the survey, we would still like to fit a model that will allow us to draw accurate conclusions about the entire sample, but we may not wish to go through the process of filling in values for the missing data. FIML estimation can help us accomplish this by using the observed responses to supplement the loss of information due to the missing responses (Enders, Dietz, Montague, & Dixon, 2006). In the end, FIML will produce unbiased and efficient estimates of the model parameters that we’re actually interested in, without ever requiring us to consider the true values of the missing data. While FIML is often the preferred method of addressing nonresponse in psychological studies, there are certain situations in which maximum likelihood methods are not applicable. For example, if a researcher wishes to calculate scale scores or parcels from incomplete data, they will find that there is no way to apply FIML estimation to those analyses without an additional loss of information (Enders, 2010). In such circumstances, it is beneficial to begin the substantive analysis with a complete data set. Thus, the analyst will be better served by a technique that fills in the missing values through imputation (Rubin, 1996). Multiple Imputation Single-imputation techniques all use some prediction mechanism (some better than others) to fill in the missing value with a plausible value. Once the missing data are thus filled in, the analysis of interest is run on the completed data set. A fundamental problem with this approach is that there is no way to quantify the additional uncertainty that results from the missingness (Rubin, 1987; Schafer, 1997). For example, a data set with 200 cases and 5% missing data

Page 768

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Full Information Maximum Likelihood

contains much more information than the same data set with 75% missing data. With a single-imputation approach, both situations will result in statistics that have similarly sized standard errors and confidence intervals. Intuitively, we can see that this should not be the case—statistics based on data with 75% missingness should have more uncertainty than those based on data with 5% missingness. This is where single imputation fails. Multiple imputation overcomes this problem through introducing a second source of variation by repeating the imputation process multiple times. In MI, a modified single imputation technique (usually some form of regression imputation) is repeated to simulate multiple plausible values for each missing data point. Because there are small differences in the values chosen to replace the missingness from one iteration of the MI algorithm to the next, the final answer is a plausible distribution for each missing value, rather than an arbitrarily chosen single value that replaces each missing datum (Rubin, 1987; Schafer, 1997). With more than one imputation (traditionally three to five, but these days more often 20 or even 100), the amount of information loss can be measured by looking at how different the imputed data sets are (Harel, 2007). If very little information is lost (e.g., if only a few values are missing, or if the missing values are highly redundant with complete values), then the imputed data sets will be highly similar to each other; that is, there will be low “between-imputation” variability. If a lot of information is lost (e.g., if a high proportion of data are missing, or if the missing values contain a lot of unique information), then the imputed data sets will differ a lot, and there will be high between-imputation variability. When doing an analysis on imputed data sets, the standard errors of the produced statistics are adjusted using the between-imputation variability—the greater the between-imputation variability, the larger the standard error becomes, to reflect the missing information. In this way, multiple imputation produces both unbiased estimates and correct standard errors. If we think of the missing data as random variables with a distribution of possible values—of which the imputed values are just one possibility from an infinitely large pool—it is natural to compare the rationale for the multiplicity of MI to the rationale for collecting large samples, in general (Rubin, 1987). For most substantive research, gathering larger samples is advisable because basing your inferences upon a larger pool of observations gives you more certainty in the veracity of your conclusions. The exact same reasoning can be applied to the repeated imputations of MI.

12:38pm

769

For the sake of argument, consider the following extreme example. Say you wish to examine the prevalence of claustrophobia among elementary school students. You could go to a local elementary school and ask a single student to rate his or her fear of confined spaces on a scale from 1 to 10. If that student gives a response of 9, and you don’t question any other children, then you are limited to the information carried by that student’s data when constructing your inference. In this situation, your only option would be to conclude that elementary school students universally suffer from severe claustrophobia. This approach is plainly ridiculous, and it is clear that sampling more students would likely alter your finding. While most researchers are naturally skeptical of statistical inferences based on very small samples, they may not automatically apply this same reasoning to missing data analysis. However, just as relying on only one child’s data to model the severity of childhood claustrophobia can lead to highly erroneous findings, so too can blindly trusting the single set of imputed values produced by a single-imputation technique (Schafer, 1997). In the preceding example, it would be advisable to sample many children and use that large sample to derive a probability distribution that describes the levels of claustrophobia among elementary school students. The same is true of imputation. It is bad practice to impute a single value and trust it without question. A much more trustworthy answer can be achieved by simulating many different values for the missing data (i.e., create multiple imputations) so that we can derive a distribution to describe the plausible values that the missing data may take in the population. This distribution of plausible values helps us understand the degree of uncertainty associated with any of the parameter estimates of a given model. At a conceptual level, the MI approach is relatively straightforward. An MI analysis is a piecemeal endeavor that is broken into three phases: the imputation phase, the analysis phase, and the pooling phase (Enders, 2010). Suppose you wish to analyze an incomplete data set Y = {Yobs , Ymis }, where Yobs is the observed portion of the data and Ymis is the missing portion. In the imputation phase, you must fill in Ymis with M different sets of plausible values (i.e., you must create M imputations). This process will result in M imputed data (1) (2) sets Y (1) = {Yobs , Yimp }, Y (2) = {Yobs , Yimp }, . . . , Y (M) = (M) } in which the original missing values of Y are {Yobs , Yimp replaced by the M sets of imputations. In the analysis phase, you will use these imputed data sets to fit M replicates of your statistical model. This step will result in M distinct ̂ (1) , Q ̂ (2) , . . . , Q ̂ (M) sets of model parameter estimates Q

Page 769

Trim Size: 8.5in x 11in

770

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

(these could be fitted regression coefficients, for example). Finally, in the pooling phase, you will combine these M sets of estimates into a single pooled set of estimates from which you will draw your conclusions. This final pooling is most commonly accomplished by applying Rubin’s Rules (Rubin, 1987). The original framework for MI was proposed by Rubin (1978) and greatly expanded by Rubin (1987), but the flexibility of the underlying concept has left the door open for the development of many different algorithms with which the imputations themselves can actually be created. Table 17.3 shows a list of currently popular software packages that can implement MI. Although there are seven distinct software packages listed in Table 17.3, the estimation methods they use to produce the imputations can be broken into three basic classes, namely: Bayesian data augmentation (or Bayesian MCMC), multiple imputation with chained equations (or sequential regression imputation), and bootstrapped EM imputation. These three approaches can be further categorized into joint modeling and fully conditional specification techniques. The joint modeling techniques (data augmentation and bootstrapped EM imputation) treat all missing data simultaneously by applying a single parametric model that describes the joint distribution of all the missing values. Fully conditional specification techniques (multiple imputation with chained equations), on the other hand, treat each individual variable sequentially and model the conditional distribution of each individual variable’s missing data separately (van Buuren & Groothuis-Oudshoorn, 2011). We will begin by discussing joint modeling techniques because they were the first to become widely available (Schafer, 1997; Schafer & Olsen, 1998), and fully conditional specification techniques can be discussed as special cases of the joint modeling approach.

𝝁 and the covariance matrix 𝚺 (user specified starting values are used in the first I-step) are used to construct the coefficients of a series of multivariate regression equations (one for each distinct pattern of missingness). A randomly drawn, multivariate normally distributed vector of residual terms is then added to each of the predicted values from these equations and the resulting quantities are used to replace the missing values. Thus, after each I-step, all missing data are filled in, and the data set contains no missingness. From the Bayesian perspective, this is equivalent to creating the imputations by randomly sampling from a posterior predictive distribution of the missing data (Enders, 2010). The second step, or posterior (P) step, uses the imputed data set from the preceding I-step to calculate updated values for 𝝁 and 𝚺. These new values for 𝝁 and 𝚺 are then used to construct the Bayesian posterior distributions for the mean vector P (𝝁|Y, 𝚺) and covariance matrix P (𝚺|Y, 𝝁 ). Finally, MCMC simulation techniques are used to randomly draw simulated values for the mean vector 𝝁* and the covariance matrix 𝚺* from these posterior distributions. These simulated values are then passed on to the next I-step, where they are used to construct updated regression coefficients for use in the next round of stochastic regression imputation. This process is repeated a large number of times (e.g., 5000 I and P step pairs) and every n-th imputed data set produced by the I-step is saved. It is important to ensure that the number of imputed data sets to discard (n – 1) is large enough that the imputed values in subsequently saved data sets are statistically independent of each other (Baraldi & Enders, 2010). When M data sets have been saved, the algorithm terminates and these saved data sets make up the final collection of multiply imputed data sets. Bootstrapped EM

Data Augmentation Data augmentation was one of the first algorithms to become popular for use in creating multiple imputations. Data augmentation is a special application of Gibbs sampling that was proposed by Tanner and Wong (1987) as a method of estimating a Bayesian posterior distribution of model parameters when the underlying data are incomplete. These parameters can then be used to construct the multivariate regression equations needed to impute the missing data. The process is iterative, and accomplished in two steps. The first step, or imputation (I) step, is basically a multivariate stochastic regression imputation. During the I-step, the most up-to-date values for the mean vector

Another popular joint modeling technique is the bootstrapped EM algorithm that was developed by Honaker and King (2010) and implemented in the R (R Core Team, 2012) package Amelia II (Honaker, King, & Blackwell, 2011) as a way to address the computational inefficiency of traditional data augmentation approaches. The bootstrapped EM algorithm is very simple at a conceptual level. First, standard bootstrapping procedures are applied and the incomplete data are resampled with replacement to create M incomplete bootstrapped samples of the same dimension as the original data set. Next, the EM algorithm is applied to each of these M samples to create M ML estimates of the mean vector 𝝁 and covariance matrix 𝚺

Page 770

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Full Information Maximum Likelihood

12:38pm

771

TABLE 17.3 Software Packages That Implement Multiple Imputation Software (citation)

Estimation technique Pros

Cons

Amelia II (Honaker, King, & Blackwell, 2011)

Bootstrapped EM Algorithm

Robust convergence Models time when imputing longitudinal data Freeware/Open Source

Only supports MVN imputation models Applies the same imputation model to all variables

mice (van Buuren & Groothuis-Oudshoorn, 2011)

Chained Equations

Accommodates mixed variable types Applies the correct imputation model to categorical and count variables Maintains factor labels for imputed values Offers numerous elemental methods Allows unique imputation models for each variable Freeware/Open Source Imputes multilevel data

Imputed values can be sensitive to the order that the variables are visited

mi (Su, Gelman, Hill, & Yajima, 2011)

Chained equations

Accommodates mixed variable types Applies the correct imputation model to categorical and count variables Allows unique imputation models for each variable Freeware/Open Source Includes many imputation diagnostics

Imputed values can be sensitive to the order that the variables are visited

Mplus (Muthen & Muthen, 2012)

Bayesian MCMC

Accommodates many different imputation models Imputes multilevel data Allows user-specified imputation models Categorical variables are imputed using probit models

Unrestricted models can exhibit poor convergence for large imputation models Applies the same imputation model to all variables Does not implement the full flexibility of the MICE approach

Chained equations

Accommodates mixed variable types Allows unique imputation models for each variable

Imputed values can be sensitive to the order that the variables are visited Does not implement the full flexibility of the MICE approach Limited elemental methods

Data augmentation

Uses one of the best studied MCMC algorithms available

Only supports MVN imputation models Applies the same imputation model to all variables

Chained equations

Accommodates mixed variable types Applies the correct imputation model to categorical variables Allows unique imputation models for each variable

Does not implement the full flexibility of the MICE approach Imputed values can be sensitive to the order that the variables are visited

SPSS: Missing Values (IBM Corp, 2012)

Chained equations

Accommodates mixed variable types Applies the correct imputation model to categorical variables Allows unique imputation models for each variable

Imputed values can be sensitive to the order that the variables are visited Does not implement the full flexibility of the MICE approach

Stata: mi (StataCorp, 2013)

Chained equations

Accommodates mixed variable types Applies the correct imputation model to categorical variables Allows unique imputation models for each variable

Imputed values can be sensitive to the order that the variables are visited

Data augmentation

Uses one of the best studied MCMC algorithms available

Only supports MVN imputation models Applies the same imputation model to all variables

SAS 9.3: Proc MI (SAS Institute, 2011)

(Table 17.1). Finally, the M estimates of 𝝁 and 𝚺 are used to carry out M multivariate stochastic regression imputations which are applied to the original incomplete data, just as in the I-step of traditional data augmentation. This process results in M replicates of the original data, each of which contain a different set of imputed values, in other words, M multiply imputed data sets.

Multiple Imputation with Chained Equations Multiple imputation with chained equations (MICE) has been gaining in popularity over recent years. It is unique among the algorithms discussed here because it is a fully conditional specification technique rather than a joint modeling technique. Therefore, MICE does not seek to treat all variables simultaneously; rather, the MICE

Page 771

Trim Size: 8.5in x 11in

772

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

approach sequentially addresses each variable individually (van Buuren, 2012). While this is an important distinction, with considerable implications for the usage of the MICE approach, the estimation process underlying MICE is essentially identical to data augmentation. The only meaningful difference is that joint modeling techniques treat the imputation as a large multidimensional problem, in which the I-step performs a multivariate stochastic regression imputation and the P-step draws a k-variate 𝝁* and k × k 𝚺* from the joint posterior distributions P (𝝁|Y, 𝚺 ) and P (𝚺|Y, 𝝁 ). The MICE I- and P-steps, on the other hand, reduce to repeatedly solving k univariate analogues of the data augmentation I- and P-steps, where k is the number of incomplete variables to be imputed. In other words, the MICE I-step imputes the missing portion of only a single variable at a time, and the MICE P-step simulates only the univariate mean 𝜇* and variance 𝜎 2* of that focal item. This approach affords a great deal of flexibility to the analyst because each variable’s imputations can be created from its own distinct regression equation (van Buuren & Groothuis-Oudshoorn, 2011). This approach also allows specification of appropriate posterior distributions for the parameters simulated in the P-step even in cases where the joint distribution of all of these parameters may not have a well-defined form (van Buuren, 2012). For example, suppose you have collected data from middle school students and you wish to analyze three variables: age, gender, and a count of aggressive outbursts in class. If there is missingness on these variables, you will be limited in your ability to treat that missingness with data augmentation or bootstrapped EM, since both techniques are based on the assumption that the incomplete data are jointly multivariate normally distributed. However, MICE allows an imputation model where age is imputed with a normal-theory regression model, gender with a logistic regression model, and the count of aggressive outbursts with a Poisson regression model. Thus, the MICE approach will produce correctly distributed imputed values for each special variable type while only requiring the user to run a single instance of the multiple imputation procedure. Rubin’s Rules Rubin’s rules are a simple set of formulae given by Rubin (1987) that allow pooling of the M distinct sets of parameter estimates derived from the analysis phase of an MI analysis. Recall that the product of the MI analysis phase ̂ (2) , . . . , Q ̂ (M) . The ̂ (1) , Q is a set of M parameter estimates Q

multiple imputation point estimate is simply the arithmetic mean of these M individual parameter estimates: 1 ∑ ̂ (m) Q , M m=1 M

Q=

(17.3)

̂ (m) is the parameter where Q is the pooled point estimate, Q estimate from the m-th imputation, and M is the number of imputations. Pooling the variance components is a little more involved than pooling the point estimates and requires three distinct steps. The pooled within-imputation variance is simply the arithmetic mean of the M parameters’ sampling variances: 1 ∑ SE 2̂ , M m=1 Q(m) M

U=

(17.4)

where U is the pooled within-imputation variance, SE 2̂

Q(m)

is the squared standard error for the parameter estimate from the m-th imputation, and M is the number of imputations. The next step is calculating the between-imputation variance. The between-imputation variance simply captures the dispersion in the imputed values across all M imputations. It is calculated via the following formula: 2

M ( ) 1 ∑ ̂ (m) B= Q −Q , M − 1 m=1

(17.5)

̂ (m) is the where B is the between-imputation variance, Q parameter estimate from the m-th imputation, Q is the multiple imputation point estimate, and M is the number of imputations. Finally, these two sources of variance must be combined into the total sampling variance. The total sampling variance is simply a weighted sum of the within and between-imputation variances: T = U + (1 + M −1 )B,

(17.6)

where T is the total sampling variance, U is the pooled within-imputation variance, B is the between-imputation variance, and M is the number of imputations. With a large number of imputations, the influence of the weighting term in equation 17.66 becomes negligible and the total sampling variance reduces to a simple sum of the within and between-imputation variance. Once the multiple imputation point estimate and total sampling variance have been calculated, they can be

Page 772

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Full Information Maximum Likelihood

combined into familiar test statistics. The MI Wald statistic is produced by combining the pooled components in the natural way: t(v)MI =

Q − Q′ , √ T

(17.7)

where Q is the multiple imputation point estimate, Q′ is the null hypothesized value of the parameter estimate, and T is the total sampling variance. This test statistic can be compared with a student’s t distribution with degrees of freedom given by the following: [ U v = (M − 1) 1 + ( ) 1 + M −1 B

]2 ,

(17.8)

where v is the degrees of freedom, U is the pooled withinimputation variance, B is the between-imputation variance, and M is the number of imputations. These components can also be combined into the following confidence interval for the multiple imputation point estimate: √ (1 − 𝛼)%CIMI = Q ± tv,1−𝛼∕2 T,

(17.9)

where 𝛼 is the probability of a type I error (i.e., significance level), Q is the multiple imputation point estimate, tv, 1−𝛼∕2 is the cutoff value on the student’s t distribution corresponding to the desired significance level and the degrees of freedom calculated via equation 17.8, and T is the total sampling variance. Convergence To ensure the quality of the imputed data sets when MCMC is used to impute data, convergence of MCMC needs to be carefully monitored. An MCMC procedure is deemed converged if the posterior distribution of the parameters in the imputation model reaches a stationary point. The convergence of MCMC ensures that the imputed data are drawn from the right posterior distribution. The iterations prior to convergence, termed burn-in iterations, need to be disregarded. Depending on the specific program one chooses, the number of burn-in iterations of MCMC is either decided by the program automatically or needs to be specified by researchers themselves. In Mplus, convergence of MCMC is monitored using a criterion termed potential scale reduction (PSR; Gelman &

12:38pm

773

Rubin, 1992). The number of burn-in iterations is then specified based on this criterion. Detailed explanation of the PSR is beyond the scope of the chapter. What applied researchers need to know is that there will be one PSR for each parameter in the imputation model at each iteration of MCMC. According to Asparouhov and Muthén (2012), MCMC is deemed as converged if the worst PSR becomes less than 1.05 or 1.1 (Asparouhov & Muthén, 2012). In addition, graphical tools such as trace plots and autocorrelation plots can be very useful in assessing MCMC convergence. Trace plots display the change of simulation draws of each of the imputation parameters over iterations. When MCMC converges, the simulation draws should show random fluctuations over iterations. Autocorrelation plots are plots of autocorrelations between the simulated draws for each of the imputation parameters against lag. The autocorrelations usually decrease to zero after a certain number of lags. The faster the autocorrelations decrease to zero, the faster an MCMC procedure converges. Due to their autocorrelations, the imputed values from successive iterations are also correlated. Because the imputed data sets need to be independent to produce correct standard error estimates, it is necessary to allow a number of iterations to lapse every time before saving an imputed data set. This strategy is termed thinning and the number of iterations between imputed data sets are termed between-imputation iterations. Usually 100 between-imputation iterations would be sufficient for most of cases. However, if one has a difficult data set with a large fraction of missing information, a larger value should be specified for the thinning interval. The autocorrelation plots can inform researchers of the appropriate number of between-imputation iterations. There are also many things researchers can do to facilitate the convergence of MCMC (Asparouhov & Muthén, 2012). First, researchers need to make sure that there are no variables in the imputation model that are highly correlated with one another. Second, try not to include variables that do not predict the missing values in the imputation model (e.g., ID). Although, including them has no harm on the parameter estimates (Collins, Schafer, & Kam, 2001), they might cause convergence problems. Third, if there are missing data indicators (binary variables created to indicate whether a score is missing or not), they should not be included in the imputation model as their correlations with the variables having missing data are not identified. Fourth, the imputation model has to be identified. For single-level data, the number of observations on each of the variables

Page 773

Trim Size: 8.5in x 11in

774

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

should be no less than the number of variables included in the imputation model. For example, if the variable with most of the missing data has 50 observations, then including more than 50 variables in the imputation model would create convergence problem. Finally, If MI does not converge and you don’t know which variable causes the problem, try to include a subset of variables and add variables sequentially into the model. Number of Imputations The number of imputations that are needed for MI to give accurate and efficient standard errors depends on the amount of missing information (Bodner, 2008; Graham et al., 2007). Although MI standard errors are less efficient than those of FIML as long as the number of imputations is finite, practically speaking the incremental gain in efficiency from more imputations drops off very quickly (von Hippel, 2005). Until recently, five imputations was considered a sufficient number of imputations to yield standard errors virtually as efficient as FIML (Rubin, 1987). More recently, 20 imputations has emerged as the new minimum and many analysts recommend producing 100 imputed data sets to ensure comparability to FIML estimates (Bodner, 2008; Enders, 2010; Graham et al., 2007; Little, 2013). A key consideration in deciding the number of imputations is the amount of missing information, which can be quantified in the fraction of missing information (FMI) statistic (Bodner, 2008; Graham et al., 2007; Savalei & Rhemtulla, 2011). Recall that when more information is missing, there is greater variability between imputations. As such, each individual imputed data set is less reliable. A lower FMI means that the missing data are highly predictable (or there are very few missing data) and thus fewer imputations are needed to produce a reliable result, whereas a high FMI indicates that more imputations are needed due to the uncertainty of the missing values (Fraley, 1999; Rubin, 1987). White, Royston, and Wood (2011) suggested a rule of thumb that the number of imputations should be at least the fraction of missing information multiplied by 1000. That is, if 50% of the information is missing (FMI = .5), then the minimum number of imputations should be 50. Unfortunately, estimates of the FMI itself can fluctuate dramatically when few imputations are used. To get a reliable estimate of the fraction of missing information from MI, several hundred imputations are needed. As such, with advances in computing making it relatively quick and easy to do many imputations and combine the results across them using automated software, we recommend using 100 imputations whenever possible.

Above, we mentioned a type of missing data that we called structural missing such as items on a questionnaire that would not be asked of certain individuals because the items do not pertain to the person (e.g., asking boys about their menstruations). Multiple imputation procedures are unable to distinguish between structural missing and the nonresponse forms of missingness that we desire to recover. In these situations, we recommend creating a flag variable to indicate true missing from structural missing. The imputed values for the structural missing can then be reset to missing by using the flag variable. PRACTICAL CONSIDERATIONS Keep in mind that the primary goal of a modern missing data procedure is to reproduce the summary parameter estimates (Rubin, 1996). In the case of FIML estimation, the goal is to provide unbiased values of the parameter estimates from the statistical model being fit to the data (Arbuckle, 1996; Enders, 2010). As mentioned previously, all observations are used to inform the model’s parameter estimates. With MI, the goal is to produce sufficient statistics that are unbiased (Rubin, 1987; 1996) and accurately reflect the statistical precision in the data. The sufficient statistics are the means, variances, and covariances among the variables (note that the range is not a sufficient statistic). The statistical model is fit at the level of the sufficient statistics and it produces unbiased parameter estimates that are equivalent to those of FIML (all other things being equal; Graham, et al., 2007). With MI, the person-level information is not part of the imputation objective (Rubin, 1996; Schafer & Graham, 2002). The imputation model does use plausible values for the cells that are missing but these values are not treated as scores for the person because many different plausible values can be used to optimize recovery of the unbiased sufficient statistics. The value assigned to a cell is more of a temporary place holder that is used to infer the population sufficient statistics. When a statistical model is fit to the aggregate data, the parameter estimates of this model will be unbiased. Imputed values for a given person, therefore, should be treated with caution. The distribution of the imputed values for a given person can be quite variable and any one imputation would not be a trustworthy estimate for that person (Schafer, 1997). It would be possible to produce an estimated value from the imputed scores that would be the mean of the imputed values with a confidence interval around this value to be used for assessment purposes, for example. A second important objective of missing data analysis is accurate quantification of the information inherent in the

Page 774

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Practical Considerations

incomplete data (Harel, 2007). The missing information principle of Orchard and Woodbury (1972) states that the information of an incomplete data set can be partitioned into two quantities: the observed information and the missing information. Modern missing data tools allow us to differentiate between these two types of information and accurately estimate their values (Savalei & Rehemtulla, 2012). While both MI and FIML quantify the data’s information, they address the problem in two different ways. MI quantifies the missing information as a function of the between-imputation variance (Rubin, 1987; Schafer, 1997). As suggested above, this way of quantifying is another reason that it is advantageous to construct a large number of imputations—doing so will produce a more accurate estimate of the missing information. FIML, on the other hand, implicitly derives a corrected estimate of the observed information by ascribing more certainty to those parameter estimates that are derived from a larger number of observed responses (Enders, 2010). Assumptions Both FIML and MI are model based in the sense that the accuracy of their results rely on the assumptions underlying their respective models. For FIML, the missing data model is the analysis model (e.g., a structural equation model or regression model). For MI, the missing data model is the imputation model, which is specified separately. In both cases, to the extent that the model is false, the results of the procedure will suffer (Little & Rubin, 2002; Schafer & Olsen, 1998). One common assumption is that the associations between variables, including the associations that account for missingness, are linear (Honaker & King, 2010). If the missingness is actually due to an interaction between two variables, for example, and that interaction term is omitted from the missing data model, the missingness will not be accounted for. In MI, new variables such as powered polynomials and product terms can be inserted into the imputation model to represent potential nonlinear effects of missingness. For example, socioeconomic status (SES) is a common linear correlate of dropout in longitudinal studies. If the rate of dropout increased in a nonlinear fashion at lower levels of SES, the squared powered term of SES (SES2 ) can be included in the imputation model to recover the nonlinearity of the MAR process (Graham, 2009; von Hippel, 2009). In FIML, these nonlinear variables may be used as auxiliary variables (see next section). Another common assumption is that of independence, which is violated when data are nested. Nested data are

12:38pm

775

common in longitudinal data and samples collected from schools, where data points are nested within individuals, within classrooms, or within schools. In these cases, if the missing data model does not take into account the nested structure of the data, the complex relations between variables within and across clusters can be obscured, resulting in incorrect estimates of between- and within-cluster variability and incorrect standard errors (Beunckens, Molenberghs, Thij, & Verbeke, 2007). When dealing with nested data and using FIML, if the nested structure is accounted for in the model (e.g., by using a multilevel model or multilevel structural equation model; Muthén & Asparouhov, 2009) maximum likelihood estimation of the model will successfully recover the relations among the variables at the levels included in the model. When using MI, it is important to remember to incorporate the nesting structure into the imputation model as well as the analysis model. If the imputation model does not account for nesting, the imputed values will underestimate the amount of variance at the individual level, leading to too-small standard errors, even if a correct multilevel model is used as the analysis model (Schafer, 1997). When more than one level of analysis is of interest, for example, if a researcher is interested in both the effects of student-level predictors such as student motivation as well as teacher-level predictors such as teacher experience, the former effects will be overestimated and the latter effects will be underestimated if the nesting is not properly modeled during multiple imputation (Reiter, Raghunathan, & Kinney, 2006). Several methods have recently been implemented in software to account for nesting in multiple imputation (e.g., mice—van Buuren & Groothuis-Oudshoorn, 2011; Pan—Schafer &Yucel, 2002; SHRIMP—Yucel & Raghunathan, 2006). Another assumption of many modern missing data treatments is that the variables are multivariate normally distributed (Allison, 2006; Demirtas, Freels, & Yucel, 2008; Enders, 2001a; Gold, Bentler, & Kim, 2003; Honaker & King, 2010; Savalei, 2010a). Only a few studies have been done on the degree to which MI and FIML are robust to violations of this assumption and whether some modern procedures are better than others when the normality assumption is violated (e.g., Savalei & Bentler, 2005; Shin, Davidson, & Long, 2010). Although some work suggests that both normal-theory MI and FIML are robust to violations of normality (Demirtas et al., 2008; Enders, 2001a; Graham & Schafer, 1999; Yuan, 2009b), further research in this area is needed. We still recommend a modern imputation approach even when the assumptions are violated because the consequences of using another method

Page 775

Trim Size: 8.5in x 11in

776

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

are graver than the lack of precision that nonnormal data might introduce (Graham, 2009; Schafer & Graham, 2002). Auxiliary Variables Auxiliary variables are those that are associated with the missing data pattern or the values of the missing responses, but are not substantively interesting to the researchers (Collins et al., 2001; Enders & Gottschall, 2011). That is, they are variables that capture a MAR missing data process. Recall that FIML and MI only result in unbiased estimates when the causes of missingness are included in the model (FIML) or the imputation model (MI). When FIML estimation is used, auxiliary variables must be included in the analysis model. Graham (2003) described a method for including auxiliary variables in any structural equation model without changing the interpretation of the substantively interesting model parameters, called the “saturated correlates” approach. In a nutshell, this approach introduces auxiliary variables into the model by allowing them to correlate with other exogenous variables (e.g., predictors) in the model, without allowing them to predict unique variance in any dependent variables. In this way, the paths of interest are not affected, but the auxiliary variables are nonetheless part of the model. Once they are part of the model, the influence of auxiliary variables on missingness is taken into account. If auxiliary variables are not used when FIML estimation is employed, the MAR process is not likely to be represented (unless the MAR-associated variables just happen to be part of the desired analysis model; Enders, 2008, 2010; Graham, 2003). Auxiliary variables are easier to deal with in MI. Here, auxiliary variables are just included in the imputation model, which typically contains the complete data set (Collins, et al., 2001; Schafer, 1997). In MI, there is no need to explicitly distinguish auxiliary variables from analysis variables—auxiliary variables can be any variables in the imputation model, whereas analysis variables are whichever variables are chosen, post-imputation, for use in the analysis. Once imputation is complete, there is no further need for auxiliary variables. Two general strategies for including auxiliary variables are recommended. The first is the inclusive strategy (Collins et al., 2001). Here, all variables on a data set that are not part of the analysis model are included in the analysis as auxiliary variables. This strategy ensures that if a MAR association is contained among the set of auxiliary variables, its impact will be represented (Collins et al., 2001; Kreuter & Olson, 2011; Yoo, 2009). The inclusive strategy will generally become intractable when too many variables are included, particularly because the auxiliary variables

themselves will often contain missing values (Enders, 2008; 2010). The large number of variables coupled with the variety of missing data patterns that occurs can lead to an insurmountable estimation problem—the software simply will not converge on a solution (Honaker & King, 2010). The restrictive strategy is an alternative approach that entails selecting variables that have an observable relationship with the missing data patterns, that is, those variables with a significant unique predictive association to be included as auxiliary variables (Enders, 2010). Van Buuren (2012) suggested including all variables that appear in the analysis model, the variables that are related to the variables having missing data, and the variables that are highly correlated with the target variables. In addition, including an incomplete auxiliary variable is beneficial, even if that variable contains MNAR missingness. However, variables that have too many missing data might introduce estimation problems, and should not be included in an imputation model. In general, the restrictive strategy tends to be less effective than the inclusive strategy (Collins et al., 2001). However, it is sometimes necessary to reduce convergence problems. Both the inclusive strategy and the restrictive strategy often ignore potential nonlinearity in the MAR association of the variables on the data set (Collins et al., 2001). To allow an effective use of the inclusive strategy and to encourage broader representation of the potential MAR process, Howard, Rhemtulla, and Little (2015) introduced the use of principal components scores as auxiliary variables. The principal components strategy relies on the characteristic of component scores as distilled sources of much of the variance contained in a set of variables. The basic strategy involves (1) doing a single imputation of the missing values using a stochastic regression imputation or EM imputation, (2) computing nonlinear variables including powered and interaction terms, and (3) reducing the resulting very large data set down to a handful of meaningful component scores. The first step uses precise predicted values as temporary substitutes for the values that are missing. The uncertainty issues for statistical inference are not a concern in this step because the predicted values are simply estimates for the calculation of the component scores. In the second step, when nonlinear terms are calculated, the strategy is to be very broad. For example, each variable could be squared and all possible two-way interaction terms among the set of variables could conceivably be calculated. Ongoing simulation work and comparative analysis on existing data sets indicates that using around 10 component scores is sufficient to capture the missing data relevant

Page 776

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Practical Considerations

information that is contained on many practical data sets (Garnier-Villarreal et al., 2013). These 10 component scores are included as the auxiliary variables in an FIML or MI analysis model. Fraction of Missing Information As suggested earlier, fraction of missing information (FMI) is probably the most useful value for quantifying the effect that missing data has on a statistical analysis. Allison (2002) gave a conceptual definition of FMI as “how much information is lost about each coefficient because of missing data” (p. 48). FMI can also be interpreted as the proportion of the total sampling variance of a focal parameter Qj that is attributable to the missing data (Rubin, 1987), in this sense it is much like an R2 statistic of the missing data (Enders, 2010). Alternatively, the FMI shows how much more variable a given parameter estimate is with missing data than it would have been had all of the data been fully observed, thereby conveying information about the statistical efficiency lost to nonresponse (Harel, 2007; Savalei & Rhemtulla, 2012). FMI has a fundamental place in any missing data analysis because it is intimately tied to many aspects of the missing data problem. The FMI directly impacts power loss due to missingness (Savalei & Rhemtulla, 2012), rates of convergence for missing data algorithms (Fraley, 1999; Harel, 2007; Schafer, 1997), and the number of imputations required for MI (Bodner, 2008; Graham et al., 2007). Although it seems natural to quantify the scope of a missing data problem by the raw percentage of cells in the data set that are missing (i.e., percent missing [PM]), FMI is actually the more appropriate metric from which to draw guidance (Harel, 2007). FMI can be viewed as showing the impact of the raw PM after accounting for the strength of the item interrelations in the observed data (Enders, 2010; Rubin, 1987). Consequently, FMI will often be lower than the simple PM because high variable correlations may mean that there is less missing information than missing data. On the other hand, particularly when the missing data mechanism is MAR, FMI can be much higher than the raw percentage of missing values (Rhemtulla, Savalei, & Little, 2014; Savalei & Rhemtulla, 2011). The formula for FMI is usually given in terms of the variance components of an MI analysis. Schafer and Graham (2002) gave the following formula for a sample estimate of the FMI for a focal model parameter Qj :

12:38pm

777

where Bj is the between-imputation variance of the focal parameter, U j is the average within-imputation variance for the focal parameter, and M is the number of imputations. With a large number of imputations the formula for FMI reduces to the approximation shown rightmost in equation 17.10 (Savalei & Rhemtulla, 2012). This approximation makes clear the conceptual parallel of FMI and the R2 statistic. Replacing Bj with between-subject variance and U j with within-subject variance gives the familiar R2 statistic from ANOVA. FMI is often discussed in the literature as a natural consequence of MI that is not easily derived from ML-based missing data tools (e.g., McKnight, McKnight, Sidani, & Figueredo, 2007; Schafer, 2001), but Savalei and Rhemtulla (2012) appealed to the missing information principle of Orchard and Woodbury (1972) to give a simple method for estimating FMI in the course of a FIML-based analysis. The computational formula given by Savalei and Rhemtulla (2012) is the following:

FMIj, FIML = 1 −

SEj,2 Comp SEj,2 FIML

.

(17.11)

where SEj,2 Comp is the squared standard error of the focal ̂ parameter Q j, Comp derived from complete-data ML estima2 tion and SEj, FIML is the squared standard error of the focal ̂ parameter Q derived from a FIML analysis. j, FIML

The natural stumbling block of this approach is the ̂ apparent inability to estimate Q j, Comp from incomplete data, but an asymptotic approximation is readily available. The method described by Savalei and Rhemtulla (2012) involves breaking the analysis into two phases. Phase 1 ̂ consists of estimating Q j, FIML by applying FIML to the incomplete data. Phase 2 consists of using ordinary ML estimation to fit an identical model to the model implied mean vector and covariance matrix produced by Phase 1. Because FIML produces model implied moments that are asymptotically equal to their complete-data analogues, the estimates derived from the Phase 2 model are asymptotically equal to the complete data ML estimates 2 2 ̂ Q j, Comp . Furthermore, because SEj, Comp and SEj, FIML are ̂ ̂ natural byproducts of estimating Q and Q , j, Comp

j, FIML

these two models produce all the necessary components of equation 17.11, and FMI is easily calculated. Assessing Model Fit with Missing Data

FMIj, MI =

(1 +

M −1 )Bj

U j + (1 + M −1 )Bj

≈

Bj U j + Bj

,

(17.10)

In many popular modeling paradigms assessing how well an estimated model recreates the relationships observed in

Page 777

Trim Size: 8.5in x 11in

778

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

the sample data (i.e., the model fit) is an integral component of the research process (Davey, Savla, & Luo, 2005). Latent variable modeling techniques, in particular, have been developed hand-in-hand with ever more sophisticated model fit indices. Consequently, the consumers of research that employs latent variable models have come to expect a complement of fit indices to be reported alongside parameter estimates and test statistics. The ability to accurately assess model fit, therefore, is of paramount importance to any researcher seeking to employ a latent variable model. Unfortunately, missing data complicate this issue. As with any other statistic we may wish to estimate, fit indices were developed under the assumption of completely observed data, and special consideration must be given to their calculation when they are derived from models fit to incomplete data (Davey et al., 2005; Savalei, 2010b). The assessment of model fit is problematic whether employing FIML or MI, but the issues inherent in the two approaches are unique, so we will address them separately below. If using FIML estimation, one of the biggest issues arises from the unique form of the likelihood function that FIML maximizes to derive its parameter estimates. Traditional, complete-data, ML estimation is customarily accomplished by minimizing the discrepancy between the observed data moments and the model implied moments via the discrepancy function given by Joreskog (1967): ( ) FML = ln |Σ| + tr SΣ−1 − ln |S| − p

(17.12)

where Σ is the model implied covariance matrix, S is the covariance matrix of the observed data, p is the total number of manifest variables, |⋅| represents the determinant of a matrix, and tr(.) is the trace of a matrix. Once FML is calculated, it can be used to compute the likelihood ratio (chi-squared) statistic in the following way: 2 𝜒ML = (N − 1)FML ,

(17.13)

where FML is the minimized discrepancy function from equation 17.12, and N is the total sample size. However, as described previously, FIML maximizes the joint case-wise loglikelihoods given by equation 17.2. Unfortunately, the FIML 𝜒 2 cannot be calculated by simply replacing FML in equation 17.13 with 𝓁FIML from equation 17.2 (even though they are analogous quantities; Enders, 2001c). Fortunately, the FIML 𝜒 2 can be computed without a great deal of additional effort on the part of the researcher. Rather than naively replacing terms in equation 17.13, the FIML 𝜒 2 is calculated by taking the difference between

the maximized loglikelihood functions associated with the researcher’s hypothesized model and a fully saturated model: [ ) )] ( ( 2 = −2 𝓁FIML 𝝁1 , 𝚺1 |Y − 𝓁FIML 𝝁0 , 𝚺0 |Y 𝜒FIML (17.14) ) ( where 𝓁FIML 𝝁1 , 𝚺1 |Y represents the loglikelihood function from equation 17.2 fit to the mean vector 𝝁1 and covariance matrix 𝚺1 implied by the ) researcher’s hypoth( esized model, and 𝓁FIML 𝝁0 , 𝚺0 |Y represents this same loglikelihood function fit to a fully saturated mean vector 𝝁0 and covariance matrix 𝚺0 . The fit of the hypothesized 2 model can then be assessed by comparing 𝜒FIML with a 2 central 𝜒 distribution with degrees of freedom equal to the difference in the number of free parameters in the hypothesized and fully saturated models. Model fit can also be judged by computing derived fit indices (e.g., CFI, 2 2 TLI, RMSEA) by substituting 𝜒FIML for 𝜒ML in their standard formulae (Wu, West, & Taylor, 2009). The formulation of equation 17.14 may not map intuitively back to the two step procedure illustrated by equations 17.12 and 17.13, but the underlying rationale ) ( becomes more clear if we consider 𝓁FIML 𝝁1 , 𝚺1 |Y to carry roughly the same information as the model-implied ) of equation 17.12 and we think of 𝓁FIML (components 𝝁0 , 𝚺0 |Y as carrying the same information as the observed data components of equation 17.12. An additional concern arises when using incremental fit indices (i.e., CFI, TLI) to assess the fit of FIML-based models that employ the saturated correlates approach to include auxiliary variables (Enders, 2010; Graham, 2003). The saturated correlates approach necessarily adds extraneous variables to the model, but the software may not recognize the unique status of the new variables. Therefore, relying on the default independence model when calculating incremental fit indices will lead to incorrectly calculated degrees of freedom and fit indices that suggest better fit than they should. Interested readers can find a more detailed discussion of this issue and a description of the procedure for correct specification of the null model in Enders (2010, pp. 137–140). Although this issue still holds, in general, several software packages have implemented the ability to compute the correct null model when employing the saturated correlates approach. If the auxiliary variables are correctly entered into the model, the Mplus software (Muthén & Muthén, 2012) will calculate the correct null model (Asparouhov & Muthén, 2008). By using the R package semTools (Pornprasertmanit, Miller, Schoemann, & Rosseel, 2013) to introduce the auxiliary variables, the incremental fit indices from models fit with the R package

Page 778

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Practical Considerations

lavaan (Rosseel, 2012) are also calculated from the correct null model. The biggest issue facing researchers who wish to assess model fit when employing MI is not calculating the 𝜒 2 statistic, but correctly pooling the inevitable replicates of the 𝜒 2 . As with any implementation of MI, the analysis phase of an MI-based latent variable modeling enterprise will result in M distinct replicates of the 𝜒 2 statistic. Unfortunately, Rubin’s rules were not created to accommodate latent variable modeling, and they cannot be used to pool the 𝜒 2 . Therefore, as one would expect, the fit indices that are derived from the 𝜒 2 statistic (e.g., RMSEA, CFI, TLI) cannot be pooled using Rubin’s Rules either. The issue of pooling MI-based fit indices has been explored in the statistical and methodological literature at length, but a widely applicable and easily implemented solution has yet to be proposed. Unfortunately, simply averaging the M 𝜒 2 statistics is an untenable solution that will lead to a pooled statistic that is much too large and will reject the hypothesized model too often (i.e., the pooled statistic will have a highly inflated type I Error rate; Asparouhov & Muthén, 2010a; Lang, 2013). One successful, albeit somewhat tedious, approach that has been implemented in several popular software packages was developed by Meng and Rubin (1992). Their approach is a multiple step procedure that centers on calculating a 𝜒 2 statistic of the form given by Equation 17.14 (i.e., the difference in the minimized likelihoods of a full and restricted model). The first step involves aggregating the M 𝜒 2 statistics and parameter estimates produced by the analysis phase of an ordinary implementation of MI: M 1 ∑ 2 𝜒2 = 𝜒 , M m=1 m

(17.15)

1 ∑̂ Q , M m=1 0m

(17.16)

1 ∑̂ Q Q1 = M m=1 1m,

(17.17)

M

Q0 =

M

̂ , and Q ̂ are the estimated 𝜒 2 statistic, null where 𝜒m2 , Q om 1m model (fully saturated) parameter estimates, and hypothesized model parameter estimates, respectively, derived from the m-th imputed data set, and M is the number of imputations. The next step is to calculate an updated 𝜒 2 statistic for each imputed data set by fixing the null and hypothesized parameter values at Q0 and Q1 , respectively, and evaluating

12:38pm

779

a special case of equation 17.14 within each of the imputed data sets: 𝜒̃m2 = −2[𝓁ML (Q1 |Y (m) ) − 𝓁ML (Q0 |Y (m) )],

(17.18)

where 𝓁ML (Q0 |Y (m) ) represents the ordinary, completedata, loglikelihood function fit to the m-th imputed data set and the aggregated point estimates given by equation 17.16, and 𝓁ML (Q1 |Y (m) ) represents this same loglikelihood function fit to the same imputed data set and the aggregated point estimates given by equation 17.17. Next, the m estimates of 𝜒̃m2 are pooled by simply taking their average across the imputed data sets: ∑̃ ̃2 = 1 𝜒2 , 𝜒 M m=1 m M

(17.19)

where 𝜒̃m2 are the updated 𝜒 2 statistics from equation 17.18, and M is the number of imputations. Finally, the adjusted 𝜒 2 statistic can be calculated as follows: 2 𝜒imp =

̃2 𝜒 , 1 + r3

(17.20)

̃2 is the pooled 𝜒 2 statistic from equation 17.19, and where 𝜒 r3 is a correction factor given by r3 =

M +1 ̃2 ), (𝜒 2 − 𝜒 (M − 1)(p1 − p0 )

(17.21)

̃2 𝜒 2 is the aggregate 𝜒 2 statistic from equation 17.15, 𝜒 2 is the pooled 𝜒 statistic from equation 17.19, p1 and p0 are the number of freely estimated parameters from the hypothesized and null models, respectively, and M is the number of imputations. 2 statistic has been shown to perform well when The 𝜒imp 2 statistic given by equation 17.14 compared to the 𝜒FIML when the amount of missing data is low or the number of imputations is large (Asparouhov & Muthén, 2010a). 2 Though the process of calculating 𝜒imp manually can be onerous, this approach is well implemented in the Mplus software package (Muthén & Muthén, 2012) as the default method of pooling 𝜒 2 statistics when fitting ML-based, single-level structural equation models. It is also available in the R package semTools (Pornprasertmanit et al., 2013) as an option for pooling 𝜒 2 statistics from structural equation models estimated with the R package lavaan (Rosseel, 2012).

Page 779

Trim Size: 8.5in x 11in

780

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

The question of how to best assess model fit with missing data has been a very active area of missing data research over the course of the last 20 years, and many interesting applications have been developed whose discussion falls outside the scope of this chapter (e.g., the application of parametric bootstrapping to the assessment of model fit by Enders, 2002; Savalei & Yuan, 2009). One such family of techniques still warrants a brief introduction, though. These are the so-called two-stage estimators that are based on the EM algorithm. The two-stage estimation approach was developed by Yuan and Bentler (2000) as a way to carry out ML-based missing data analysis under difficult situations, such as under nonnormal data distributions. Before Graham’s (2003) saturated correlates approach was developed and implemented, a two-stage approach was the only way to get auxiliary variables into an FIML analysis. It has remained a fruitful starting point for several recent advances in assessing model fit with missing data. For example, Cai and Lee (2009) adapt the two-stage estimation approach to produce a correctly distributed 𝜒 2 statistic by incorporating the Meng and Rubin (1991) supplemental EM algorithm, and Savalei and Bentler (2009) derived their own correction to produce accurate standard errors and 𝜒 2 statistics, which they show to perform comparably to FIML estimates. Savalei and Falk (2014) showed that this method can perform better than FIML when data are nonnormally distributed. One downside of approaches like Meng and Rubin’s (1992) technique and the corrected versions of the Yuan and Bentler (2000) two-stage estimator is that they tend to be difficult to implement without good software integration. To mitigate this difficulty certain authors have explored the possibility of pooling the imputed data structures prior to the analysis phase of MI. Yet, all such studies have found that the 𝜒 2 statistics derived from these pooled data structures required a post hoc correction before they followed a 𝜒 2 distribution (Lang, 2013; Lee & Cai, 2012). Lee and Cai (2012) derived such a correction by adapting Proposition 4 from Browne (1984) to produce a test statistic TMB that asymptotically follows a central 𝜒 2 distribution. They also provided a SAS Macro that can be used to implement their approach software. Lang and Little (2014) proposed an approach that was designed to be easily implemented with any statistical analysis software without relying on specialized code. Their approach (which we dub the supermatrix technique) is essentially an expansion of the initial aggregation phase of the Lee and Cai (2012) two-stage approach. Using a Monte Carlo simulation study they provided empirical evidence to suggest that accurate hypothesis tests can be

accomplished by appealing to nested-model 𝜒 2 difference tests without the need for post hoc correction to the component 𝜒 2 statistics (even though the component 𝜒 2 statistics are biased measures of model fit when considered individually). To illustrate how the supermatrix technique is implemented, consider a hypothetical missing data analysis in which you begin with an incomplete N × p data set. Application of the supermatrix technique can then be broken into four basic steps: 1. Create a set of M multiple imputations using any MI algorithm. 2. Stack the M imputed data sets into an aggregate MN × p data frame. 3. Calculate a single p × p covariance matrix from the stacked data from step 2. 4. Use the covariance matrix from step 3 and the original sample size N as the sufficient statistics and sample size, respectively, for subsequent model fitting. Figure 17.2 gives a graphical summary of the supermatrix approach for a hypothetical example in which the original incomplete data consists of four observations of three variables. The results of a Monte Carlo simulation study conducted by Lang (2013) suggested that fit indices derived from structural equation models fit using the supermatrix as sufficient statistics will have inflated type I Error rates (i.e., they will overreject well-fitting models) and should not be used to assess overall model fit. However, simulation results from Lang and Little (2014) showed very good performance for nested-model Δ𝜒 2 statistics computed from parent and child models that are both fit to a supermatrix. When Δ𝜒 2 statistics were calculated from parent and child models fit to supermatrices, Lang and Little (2014) found that these Δ𝜒 2 statistics were nearly identical to analogous Δ𝜒 2 statistics calculated from parent and child models fit to fully observed data. Mediation Analysis Mediation is often of central interest to researchers in social and behavioral science research. Mediation is said to occur when the effect of an independent variable (X) on a dependent variable (Y) is transmitted by a third variable (M) (Baron & Kenny, 1986; Sobel, 1982). M is termed a mediator as it mediates the effect of X on Y. A mediating effect (also termed an indirect effect) can be quantified by and tested through the product of the two

Page 780

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Practical Considerations

X

Y

Z

x1

•

•

•

y2

z2

x3

y3

•

x4

•

X1

Y1

x1 y1

1) Create M imputed data sets

(1)

x2(1) y2 x3

z4

Z1 z1(1) z2

y3 z3(1)

x4 y4(1) z4 X x1

Figure 17.2

Y

X2

Y2

Z2

XM YM

x1 y1(2) z1(2) x2(2) y2 x3

y3

z2

12:38pm

781

ZM

x1 y1(M) z1(M) •

•

•

z3(2)

x2(M) y2 x2

x4 y4(2) z4

z2

y2 z2(M)

x4 y4(M) z4

Z

y1(1) z1(1)

x2(1)

y2

z2

x3

y3

z3(1)

x4

y4(1)

z4

x1

y1(2) z1(2)

x2(2)

y2

z2

x3

y3

z3(2)

x4

y4

• • •

x1

• • • y1(M)

• • • z1(M)

x2(M)

y2

z2

x3

y3

z3(M)

x4

y4(M)

z4

(2)

2) Stack all M imputed data sets into a single data frame

z4

X 3) Compute a single covariance matrix from the aggregate data frame

Y

X

σ2x

Y

σY,X

σ2Y

Z

σZ,X

σZ,Y

Z

σ2Z

A graphical representation of the supermatrix technique as applied to a data set with four observations and three variables.

Source: From Lang and Little (2013). Used with permission.

direct effects (denoted by ab): (1) the effect of X on M (a effect); and (2) the effect of M on Y controlling for X (b effect). A well-known strategy to test ab is to empirically establish the sampling distribution of ab through nonparametric bootstrapping and then use the confidence interval computed from the sampling distribution to test the null hypothesis that ab = 0 (MacKinnon, Lockwood, & Williams, 2004). A nonparametric bootstrapping procedure starts with drawing a large number of samples with replacement (e.g., 1,000) with the same sample size of the original data, assuming that the original sample is the population (Efron, 1979; Efron & Tibshirani, 1994). A mediation model is then fit to each of the bootstrap samples, resulting in a large number of point estimates of ab, which form the empirical sampling distribution of ab. When there are missing data, nonparametric bootstrapping can be combined with FIML or MI to test mediation effect (Wu & Jia, 2013; Zhang & Wang, 2013). To combine nonparametric bootstrapping with FIML, one needs to bootstrap first and then run the target analysis using FIML for each bootstrapped sample (denoted by BOOT[FIML]). Many software packages such as Mplus and EQS have automated this procedure. To combine bootstrapping with

MI, two procedures have been proposed. One is to bootstrap first and then run MI for each bootstrap sample. This procedure is computationally intensive because it requires running MI for each bootstrap sample. In other words, if there are 1,000 bootstrapped samples, the MI needs to be run 1,000 times. This large number of runs would impose a tremendous amount of computational burden. To mitigate this problem, Wu and Jia (2013) suggested running MI first to obtain a number of imputed data sets and then running the bootstrap procedure for each imputed data set (denoted by MI[BOOT]). Because MI(BOOT) requires only 1 run of MI, it is more computationally efficient than BOOT(MI). Wu and Jia (2013) showed that MI(BOOT) performed comparably to BOOT(FIML) when data are normally distributed and missing data are ignorable. However, one should be cautious about using either procedure to test mediation as MNAR, nonnormality, or both can distort the point and interval estimates of a nonzero ab for both procedures. In addition to nonparametric bootstrapping, Bayesian estimation methods are well suited to mediation analysis and can be readily extended to treat missing data (Biesanz, Falk, & Savalei, 2010; Enders, Fairchild, & MacKinnon,

Page 781

Trim Size: 8.5in x 11in

782

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

2013; Yuan & MacKinnon, 2009). The Bayesian method generates a posterior distribution of ab (i.e., a distribution of ab conditional on the observed data) based on which the statistical inference of ab can be made. However, past research on Bayesian approach is limited to normal data. More studies need to be conducted to examine the performance of the Bayesian approach when the normality assumption is violated and to compare it with the other strategies. When to Use FIML Versus MI FIML and MI provide comparable point and standard error estimates given that (1) the multivariate normality assumption is satisfied, (2) the missing data are ignorable (i.e., MCAR and MAR), (3) missing data model is the same, and (4) the number of imputations in MI approaches infinity (Collins, Schafer, & Kam, 2001; Graham et al., 2007; Little & Rubin, 2002; Schafer, 1997; Schafer & Graham, 2002). However, they have their own pros and cons in practice. Some of these have been mentioned in the previous sections. Here we provide a more detailed comparison of the two techniques. There are many reasons researchers might prefer FIML. First, FIML is more efficient than MI. The standard error estimates from MI will generally be larger than those from FIML although the difference is trivial when the number of imputations is large. Second, FIML is also easier to implement because it is a one-step approach and available in most of the commercial software packages (e.g., Mplus, LISREL, EQS, AMOS, SAS PROC CALIS, and lavaan in R). In contrast, MI involves multiple steps that have not been fully automated. In addition, implementation of MI requires many decisions. As described before, researchers might need to decide on an appropriate number of imputations, the correct imputation model (with the relationships among the variables correctly specified), and an appropriate number of burn-in and between-imputation iterations. Finally, it is also easier to obtain model fit indices with FIML as described previously. However, there are benefits of using MI. Certain imputation algorithms and imputation models might provide a better treatment of categorical variables and nonparametric relationships among the variables (Asparouhov & Muthén, 2010a; White, Royston, & Wood, 2011). In addition, MI creates complete data sets. As a result, statistical methods that are only implemented for use with complete data can be applied. Under certain circumstances, MI has to be adopted (Enders, 2010; Gottschall, West, & Enders, 2012). One circumstance occurs when there are item-level missing data. To illustrate, suppose X, Y, and M in a

mediation model are scale scores that are computed by averaging multiple item scores. If some of the items have missing data, the best way to handle the missing data would be imputing them at the item level and then calculating the scale scores from the imputed item scores (Gottschall, West, & Enders, 2012). By doing this item-level imputation, the observed items’ scores are used to recover the missing information. FIML is inferior in this case as it relies on the analysis model in which the item scores are not directly included, resulting in underuse of the available information. Another circumstance in which MI may work better is when the model contains categorical dependent variables. Though FIML can be used with categorical data, when there are more than two categories or the model is complicated, this becomes extremely computationally intensive. Instead, the most common estimation technique is to use a diagonally weighted least squares estimator on the matrix of polychoric correlations (this is called “WLSMV” in Mplus, where it is the default estimator for categorical variables). Because least squares estimators cannot handle missing data efficiently, MI allows the researcher to deal with the missingness first (e.g., by using MICE to impute categorical variables) and use the least-squares estimator to analyze the model on the imputed data sets (Asparouhov & Muthén, 2010a, 2010b). REVIEW OF MISSING DATA PRACTICES IN PSYCHOLOGICAL RESEARCH In his introduction to the special section of Psychological Methods on missing data analysis, West (2001) noted that a great many methodological advances had been made in the treatment of missing data analysis in the preceding 25 years. In an interview shortly thereafter, West predicts that increasing implementation of principled missing data tools will be hallmark of psychological research in the early twenty-first century (Azar, 2002). Many of the advances that West (2001) was referencing had, by the turn of the twenty-first century, yet to make their way into practice in the applied social science literature. The American Psychological Association was clearly cognizant of the need for improved treatment of missing data. This concern is evident when reviewing the recommendations of the Task Force on Statistical Inference published in 1999. One of the key recommendations of this task force was that authors publishing in psychology journals should accurately report on the missingness in their studies so that its impact on the conclusions can be judged transparently (Wilkinson & Task Force on Statistical Inference, 1999). The committee also went on to advise

Page 782

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Review of Missing Data Practices in Psychological Research

that researchers adopt optimal missing data treatments and eschew antiquated ad hoc approaches. They condemned the use of deletion methods, in particular, by stating, “The two popular methods for dealing with missing data that are found in basic statistics packages—listwise and pairwise deletion of missing values—are among the worst methods available for practical applications” (p. 598). So how prophetic was West’s claim? Have there been considerable improvements in missing data practices since 2000? Unfortunately, even in the shadow of an ever growing monolith of evidence that ad hoc missing data tools are universally inferior to modern principled techniques, psychological researchers have been relatively slow to adopt more sophisticated missing data tools. Reviews of the missing data practices employed in psychological studies published in 1999, 2003, and 2012 have all found that ad hoc techniques, and deletion based techniques in particular, remain the most prevalent treatment for missing data (Bodner, 2006; Little et al., 2014; Peugh & Enders, 2004). However, all hope is not lost, because there are signs that the field is on a trajectory of improvement. Recognizing that the aforementioned recommendations of the Task Force on Statistical Inference represented a turning point for missing data analysis in the social sciences, Bodner (2006) and Peugh and Enders (2004) reviewed the missing data practices reported in samples of journal articles published in 1999. These studies’ findings were convergent and bleak. Perhaps most troubling was the very high frequency with which missing data were not reported at all (i.e., the presence of missingness often had to be inferred from discrepancies in the degrees of freedom reported for test statistics). Additionally, for those cases in which the missing data treatment could be elucidated, a principled technique was never used, and deletion-based approaches were, by far, the most frequently applied treatments. To judge the improvement in missing data analysis in the years following the recommendations of the Task Force on Statistical Inference, Peugh & Enders (2004) also reviewed the missing data practices reported in a sample of journal articles published in 2003. They found a discernible improvement in missing data practices, especially in terms of reporting practices. This second review found much more accurate reporting of missing data (i.e., a higher proportion of authors explicitly acknowledged the presence of missingness in their studies). They also note a higher frequency of authors explicitly discussing the assumptions underlying the missing data analysis (e.g., actively testing the MCAR assumption). Unfortunately, ad hoc approaches were still the norm, and deletion

12:38pm

783

techniques remained the most common missing data treatment. However, they did record several instances of principled techniques being employed, whereas there had not been any use of principled approaches in their review of journal articles from 1999.3 Finally, Little et al. (2014) conducted a comprehensive review of the missing data practices reported in empirical studies from the 2012 volume of the Journal of Pediatric Psychology and found further evidence for a trajectory of improvement in missing data practices. The improvement in reporting practices vis-à-vis those found by Bodner (2006) and Peugh and Enders (2004) was considerable. The majority of the articles examined transparently acknowledged the presence of missingness in their studies and explicitly described their treatment thereof. There was also a considerable increase in the usage of principled missing data tools (especially FIML) relative to the rates reported by Bodner (2006) and Peugh and Enders (2004). A sizable minority of the articles examined employed FIML (or some variant of FIML) to treat the missingness in their studies. Unfortunately, ad hoc approaches were still the most frequently employed techniques, and several papers mixed ad hoc approaches with principled techniques (e.g., listwise deletion of attrition and imputation of the remaining arbitrary missingness). An interesting finding of the Little et al. (2014) review was that every study that employed FIML either fit a structural equation model using Mplus (Muthén & Muthén, 2012) or it fit a multilevel model. Because FIML estimation is the default option when analyzing incomplete data in Mplus, and multilevel modeling relies on FIML (or some specialized variant of FIML) to accommodate the inherently unbalanced data for which it is designed, the 3

Principled techniques are defined as any missing data technique that treats the nonresponse by considering the underlying distribution of the missing data and either fills in these missing values via predictive models constructed to produce predicted values that conform to that underlying distribution (in the case of MI) or partition the missingness out of the likelihood function by appealing to the Rubin (1976) partitioning of the complete data distribution, which is only possible when explicitly considering the distribution of the missing values (as in EM and FIML). These principled techniques are usually considered in contrast to ad hoc approaches that are so called because they are simple one-off convenience solutions to the missing data that basically ignore any missing data that ever existed. In the strictest sense principled and ad hoc are not mutually exclusive. You can have a principled ad hoc approach if it is built up from initial consideration of the distribution of the missing values, but it is specialized to a specific problem and not generalizable to any other analysis.

Page 783

Trim Size: 8.5in x 11in

784

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

greater usage of principled techniques observed by Little et al. (2014) may not represent a meaningful improvement in practice. Although, this is a pessimistic outlook, the fact that all of the reported FIML analyses were conducted by naively employing a restrictive strategy (i.e., they did not include auxiliary variables in the FIML model) lends credibility to the suggestion. Judging by the patterns of missing data practices observed in the reviews presented by Bodner (2006), Little, et al. (2014), and Peugh and Enders (2004) it is apparent that missing data are becoming a more active consideration for researchers as time goes on. Similar findings are reported by Wood, White, and Thompson (2004), who employed similar methodology to analyzed the missing data practices reported in a sample of studies published in the 2001 volume of several top-tier medical journals. However, the findings of Little et al. (2014) suggest that many applied researchers in psychology are still inadequately informed when it comes to optimal missing data treatments, and many decisions in a missing data analysis are guided by software defaults. While optimal missing data tools are becoming better integrated into popular software every year (Enders, 2010), the findings discussed above suggest that there is still a need for education on how to best implement these tools as they become available.

WHY CHANGE? Too often, entreaties to employ modern approaches to address missing data or to incorporate planned missingness as a design feature are met with an unfounded fear or unwarranted hubris. As detailed in Table 17.1, the traditional methods of handling missing data are rife with error and bias. Maintaining a staunch commitment to the practices of old is an untenable position to take (Wilkinson & Task Force on Statistical Inference, 1999). The only time that some of these practices are warranted is when the missing data are MCAR and power is not an issue (Enders, 2010; Little & Rubin, 2002), or when the amount of missing data is very small (e.g., less than 5%; Graham, 2009). Otherwise, invalid conclusions will result. Fundamentally, failing to address the bias created by missing data represents a miscarriage of social justice. Because results are biased when traditional approaches are employed, policy recommendations based on the analyses of faulty data lead to failures in providing accurate guidance for stakeholders and decision makers. Efforts to simplify analyses by shortcutting the appropriate ways of addressing missing data are unconscionable. Baseless fears

or unjustified hubris are not legitimate excuses to avoid appropriately treating the missing data. Rather than ignoring the issue in the abject desire that the bad consequences of missing data won’t be too bad, we should aggressively address the mechanisms of missing data to minimize their undermining influences (Wilkinson & Task Force on Statistical Inference, 1999). In this vein, thoroughly examining and reporting on potential predictors of missingness would benefit future research because investigators could identify potential predictors of missingness to be included into a data-collection protocol. Such information could be added to the supplemental materials for a published paper and posted on a reliable web page. Such virtual repositories would provide a wealth of information for future study planning.

PLANNED MISSING DATA DESIGNS The preceding discussion of missing data has centered on the unplanned missingness that inevitably occurs in human science research. We now turn our attention to planned missing data designs and the recent research that is still emerging in this area (e.g., Graham, Hofer, & MacKinnon, 1996; Graham, Taylor, & Cumsille, 2001; Graham, Taylor, Olchowski, Cumsille, 2006; Jia et al., 2014; Jorgensen et al., 2014; Rhemtulla, Jia, Wu, & Little, 2014, Rhemtulla, Savalei, et al., 2014). Planned missing data designs have been around for some time but maybe not obviously recognized as such. The Solomon four-square design, for example, is a planned missing data design (see Table 17.4, Groups 1–4; see also Iacobucci, 1995). This design is often used with classical missing data handling procedures. As such, researchers have not taken full advantage of it. If classical analyses are used, it is a very underpowered design. If modern missing data handling procedures are used, on the other hand, it becomes a very powerful pre–post TABLE 17.4 The Solomon Four-Group Design as a Planned Missing Design with the Addition of Two Additional Groups for a Complete Planned Missing Design Group 1 2 3 4 5 6 1O

Pretest1

Treatment/intervention

Posttest1

O1 O3 Planned missing Planned missing O7 O8

X

O2 O4 O5 O6 Planned missing Planned missing

X X

reflects the observed variables, including all demographics and outcome scores and X represents a treatment or intervention that has been implemented. The subscripts differentiate the different groups at the different testing occasions.

Page 784

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Planned Missing Data Designs

experimental design. The participants who do not receive the pretest scores are randomly assigned to be in Group 3 and Group 4 (Table 17.4). These missing pretest scores are imputed if a modern missing data approach is used, thereby increasing the power of the overall design. A fully planned missing data design might also add two additional groups (Groups 5 & 6 in Table 17.4) to allow for more power at less cost. Planned missing data designs leverage the capabilities of the modern missing data handling procedures to minimize the loss in power that traditional approaches suffer. Planned missing data designs also are used because they can substantially reduce costs and, in many cases, increase validity by reducing participant burden/fatigue and minimize other threats to validity, such as test reactivity and unplanned missing data (Harel, Stratton, & Aseltine, 2011; Popham, 1993). Planned missing data designs are most useful in the context of large sample research involving latent variable SEM techniques that extend easily to longitudinal designs (Little, 2013). In the following, therefore, we will assume that multiple indicators of constructs are generally used, reasonably large sample sizes are available (e.g., > 140), and that latent variable SEM or one of its variants will be used (see Little, 2013, for a discussion of the various options). This latter assumption is not essential for using all of the planned missing designs we will discuss but it is for some of them. Harel et al. (2011) investigated whether planned missingness can be used to mitigate test reactivity in a pre-post control group experimental design. In a suicide prevention assessment trial, students received either all or a subset of items from the Signs of Suicide 17-item questionnaire at pretest and the complete questionnaire 3 months postintervention. Three notable findings arose: first, students who were given more items to complete were less likely to complete all of them: 27% of those who were given all 17 items failed to answer all of them, compared to 14% who were given just 6 items and 18% who were given 8 items. Second, students who were given the full questionnaire at baseline were three times more likely to skip the posttest assessment entirely (21% compared with 7% for those who received a truncated questionnaire at baseline). Third, of those students who attempted the posttest questionnaire, those who received the full questionnaire at baseline were again less likely to complete all items at posttest (25% of these students skipped items compared with 15% of students who received a truncated questionnaire at baseline), despite the fact that all students received the full 17-item questionnaire at posttest. These results suggest that planned missing designs can significantly reduce rates

12:38pm

785

of unplanned missing data—both item nonresponse and attrition—by substantially reducing participant burden. In addition to reducing the rates of unplanned missingness, Harel et al. (2011) showed that the planned missing strategy alleviated the effects of test reactivity. Test reactivity occurs when persons assigned to the control group get better simply by virtue of being tested with protocols that clearly indicate the purpose of a study. With the planned missing protocol used by Harel et al., those participants who received only some of the signs of suicide items did not react to the protocol, whereas those who received the whole questionnaire showed changes at post-test that obscured the effect of the intervention. The Harel et al. (2011) study is a unique application of a planned missing data design that highlights the utility of such designs. In the following, we will discuss more standard applications of planned missing data designs and extend them to the case of longitudinal data collections: (1) multiform designs, which focus on shortening the questionnaire protocol; (2) the two-method design, which focuses on leveraging high-quality measures to increase power; and (3) wave-missing designs, which focus on reducing the costs of longitudinal data collection. Multiform Planned Missing Protocols The simplest multiform example is the three-form planned missing data design, an instance of which is presented in Table 17.5. Multiform designs have been described in various literatures under various names and with various analysis models (e.g., Bunting & Adamson, 2000; Bunting, Adamson, & Mulhall, 2002; Dings, Childs, & Kingston, 2002; Gonzalez & Eltinge, 2008; Mislevy, Beaton, Kaplan, & Sheehan, 1992; Raghunathan & Grizzle, 1995; Shoemaker, 1973; Wacholder, Carroll, Pee, & Gail, 1994). As seen in the example of a three-forms design in Table 17.5, items are divided into four different item sets and then assembled to create three questionnaire protocol forms. Participants are randomly assigned to fill out one of the three forms. The items that are not included in a form are thereby missing completely at random (i.e., MCAR) because the investigator controls, through random assignment, which variables are missing for each person. TABLE 17.5 Example of a Three-Form Planned Missing Design Form Common Set X Variable Set A Variable Set B Variable Set C 1 2 3

Complete Complete Complete

Complete Complete Missing

Complete Missing Complete

Missing Complete Complete

Page 785

Trim Size: 8.5in x 11in

786

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

Multiform questionnaire protocols are a way to create short forms that fit within a specified time frame, to reduce overall costs, and to minimize respondent burden and fatigue. Each participant receives a questionnaire protocol that can be substantially shorter than the full set of protocol questions. The benefit of the multiform design is that it leverages the capabilities of the modern missing data treatments to maintain the information (both reliability and validity) that is inherent in the full set of questions. Moreover, the planned missing protocol has the potential to increase validity by reducing burden, fatigue, unplanned missing, and test reactivity (e.g., Harel et al., 2011) as well as practice and retest effects with longitudinal data (Beglinger et al., 2005; Jorgensen et al., 2013; Salthouse & Tucker-Drob, 2008). These many benefits notwithstanding, planned missing protocols can be less efficient than complete case protocols. Recall that efficiency, which is related to power, reflects the size of the standard error for a given parameter estimate. The power loss associated with the planned missing protocol can be offset by an increasing the sample size to produce the same efficiency as the design would have had with complete data. The extent to which parameter efficiency suffers depends on the pattern of missing data, the parameter type, and the strength of correlations between variables (Garnier-Villaréal et al., in press; Rhemtulla, Jia, et al., 2014; Rhemtulla, Savalei, et al., 2014). One salient recommendation that has come out of our recent research on optimal planned missing data designs is that efficiency is maximized when groups of highly correlated items (e.g., items from the same scale) are split across item sets, such that every participant answers a proportion of the items on every scale. Intuitively, this makes sense—once a participant has responded to four items on a six-item conscientiousness scale, four items on a six-item neuroticism scale, and four items on a six-item extraversion scale, the incremental benefit of responding to another two items on each of those scales is small. In contrast, if a participant responds to all six items on the conscientiousness scale and the neuroticism scale, and zero items on the extraversion scale, even though in both cases the participant has answered 12/18 questions, more information is missing when a whole scale is absent. In the first case, we have a pretty good estimate of that participant’s personality along all three dimensions; in the second case, we have a slightly better estimate of personality along two dimensions but no estimate at all on the third. As such, planned missing designs where scales are split across sets result in much more powerful analyses.

Similarly, having indicators of constructs in the X block minimizes the reduction in efficiency for many parameters. As with most complex modeling endeavors, we recommend careful planning and preliminary simulation runs to identify any potential weakness of a planned missing data design. Even if poorly constructed, however, a planned missing data design brings more benefits than costs when compared with the business as usual approach that permeates current literature. As with many statistical techniques, the details of how to create a multiform design can be more cumbersome than a first glance may suggest. A multiform design is no different. A number of important issues need to be considered when using a multiform design in general. We will discuss these issues in the context of the three-form design. By way of example, we will walk through the creation of three forms of the Inventory of the Forms and Functions of Aggression (iFFA) survey (Little, Jones, Henrich, & Hawley, 2003). The iFFA measures six constructs of aggression each with six items: Pure Overt, Instrumental-Overt, Reactive-Overt, Pure Relational, Instrumental-Relational, and Reactive-Relational. The iFFA is a self-report battery that is age appropriate from elementary age to late adolescence. For this hypothetical instrument, we will also collect demographic information and an important screening item for suicide ideation—“During the past 3 months, did you ever seriously consider attempting suicide (yes or no)?”—as well as self-reports reports on truancy, detentions, and visits to the principal’s office. As shown in Table 17.5, the three-form design consists of an X block of items, which is administered to all participants. In creating a three-form protocol from these variables, all singly indicated constructs (i.e., constructs that are measured by just one item, including truancy frequency, detention frequency, and number of principal visits), demographic variables (age, gender, race/ethnicity), and any critical screening variables (suicide ideation) are assigned to the X block. For the latent constructs with multiple indicators, the items for each construct would be assigned based on the item characteristics. Indicators of constructs can be classified as (1) parallel, (2) tau-equivalent, or (3) congeneric. Parallel items are psychometrically identical in their measurement of a construct (as can be found in abilities testing, for example). In the behavioral and social sciences, constructs are more likely to be either tau-equivalent or congeneric. With tau-equivalent items, each items is about equally good at representing the construct (i.e., each item loads on the construct to about the same degree). With congeneric items, some items are stronger indicators of the construct (i.e., have high

Page 786

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Planned Missing Data Designs

loadings) then other items of the construct (i.e., items with lower loadings). The iFFA is a battery with tau-equivalent indicators. A loneliness scale with items such as “I feel lonely” and “I feel like no one likes me” would likely have congeneric qualities with the “I feel lonely” item being a marker variable of the construct loneliness and the “I feel like no one like me” would be a lower loading (yet important) indicator of the construct. With this slightly tangential discussion in mind, items representing the same construct would be spread among the X, A, B, and C blocks of times. If the items are tau-equivalent the process can be essentially random. With six items for a construct, two of the four blocks of items would have two items assigned to them. The key here would be to balance the number of items in the A, B, and C blocks to ensure about equal numbers so that each of the three forms is of equal length. If one of the forms is longer than the other two, unplanned missing data might emerge at a greater frequency for the longer form. A set of dummy codes representing the assigned form could be used in the imputation model to account for this feature if it were to occur. If items are congeneric and one item can be identified as the strongest representative of the construct, it would be assigned to the X block. The remaining items would be distributed as evenly as possible among the A, B, and C blocks. As mentioned previously, when building a multiform protocol, the efficiency of the parameters estimated from the resulting data set is increased if the between block correlations are high. Spreading items of the same construct across the item blocks helps ensure that the between block correlations are high. Putting all the items of a given construct into the same item block will increase the within block correlations but will have a tendency to reduce the between block correlations, which can lower efficiency. Sometimes, questions are blocked or created as a testlet. Here, the set of items in a block or testlet cannot be spread among the different item sets. In this case, the block of items or testlet must be assigned to one of the items sets. If only one block or testlet is involved it would be best placed in the X set. If two or more of them exist, then they could be assigned to different blocks of items. Extending the three-form design longitudinally is also a matter for consideration. As mentioned, the three-form planned missing data design allows researchers to measure more items from more participants on more occasions using the same budget as a complete data design. It also reduces burden, fatigue, and response reactivity. After randomly assigning participants to complete one of three forms on the first occasion, for subsequent measurements,

12:38pm

787

one might assign participants to complete the same form, a different form that they haven’t seen yet, or use random assignment again. Retest effects are expected when making repeated measurements, and assigning different forms over time would decrease the number of variables that participants see on consecutive occasions, thus decreasing the degree to which retest effects manifest in variables not included in the X block. Using a Monte Carlo simulation approach, Jorgensen et al. (2014) compared the different assignment strategies in the context of autoregressive latent-variable panel models, a commonly used type of longitudinal model. Data were simulated based on a fixed, moderately large sample size using a range of population parameter values, and certain patterns (corresponding to each of the three methods of assigning forms) of complete data were replaced with missingness, replacing either 11% or 22% of complete data with planned missingness. The results indicated negligible differences between assignment methods only in the absence of retest effects, which is not a likely scenario in practice. When they introduced practice effects in the form of small (Cohen’s d = 0.1) mean increases for items seen on consecutive occasions, strong invariance models (with intercepts constrained to equality across occasions) showed an increase in the latent mean, although the population mean was zero. This bias manifested less when forms were randomly assigned on each occasion, but not nearly so much as when different forms were assigned systematically, such that no one would see the same form twice until seeing all other forms first. Reduction of bias due to retest effects makes it preferable to systematically assign different forms over time. Two-Method Planned Missing Design In many studies, researchers are put in the difficult situation of choosing between two methods for measuring a critical construct. Often, one method, such as blood cortisol levels for measuring stress, is a highly valid but expensive tool whereas a second method, such as a self-report battery of perceived stress, is a less valid but easily (inexpensively) administered tool. In situations like this, the expensive measure is considered the gold standard for assessing a construct, but the costs of using it are prohibitive. If the expensive measure is chosen, the costs are often reduced by reducing the sample size, which typically results in an underpowered or highly simplified study. In contrast, if the inexpensive measure is chosen, the increases in sample size and design complexity are offset by the fact that the inexpensive measure contains systematic bias (e.g., due to the self-report nature of the measure).

Page 787

Trim Size: 8.5in x 11in

788

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

This scenario does not have to be an either–or choice. Instead, the two-method planning missing data design provides an ingenious way to leverage the capabilities of both modern missing data treatments and latent variable structural equation modeling. When two methods for measuring the same underlying construct exist (or can be developed), the two-method design allows for larger sample sizes while maintaining the high validity of using only an expensive gold standard tool. Table 17.6 provides a list of candidate constructs for which two methods exist. From physiological indices, to classroom observations, to clinical diagnoses, many constructs have expensive gold standard ways of measuring them. Many of these constructs also have easily administered alternatives. In any of these situations, the two-method design can be implemented. One key to the two-method planned missing data design is the use of latent variable structural equation modeling. Figure 17.3 shows a couple of examples of two-method construct representations. As shown in Figure 17.3, the two-method planned missing data design relies on the bifactor model and the common factor model to represent

the construct of interest as well as a bias factor. It also relies on the ability to represent both the gold standard measure and the biased measure with multiple indicators. Using multiple indicators is essential to disentangle the reliable from the unreliable variance and then to further disentangle the biased information from the unbiased information about the focal construct of interest. A second key to the two-method planned missing data design is that all participants are assigned to receive the biased measure. For the unbiased, gold standard measure, only a subset of participants are randomly assigned to be assessed. The ratio of expensive to cheap observations depends on the degree of correlation between the two methods of measuring the construct and on the degree of reliability of the two measures. Graham et al. (2006) provided a straightforward way to calculate the ratio that hits the sweet spot in terms of power and cost. Generally speaking, the amount of planned missing data for this design can be quite high if the biased measure is not too biased and both measures are measured reasonably reliably.

TABLE 17.6 Possible Candidate Constructs for Use with the Two-Method Planned Missing Data Design Construct

Expensive/Gold Standard

Inexpensive/Biased Measure

Stress Intelligence Smoking cessation Student/teacher behavior Nutrition Adiposity

Cortisol assays Wechsler WAIS assessments Blood draws Classroom observations Face-to-face-interview Densitometry, hydrostatic weighing, or dual energy x-ray absorptiometry

Self-report perceived stress Multiple-choice test of ability Self-report Self-, Peer-, or Teacher-report Self-report survey Body mass index

(A) smoking Ceassation as an example

(B)Stress as an example

Bias in the Self Report Measure

Self-report Parcel 1

Bias in the Self Report Measure

Self-report Parcel 2

Smoke Cessation Unbiased!

Cotinine

CO2

Self-report Parcel 1

Self-report Parcel 2

Cortizol Time 1

Cortizol Time 2

Actual Stress Unbiased!

Figure 17.3 Two variations of a two-method planned missing model. (Note: Only a subset of individuals are randomly selected to receive the expensive measure.)

Page 788

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

Planned Missing Data Designs

12:38pm

789

design is the accelerated longitudinal design (Table 17.7). In this design, the missing information is not randomly assigned but, instead, is a feature of the cohort from which participants are selected. Such missing data, however, are due to a predictable MAR process and can therefore be recovered to provide unbiased parameter estimates. In the following, we discuss three wave missing designs: the developmental lag missing design introduced by McArdle and Woodcock (1997); the accelerated longitudinal design originally introduced by Bell (1953); and the growth curve wave missing design introduced by Graham and colleagues (2001). McArdle and Woodcock (1997), for example, introduced a planned missing data design that allowed them to examine memory loss and retest effects over 8 days. Their design involved only two measurement occasions for each participant but the interval of time between the two measurements ranged across participants. Putting the responses into bins of baseline and six successive waves, they used growth curve modeling procedures to represent a slope construct of change over time and a separate retest construct. In one model, the retest effect was estimated to be a constant value regardless of the length of time between measurement occasions. In a second model, the retest effect was specified to be an exponentially declined effect. This exponentially declining retest construct fit the data better than the constant effect model, suggesting that retest effects are more pronounced with shorter intervals of measurement and less pronounced as the time span between occasions gets longer. Bell (1953) introduced the accelerated longitudinal design as a way to estimate age trends from cross-sequential data. Figure 17.4 shows the three sequential data designs

Obtaining an optimal number of indicators for each construct is not difficult if one considers the use of parceling techniques to provide the correct level of aggregation needed to create multiple indicators for each construct. Although the use of parcels is somewhat a matter of debate (and the nature of the debate is beyond the scope of this chapter), the ways parcels can be used for purposes of the two-method planned missing data design are straightforward (see, e.g., Little, Cunningham, Shahar, & Widaman, 2002; Little, Rhemtulla, Gibson, & Schoemann, 2013). To extend the two-method planned missing data design to longitudinal data collection procedures, we need to discuss some nuances for how it should be implemented. In a simulation study of the power and efficiency of the two-method planned missing data design, GarnierVillarréal et al. (2013) and Rhemtulla and Little (2013) provide a number of useful recommendations for extending this model to the longitudinal case. Wave-Missing Designs A number of wave-missing designs are available to developmental researchers. These designs and their variants have different nuances and considerations that must be considered to implement them well. Some of the wave-missing designs involve random assignment and thus make the missing information MCAR. Recall that MCAR missing (missing completely at random) is the hallmark of planned missing data designs because there is no introduction of bias with MCAR missingness, only a decrease in power. As such, these designs provide tremendous potential. Some wave-missing designs, however, are not based on random assignment to missing data pattern. One such

TABLE 17.7 Converting a Cohort-Sequential Data Set into an Accelerated Longitudinal Design Age in years; months for each cohort at each assessment Age/cohort

0 Months

4 Months

8 Months

12 Months

16 Months

20 Months

11;0 12;0 13;0 14;0 15;0 16;0

11;4 12;4 13;4 14;4 15;4 16;4

11;8 12;8 13;8 14;8 15;8 16;8

12;0 13;0 14;0 15;0 16;0 17;0

12;4 13;4 14;4 15;4 16;4 17;4

12;8 13;8 14;8 15;8 16;8 17;8

Age 11 years Age 12 years Age 13 years Age 14 years Age 15 years Age 16 years

Full span of the ages covered 11;0 11;4 11;8 12;0 12;4 12;8 13;0 13;4 13;8 14;0 14;4 14;8 15;0 15;4 15;8 16;0 16;4 16;8 17;0 17;4 17;8 Age 11 years Age 12 years Age 13 years Age 14 years Age 15 years Age 16 years

W1

W2

W3

W4 W1

W5 W2

Source: Little (2013). Used with permission.

W6 W3

W4 W1

W5 W2

W6 W3

W4 W1

W5 W2

W6 W3

W4 W1

W5 W2

W6 W3

W4 W1

W5 W2

W6 W3

W4

W5

W6

Page 789

Trim Size: 8.5in x 11in

790

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

Gen-Yers

Gen-Xers

Cohort Time of Measurment (Birth Year) 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 26 23 24 25 27 28 29 22 31 30 1975 21 32 33 34 35 25 22 23 24 26 27 28 21 30 29 1976 20 31 32 33 34 24 21 22 23 25 26 27 20 29 28 1977 19 30 31 32 33 23 21 22Cohort-Sequential 24 25 26 20 19 28 27 1978 18 29 30 31 32 22 19 20 21 23 24 25 18 27 26 1979 17 28 29 30 31 17 21 18 19 20 22 23 24 26 25 1980 16 27 28 29 30 16 20 17 18 19 21 22 23 25 24 1981 15 26 27 28 29 16 19 17 18 20 21 22 15 24 23 1982 14 25 26 27 28 16 18 17 19 20 21 15 14 Time 23 22 1983 13 24 25 26 27 16 17 15 18 19 20 14 13 22 21 1984 12 Sequential 23 24 25 26 16 14 15 17 18 19 13 12 21 20 1985 11 22 23 24 25 16 13 14 15 17 Cohort-Sequential 18 12 11 20 19 1986 10 21 22 23 24 16 15 12 13 14 17 11 10 19 18 1987 9 20 21 22 23 16 14 11 12 13 15 10 9 18 17 1988 8 19 20 21 22 16 13 10 11 12 14 15 9 8 17 1989 7 18 19 20 21 16 12 9 10 11 13 14 15 8 7 1990 6 17 18 19 20

Figure 17.4 The three traditional sequential designs used to disentangle age, cohort, and time of measurement effects. Source: Little (2013). Used with permission.

that are used to disentangle two of the three confounded effects in longitudinal data. Specifically, age, cohort, and time of measurement cannot be separated in a single cohort longitudinal study. By broadening the measurement frame, however, two of the three effects can be adequately disentangled (see Little, 2013; Wohwill, 1973). With the cross-sequential design (which is often mislabeled the cohort-sequential design), it is possible to realign the data set to represent the ages of the participants regardless of the cohort or wave at which the information was assessed. Figure 17.4 shows how the cross-sequential data are transformed into an accelerated longitudinal design. The final design in this section is the growth curve planned missing design. Rhemtulla, Jia, et al. (2014) examined the performance of three types of planned missing designs in multivariate latent growth models, including a three-form design (where participants are assigned to be missing a subset of items at each occasion), a wave-missing design (where participants are assigned to be missing a subset of measurement occasions), and a design that combined both types of missingness. They found that the three-form design resulted in high convergence rates, little bias, and high efficiency relative to complete data for almost all parameter types. In contrast, the wave-missing and combined designs resulted in biased parameter estimates at sample sizes smaller than 500, and dramatically lower power for many of the structural parameters, including the variance and covariances among latent slopes. Based on these results, they recommended that wave missing designs are only well suited to latent growth curve models when the sample is very large (e.g., N > 1000) or effect sizes are large.

In contrast, three-form missingness is an easy and effective way to achieve the advantages of a planned missing design without sacrificing accuracy or precision.

CONCLUSIONS As should be clear, missing data should no longer be considered a bane to researchers. Modern techniques, when applied in an appropriately principled manner, will yield unbiased and suitably powered parameter estimates for any statistical model chosen. Accurate inferences are possible even under conditions of selective nonresponse (i.e., when the missing data mechanism is MAR). To optimize the ability to recover from the losses associated with unplanned missing data, we admonish researchers to identify and measure all known and likely predictors of nonresponse. To facilitate this practice, researchers should carefully scrutinize extant data sets to identify variables that are predictive of the unplanned missing data. The details of such an analysis can easily be included as an online appendix for interested researchers to access. This online appendix can also contain summary reports of the fraction of missing information in the means and variances of the variables on the data set (perhaps in the form of a histogram of the FMI values of the means and the FMI values for the variances). As well as provide a detailed log of the steps and options that were selected in treating the missing data. With the ease and capacity of web repositories, we can easily be much more transparent in sharing information about a given data set. Another admonishment that we offer is to leverage the advantages of using planned missing data designs in

Page 790

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

References

conjunction with modern treatments and latent variable analyses. The clear gains in cost savings and the potential gains in validity are both sufficient reasons to employ these designs. We acknowledge that the details of how to properly implement planned missing data designs can be more cumbersome than the traditional collect-everythingfrom-everyone approach. The rapidly expanding literature in this area provides the needed details for proper guidance in constructing data collection protocols with planned missing elements (as new works emerge, they will be added to our online bibliography at immap.educ.ttu.edu). We see the benefits of these designs in terms of the costs, veracity, and verisimilitude of research enterprise as far outweighing the time and thought needed to optimally include planned missing data designs in all future studies. In fact, planned missing elements can also be implemented in active longitudinal studies going forward. REFERENCES Allison, P. D. (2002). Missing data. Thousand Oaks, CA: Sage Publications. Allison, P. D. (2006). Multiple imputation of categorical variables under the multivariate normal model. Retrieved from http:// www.statisticalhorizons.com Allison, E. R., Goodson, P., & Neilands, T. B. (2008). Out of sight, not out of mind: Strategies for handling missing data. American Journal of Health Behavior, 32(1), 83–92. Anderson, T. W. (1957). Maximum likelihood estimates for a multivariate normal distribution when some observations are missing. Journal of the American Statistical Association, 52(278), 200–203. doi:10.1080/01621459.1957.10501379 Arbuckle, J. L. (1996). Full information estimation in the presence of incomplete data. In G. A. Marcoulides & R. E. Schumacker (Eds.), Advanced structural equation modeling (pp. 243–277). Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Asparouhov, T., & Muthén, B. O. (2008). Auxiliary variables predicting missing data. Technical report. Retrieved from http://www .statmodel.com/do Asparouhov, T., & Muthén, B. (2010a). Multiple imputation with Mplus. Technical report. Retrieved from http://www. statmodel.com Asparouhov, T., & Muthén, B. (2010b). Weighted least squares estimation with missing data. Technical report. Retrieved from http://www.statmodel.com/download/GstrucMissingRevision.pdf Azar, B. (2002). Finding a solution for missing data. Monitor on Psychology, 33, 70. Baraldi, A. N., & Enders, C. K. (2010). An introduction to modern missing data analyses. Journal of School Psychology, 48, 5–37. doi: 10.1016/j.jsp.2009.10.001 Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic and statistical considerations. Journal of Personality and Social Psychology, 51, 1173–1182. doi: 10.1037/0022-3514.51.6.1173 Beale, E. M. L., & Little, R. J. A. (1975). Missing values in multivariate analysis. Journal of the Royal Statistical Society, Series B, 37, 129–145. Beglinger, L. J., Gaydos, B., Tangphao-Daniels, O., Duff, K., Kareken, D. A., Crawford, J., . . . Siemers, E. R. (2005). Practice effects and the

12:38pm

791

use of alternate forms in serial neuropsychological testing. Archives of Clinical Neuropsychology, 20, 517–529. doi: 10.1016/j.acn.2004.12.003 Bell, R. Q. (1953). Convergence: An accelerated longitudinal approach. Child Development, 24(2), 145–152. doi: 10.2307/1126345 Beunckens, C., Molenberghs, G., Thij, H., & Verbeke, G. (2007). Incomplete hierarchical data. Statistical Methods in Medical Research, 16, 457–492. doi: 10.1177/0962280206075310 Biesanz, J. C., Falk, C., & Savalei, V. (2010). Assessing mediational models: Testing and interval estimation for indirect effects. Multivariate Behavioral Research, 45, 661–701. doi: 10.1080/00273171.2010.498292 Bodner, T. E. (2006). Missing data: Prevalence and reporting practices. Psychological Reports, 99, 675−680. doi: 10.2466/PR0.99.3.675-680 Bodner, T. E. (2008). What improves with increasing missing data imputations? Structural Equation Modeling, 15, 651–675. Brown, C. H. (1983). Asymptotic comparison of missing data procedures for estimating factor loadings. Psychometrika, 48, 269–291. doi: 10.1007/BF02294022 Brown, M. B. (1974). Identification of the sources of significance in two-way tables. Applied Statistics, 23, 405–413. doi: 10.2307/2347132 Brown, R. L. (1994). Efficacy of the indirect approach for estimating structural equation models with missing data: A comparison of five methods. Structural Equation Modeling, 1, 287–316. doi: 10.1080/ 10705519409539983 Browne, M. W. (1984). Asymptotically distribution-free methods for the analysis of covariance structures. British Journal of Mathematical and Statistical Psychology, 37, 62–83. doi: j.2044-8317.1984.tb00789.x Bunting, B., & Adamson, G. (2000). Assessing reliability and validity in the context of planned incomplete data structures for multitrait-multimethod models. Developments in Survey Methodology, 15, 37–53. Bunting, B. P., Adamson, G., & Mulhall, P. K. (2002). A Monte Carlo examination of an MTMM model with planned incomplete data structures. Structural Equation Modeling, 9, 369–389. doi: 10.1207/S15328007SEM0903_4 Cai, L., & Lee, T. (2009) Covariance structure model fit testing under missing data: An Application of the supplemented EM algorithm. Multivariate Behavioral Research, 44, 281–304. doi: 10.1080/ 00273170902794255 Carter, W. H., Jr., & Meyers, R. H. (1973). Maximum likelihood estimation from linear combinations of discrete probability functions. Journal of the American Statistical Association, 68, 203–206. doi: 10.1080/01621459.1973.10481366 Chen, T., & Fienberg, S. (1976). The analysis of contingency tables with incompletely classified data. Biometrics, 32, 133–144. doi: 10.2307/ 2529344 Cheung, M. (2007). Comparison of methods of handling missing time-invariant covariates in latent growth models under the assumption of missing completely at random. Organizational Research Methods, 10, 609. doi: 10.1177/1094428106295499 Collins, L. M., Schafer, J. L., & Kam, C. M. (2001). A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychological Methods, 6, 330−351. doi: 10.1037/1082-989X.6.4.330 Davey, A., & Savla, J. (2010). Statistical power analysis with missing data. A structural equation modeling approach. New York, NY: Routledge. Davey, A., Savla, J., & Luo, Z. (2005). Issues in evaluating model fit with missing data. Structural Equation Modeling, 12, 578–597. doi: 10.1207/ s15328007sem1204_4 Demirtas, H., Freels, S. A., & Yucel, R. M. (2008). Plausibility of multivariate normality assumption when multiply imputing nonGaussian continuous outcomes: A simulation assessment. Journal

Page 791

Trim Size: 8.5in x 11in

Cicchetti

792

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

of Statistical Computation and Simulation, 78, 69–84. doi: 10.1080/ 10629360600903866 Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, Series B, 39, 1–38. Dings, J., Childs, R., & Kingston, N. (2002). The effects of matrix sampling on student score comparability in constructed-response and multiple-choice assessments. Washington, DC: Council of Chief State School Officers.

Enders, C. K., Dietz, S., Montague, M., & Dixon, J. (2006). Modern alternatives for dealing with missing data in special education research. In T. E. Scruggs & M. A. Mastropieri (Eds.), Advances in learning and behavioral disabilities (Vol. 19, pp. 101–130). New York, NY: Elsevier. Enders, C. K., & Gottschall, A. C. (2011). Multiple imputation strategies for multiple group structural equation models. Structural Equation Modeling, 18, 35–54. doi: 10.1080/10705511.2011.532695

Efron, B. (1979). Bootstrap methods: Another look at the jackknife. Annals of Statistics, 7, 1–26. doi: 10.1214/aos/1176344552

Enders, C. K., & Peugh, J. L. (2004). Using an EM covariance matrix to estimate structural equation models with missing data: Choosing an adjusted sample size to improve the accuracy of inferences. Structural Equation Modeling, 11, 1–19. doi: 10.1207/S15328007SEM1101_1

Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. New York, NY: Chapman & Hall/CRC.

Finkbeiner, C. (1979). Estimation for the multiple factor model when data are missing. Psychometrika, 44, 409–420. doi: 10.1007/BF02296204

Enders, C. K. (2001a). The impact of nonnormality on full information maximum likelihood estimation for structural equations models with missing data. Psychological Methods, 6, 352–370. doi: 10.1037/ 1082-989X.6.4.352

Fraley, C. (1999). On computing the largest fraction of missing information for the EM algorithm and the worst linear function for data augmentation. Computation Statistics & Data Analysis, 31, 13–26. doi: 10.1016/S0167-9473(99)00003-1

Enders, C. K. (2001b). The performance of the full information maximum likelihood estimator in multiple regression models with missing data. Educational and Psychological Measurement, 61, 713–740. doi: 10.1177/0013164401615001

Garnier Villarreal, M., Dowsett, C., & Little, T. D. (2013). Using component scores as auxiliary variables: How many components are needed? Manuscript in preparation.

Enders, C. K. (2001c). A primer on maximum likelihood algorithms available for use with missing data. Structural Equation Modeling, 8(1), 128–141. doi: 10.1207/S15328007SEM0801_7 Enders, C. K. (2002). Applying the Bollen-Stine bootstrap for goodness-of-fit measures to structural equation models with missing data. Multivariate Behavioral Research, 37(3), 359–377. doi: 10.1207/ S15327906MBR3703_3 Enders, C. K. (2003). Using the expectation maximization algorithm to estimate coefficient alpha for scales with item-level missing data. Psychological Methods, 8(3), 322–337. doi: 10.1037/1082-989X.8.3.322 Enders, C. K. (2004). The impact of missing data on sample reliability estimates: Implications for reliability reporting practices. Educational and Psychological Measurement, 64, 419–436. doi: 10.1177/ 0013164403261050 Enders, C. K. (2006). A primer on the use of modern missing-data methods in psychosomatic medicine research. Psychosomatic Medicine, 68, 427−436. doi: 10.1097/01.psy.0000221275.75056.d8 Enders, C. K. (2008). A note on the use of missing auxiliary variables in full information maximum likelihood-based structural equation models. Structural Equation Modeling, 15(3), 434–448. doi: 10.1080/10705510802154307 Enders, C. K. (2010). Applied missing data analysis. New York, NY: Guilford. Enders, C. K. (2011). Missing not at random models for latent growth curve analysis. Psychological Methods, 16, 1–16. doi: 10.1037/ a0022640 Enders, C. K. (2013). Dealing with missing data in developmental research. Child Development Perspectives, 7, 27–31. doi: 10.1111/ cdep.12008 Enders, C. K., & Bandalos, D. L. (2001). The relative performance of full information maximum likelihood estimation for missing data in structural equation models. Structural Equation Modeling, 8, 430–457. doi: 10.1207/S15328007SEM0803_5 Enders, C. K., Fairchild, A. J., & MacKinnon, D. P. (2013). Using Bayesian parameter simulation to evaluate mediation effects with missing data. Multivariate Behavioral Research, 48, 340–369. doi: 10.1080/00273171.2013.784862

Garnier Villarreal, M., Rhemtulla, M., & Little, T. D. (2014). Two-method planned missing designs for longitudinal research. International Journal of Behavioral Development, 38, 411–422. doi: 10.1177/ 0165025414542711 Gelman, A., and Rubin, D. B. (1992), Inference from iterative simulation using multiple sequences. Statistical Science, 7, 457–472. doi: 10.1214/ss/1177011136 Gleason, T. C., & Staelin, R. (1975). A proposal for handling missing data. Psychometrika, 40, 229–252. doi: 10.1007/BF02291569 Gold, M. S., & Bentler, P. M. (2000). Treatments of missing data: A Monte Carlo comparison of RBHDI, iterative stochastic regression imputation, and expectation-maximization. Structural Equation Modeling, 7, 319–355. doi: 10.1207/S15328007SEM0703_1 Gold, M. S., Bentler, P. M., & Kim, K. H. (2003). A comparison of maximum-likelihood and asymptotically distribution-free methods of treating incomplete nonnormal data. Structural Equation Modeling, 10, 47–79. doi: 10.1207/S15328007SEM1001_3 Gonzalez, J. M., & Eltinge, J. L. (2008). Adaptive Matrix Sampling for the consumer expenditure quarterly interview survey. Proceedings of the Section on Survey Research Methods, American Statistical Association, 2081–2088. Gottschall, A. C., West, S. G., & Enders, C. K. (2012). A comparison of item-level and scale-level multiple imputation for questionnaire batteries. Multivariate Behavioral Research, 47(1), 1–25. doi: 10.1080/ 00273171.2012.640589 Graham, J. W. (2003). Adding missing-data-relevant variables to FIML-based structural equation models. Structural Equation Modeling, 10, 80–100. doi: 10.1207/S15328007SEM1001_4 Graham, J. W. (2009). Missing data analysis: Making it work in the real world. Annual Review of Psychology, 60, 549–576. doi: 10.1146/ annurev.psych.58.110405.085530 Graham, J. W. (2012). Missing data: Analysis and design. New York, NY: Springer. doi:10.1007/978-1-46144018-5_1 Graham, J. W., Cumsille, P. E., & Shevock, A. E. (2012). Methods for handling missing data. In J. A. Schinka & W. F. Velicer (Eds.), Handbook of Psychology, Research methods in psychology (2nd ed.,Vol. 2, pp. 109–141). New York, NY: John Wiley & Sons. doi: 10.1002/9781118133880.hop202004

Page 792

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

References

12:38pm

793

Graham, J. W., Hofer, S. M., Donaldson, S. I., MacKinnon, D. P., & Schafer, J. L. (1997). Analysis with missing data in prevention research. In K. J. Bryant, M. Windle, & S. G. West (Eds.), The science of prevention: Methodological advances from alcohol and substance abuse research (pp. 325−366). Washington, DC: American Psychological Association. doi: 10.1037/10222-010

Howard, W., Rhemtulla, M., & Little, T. D. (2015). Using principal components as auxiliary variables in missing data estimation. Multivariate Behavioral Research, 50, 285–299. doi: 10.1080/00273171 .2014.999267

Graham, J. W., Hofer, S. M., & MacKinnon, D. P. (1996). Maximizing the usefulness of data obtained with planned missing value patterns: An application of maximum likelihood procedures. Multivariate Behavioral Research, 31, 197–218. doi: 10.1207/s15327906mbr3102_3

IBM Corp. (2012). IBM SPSS Statistics for Windows, Version 21.0. Armonk, NY: IBM Corp.

Graham, J. W., Hofer, S. M., & Piccinin, A. M. (1994). Analysis with missing data in drug prevention research. In L. M. Collins & L. Seitz (Eds.), Advances in data analysis for prevention intervention research (pp. 13–63). Washington, DC: National Institute on Drug Abuse. Graham, J. W., Olchowski, A. E., & Gilreath, T. D. (2007). How many imputations are really needed? Some practical clarifications of multiple imputation theory. Prevention Science, 8, 206−213. doi: 10.1007/s11121-007-0070-9 Graham, J. W., & Schafer, J. L. (1999). On the performance of multiple imputation for multivariate normal data with small sample size. In R. H. Hoyle (Ed.), Statistical strategies for small sample research (pp. 1–29). Thousand Oaks, CA: Sage.

Iacobucci, D. (1995). The analysis of variance for unbalanced data. Marketing Theory and Applications, 6, 337–343.

Jamshidian, M., & Bentler, P. M. (1999). ML estimation of mean and covariance structures with missing data using complete data routines. Journal of Educational and Behavioral Statistics, 24, 21–41. doi: 10.3102/10769986024001021 Jia, F., Moore, E. W. G., Kinai, R., Crowe, K. S., Schoemann, A. M., & Little, T. D. (2014). Planned missing data design on small sample size: How small is too small? International Journal of Behavioral Development, 38, 435–452. doi: 10.1177/0165025414531095 Jöreskog, K. G. (1967). Some contributions to maximum likelihood factor analysis. Psychometrika, 32(4), 443–482. doi: 10.1007/BF02289658 Jorgensen, T. D., Schoemann, A. S., McPherson, B., Rhemtulla, M., Wu, W., & Little, T. D. (2013). Optimal assignment methods in three-form planned missing data designs for longitudinal panel studies. International Journal of Behavioral Development, 38, 397–410. doi: 10.1177/0165025414531094

Graham, J. W., Taylor, B. J., & Cumsille, P. E. (2001). Planned missing data designs in the analysis of change. In L. M. Collins & A. G. Sayer (Eds.), New methods for the analysis of change (pp. 335–353). Washington, DC: American Psychological Association. doi: 10.1037/10409-011

Kenward, M. G., Lesaffre, E., & Molenberghs, G. (1994). An application of maximum-likelihood and generalized estimating equations to the analysis of ordinal data from a longitudinal-study with cases missing at random. Biometrics, 50(4), 945–953. doi: 10.2307/2533434

Graham, J. W., Taylor, B. J., Olchowski, A. E., & Cumsille, P. E. (2006). Planned missing data designs in psychological research. Psychological Methods, 11, 323−343. doi: 10.1037/1082-989X.11.4.323

Kenward, M. G., & Molenberghs, G. (1998). Likelihood based frequentist inference when data are missing at random. Statistical Science, 13, 236–247. doi: 10.1214/ss/1028905886

Haitovsky, Y. (1968). Missing data in regression analysis. Journal of the Royal Statistical Society, Series B, 30, 67–82.

Kim, K. H., & Bentler, P. M. (2002). Tests of homogeneity of means and covariance matrices for multivariate incomplete data. Psychometrika, 67, 609–624. doi: 10.1007/BF02295134

Harel, O. (2007). Inferences on missing information under multiple imputation and two-stage multiple imputation. Statistical Methodology, 4, 75–89. doi: 10.1016/j.stamet.2006.03.002 Harel, O., Hofer, S., Hoffman, L., Pedersen, N., & Johansson, B. (2007). Population inference with mortality and attrition in longitudinal studies on aging: A two-stage multiple imputation method. Experimental Aging Research, 33, 187–203. doi: 10.1080/03610730701239004 Harel, O., Stratton, J., & Aseltine, R. H. Jr. (2011). Designed missingness to better estimate efficacy of behavioral studies. Technical Report 11-15. Storrs, CT: Department of Statistics, University of Connecticut. Hartley, H. O. (1958). Maximum likelihood estimation from incomplete data. Biometrics, 14, 174–194. doi: 10.2307/2527783

Kreuter, F., & Olson, K. (2011). Multiple auxiliary variables in nonresponse adjustment. Sociological Methods & Research, 40(2), 311–332. doi: 10.1177/0049124111400042 Lang, K. M. (2013). Streamlining missing data analysis by aggregating multiple imputations at the data level: A Monte Carlo simulation to assess the tenability of the supermatrix approach. Unpublished master’s thesis, University of Kansas, Lawrence, KS. Lang, K. M., & Little, T. D. (2014). The supermatrix technique: A simple framework for hypothesis testing with missing data. International Journal of Behavioral Development, 38, 461–470. doi: 10.1177/ 0165025413514326

Hartley, H. O., & Hocking, R. R. (1971). The analysis of incomplete data. Biometrics, 27, 783–823. doi: 10.2307/2528820

Lauritzen, S. (1995). The EM algorithm for graphical association models with missing data. Computational Statistics & Data Analysis, 19, 191–201. doi: 10.1016/0167-9473(93)E0056-A

Healy, M., & Westmacott, M. (1956). Missing values in experiments analysed on automatic computers. Applied Statistics, 5, 203–206. doi: 10.2307/2985421

Lee, S.-Y. (1986). Estimation for structural equation models with missing data. Psychometrika, 51, 93–99. doi: 10.1007/BF02294002

Hofer, S. M., & Hoffman, L. (2007). Statistical analysis with incomplete data: A developmental perspective. In: T. D. Little, J. A. Bovaird, & N. A. Card (Eds.), Modeling ecological and contextual effects in longitudinal studies of human development (pp. 13–32). Mahwah, NJ: LEA. Honaker, J., & King, G. (2010). What to do about missing values in time-series cross-section data. American Journal of Political Science, 54(2), 561–581. doi: 10.1111/j.1540-5907.2010.00447.x Honaker, J., King, G., & Blackwell, M. (2011). Amelia II: A program for missing data. Journal of Statistical Software, 45, 1–47. Retrieved from http://www.jstatsoft.org/v45/i07

Lee, T., & Cai, L. (2012). Alternative multiple imputation inference for mean and covariance structure modeling. Journal of Educational and Behavioral Statistics, 37(6), 675–702. doi: 10.3102/1076998612458320 Little, R. J. A. (1988). A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association, 83, 1198–1202. doi: 10.1080/01621459.1988.10478722 Little, R. J. A., & Rubin, D. B. (1989). The analysis of social science data with missing values. Sociological Methods and Research, 18, 292–326. doi: 10.1177/0049124189018002004 Little, R. J. A., & Rubin, D. B. (2002). Statistical analysis with missing data. Hoboken, NJ: John Wiley & Sons.

Page 793

Trim Size: 8.5in x 11in

794

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

Little, R. J. A., & Schenker, N. (1995). Missing data. In G. Arminger, C. C. Clogg, & M. E. Sobel (Eds.), Handbook of statistical modeling for the social and behavioral sciences (pp. 39–75). New York, NY: Plenum. doi: 10.1007/978-1-4899-1292-3_2 Little, T. D. (2013). Longitudinal structural equation modeling. New York, NY: Guilford. Little, T. D., Cunningham, W. A., Shahar, G., & Widaman, K. F. (2002). To parcel or not to parcel: Exploring the question, weighing the merits. Structural Equation Modeling, 9(2), 151–173. doi: 10.1207/ S15328007SEM0902_1 Little, T. D., Howard W. J., McConnell, E. K., & Stump, K. N. (2011). Missing data in large data projects: Two methods of missing data imputation when working with large data projects. Department of Psychology, University of Kansas, KUant Guides. Retrieved from www.quant.ku.edu/resources/guides.html Little, T. D., Jones, S. M., Henrich, C. C., Hawley, P. H. (2003). Disentangling the “whys” from the “whats” of aggressive behavior. International Journal of Behavioral Development, 27(2), 122–133. doi: 10.1080/ 01650250244000128 Little, T. D., Jorgensen, T. D., Lang, K. M., & Moore, E. W. G. (2014). On the joys of missing data. Journal of Pediatric Psychology, 39, 151–161. doi: 10.1093/jpepsy/jst048 Little, T. D., Rhemtulla, M., Gibson, K., & Schoemann, A. M. (2013). Why the items versus parcels controversy needn’t be one. Psychological Methods, 18, 285–300. doi: 10.1037/a0033266 Little, T. D., Schnabel, K. U., & Baumert, J. E. (Eds.). (2000). Modeling longitudinal and multilevel data: Practical issues, applied approaches, and specific examples. Mahwah, NJ: Lawrence Erlbaum Associates. Longford, N. T. (2005). Missing data and small-area estimation: Modern analytical equipment for the survey statistician. New York, NY: Springer. McArdle, J. (1994). Structural factor analysis experiments with incomplete data. Multivariate Behavioral Research, 29, 409–454. doi: 10.1207/s15327906mbr2904_5 McArdle, J., & Hamagami, F. (1992). Modeling incomplete longitudinal and cross-sectional data using latent growth structural models. Experimental Aging Research, 18, 145–166. doi: 10.1080/03610739208253917 McArdle, J. J., & Hamagami, F. (2001). Linear dynamic analyses of incomplete longitudinal data. In L. Collins & A. Sayer (Eds.), Methods for the analysis of change (pp. 139–175). Washington, DC: American Psychological Association. McArdle, J. J., & Woodcock, R. W. (1997). Expanding test–retest designs to include developmental time-lag components. Psychological Methods, 2, 403–435. doi: 10.1037/1082-989X.2.4.403 MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence limits for the indirect effect: Distribution of the product and resampling methods. Multivariate Behavioral Research, 39(1), 99–128. doi: 10.1207/s15327906mbr3901_4 McKnight, P. E., McKnight, K. M., Sidani, S., & Figueredo, A. J. (2007). Missing data: A gentle introduction. New York, NY: Guilford. Meng, X.-L., & Rubin, D. B. (1991). Using EM to obtain asymptotic variance-covariance matrices: The SEM algorithm. Journal of the American Statistical Association, 86, 899–909. doi: 10.1080/01621459 .1991.10475130 Meng, X.-L., & Rubin, D. B. (1992). Performing likelihood ratio tests with multiply-imputed data sets. Biometrika, 79(1), 103–111. doi: 10.1093/ biomet/79.1.103 Mislevy, R. J., Beaton, A. E., Kaplan, B., & Sheehan, K. M. (1992). Estimating population characteristics from sparse matrix samples of

item responses. Journal of Educational Measurement, 29, 133–161. doi: 10.1111/j.1745-3984.1992.tb00371.x Muthén, B., & Asparouhov, T. (2009). Multilevel regression mixture analysis. Journal of the Royal Statistical Society, Series A, 172, 639–657. doi: 10.1111/j.1467-985X.2009.00589.x Muthén, B., & Asparouhov, T. (2012). Bayesian structural equation modeling: A more flexible representation of substantive theory. Psychological Methods, 17(3), 313–335. doi: 10.1037/a0026802 Muthén, B., Kaplan, D., & Hollis, M. (1987). On structural equation modeling with data that are not missing completely at random. Psychometrika, 51, 431−462. doi: 10.1007/BF02294365 Muthén, L. K., & Muthén, B. O. (2012). Mplus user’s guide (7th ed.). Los Angeles, CA: Muthén & Muthén. Nakagawa, S., & Freckleton, R. P. (2010). Model averaging, missing data and multiple imputation: A case study for behavioral ecology. Behavioral Ecology and Sociobiology, 65, 103–116. doi: 10.1007/ s00265-010-1044-7 . Orchard, T., & Woodbury, M. A. (1972). A missing information principle: Theory and applications. Paper presented at the sixth Berkeley symposium on mathematical statistics and probability, University of California, Berkeley, CA. Peugh, J. L., & Enders, C. K. (2004). Missing data in educational research: A review of reporting practices and suggestions for improvement. Review of Educational Research, 74, 525−556. doi: 10.3102/ 00346543074004525 Popham, W. J. (1993). Circumventing the high costs of authentic assessment. Phi Delta Kappan, 7, 470–473. Pornprasertmanit, S., Miller, P., Schoemann, A. M., & Rosseel, Y. (2013). semTools: Useful tools for structural equation modeling. R package version 0.4-0. Retrieved from http://cran.r-project.org/web/packages/ semTools/index.html R Core Team. (2012). R: A language and environment for statistical computing (version 2.15.2) [Computer software]. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from http://www .R-project.org/ Raghunathan, T. E. (2004). What do we do with missing data? Some options for analysis of incomplete data. Annual Review of Public Health, 25, 88−117. doi: 10.1146/annurev.publhealth .25.102802.124410 Raghunathan, T. E., & Grizzle, J. E. (1995). A split questionnaire survey design. Journal of the American Statistical Association, 90, 54–63. doi: 10.1080/01621459.1995.10476488 Reiter, J. P., Raghunathan, T. E., & Kinney, S. (2006). The importance of the sampling design in multiple imputation for missing data. Survey Methodology, 32(2), 143–150. Rhemtulla, M., Jia, F., Wu, W., & Little, T. D. (2014). Planned missing designs to optimize the efficiency of latent growth parameter estimates. International Journal of Behavioral Development, 38, 423–434. doi: 10.1177/0165025413514324 Rhemtulla, M. & Little, T. D. (2013). Planned missing data designs for longitudinal organizational research. In M. Raukko & E. Paavilainen-Mäntymäki (Eds.), Handbook of longitudinal research methods in organisation and business studies (pp. 56–77). Northampton, MA: Edward Elgar Publishing. doi: 10.4337/9780857936790 .00012 Rhemtulla, M., Savalei, V., & Little, T. D. (2014). On the asymptotic relative efficiency of planned missingness designs. Psychometrika, Advance online publication. doi: 10.1007/s11336-014-9422-0

Page 794

Trim Size: 8.5in x 11in

Cicchetti

c17.tex printer - Volume I - 12/10/2015

References

12:38pm

795

Rosseel, Y. (2012). lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48, 1–36. Retrieved from http://lavaan.ugent.be/

Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the art. Psychological Methods, 7, 147–177. doi: 10.1037/ 1082-989X.7.2.147

Rovine, M. J. (1994). Latent variables models and missing data analysis. In A. von Eye & C. C. Clogg (Eds.), Latent variables analysis: Applications for developmental research (pp. 181–225). Thousand Oaks, CA: Sage.

Schafer, J. L., & Olsen, M. K. (1998). Multiple imputation for multivariate missing-data problems: A data analyst’s perspective. Multivariate Behavioral Research. Special Issue: Innovative Methods for Prevention Research, 33(4), 545–571. doi: 10.1207/s15327906mbr3304_5

Rubin, D. B. (1976). Inference and missing data. Biometrika, 63, 581–592. doi: 10.1093/biomet/63.3.581 Rubin, D. B. (1978). Multiple imputations in sample surveys—A phenomenological bayesian approach to nonresponse. Proceedings of the Survey Research Methods Section of the American Statistical Association, 30–34. Rubin, D. B. (1987). Multiple imputation for nonresponse in surveys. New York, NY: Wiley. doi: 10.1002/9780470316696 Rubin, D. B. (1991). EM and beyond. Psychometrika, 56, 241–254. doi: 10.1007/BF02294461 Rubin, D. B. (1996). Multiple imputation after 18+ years. Journal of the American Statistical Association, 91, 473−489. doi: 10.1080/ 01621459.1996.10476908 Salthouse, T., & Tucker-Drob, E. (2008). Implications of short-term retest effects for the interpretation of longitudinal change. Neuropsychology, 22, 800–811. doi: 10.1037/a0013091 SAS Institute (2011). SAS/STAT software, version 9.3. Cary, NC: SAS Institute Inc.

Schafer, J. L., & Yucel, R. M. (2002). Computational strategies for multivariate linear mixed models with missing values. Journal of Computational and Graphical Statistics, 11, 421–442. Shin, T., Davidson, M. L., & Long, J. D. (2010). Effects of missing data methods in structural equation modeling with nonnormal longitudinal data. Structural Equation Modeling, 16, 70–98. doi: 10.1080/ 10705510802569918 Shoemaker, D. M. (1973). Principles and procedures of multiple matrix sampling. Cambridge, MA: Ballinger Publishing. Sinharay, S., Stern, H. S., & Russell, D. (2001). The use of multiple imputation for the analysis of missing data. Psychological Methods, 6, 317−329. doi: 10.1037/1082-989X.6.4.317 Sobel, M. E. (1982). Asymptotic intervals for indirect effects in structural equations models. In S. Leinhart (Ed.), Sociological methodology 1982 (pp. 290–312). San Francisco, CA: Jossey-Bass. Song, X. Y., & Lee, S. Y. (2002). Analysis of structural equation model with ignorable missing continuous and polytomous data. Psychometrika, 67, 261–288. doi: 10.1007/BF02294846 StataCorp. (2013). Stata Statistical Software: Release 13. College Station, TX: StataCorp LP.

Savalei, V. (2010a). Small sample statistics for incomplete nonnormal data: extensions of complete data formulae and a Monte Carlo comparison. Structural Equation Modeling, 17, 241–264. doi: 10.1080/ 10705511003659375

Su, Y., Gelman, A., Hill, J., & Yajima, M. (2011). Multiple imputation with diagnostics (mi) in R: Opening windows into the black box. Journal of Statistical Software, 45(2), 1–31.

Savalei, V. (2010b). Expected versus observed information in SEM with incomplete normal and nonnormal data. Psychological Methods. doi: 10.1037/a0020143

Tanner, M. A., & Wong, W. H. (1987). The calculation of posterior distributions by data augmentation. Journal of the American Statistical Association, 82, 528–540. doi: 10.1080/01621459.1987.10478458

Savalei, V., & Bentler, P. M. (2005). A statistically justified pairwise ML method for incomplete nonnormal data: A comparison with direct ML and pairwise ADF. Structural Equation Modeling, 12, 183–214. doi: 10.1207/s15328007sem1202_1

Van Buuren, S. (2012). Flexible imputation of missing data. Boca Raton, FL: CRC Press. doi: 10.1201/b11826

Savalei, V., & Bentler, P. M. (2009). A two-stage approach to missing data: Theory and application to auxiliary variables. Structural Equation Modeling, 16, 477–497. doi: 10.1080/10705510903008238 Savalei, V., & Falk, C. F. (2014). Robust two-stage approach outperforms robust FIML with incomplete non-normal data. Structural Equation Modeling, 21, 280–302. doi: 10.1080/10705511.2014.882692 Savalei, V., & Rhemtulla, M. (2011). Some explorations of the local and global measures of missing information. Paper presented at the the 76th annual and 17th international meeting of the Psychometric Society, Hong Kong. Savalei, V., & Rhemtulla, M. (2012). Teacher’s corner: On obtaining estimates of the fraction of missing information from FIML. Structural Equation Modeling, 19, 477–494. doi: 10.1080/10705511.2012.687669 Savalei, V., & Yuan, K. (2009). On the model-based bootstrap with missing data: Obtaining a p-value for a test of exact fit. Multivariate Behavioral Research, 44, 741–763. doi: 10.1080/00273170903333590 Schafer, J. L. (1997). Analysis of incomplete multivariate data. New York, NY: Chapman Hall. Schafer, J. L. (2001). Multiple imputation with PAN. In L. M. Collins & A. G. Sayer (Eds.), New methods for the analysis of change (pp. 357–377). Washington, DC: American Psychological Association. doi: 10.1037/10409-012

Van Buuren, S., & Groothuis-Oudshoorn, K. (2011). MICE: Multivariate imputation by chained equations in R. Journal of Statistical Software, 45(3), 1–67. Von Hippel, P. T. (2005). How many imputations are needed? A comment on Hershberger and Fisher (2003). Structural Equation Modeling, 12, 334–335. doi: 10.1207/s15328007sem1202_8 Von Hippel, P. T. (2009). How to impute interactions, squares, and other transformed variables. Sociological Methodology, 39, 265–291. doi: 10.1111/j.1467-9531.2009.01215.x Wacholder, S., Carroll, R. J., Pee, D., & Gail, M. H. (1994). The partial questionnaire design for case-control studies. Statistics in Medicine, 13, 623–634. doi: 10.1002/sim.4780130523 West, S. G. (2001). New approaches to missing data in psychological research: Introduction to the special section. Psychological Methods, 6(4), 315–316. doi: 10.1037/1082-989X.6.4.315 White, I. R., Royston, P., & Wood, A. M. (2011). Multiple imputation using chained equations: Issues and guidance for practice. Statistics in Medicine, 30(4), 377–399. doi: 10.1002/sim.4067 Wilkinson, L., & Task Force on Statistical Inference (1999). Statistical methods in psychology journals: Guidelines and explanations. American Psychologist, 54(8). 594–604. doi: 10.1037/0003-066X.54.8.594 Wohwill, J. F. (1973). The study of behavioral development. New York, NY: Academic Press. Wood, A. M., White, I. R., & Thompson, S. G. (2004). Are missing outcome data adequately handled? A review of published randomized

Page 795

Trim Size: 8.5in x 11in

796

Cicchetti

c17.tex printer - Volume I - 12/10/2015

12:38pm

Missing Data

controlled trials in major medical journals. Clinical Trials Review, 1, 368−376. doi: 10.1191/1740774504cn032oa Wothke, W. (2000). Longitudinal and multi-group modeling with missing data. In T. D. Little, K. U. Schnabel, & J. Baumert (Eds.), Modeling longitudinal and multiple group data: Practical issues, applied approaches and specific examples (pp. 219–240). Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Wu, W., West, S. G., & Taylor, A. B. (2009). Evaluating model fit for growth curve models: Integration of fit indices from SEM and MLM frameworks. Psychological Methods, 14(3), 183–201. doi: 10.1037/ a0015858 Wu, W., & Jia, F. (2013). A new procedure to test mediation with missing data through nonparametric bootstrapping. Multivariate Behavioral Research, 48, 663–691. doi: 10.1080/00273171.2013.816235 Yoo, J. E. (2009). The effect of auxiliary variables and multiple imputation on parameter estimation in confirmatory factor analysis. Educational and Psychological Measurement, 69(6), 929–947. doi: 10.1177/ 0013164409332225

Yuan, K. (2009a). Identifying variables responsible for data not missing at random. Psychometrika, 74, 233–256. doi: 10.1007/s11336-008-9088-6 Yuan, K. (2009b). Normal distribution based pseudo ML for missing data: With applications to mean and covariance structure analysis. Journal of Multivariate Analysis, 100, 1900–1918. doi: 10.1016/j.jmva .2009.05.001 Yuan, K.-H., & Bentler, P. M. (2000). Three likelihood-based methods for mean and covariance structure analysis with nonnormal missing data. Sociological Methodology, 30, 165–200. doi: 10.1111/0081-1750 .00078 Yuan, Y., & MacKinnon, D. P. (2009). Bayesian mediation analysis. Psychological Methods, 14(4), 301–322. doi: 10.1037/a0016972 Yucel, R. M., & Raghunathan, T. (2006). Sequential hierarchical regression imputation (SHRIMP). Paper presented at the annual meeting of the American Statistical Association, Seattle, WA. Zhang, Z., & Wang, L. (2013). Methods for mediation analysis with missing data. Psychometrika, 78(1), 154–184. doi: 10.1007/s11336012-9301-5

Page 796

CHAPTER 18

Person-Oriented Approaches G. ANNE BOGAT, ALEXANDER VON EYE, and LARS R. BERGMAN

CHAPTER OVERVIEW 797 DESCRIPTION OF THE VARIABLE- AND PERSON-ORIENTED APPROACHES 799 Variable-Oriented Approach 800 Person-Oriented Approach 804 THREE EARLY PROTAGONISTS OF THE PERSON-ORIENTED APPROACH 812 William Stern 813 Kurt Lewin 814

Jack Block 815 STATISTICAL APPROACHES 818 Methods of Person-Oriented Data Analysis 818 Structural Equation Modeling 819 Log-Linear Modeling 822 Cluster Analysis 826 CONCLUSION AND FUTURE DIRECTIONS 837 REFERENCES 839

CHAPTER OVERVIEW

oriented research also assumes that there are universal laws involving human behavior, thus, information is not lost when specific variable relationships that are part of a larger process are examined and when data are aggregated across individuals. In contrast, person-oriented approaches focus on patterns or profiles that describe individuals or groups of individuals, the constancy and change of these patterns or profiles, and whether or how these patterns or profiles relate to other profiles or patterns within as well as events outside the individual. Kuhn (1962/2012) argued that it is inevitable that paradigms2 become entrenched. Scientific training focuses on inculcating the values, theory, and methods of a specific paradigm; in essence, students are schooled in anticipation of joining a specific scientific community whose members were also indoctrinated in the paradigm. Thus, the basic building blocks upon which the student and then the scientist conduct research are unlikely to be questioned. In fact, Kuhn argued “that commitment and the apparent consensus it produces are prerequisites for normal science, i.e.,

The variable-oriented approach has been the dominant paradigm in psychological research for most of its history. Person-oriented approaches have been advocated strongly at several junctures in this history; for example, at the turn of the twentieth century by James, Dewey, and Stern, in the 1930s by Lewin and Allport, in the 1960s by Block, and in the 1980s to the present by Magnusson, Bergman, and von Eye. Yet, even though there has been a noticeable increase in the use of person-oriented approaches, the majority of published research is still variable-oriented. By variable-oriented approaches we mean theoretical and statistical approaches that discuss or demonstrate relationships among variables. For example, when we find that childhood depression is positively related to maternal depression, we are discussing the relationship between two variables. This may say little about a specific child. Assuming a moderate positive correlation, there will be many children whose childhood depression and maternal depression scores are far from the regression line, for instance, being fairly high in childhood depression and fairly low in maternal depression, contrary to the positive linear relationship indicated by the correlation coefficient. In developmental research, variable-oriented research focuses on constancy and change among the variables. Variable-

2

In the 1969 postscript to his book, Kuhn noted that paradigm had two meanings: (1) “the entire constellation of beliefs, values, techniques, and so on shared by the members of a given community,” a meaning Kuhn believed was better exemplified by the term disciplinary matrix; and (2) exemplars, examples to which scientists refer to understand the common knowledge base of a discipline. In this chapter, the first meaning of paradigm is used.

1

Color versions of Figures 18.1, 18.4–18.7 are available at http://onlinelibrary.wiley.com/book/10.1002/9781118963418 797

798

Person-Oriented Approaches

for the genesis and continuation of a particular research tradition” (p. 11). This commitment and consensus to the dominant paradigm are two of the major reasons that paradigm shifts occur slowly. For example, there is no doubt few, if any, undergraduate or graduate students understand that the science they are being taught uses a variable-oriented paradigm. Without specifically defining the paradigm, the presentation of material in textbooks assumes the validity and worth of research that elaborates the relationships among variables. Statistics courses, beginning at the undergraduate level, further instill the variable-oriented approach, focusing on methods that aggregate individuals (e.g., analysis of variance [ANOVA], regression) and often make questionable assumptions about linearity, data distributions, and random error. The dominant paradigm finds a way to integrate anomalous findings until the moment when it cannot. Kuhn (1962/2012) noted that over time a crisis is reached that makes a new paradigm more acceptable. At present, psychology is dominated by the variable orientation. There does not yet appear to be a crisis in psychological research that necessitates the abandonment of the variable-oriented approach or that propels the person-oriented approach to greater ascendance. However, at the same time, there is a slow but steady accretion of person-oriented research in various fields of psychology, such as developmental and developmental psychopathology (e.g., Martinez-Torteya, Bogat, von Eye, & Levendosky, 2009), clinical (e.g., Lundh, Saboonchi, & Wangby, 2008), personality (e.g., Laible, Carlo, Panfile, Eye, & Parker, 2010), and industrial organizational (e.g., Foti, Thompson, & Allgood, 2011), and in other disciplines such as education (e.g., Boscardin, Muthén, Francis, & Baker, 2008), criminal justice (e.g., Sampson & Laub, 2005), kinesiology (e.g., Sjoquist et al., 2010), medicine (e.g., Axen et al., 2011), public health (e.g., Lennon, McAllister, Kuang, & Herman, 2005), and sociology (e.g., Kim & Fram, 2009). There is also a burgeoning focus on complementary theory and methods to the person-oriented approach, including idiographic science (e.g., Molenaar, 2004, 2007, 2010). In Kuhnian terms, the person-oriented approach is an alternative paradigm, because “it regularly raises questions that cannot be resolved by the criteria of normal science” (Kuhn, 1962/2012, p. 109). If a paradigm shift from variable- to person-oriented approaches occurs, or if the person-oriented approach becomes a promising alternative to the variable-oriented approach, it will be because the person-oriented approach affects a change in what are considered legitimate problems to solve as well as what are acceptable solutions to address those problems.

The type versus trait debate in personality is a good example of how difficult it is to affect a paradigm shift. Research on personality has centered for years on a specific question: “Should the field conceptualize personality in terms of the individual attributes that vary across people or the overarching configuration of attributes that defines each person?” (Donnellan & Robins, 2010, p. 1072). The trait approach, consistent with a variable-oriented approach, searches for universals based on the presence of attributes that vary across people. The five-factor model (FFM; McCrae & Alik, 2002) is the most widely used trait approach. The type approach, consistent with the person-oriented approach, focuses on understanding the personality of the individual. There is strong research evidence for three personality types: ego resilients, overcontrollers, and undercontrollers (Robins, John, Caspi, Moffitt, & Stouthamer-Loeber, 1996). Various rationales have been offered against types and for traits. One is that the three types have not been found in all samples (e.g., Costa, Herbst, McCrae, Samuels, & Ozer, 2002). Even Donnellan and Robins (2010), who found merit in types and defend the person-oriented approach, suggested that the inability to replicate the three types in every sample may be the result of, among other factors, sampling error. Building on this reasoning, Costa et al. (2002) argued that the failure to replicate the three personality types in every sample indicates the superiority of the trait (FFM) approach. Both assertions disregard research that the five factors also have not been replicated in every sample (e.g., Vassend & Skrondal, 2011). These statements reflect how deeply ingrained the variable-oriented approach is in scientific thinking, or, stated another way, to most psychologists, the variable-oriented approach is current scientific thinking. When scientists search for universals and cannot consistently find them, they conclude that there is a flaw in their theory or in the specific methods they used (e.g., sampling error). The paradigm—that is, the overarching perspective that guides their specific theories and research methods and underpins these inconsistent findings—is rarely considered. Using the person-oriented approach as a paradigm, finding or not finding the three personality types in every sample does not mean that these types are wrong, inconclusive, or not useful. It simply means that certain profiles are good descriptors of only certain individuals. Individuals or groups of individuals who do not fit into one of the three types are not exemplars of theory or methods failure, but, rather, may represent individuals or groups that occur at low frequency within certain samples. That is, there is likely to be heterogeneity in samples. The tension between a variable- and person-oriented approach is also apparent in psychopathology research

Description of the Variable- and Person-Oriented Approaches

and practice. To illustrate, we focus on the issue of heterogeneity vs. homogeneity in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM; American Psychiatric Association, 2013). The categorical classification of the DSM, based on a medical diagnosis approach, is structured so that mental disorders are defined by the presence or absence of specific symptoms. Advantages to a categorical diagnostic classification system include providing a common language among mental health professionals that improves communication as well as scaffolding that allows researchers and practitioners to determine the reliability and validity of their clinical decisions (e.g., Maughan, 2005). The DSM categories are based on empirical research that finds general patterns of specific behaviors/symptoms across individuals who meet a specific diagnosis. Thus, the categories consist of criteria that, in general, are associated with a particular disorder. The authors of the DSM acknowledge that there will be heterogeneity within diagnostic categories. However, the diagnostic categories start with the premise that heterogeneity of symptom features is not the most salient feature in understanding a specific type of disorder. Ignoring heterogeneity or subgroups within diagnostic categories may have significant implications for both practice and research. For example, in randomized clinical trials of empirically supported treatments, diagnoses are made before clients/participants are enrolled. If the presence of subgroups is not assessed, the results of the clinical trial may be less clearly understood. We know that treatments do not have salutary effects on all participants and that there is a range of drop out in clinical trials. Certainly there are many factors affecting these two outcomes, but individual differences are no doubt one of them. In the 1960s, psychotherapy researchers (e.g., Kiesler, 1966; Strupp & Bergin, 1969) called for methods to elaborate specific therapist behaviors that would help specific clients and produce specific types of behavior change—a charge consistent with a person-oriented approach. Kiesler challenged the notion that all therapists, clients, and treatments were alike, and he labeled the desire of researchers to treat them as such uniformity myths. One possible solution to the problems with heterogeneity identified by Kiesler and others would have been to conceptualize psychotherapy and psychotherapy research from a person-oriented approach. Such a perspective would have, in part, focused on understanding patterns of clients, therapists, treatments, and behaviors that led to successful outcomes. But the variable-oriented approach was and is the dominant paradigm; thus, it framed the concern with heterogeneity identified by Kiesler and others.

799

The effect of the criticism was that many psychotherapy researchers attempted to reduce heterogeneity by creating more homogenous treatments (e.g., manualized treatments) and more homogeneous patient groups exposed to treatment (e.g., DSM diagnosable conditions, such as depression). The attempt to create homogeneous patient groups may have been illusory, given the inherent heterogeneity in the diagnostic categories3 . And, importantly, the creation of homogeneous groups and the subsequent use of variable-oriented methods to analyze data did not answer the basic criticism of the field, which suggested that understanding the effects of specific treatments and therapists on specific patients should be the focus of researchers’ efforts. It is easy to see how a person-oriented approach to psychotherapy outcome research would have led the field in different directions, focusing on patterns and profiles of therapist, client, and treatment factors that are related to change. The purpose of this chapter is to elucidate the personoriented approach. To do this, we organize the chapter into four sections. In the first section, we describe the tenets and criticisms of the variable- and person-oriented approaches. In the second section, we provide a historical context for the modern formulation of the person-oriented approach by focusing on the work of three early protagonists who advocated this approach at different times during the twentieth century (Stern, Lewin, and Block). In the third section, we discuss three methods of data analysis (structural equation modeling, log-linear modeling, and cluster analysis) to determine whether and how they can be used in the context of person-oriented research. Finally, the last section concludes with a summary and future directions for person-oriented research.

DESCRIPTION OF THE VARIABLE- AND PERSON-ORIENTED APPROACHES Kuhn (1962/2012) used the term paradigm and later disciplinary matrix to refer to theory and methods of a normal science. In discussing the person orientation, Sterba and Bauer (2010) use the term approach, meaning an overall rubric that encompasses theoretical principles and the methods used to elucidate those principles. Regardless of the nomenclature, the point is that any scientific

3 There is a movement in psychopathology research away from taxonomies and toward dimensional conceptualizations of mental health disorders. However, the dimensional approaches are variable oriented.

800

Person-Oriented Approaches

perspective cannot divorce theory and methods, as they are intimately tied to and inform one another. Theory and data analytic techniques from one paradigm can reach conclusions only consonant with that paradigm (Cairns, 1983; Magnusson, 1985, 1995). We will elucidate this essential quality of paradigms as we describe both the variable- and person-oriented approaches. Variable-Oriented Approach The dominance of the variable-oriented approach in scientific research makes it difficult to explain or define it—the variable-oriented approach is likely to be considered science itself, and not just a paradigm guiding scientific inquiry (Bergman & Andersson, 2010). We discuss two broad assumptions that characterize the variable-oriented approach. First, that objects are best understood by focusing on their component parts in isolation is a foundation of the variable-oriented approach (Kelso, 2000). The variableoriented approach is based on the premise that examining one aspect of the whole is thought to reflect the inherent properties of the whole, and that characteristics of the entire entity “have properties irreducible to the individual-level yet nevertheless affecting individual-level phenomena” (Subramanian, Jones, Kaddour, & Krieger, 2009, p. 350). A variable is considered important to the extent that individuals are the same or different on the specific dimension it measures. The focus is on individual variables, not patterns or profiles of variables. Bergman and Andersson (2010) also noted that when statistical analyses of the variables are undertaken, simplifying assumptions often are made about the structure of the variables. Second, the variable-oriented approach focuses on finding universal laws that describe and predict human behavior. As theory and hypotheses are driven by this assumption, researchers expect to find homogeneity across the samples or populations studied. To the extent that these universals are not found within a specific sample or population, the theory or methods are considered flawed (see earlier discussion of the type vs. trait controversy). When heterogeneity is considered, it is often based on subgroups defined by the researcher that derive from a priori assumptions that those subgroups differ on relevant factors (e.g., males vs. females; schizophrenics vs. nonschizophrenics); these assumptions may or may not be valid. The focus on universals often means that assumptions underlying widely used statistical analyses are ignored. For example, aggregation of data, which we discuss at length

later in this section, is a widely used technique. Because, in the search for universal laws, the variable-oriented approach focuses on the importance of interindividual differences rather than intraindividual differences, aggregation is considered appropriate. “Theories are often formalized in terms of causal relationships between constructs which are regarded as dimensions on which a single individual may vary across time and on which different individuals vary” (Bergman & Andersson, 2010, p. 155). Related to aggregation is the problem of outliers. Variable-oriented researchers are usually willing to eliminate or replace them. Outliers are, by definition, statistically rare. There are different types of outliers, including distance outliers (those far from the mean), leverage outliers (those that have undue influence vis-à-vis other scores on specific parameters), and inliers (scores close to other scores but which still have a proportionally extreme effect on specific parameters). Statistical methods have been proposed for dealing with outliers, including considering the outlier scores missing and imputing values to replace the deleted scores. The assumption is that outliers represent error rather than true values. This assumption of outliers as error fails to consider that (1) outliers may be members of a different population than the majority of individuals in the sample, and (2) the study of this different population (regardless of how small it is) vis-à-vis the other population may be important (von Eye et al., 2015). In addition, there is a form of aggregation that occurs in calculating the value of a specific variable. Most variables are scored using a cumulative technique where the more answers an individual gives the more of that variable (e.g., depression) the individual is assumed to have. This is in contrast to other techniques, such as differential scale measurement, where an individual’s pattern on the various questions that constitute the variable might be considered; for example, ogive rules, where extreme responses are given more weight than responses that approximate the median (Loevinger & Wessler, 1970). Although not typically used in person-oriented research, approaches such as item response theory (IRT) can be adapted (see von Eye et al., 2015). In addition, commonly employed statistical analyses often assume that relationships among variables are linear and that “errors are random, follow a normal distribution, and are not related to the true scores” (Bergman & Andersson, 2010, p. 156). And, finally, factors that might affect an individual’s score on the key variables are considered nuisance variables that confound understanding of the key variables; thus, they are typically controlled in the analyses. The problems with these assumptions are

Description of the Variable- and Person-Oriented Approaches

discussed in more detail below in the section on limitations of the variable-oriented approach. Like any scientific paradigm, the variable-oriented approach has legitimate advantages. For example, Kuhn (1962/2012) argued that the areas of inquiry toward which a paradigm directs scientists “leads to a detail of information and to a precision of the observation-theory match that could be achieved in no other way” (p. 64). That is, in the social sciences, variable-oriented theories and methods identify, distinguish, and classify the component parts of the object under investigation (e.g., the child, the family, the school classroom). The variable-oriented approach provides the researcher with objectivity as a result of “clearly defined measurements, the additional information value that comes out of good scales, the usefulness of inferential statistics based on variables, the strength of model testing, and the possibilities for causal inferences that may arise from the study of relationships with proper control for confounders . . . ” (Bergman & Andersson, 2010, p. 156). The variable-oriented approach, through its focus on understanding the relationships among precisely defined variables, also can aid identification of mechanisms that occur in the overall process under investigation (Magnusson & Törestad, 1993). Some argue that the variable approach provides the foundation on which science rests—without it we could not move beyond basic descriptive information or idiographic approaches (e.g., Mandara, 2003). As with any paradigm, there are situations or circumstances about which the variable-oriented approach provides completely accurate data and conclusions. Even Jack Block, a major proponent of the person-oriented approach, whom we shall discuss in detail later, suggested that sometimes the variable-oriented approach was appropriate. Regarding data analysis, the person-oriented approach may not be needed if (1) the variable-oriented approach provides sufficient explanation of the theory or hypothesis tested (e.g., in those cases in which the theory tested and data analysis conducted are focused only on making statements about the aggregate, and not the individual), and (2) the process studied over time is ergodic; that is, the process is stationary and inter- and intraindividual change are identical (Molenaar, 2004).4 In a Kuhnian sense, it is only through the refinement of the variable-oriented approach that psychologists can ascertain anomalies; that is, we see the novel only when we know what to expect and when the expected does not occur. 4

However, Molenaar argued that ergodicity can almost never exist in development.

801

Limitations of the Variable-Oriented Approach Two overall disadvantages with the variable-oriented approach have been articulated. The first concerns assumptions of variable-oriented data analytic techniques. A major problem in this regard is the logical error of using variableoriented data to make conclusions about individuals or subgroups. Leo Tolstoy’s famous opening line of Anna Karenina captures this problem: “Happy families are all alike; every unhappy family is unhappy in its own way.” The first phrase is variable-oriented—if you’ve seen one happy family you’ve seen them all; whereas the second phrase is person-oriented—unhappy families find myriad ways to make themselves miserable. Perhaps Tolstoy knew very few happy families, and his assessment was that they all had similar traits; for example, perhaps all of these families had a high level of involvement and the children and parents were typically happy. Researchers using the variable-oriented approach might draw the same conclusion as Tolstoy by assessing numerous families, measuring family involvement and happiness, and then examining the covariance between the two variables. Such research tells us something about the aggregate—how the variables involvement and happiness are generally related, but it tells us little about a specific family. Although it would be easy for researchers if all happy families were alike, there is no indication that this is the case. Involvement is likely only one of several variables that are associated with family happiness. An ecological fallacy exists when “aggregated data are analyzed and the results are assumed to apply to relationships at the individual level” (Everitt, 1998, p. 110). Robinson (1950) first demonstrated the ecological fallacy by examining the data from the 1930 U.S. census. Individual correlations indicated a negative correlation (r = −0.11) between illiteracy and nativity (i.e., whether one was foreign-born or not), suggesting that the foreign-born were less literate than native born Americans. However, when the same correlation was calculated at the state level, the correlation was r = .53, suggesting just the opposite— that foreign-born individuals were more literate than the American born. Another sociologist, in a discussion of proper replication as it related to Durkheim’s suicide data and his methods, coined the term ecological fallacy, defined as “the invalid transfer of aggregate results to the individual” (Selvin, 1958, p. 613). In the field of epidemiology, there has been a lively debate disputing Robinson’s (1950) admonition that the aggregate may not apply to the individual case. Subramanian et al. (2009) summarized some of the critiques, including searching for instances where there might be

802

Person-Oriented Approaches

correspondence between aggregate- and individual-level associations, reducing bias in estimates of aggregate models so they more closely approximate the individual, demonstrating that the individual-level relationships among variables can be as error-filled as the aggregate level relationships among those variables, and elucidating “solutions that approximate individual-level relationship based on the information that bounds them at the aggregate level” (p. 343). Finally, the critique extends to the term ecological fallacy itself, which some researchers feel diminishes the importance of the ecological and inappropriately privileges individualism; Susser (1994) indicated the better label might be aggregative fallacy. From a Kuhnian perspective, these attempts to discredit the general assertion—that the aggregate does not always apply to the individual—speak to how strongly the variable-oriented approach is held in the field of epidemiology (and elsewhere) and exemplifies Kuhn’s (1962/2012) principle that until a paradigm shift occurs, scientists will continue to fit anomalous data into the dominant paradigm. Regardless of the critiques suggesting that applying the aggregate to the individual may not be problematic, much research since Robinson (who demonstrated the fallacy using bivariate correlations) has indicated that the problem occurs frequently and the same fallacy can exist for any parameter that is estimated at the individual and aggregate level (e.g., Schmitz, 2000; von Eye & Bergman, 2003; von Eye & Bogat, 2006). von Eye and colleagues (2003, 2015) provide a striking example of the ecological fallacy in the development of alcoholism (Perrine, 1995). The data came from a sample of male adults, collected over the course of 3 years, who provided daily information regarding how much alcohol they had consumed the day prior to the interview. In one article (von Eye & Bergman, 2003), the first four participants in the data set were examined (respondents 3000, 3004, 3005, and 3007). Autocorrelograms, a visual representation of the relationship of each participant’s scores with each other as a function of time, were created for each individual as well as an aggregate for the four participants. As the authors noted, the aggregate autocorrelogram indicated that the group had a specific pattern of drinking behavior that consisted of high and positive correlations during weekdays. Unfortunately, this pattern did not represent any of the individual participants’ drinking patterns. For example, one participant demonstrated a fairly erratic pattern that was not predictable day-to-day or over time. Another participant’s pattern was also erratic, yet it was possible to predict his day-to-day consumption even if his consumption could not be predicted over a longer time

period. In a more recent chapter, von Eye and colleagues (von Eye, Bergman, & Hsieh, 2015) demonstrated that these results were not a coincidence of subject selection. They chose the next four participants in the data set, conducted similar analyses, and found similar results; that is, the aggregate autocorrelation plot does not represent any of the individual participants’ plots. Schmitz (2000) developed four theorems, based on methods characteristics, that highlight problems of aggregating data and then applying it to the individual. The first theorem is that individual trajectories cannot be determined from aggregated trajectories. Schmitz’s second theorem is that cross sectional and interindividual relationships cannot be used to describe processes within the individual. The third focuses on asynchronous relationships. In longitudinal research, cross-lagged panel analyses may not capture the relationships among the variables that are true of any given individual. Finally, the fourth theorem suggests that the sequence of aggregating and prediction should start with the individual, not the group. The variable-oriented approach, which generalizes from the aggregate to the individual, assumes that the data are perfect (e.g., aggregate- and individual-level data have the same parameter estimates). However, we do not live in a perfect world, and, because of this, the researcher is faced with two general problems when he or she generalizes from the aggregate to the individual. First, data in each individual case do not always speak to the same effect. Individual- and aggregate-level information can only be equal if the parameter estimates are unbiased; that is, correctly specified. Second, aggregating is often confounded with unmeasured variables. Lurking variables are those omitted from the design or data analysis that might lead the researcher to unknowingly make false or misleading conclusions. Lurking variables might also underlie spurious relations among other variables as well as mask heterogeneity in a population or sample (von Eye et al., 2015). There is, as yet, no perfect answer to problems with aggregation, although sampling, design, and data analysis solutions have been proposed (Subramanian et al., 2009; von Eye & Bogat, 2006). In addition to problems with aggregation, four other methodological problems often occur with the variableoriented approach (Bergman & Andersson, 2010). First, rarely is it the case that models generated using all subjects in a sample or population describe specific individuals or subgroups. Researchers seldom decompose these models to examine possible subgroup differences. Variable-oriented solutions to the problems of subgroup differences (e.g., adding exogenous variables such as gender to a structural

Description of the Variable- and Person-Oriented Approaches

equation model) are usually not true resolutions. For example, the inclusion of an exogenous variable, such as socioeconomic status, in a structural equation model does not address whether the relationships among the variables are similar or different among subgroups that differ on socioeconomic status. The second issue is that model fit is often considered a statement about the fit of the model to the data, but this is not always the case. Bergman and Andersson (2010) noted that the accuracy of model fit depends on the veracity of the underlying assumptions. Most models create a data summary matrix based on pairwise linear relationships, which cannot account for the presence of higher order interactions or nonlinear relationships. Models are based on the goodness of the data of which they are composed. This means that, fairly often, the model fit only tells us about “the fit to a flawed summary of the data” (p. 156). Meehl’s paradox (1950) demonstrated there are occasions in which no significant main effects or lower order interactions are present, and the highest order effect explains all the variability (Krauth & Lienert, 1973; von Eye, 2002). Thus, omitting higher order interactions in a specific model will sometimes completely misrepresent the data and the underlying relationships among the variables. Third, Toomela (2009) argued that post–World War II psychological science is built on the assumption that cause–effect relationships are linear; therefore, standard data analytic techniques, which assume linear relationships among variables, are considered sufficient to answer most research questions. Unfortunately, it is unlikely that most human behavior is linear. For example, two variables may have a relationship with each other, but only under certain conditions. Interventions might have a negative effect on a participant early on, but later show a positive effect; for other participants, the positive effect may occur early and then erode and become negative over time. There are numerous examples of how development is typically not linear. For example, Sampson and Laub (2005) discussed turning points; that is, rare events in the life of an individual that lead to a cessation in the trajectory of crime. And Fogel, Garvey, Hsu, & West-Stroming (2006) noted a bridging frame that explains how mother-infant communication followed an inverted U-shaped curve as it moved from a dyadic mode (mother–infant) to a triadic guided mode (mother shows infant toy) to a triadic not-guided mode (mother shows infant toy and infant plays with it) in the first months of life. It is also a basic assumption in most statistical analyses that errors are random and normally distributed. In addition, random effects are assumed to be independent of

803

model covariates. These assumptions are unlikely to hold and can have serious consequences for understanding data. For example, statistical models make implicit assumptions about the variables we do not measure or include—that they are equal across individuals or situations (Ram & Nesselroade, 2007). For example, psychological research on depression assumes that environmental factors are not an important influence on individuals’ depressive symptoms. Environmental factors may not be measured or various aspects of the environment (e.g., socioeconomic status) may be controlled, thus ignoring epidemiological research indicating that, in the United States, neighborhood characteristics are associated with depression, independent of individual characteristics (e.g., see Kim, 2008, for a review). Before closing this section, we discuss two metatheoretical disadvantages of the variable-oriented approach: (1) whether developmentalists are, in fact, concerned with understanding individual behavior, and (2) if they are, what is the best way to accomplish this goal. First, in some fields of research, it may not be necessary or important to understand the individual. One of the points made by Subramanian et al. (2009) is that epidemiologists and sociologists are not always interested in the individual. For example, an individual’s income may not be the same (it might be higher or lower) as the average income in the neighborhood or census tract where he or she lives. However, the effect of the overall neighborhood aggregation of SES may play an important role in individual behavior, or it, by itself, may be the focus of interest. However, studies in developmental psychopathology are interested in individual behavior. The second metatheoretical issue concerns systems theory. If systems theory is, at present, the most accurate way to describe human behavior, then it is not possible to understand an individual by studying aspects of the individual in isolation; we can only understand an individual by studying multiple aspects of the person and his or her environment and how these aspects relate to each other. Magnusson and Törestad (1993) noted that to understand an individual’s dynamic psychological functioning, at a minimum, we must understand his or her thoughts and actions as an intentional, active being; his or her biological systems, including genetics; and his or her interface with the social environments as well as how all of these factors affect each other in the present and over time. This relates to Lerner and Busch-Rossnagel’s (1981) conception of the individual as someone who generates his or her development. Thus far, the variable-oriented approach has not been especially effective at elucidating humans as dynamic systems.

804

Person-Oriented Approaches

Person-Oriented Approach The modern-day person-oriented approach derives from a zeitgeist shift from research on development and developmental psychopathology using an individualistic perspective to research that considers the importance of a holistic, dynamic system in studying the individual (e.g., Bergman & Magnusson, 1997; Cicchetti & Scheider-Rosen, 1986; Magnusson, 1988). Bergman and Magnusson (1997) argued that the current emphasis on holistic-dynamic systems theory derives from three sources: (1) the scientific movement to integrate our understanding of biological processes with behavioral, cognitive, and environmental factors; (2) recent advances in methods focused on nonlinear dynamic systems; and (3) a renewed interest in longitudinal research. Dynamic systems5 stand in contrast to static, linear models (e.g., Bergman & Andersson, 2010). “The problem of reconciling phenomena that at the macroscopic level appear ordered, irreversible and determined with phenomena that at the microscopic level appear variable, reversible and stochastic lies at the heart of the DSP [dynamic systems perspective] and self-organization . . . ” (Witherington, 2007, p. 134). Dynamic systems theory and models focus attention on understanding change and when and how something new results from that change. Whatever process or variable is studied, at a particular moment, that process or variable resulted from an assemblage of dynamic components that formed, through nonlinear processes, into patterns. Dynamic systems are characterized by moments of equilibrium and moments of disequilibrium; at any given moment, the system can stay stable or become unstable (Kelso, 1995). Bergman and Magnusson (Bergman & Magnusson, 1991, 1997; Magnusson, 1988) described five tenets of the person-oriented approach. These have been further expanded and elaborated by von Eye and colleagues (e.g., von Eye & Bogat, 2006; von Eye & Bergman, 2003). We start with the original five (Bergman & Magnusson, 1997, all quotes p. 293). 1. “The process is partly specific to the individual.” Sterba and Bauer (2010) labeled this the individual 5

There are at least two competing metatheoretical approaches to dynamic systems: contextualism and organicism–contextualism (see Witherington, 2007, for a discussion). It is beyond the scope of this chapter to discuss these. In the discussion of dynamic systems in this chapter, we are referring to the organicism– contextualist metatheory, which best complements the personoriented approach.

specificity principle. Although it is tempting to look for universals that describe human behavior, it is unlikely that many universals, at least as related to complex behavior, exist. Even as related to brain function, where we are tempted to state unequivocally that a specific function is controlled by a specific part of the brain, we know that in some individuals, other parts of the brain can take over the same function. There will not be a universal that fully describes every individual. Even researchers taking a variable-oriented approach recognize this problem. For example, Morey and Hopwood (2009) employed item response theory to differentiate an individual’s development along an alcohol dependence continuum to aid understanding of the effectiveness of a treatment protocol. Such efforts attempt to increase predictive power by creating greater homogeneity, when there is known or suspected heterogeneity in the sample or population. But the focus on prediction can interfere with progress in understanding certain kinds of human phenomena such as development, personality, or psychopathology (Magnusson & Törestad, 1993). Prediction can be a goal, a “conceptual and methodological tool” (Magnusson & Törestad, 1993, p. 442), or both. If prediction is a goal, then it takes front and center in our scientific theory and methods. However, if we consider prediction a tool in our research arsenal, sometimes aligned with prediction as a goal, then it can help to elucidate aspects of the phenomena we wish to understand. Magnusson and Törestad (1993) expressed concern that the current focus on prediction, solely as a goal, and its attendant research methods, has led psychology to understand mechanisms, but not processes. Importantly, individual specificity implies that there are real individual differences that are not the result of errors in measurement or sampling. These individual differences may appear in different samples or within subgroups within one sample. von Eye et al. (2015) noted that outliers (or inliers), typically excluded from variable-oriented approaches, may, in some cases, typify this individuality. A variable-oriented approach, which often uses linear models to determine the relationship among variables across a sample or population, cannot describe how those variables relate in the life of a specific individual or subgroups of individuals (e.g., Magnusson & Törestad, 1993). 2. “The process is complex and is conceptualized as involving many factors that interact at various levels which may be mutually related in a complicated manner.” This tenet focuses on the complexity of human development and functioning; Sterba and Bauer (2010) labeled this principle complex interactions. Magnusson & Törestad (1993)

Description of the Variable- and Person-Oriented Approaches

labeled this multidetermination; that is, scientific understanding of the human system will be determined by multiple factors or conditions. It is unlikely that one or two variables and their interactions could describe an individual sufficiently. The complexity of human psychological functioning has been compared to general systems theory with its emphasis on dynamic, nonlinear, complex processes (e.g., Magnusson & Törestad, 1993). The separation of specific aspects of a human process or the assumption that a linear relationship is present among several aspects cannot capture human behavior at any given time or its course and development over time. Numerous writers have noted the importance of examining the whole person as a system with the underlying assumption that the whole is more than the sum of its parts. Human processes involve transactions among psychological, biological, and environmental factors. Recent theoretical and empirical work has begun to integrate these. For example, adaptive calibration models suggest that children react to specific environments based on biological responsivity (Del Giudice, Ellis, & Shirtcliff, 2011; Del Giudice, Hinnant, Ellis, & El-Sheikh, 2012; Ellis, Del Giudice, & Shirtcliff, 2013). Vedeler and Garvey (2009) proposed an integration of epigenesis and dynamic systems; in other words, human functioning at all levels (gene to culture) is understood as a dynamically evolving system. 3. “There is a meaningful coherence and structure (a) in individual growth and (b) in differences between individuals’ process characteristics. The lawfulness of the processes, within functionally organized structures, is reflected in the development and functioning of all subsystems as well as in the functioning of the organized totality.” Sterba and Bauer (2010) labeled this principle interindividual differences in intraindividual change. The dynamic, holistic perspective suggests that the organism consists of coherent processes (e.g., biological, psychological). This integration occurs with lower order processes, which then coordinate with higher order processes. However, there may well be individual differences in how various systems are organized and integrated (Magnusson & Törestad, 1993). From a developmental perspective, the organization and integration of these systems may differ, as some may come online earlier or later in particular individuals. These differences in development can have important implications. For example, prior to puberty, genes account for essentially 0% of the variance in disordered eating, but genetic effects become important (>50%) in mid-to-late adolescence (e.g., Klump, Burt, McGue & Iacono, 2007);

805

the genetic effects are quite specific—they only occur midpuberty (Klump, Keel, Sisk, & Burt, 2010) and only for girls (Klump et al., 2012). 4. “Processes occur in a lawful way within structures that are organized and function as patterns of operating factors, where each factor derives its meaning from its relations to the others.” This tenet focuses on interdependence (Magnusson & Törestad, 1993), and it may be one of the most important tenets of the person-oriented approach. Sterba and Bauer (2010) labeled this principle pattern summary; von Eye et al. (2015) suggest a revised label that focuses on the link between theory and methods— principle of patterns as units of analysis. This tenet also relates to Sterba and Bauer’s (2010) additional sixth principle—holism (generated from other writings on the person-oriented approach; Bergman & Magnusson, 1997; von Eye & Bergman, 2003; von Eye & Bogat, 2006) wherein our understanding of any given factor is based on the relationship it has to other factors. One criticism of the person-oriented approach is that there is nothing inherently distinctive about it. In other words, if person-oriented methods have, at their core, variables, then how is this approach different? The difference lies in the last phrase of the aforementioned principle. Variables are not of interest in and of themselves; they have meaning only insofar as they are one indivisible part of a variable pattern (Bergman & Magnusson, 1997). Because any one variable cannot describe an individual or process—in fact, specific variables can have meaning for one person but not for others—the person-oriented researcher focuses on patterns, profiles, or configurations of these variables as they relate to each other and other profiles. If the profiles or patterns are meaningful, they should provide more information than a variable-oriented approach (see, e.g., Martinez-Torteya et al., 2009, on resilience). 5. “Although there is, theoretically, an infinite variety of differences with regard to process characteristics and observed states at a detailed level, at a more global level there will often be a small number of more frequently observed patterns (‘common types’). The assumption applies both intraindividually (viewed over time for the same person) and interindividually (for different individuals at the same time or over time).” This tenet clarifies the concept of individual and individuality. The person approach holds that no individual is identical to another. However, as von Eye et al. (2010) noted, some differences can be ignored, as they are not important, are not interpretable, or are not significant in all contexts or specific contexts (e.g., a GRE score of 670 vs. 671). Because

806

Person-Oriented Approaches

measurement error is inherent in any variable or process we measure, small differences are usually meaningless. These inherent properties of measurement mean that there will not be a limitless number of meaningful subgroups in any sample or population. Sterba and Bauer (2010) labeled this principle pattern parsimony. Patterns involve some level of coordination and synergy; thus, it will not be necessary to study all the components of an action (or emotion, thought, etc.). For example, Kelso (1995) proposed that dynamic systems are characterized by dissipation and disequilibrium (the analogy is to thermal dynamics). Because of this, he argued that only some factors contribute to a specific behavior. He gives the example that when a baby speaks “ba” there are 36 separate muscles involved; we do not usually need to understand the coordination of this musculature to study phonemes. In translating this tenet to research methods, consideration of the search for the precise number of subgroups in the data is important. For example, if cluster analysis is to be the statistical method used, Bergman and collaborators (Bergman et al., 2003) recommended having a sample size large enough to generate about 6–10 clusters. von Eye and Bergman (2003) suggest that known effect sizes should be taken into account in making groups, and a statistical base model, based on theoretical considerations, should also provide guidance. Three assumptions, relevant to this tenet and the search for subgroups, have been advanced (Bogat et al., 2005; von Eye & Bogat, 2006). We explain each of these, and give examples where appropriate. The first is that “a sample is analyzed under the assumption that it was drawn from more than one population” (von Eye & Bogat, 2006, p. 394). There are a few ways the researcher can address this assumption. One is to determine a priori groups (e.g., younger vs. older children, males vs. females). This approach may have merit if the a priori groups are homogeneous relative to the variables under consideration. Theory often guides the generation of hypotheses that predict a priori group differences; however, often the research on which those hypotheses is based is variable-oriented, and the prior researchers may not have searched for heterogeneity within the groups. If a priori groups are used in data analysis, the researcher should establish that the groups are homogeneous by, for example, determining that there are not outliers in the groups, validating groups on variables in the space of variables not used to make the groupings, and ensuring that dimensional identity (see Tenet 6) can be demonstrated. In other words, typical subgroups (e.g., age, gender) may not be homogeneous groups; these variables may only be one component of a pattern that defines a more consistent subgroup.

Another way to form groups is to decompose data using techniques such as cluster analysis, latent class analysis, or other methods of finite mixture decomposition. One advantage of these approaches is that they can generate groups of different sizes, including groups that may include only a few individuals; from a person-oriented perspective, this flexibility is important (von Eye, Bogat, & Rhodes, 2006). An example of this type of group formation was undertaken by Bogat et al. (2012). The relationship between stress and HPA axis dysregulation is multiply determined, affected by characteristics of the stress itself as well as the psychological consequences of the stress (see Miller et al., 2007, for a review). Research to date has typically examined one or two aspects of the stress or the psychological response to it and, predictably, findings are mixed as to whether these factors are associated with dysregulation at all or hyper- or hypo-corticolism. Bogat et al. (2012) tested whether individuals have heterogeneous responses to stress and whether profiles of women who differed on specific variables might better elucidate the relationship between stress and HPA axis dysregulation. Data from a cross sectional study of mothers (n = 182) experiencing (or not experiencing) intimate partner violence were analyzed. Initial attempts to fit the psychological, cortisol, and intimate partner violence data using structural equation modeling were unsuccessful. Thus, latent profile analysis was conducted using scores on pregnancy intimate partner violence, postpartum intimate partner violence, depression, anxiety, PTSD symptoms, the cortisol awakening response slope (awakening to 30 minutes postawakening), and the morning to evening cortisol slope (30 minutes postawakening to 5 p.m.). Models with one to six latent profiles were estimated using Mplus6. Fit indices (AIC and BIC) suggested that the model with five latent profiles, each of a different size, was the best fitting model (Figure 18.1). Latent Profile 1 (n = 51) included women with significant levels of depression and anxiety, moderate levels of IPV during pregnancy, and lack of the expected cortisol awakening response. Women in Latent Profile 2 (n = 85) were characterized by relatively healthy adaptation, with low levels of IPV during pregnancy and postpartum, the lowest levels of depressive, anxiety, and PTSD symptoms, and the expected pattern of diurnal cortisol secretion. Latent Profile 3 (n = 11) consisted of women with high levels of IPV during pregnancy and postpartum with moderate levels of depression, anxiety, and PTSD, as well as dampened cortisol levels in the morning with a flattened diurnal. Latent Profile 4 (n = 3) consisted of women with the most severe violence scores at both pregnancy and

Description of the Variable- and Person-Oriented Approaches

807

140 120 100 80

Pregnancy IPV Postpartum IPV

60

Depression

40

Anxiety 20

PTSD

0 ‒20

CAR* 1 (n=51)

2 (n=85)

3 (n=11)

4 (n=3)

5 (n=32)

Cortisol Evening Slope*

‒40 ‒60 ‒80 *Cortisol slopes are multiplied by 10 to facilitate graphic representation

Figure 18.1 Five biopsychological profiles of women experiencing intimate partner violence. See footnote 1.

postpartum, high levels of psychopathology, no cortisol awakening response; and a normal evening slope. Latent Profile 5 (n = 32) included women with moderate levels of IPV at pregnancy and postpartum, the highest levels of psychopathology but who showed a cortisol diurnal pattern similar to that of the healthy group. A third way to form groups is to ascertain patterns of function or development that characterize certain individuals (e.g., changers vs. nonchangers). An example of this type of grouping was research that identified trajectories of early reading difficulties among children studied from kindergarten to Grade 2 (Boscardin, Muthén, Francis, & Baker, 2008). Growth modeling with mixture components (e.g., Muthén, 2000, 2001; Muthén & Muthén, 2000), using 12 data points, was used to create the groups. This method was chosen, in part, because it allowed for the possibility of heterogeneous subgroups within the model. The number of classes for phonological awareness in kindergarten and word recognition development in Grades 1 and 2 were modeled separately. When these models were combined, the authors allowed for the possibility that the children might stay in the same class or change class membership over time. The final model was composed of 10 classes of students with different trajectories related to reading prerequisite skills and later deficits in reading development. Finally, groups can be formed by asking “whether groups of cases exist that deviate systematically from

predictions that can be made based on parameters that had been estimated under the assumption that only one population exists” (von Eye & Bogat, 2006, p. 394). An example of forming groups in this manner was research conducted by Bogat, Levendosky, DeJonghe, Davidson, and von Eye (2004). The participants (n = 163) were part of a larger, prospective longitudinal study examining risk and resilience among women and children experiencing intimate partner violence. This particular research examined the temporal effects of intimate partner violence on women’s mental health using women’s data from three time points: the last trimester of pregnancy (Time 1), the child’s first birthday (Time 2), and the child’s second birthday (Time 3). The authors asked whether there were groups of women that could be formed based on temporal experiences of intimate partner violence that would differ from the sample as a whole. The women were grouped into eight patterns of intimate partner violence based on their reports at each of the three time periods (e.g., 000 = no violence at any time; 110 = violence at time 1 and time 2). An 8 (patterns) × 3 (observation points) mixed effect repeated measures ANOVA was performed for each of the dependent measures (self-esteem, anxiety, depression, and posttraumatic stress symptoms), decomposing them into linear and quadratic orthogonal polynomials; this allowed for the examination of different mental health trajectories. Here,

808

Person-Oriented Approaches

we focus on whether there were group differences in the time × pattern interactions. There were for self-esteem and depression but not for anxiety and posttraumatic stress disorder symptoms. For example, regarding self-esteem, across all participants, self-esteem increased during Time 1 and Time 2 and then tapered off at Time 3. There were, however, group differences regarding the quadratic polynomials. The no violence group (i.e., 000) showed a steady, linear increase; the chronic violence group (i.e., 111) showed an increase followed by a decline; and the group with only Time 2 violence (i.e., 010) showed high levels of self-esteem, which then tapered off. For depression, across all participants, symptoms declined over time. However, there was a significant group × quadratic effect for three groups (chronic violence, 111; Time 2 violence only, 010; Time 3 violence only, 001), indicating an increase in depressive symptoms at Time 3. Thus, examining mental health trajectories over the course of more than two years indicated that the chronicity and timing of intimate partner violence mattered; that is, in this sample, there were subgroup differences based on intimate partner violence experiences. The second assumption relevant to this tenet is that “attempts are made to establish external validity of the groupings.” If methods such as those suggested above are used (e.g., cluster analysis, latent class or profile analysis), a relatively small number of groups are generated with group-specific profiles. External validity can be established by comparing these groups on variables not used to constitute the group profiles. Different techniques for group comparisons are available. For example, in latent profile analysis, covariates can be included within the model to determine uncertainty within profiles (Roeder, Lynch, & Nagin, 1999). Another approach, to validate latent classes, involves rerunning the final latent class model by adding additional variables, on which groups are expected to differ. One can then conduct difference tests to compare models in which those additional variables are allowed to vary across the classes with models that assume the additional variables do not vary (e.g., see Ansara & Hindin, 2010). Finally, the third assumption is that “the groups are interpreted based on theory.” It is not difficult to find groupings of individuals in a data set, but whether these groups are distinct in important ways only can be ascertained relative to theory. Bergman, Magnusson, and El-Khouri (2003) indicated the importance of either a priori meaning of the groups or the ability to interpret the groups based on extant theory. A recent paper, examining profiles of behavioral and adrenocortical functioning among 7-month-olds,

illustrates these three assumptions concerning groups and the person orientation (Towe-Goodman et al., 2012). The authors employed latent profile analysis to generate profiles of infants, covariates were used to test the validity of the profiles, and theory was used to interpret the results. We briefly describe the rationale for the hypotheses, the findings regarding the profiles, and the interpretation of one aspect of the findings. Understanding relationships among the biological and behavioral indicators of coping with stress is important as problems in these domains in early childhood are associated with vulnerability or resilience to stress across the life span (Davies, Sturge-Apple, Cicchetti, & Cummings, 2007; Gunnar & Quevedo, 2007). However, little is known about the coordination of these systems in infancy, and research has failed to find consistent relationships between these systems (e.g., Keenan, Grace, & Gunthorpe, 2003; Ramsay & Lewis, 2003). The lack of reliable findings suggested to Towe-Goodman et al. that there might be patterns or configurations of adrenocorticol and behavioral stress responses among infants exposed to interparental conflict. This was the case; a four-class solution provided the best fit to the data. Compared with the reference group (“low reactors”), children who were in the “high cortisol reactivity, moderate negative behavior” group were more likely to have higher negative affect and to live in households where conflict occurred as well as less likely to demonstrate successful self-regulatory behaviors under stress. The findings were interpreted, in part, based on the emotional security hypothesis (Davies & Forman, 2002). Children might maintain emotional security in households with high levels of parent conflict by minimizing overt displays of distress, even when they are experiencing high physiological arousal. This behavioral suppression may result from learning that it is dangerous to draw attention to themselves, as this might lead the parents to aggress toward them. These are the basic tenets of the person orientation. Other important additions to these tenets have been suggested; we discuss two of them. 6. Dimensional identity. There have been developing statements about the meaning of dimensional identity. The expression originated with Schmidt (1977), who postulated that scales used for comparison be commensurable. Von Eye and Bergman (2003) suggested that scales used for inter- or intraindividual comparisons should also have the same dimensional structure and be invariant over time. If dimensional identity is not given, scales cannot be used for comparison. The result that dimensional identity does not exist is of interest in itself. It points to the possibility

Description of the Variable- and Person-Oriented Approaches

that multiple populations exist. In addition, this result indicates where these populations possibly differ from each other. 7. The ecological context of human behavior. Three additions to the original person-oriented tenets, relating to the ecological context, were developed by Bogat (2009), using the basic structure and language of the original tenets put forward by Bergman and Magnusson (1997). 1. The structure and dynamics of individual behavior are, at least in part, specific to the environment in which the individual lives and works. 2. There is lawfulness and structure to (a) intrasystemic constancy and change and (b) intersystemic differences in constancy and change. These processes can be described by patterns of the involved factors. 3. Validity is specific to individuals and environments. In discussing the individual as a holistic, dynamic being, Bergman and Magnusson (1997) emphasized that an individual’s behavior can be influenced by factors from the neuronal to the cultural. The setting in which an individual works, lives, or loves is as much a factor in his or her development as cognitive, biological, and psychological processes. The effects of context can be dramatic. For example, marriage dampens criminal behavior, an effect that seems to be contextual rather than based on an individual personality characteristic (Horney, Osgood, & Marshall, 1995; Sampson, Laub, & Wimer, 2006). Bronfenbrenner and Morris (1983) defined context as “any event or condition outside the organism that is presumed to influence, or be influenced by, the person’s development” (p. 359). Within this broad definition, Bronfenbrenner (1977) elaborated four nested levels of the environment: (1) the microsystem (social settings in which the individual is embedded, such as the family); (2) the mesosystem (the connections between the various microsystems, such as the family’s relationship with the church); (3) the exosystem (larger structures that subsume the micro and mesosystems, such as neighborhoods); and (4) the overarching macrosystem (i.e., culture and its various components, such as political systems). Later formulations included a fifth level—the chronosystem— which described how interactions within or among the four other levels, were influenced by time (Bronfenbrenner, 1986a, 1986b). This five-level, broad contextual rubric helps focus the researcher’s attention on the wide array of contexts that can affect an individual; however, which specific contexts should be measured or included in a specific research study are less obvious. Bronfenbrenner’s contextual levels focus on the social environment; however, others have argued that

809

the physical environment might also play an important role in child development (e.g., Widaman, 2007). Unfortunately, of all the components of the personoriented approach, the contextual component is probably the least developed and understood. In the limitations section, we explain why this might be the case. Limitations of the Person-Oriented Approach Here we discuss three limitations to the person-oriented approach: (1) generating testable hypotheses; (2) conducting pattern analyses; and (3) measuring and analyzing the contexts in which human behavior occurs. The first limitation is that the person-oriented approach does not lead easily to testable hypotheses. Bergman and colleagues (Bergman & Anderson, 2010; Bergman & Trost, 2006) argued that, in part, hypothesis generation is difficult because the person-oriented approach “is quite general and its translation into specific theory for a scientific problem can be difficult” (p. 158). On this point, we take a slightly different perspective. If the person-oriented and variable-oriented approaches are considered general paradigms for scientific inquiry, then, in and of themselves, they do not lead to specific testable hypotheses. Rather, they provide the structure under which theories and hypotheses are generated. However, even so, hypotheses from a person-oriented approach may be difficult to generate because there has not been enough attention to the match between theory and methods. Sterba and Bauer (2010) rightly object to the casual use of the term person-oriented any time individuals are grouped. They emphasize the importance of understanding what person-oriented methods can be used to test what principles of the person orientation. For example, they propose a matrix in which the person-oriented principles are crossed with four person-oriented methods (i.e., less restrictive variable-oriented approaches, classification, hybrid classification, and single subject). They conclude that the correspondence between principle and method may sometimes be a match (e.g., pattern summary × single subject), sometimes a mismatch (e.g., pattern summary cannot be tested with less restrictive variable-oriented methods), or sometimes a limited or conditional match (e.g., individual specificity is only conditionally testable with hybrid classification models). Of course, as statistical techniques evolve, these correspondences may change [e.g., in the same special issue in which Sterba & Bauer’s article appeared, Molenaar (2010) argued that all person-oriented principles can be tested with dynamic factor analysis]. Generating testable hypotheses using a person-oriented approach may also be difficult because most extant research

810

Person-Oriented Approaches

is variable-oriented. As Bergman and Andersson (2010) note, it is often difficult to use variable-oriented findings to develop person-oriented theory and methods. However, sometimes when an accrual of variable-oriented research on a specific topic generates contradictory findings in different samples or when different variables are included in the analyses, a reconceptualization of the research using person-oriented theory and methods may be beneficial and advance the science (e.g., the infant and adult cortisol research, cited earlier). The second limitation involves how to conduct data analysis of patterns (Bergman & Andersson, 2010). Personoriented theory focuses on the importance of understanding patterns; however, the statistical techniques for analyzing patterns are underdeveloped. For example, Bergman and Andersson (2010) noted that many current data analytic techniques to determine patterns depend on dissimilarity matrices (e.g., cluster analysis), yet there is not widespread agreement on what the appropriate measure of dissimilarity should be. Configural frequency analysis is a particularly promising technique because it creates all possible patterns within the variable space; the researcher can then identify which patterns occur more or less frequently than expected under a specific hypothesis. In addition to understanding patterns, in and of themselves, there is the additional problem of modeling dynamic processes in those patterns that occur over time. Kelso (1995) noted that the principles of dynamic pattern formulation confront us with two problems. The first is the problem of complexity of substance—how a pattern is constructed from a large number of components. The second is the problem of pattern complexity—how patterns accommodate complexity in whatever function they ultimately serve (e.g., by changing, by selection, by bifurcation). In the field of developmental psychopathology (as in most fields), we are far from understanding how patterns form and how they operate. However, we can assume that the relationship of the components of patterns as well as the relationship of patterns with other patterns is most often nonlinear. Thus, new data analytic techniques will be necessary to capture these processes. Dynamic systems modeling approaches, focused on recursive differential and difference equations, show some promise, and there has been, of late, much interest in this strategy (e.g., Kunnen, 2012; Valsiner, Molenaar, Lyra, & Chaudhary, 2009). Other approaches, for example, state-space nonstationary time series modeling (e.g., Molenaar, Sinclair, Rovine, Ram, & Corneal, 2009), which puts no constraints on the parameter estimates, has been used in idiographic approaches.

Analyzing patterns involves decisions about which variables are essential to constitute the patterns, since we cannot measure everything in every research study. In discussing patterning of contextual influence, Widaman (2007) suggested three contextual patterns to consider. Big Bang events, which are likely to be one-time, unrepeatable, stressful life events (e.g., death of a parent), are known to have a strong influence on development, but only in the context of other factors (e.g., age of child, socioeconomic status). Repetitious, smaller events, as the name implies, occur frequently, over the course of time. Big Bang events correspond to major life stressors; repetitious smaller events correspond to daily hassles. Finally, there are intermittent but salient events (e.g., inconsistent parenting behavior). The third limitation concerns context. What is context and how should it be measured? What contexts are most relevant to assess in the field of developmental psychopathology? These are not simple questions, and it is beyond the purview of this chapter to answer them fully. However, here we address several problems with the conceptualization and measurement of context. One problem faced by the researcher is how to define context in terms of how it differs from as well as might be dependent on individual-level data. For example, Oakes (2009), cited the work of Coleman (1990), who put forward a simple model of how social change occurs: society is composed of individuals and individuals constitute society. That is, “Institutions and other social phenomena play a key role in analyses, but these phenomena must be grounded to the activity of individuals. . . . Group-level phenomena are never simple aggregations but rather complex dynamic and multilevel phenomena” (Oakes, 2009, p. 364). There are conceptual as well as statistical problems associated with determining and measuring context. Sometimes context is “defined purely on the aggregate” as in whether or not there is a Head Start program in a specific community; sometimes context is a summary statistic of an aggregate of individuals (e.g., prevalence or incidence data, neighborhood socioeconomic status; Greenland, 2001). Greenland (2001) offered an example from the field of epidemiology regarding how aggregate measures can introduce problems of cross-level confounding when measuring the effect of context. Table 18.1 (from Greenland, 2001) presents data for two areas, A and B (which differ on a contextual variable, such as a law or social program), and their influence on the rate of a health outcome (e.g., smoking). The rates in Panel 1 indicate no contextual effect; in essence

Description of the Variable- and Person-Oriented Approaches

811

TABLE 18.1 An Example Demonstrating the Complete Confounding of Contextual and Individual Effects in Ecologic Data Area A

Area B

X=1

X=0

Total

X=1

X=0

Total

? 60 ?

? 40 ?

560 100 5.6

? 40 ?

? 60 ?

560 100 5.6

2. Possibility 1 (RRx = 2, RRA = 7/8) Y=1 N Rate

420 60 7.0

140 40 3.5

560 100 5.6

320 40 8.0

240 60 4.0

560 100 5.6

3. Possibility 2 (RRX = 1/2, RRA = 8/7) Y=1 N Rate

240 60 4.0

320 40 8.0

560 100 5.6

140 40 3.5

420 60 7.0

560 100 5.6

1. Ecologic (marginal) data Y=1 N Rate

Note: The ecologic data cannot identify the effect of group (A versus B) on the rate of the outcome Y = 1 when only a marginal summary of the individual-level covariate X is available. N = denominator (in thousands of person-years); RRA and RRX are the rate ratios for the true effects of A versus B and of X = 1 versus X = 0, respectively. Source: Greenland, S. (2001). Ecologic versus individual-level sources of bias in ecologic estimates of contextual health effects. International Journal of Epidemiology, 30, 1343–1350. Reprinted with permission from Oxford University Press.

RRA = 1 (the true effects of A vs. B). Because the presence of a risk factor (X) is unknown, it cannot affect RRA . Panels 2 and 3 show circumstances under which the same result as Panel 1 can be obtained, even though risk and context are different. For example, in Panel 2, area A has a positive contextual difference (RRA = 7/8; that is 1). Panel 3 demonstrates how this confounding can take place when the situation is reversed; that is, when the contextual difference is negative (RRA = 8/7) and there is a higher prevalence of a beneficial factor (RRX = 1/2). Greenland argued that in some cases, the differences in risk factors between areas A and B might be negligible. However, he noted that in other cases, the differences might be substantial, for example, in the case of alcohol consumption. At moderate levels alcohol consumption may be protective, whereas at high levels it may be detrimental. Thus, without knowing more about the distribution of moderate and heavy drinkers within an area, the direction or the strength of confounding of RRX on RRA would be difficult to determine. Another problem is that the meaning of context may vary over time. We noted earlier that marriage itself is associated with decreased criminal behavior. However, Gottfredson (2005) suggested that marriage is an age-dependent variable, and the meaning of marriage will differ accordingly. “At very young ages, marriage may

index low self-control . . . , during the early to late thirties higher levels [of self-control] . . . And repeated marriage may signify low rather than high self-control” (p. 51). This presents a very specific problem of dimensional identity, which may have strong effects on the data, and, thus, the conclusions that are drawn from them. Although developmental psychology and developmental psychopathology researchers study contextual factors (e.g., parenting, SES, prenatal programming), the focus is usually on a specific variable; rarely is the focus on the associations among multiple factors and levels in which individuals exist. For example, of the articles published in psychology from 2000–2005, at least one-third that had a significant focus on race, did not have a significant focus on socioeconomic status (Carlson, 2006, as cited in Cauce, 2008). Culture, race, and ethnicity are important contextual factors, yet, at present, we understand little about how they transact with individuals. The field of developmental psychopathology often focuses on identifying early risk factors that predict later problem behaviors. Sampson and Laub (2005) argued that sometimes the search for childhood risk factors may be misguided, especially when the focus is on “early childhood attributes that are presumed to be stable over the life course” (p. 40). Their research, examining criminal behavior among males from age 7 to 70 (Laub & Sampson, 2003), found that risk factors generally did not predict future behavior. When the sample was divided into

812

Person-Oriented Approaches

high- and low-risk groups (based on early detrimental family and individual factors), both groups had identical trajectories of criminal behavior—criminal behavior peaked at age 15 and then desisted over time to negligible levels. Decomposing the sample another way, a five-group solution was found; each group had a slightly different trajectory of peak and drop off of criminal activity. However, similar to the two-group solution, criminal activity declined for all groups over time. In addition, the authors found no evidence that two childhood risk factors (parental deviance and mental health) differentiated the five groups in meaningful ways. One solution to the focus on risk perspective is for person-oriented research to examine positive development. For example, Richard Lerner’s (e.g., Lerner, Taylor, & von Eye, 2002; Lerner, von Eye, Lerner, & Lewin-Bizan, 2009) positive youth development model focuses on strengths, including those of the adolescent (school engagement, self-regulation, and hope about the future) and the environment (social networks, individuals, institutions, and resources). Healthy development (as exemplified by competence, confidence, character, caring, and connection) occurs as a result of positive synergy between youth and environmental assets. This section has elucidated the principles of both the variable- and person-oriented approaches as well as outlined limitations of each. In the following section, we discuss the genesis of many of the tenets of the person-oriented approach by focusing on the theories and research of three seminal psychologists. THREE EARLY PROTAGONISTS OF THE PERSON-ORIENTED APPROACH In this section, we discuss the work of William Stern, Kurt Lewin, and Jack Block, three key personality theorists and researchers, whose ideas presaged the modern-day person-oriented approach. In examining their work, it is evident that the field of personality has debated the merits of some variant of the person-oriented vs. the variableoriented approach, since its inception (e.g., holistic vs. individualistic; idiographic vs. nomothetic; traits vs. types). We recognize that there are other important figures (e.g., Allport, Binet, Brunswik, Cairns); however, the scope of this chapter does not allow us to cover the entire history of the field. The three protagonists worked during the ascendance of the variable-oriented approach. Various historical explanations have been offered as to why the variable-oriented approach came to dominate psychological science. One

idea is that a consensus was reached in psychology regarding which statistical methods were scientific. For example, Stam (2010) focused on the Galtonian view in psychology, where “numbers . . . represent some empirical reality” (p. 147, italics in the original). He argued that this focus moved psychologists away from people and toward characteristics of tests, as aggregation occurred hand-in-hand with this representation. Lamiell (2009) discussed attempts to integrate experimental and correlational methods, which resulted in neo-Galtonian notions that “the statistical concepts and methods common to both forms of inquiry—involving the analysis of means, variances, and covariances (i.e., correlations)—made the merger of these two research programs both possible and eminently sensible” (p. 36). Another idea, advanced by Toomela (2009), suggests that causality, understanding, and explanation, after World War II, focused almost exclusively on linear models. In doing so, psychology adopted simple, mechanistic models and rejected more complicated models for understanding human behavior (see also Stam, 2010). Magnusson (2000) focused on the move to a stimulus-response paradigm in the 1950s and 1960s, which focused on molar behaviors rather than a holistic, integrated formulation of human behavior. As we mentioned earlier, Kuhn (1962/2012) argued that paradigms stay dominant, in part, because researchers are able to integrate anomalous findings into these paradigms. This certainly occurred with the personand variable-oriented approaches. Lamiell (2003) provided an example in his discussion of Allport and McClelland’s perspectives on the findings of Hartshorne, May, and Shuttleworth (the Character Education Inquiry; Hartshorne & May, 1928, 1929; Hartshorne, May, & Shuttleworth, 1930). In this classic research, numerous components of specific child traits (e.g., helpfulness, dishonesty) were measured across settings; however, the correlations among these components were negligible, leading Hartshorne and May (1928) to conclude that these were not traits, but simply habits. Allport (1937), attempting to integrate nomothetic and idiographic approaches, suggested that the findings proved “only that children are not consistent in the same way, not that they are inconsistent with themselves” (p. 250). In other words, there might be unmeasured traits influencing the results (an early example of the lurking variable argument), and/or that individual children might have unique motives for the same behavior (e.g., one child might steal because of a bravado trait, another because of a social inferiority trait). Alternatively, as Lamiell (2003) pointed out, McClelland (1951), defending the nomothetic tradition, argued that the research had methodological

Three Early Protagonists of the Person-Oriented Approach

flaws; the children were too young to demonstrate traits (which were learned from past experiences), a significant methodological problem.6 The three historical figures we discuss here were immersed in the debates about the prominence the individual should have in psychological science. They all came down on the side of the individual. William Stern William Stern (1871–1938) is well known in the field of psychology for specific contributions (e.g., developing the intelligence quotient), yet he is almost unknown as relates to his philosophical writings on critical personalism. In fact, Lamiell (e.g., 2009) suggested that Allport might have been one of the few psychologists who valued Stern’s theoretical contributions.7 His lack of renown is unfortunate, given that Stern’s ideas foreshadow important principles of the person-oriented approach. Stern began his career by developing the field of differential psychology; that is, the study of individual differences. to study individual differences, differential psychologists developed tests and compared individuals and/or groups on these tests. For example, as is well known, in the late nineteenth and early twentieth century, Binet and Cattell were developing intelligence tests and administering them to different individuals and populations. Stern’s book on differential psychology first appeared in 1900, with a revision in 1911. He was interested in specifying differences between individuals and groups, ascertaining what caused these differences, and how the variations were demonstrated; however, although the book dealt with both inter- and intraindividual differences, Stern’s emphasis was interindividual differences (Lamiell, 2010a). Stern became concerned with the testing movement (especially America’s obsession with intelligence testing and its reliance on norms determined through aggregation), and ultimately repudiated his own contributions to the field (Lamiell, 2010a). In the process, Stern embraced his interests in metaphysics and philosophy, articulating his theory of critical personalism—“a genuine differential psychology” (Stam, 2010, p. 149). For Stern (1917/2009), the individual was a unitas multiplex; that is, multiple components that are organized 6

The reader may be interested in a more recent exemplification of fitting anomalous data into the dominant paradigm of personality research (Borkenau & Ostendorf, 1998). 7 Over the last 30 years, James Lamiell has, nearly single-handedly, brought Stern’s ideas to an American audience.

813

in a coherent, and purely teleological, manner. The idea of the individual as unitas multiplex augurs Bergman and Magnusson’s (1997) second and third tenets concerning the complexity of process and the meaningful coherence in development and functioning. Stern was concerned that psychology had focused primarily on elucidating the multiple components of human experience at the expense of understanding the “unified individual.” Stern noted that simply aggregating all the components of the individual did not, in and of itself, explain higher-order processes. The unity of the individual was to be understood on four levels: “phenomena (mental experiences), acts (doings), dispositions (strivings, capabilities), and subjects (I)” (Stern, 1917/2009, p. 115, n.b. all quotes below from this source). Although heuristically, one could look at each in isolation, none could be understood fully without an appreciation of how their interrelationships created a totality. By phenomena, Stern meant individual self-perceptions, such as ideas or feelings. Acts involved the interconnections and unifications of self-perceptions that occurred within a short time frame or within the moment. However, such a temporal framework, if taken in isolation, provides a misleading picture of the individual. Stern extended the time frame by considering dispositions, enduring qualities or attributes of the individual that predispose him/her to various perceptions. The relationship between these three components is as follows: Acts mediate the relationship between dispositions and perceptions. For example, Stern noted that intelligence “is not the power to set into course thoughts or thought sequences of a particular kind, but instead the disposition to adjust oneself through acts of thought to new situations and demands” (p. 122). The fourth component was the I—the unification of the dispositions. Stern defined the I as existence itself. “The I experiences phenomena, executes acts, [and] owns dispositions” (p. 124). The I included both the psychological and physical aspects of the person. From these four components, Stern defined personality as “an entity that we encounter as a living whole which, that in and of itself strives for and is capable of realizing the enduring purposes of that whole” (p. 131). This definition puts front and center Stern’s focus on the teleology of the personality system. The personality was both selfand other-oriented (autotelie and heterotelie, respectively), and worked to integrate these two purposes, even though they were sometimes in opposition. Stern referred to this as introception. In addition, personality was not fixed by the four components, they simply potentiated the individual’s transactions with the environment. Here Stern’s writings

814

Person-Oriented Approaches

relate to the ecological context tenet of the person-oriented approach. The transaction of the person and the environment, he termed convergence. In Stern’s view, an individual’s goal focus was insufficient; it had to be augmented by the environment. The environment would influence the individual, but the individual’s specific dispositions would influence his/her experience with the environment. It is apparent that Stern’s theories relate closely to the tenets of the person-oriented approach. In addition, his delineation of research methods (separating, for instance, research on the individual and research on groups of individuals), also echoes person-oriented approaches. As Lamiell (2010b) noted, Stern believed in large-scale research and the use of modern statistical methods; at the same time, he understood that there was a distinction between knowledge about an individual and knowledge collected across a group regarding specific attributes that individuals might have. This concern presaged the dilemma of the ecological fallacy, discussed earlier, and specifically relates to Jack Block’s writings decades later that the search for universal traits is problematic because they say little about the individual. Kurt Lewin Lewin (1890–1947) and Stern knew of each other’s work (Stam, 2010). Although their theories diverged in important ways, they shared a concern regarding psychology’s misguided focus on aggregation and the search for universals and lawfulness. Lewin (1931) believed that psychology was tethered, in problematic ways, to Aristotelian logic. He was concerned that psychology focused on regularity, specifically in the sense of frequency. Thus, psychologists were primarily concerned with lawfulness (regularities across individuals) and unconcerned with the individual case. In fact, “the individual event seems to him [the psychologist] fortuitous, unimportant, scientifically indifferent” even though it may have “critically determined the destiny of the person involved” (p. 151). Lewin argued that the focus on lawfulness prevented psychology from investigating large swaths of human experience (e.g., will and affect), because these types of experiences were likely to occur less frequently than other experiences. As is obvious, Lewin, like Stern, was concerned that psychologists discounted the importance of the individual, the first tenet of the person-oriented approach. According to Lewin, in their quest to discover universal laws of human behavior, psychologists established groupings of convenience, leading to incomplete, mechanistic understandings. For example, he wrote:

Whatever is common to children of a given age is set up as the fundamental character of that age. The fact that three-year-old children are quite often negative is considered evidence that negativism is inherent in the nature of three-year-olds, and the concept of a negativistic age or stage is then regarded as an explanation (though perhaps not a complete one) for the appearance of negativism in a given particular case. (p. 153)

Again, these sentiments reflect tenets of the person-oriented approach, including avoiding the ecological fallacy, the importance of creating meaningful groups, and a focus on understanding process rather than simply mechanisms. Lewin also felt that psychology’s concentration on classes and frequency of behavior led to the use of statistics that obscured individual differences. Although psychology was preoccupied with developing fundamental principles, in practice, Lewin believed, research described averages across individuals. “It is forgotten that there is just no such thing as an ‘average situation’ any more than an average child” (p. 172). Lewin’s critique of the use of averages in statistical methods focused on the lack of precision that was inherent in this approach. He wrote: “The demands of psychology upon the stringency of its propositions go no farther than to require a validity ‘in general,’ or ‘on the average,’ or ‘as a rule.’ . . . psychology does not regard exceptions as counter-arguments so long as their frequency is not too great” (p. 156, italics in the original). Here again, Lewin’s ideas foreshadow the importance of avoiding the ecological fallacy and the need to focus on individual behavior. Lewin (1931) anticipated that psychology would move from an Aristotelian to a Galilean concept formation; he based this idea on Ernst Cassirer’s analysis of t he fields of mathematics and physics. In part, a Galilean formulation would lead psychologists to develop theories that provided complete, not partial, explanations for the phenomena under study. Central to these theories was the notion that the causes of human behavior were multiply determined, and did not rest exclusively within the individual. He noted that numerous factors might affect an individual, including his or her personality, the groups within which an individual was embedded, culture, and economic forces. Lewin (1939) was greatly influenced by the Gestalt psychologists; however, he felt that the whole and its individual parts were of equal importance. “The whole is not ‘more’ than the sum of its parts, but it has different properties. The statement should be: ‘The whole is different from the sum of its parts’” (p. 885). Lewin’s theory of multicausality ultimately evolved into his “field theory,” which proposed that human behavior resulted from the interactions among many different

Three Early Protagonists of the Person-Oriented Approach

systems; that is, a dynamic field. Lewin’s (1936) famous formula, expressed in mathematical terms, stated that behavior is a function of the person and the environment: B = F (P, E). Lewin argued that the focus for psychological research was on the life space, which consisted of the needs of the individual as well as the environment in which he or she existed [B = F (P, E) = F(L Sp)] (Lewin, 1939). The focus on multicausality and the inclusion of the environment, as well as the person, in theory and method, are preludes to the second and third tenets of the person-oriented approach—that processes involve many interacting factors, and that within these interactions there is coherence and lawfulness. Lewin’s work also reflects the fourth tenet with its focus on the importance of understanding each factor only in relation to other factors. Lewin (1931) was also an experimentalist, who believed in creating ideal or paradigmatic situations rather than simply recreating conditions that occurred frequently. The general validity, for example, of the law of movement on an inclined plane is not established by taking the average of as many cases as possible of real stones actually rolling down hills, and then considering this average as the most probable case. It is based rather upon the ‘frictionless’ rolling of an “ideal” sphere down an ‘absolutely straight’ and hard plane, that is, upon a process that even the laboratory can only approximate, and which is most extremely improbable in daily life. (p. 161)

For example, Lewin, Lippitt, and White (1939) studied aggressive behavior by organizing clubs of 10-year old boys that were guided by three types of leaders (i.e., laissez-faire, authoritarian, and democratic). Process data were collected, including a written verbatim account of all conversations, visual recordings, and moment-by-moment structural, quantitative, and interpretative accounts of group process. The authors found that most of the acts of aggression occurred during emotion-laden moments (e.g., criticism of the group from an outsider). These acts of aggression seemed to be spontaneous acts resulting from a build-up of tensions based on leadership style and the restriction in “space of free movement” that resulted from that style. Other factors considered important were the “rigidity of the group structure,” which provides a restraining force but also increases tension within the individual, and whether the child came from a culture where aggression was an accepted way of acting. This is necessarily a very brief review of Lewin’s theory. We do not discuss the aspects of his theory that are, at present, most well known, including action research, group process, and leadership styles. What is most

815

interesting to us is that the criticisms of psychology that Lewin raised, which are similar to those raised by advocates of the person-oriented approach, are largely forgotten. We now turn our attention to Jack Block. Jack Block Block (1924–2010) is often rightly cited as one of the fathers of the modern person-oriented approach. He is perhaps most widely known today as a contrarian (a label he gave himself) who offered strong critiques of the Five-Factor Model (FFM) of personality (e.g., Block, 1995, 2001, 2010), as it came to dominate the theory and research of personality. Although researchers have acknowledged the validity of some of his critiques of the FFM (e.g., McAdams & Walden, 2010), the overarching concern in which the specific critiques are subsumed—that variables, rather than people, had become the focus of personality research—is often overlooked or downplayed. Block began his career at a time when the debate about how personality was to be defined (whether there were universal truths about personality as well as which statistical techniques provided the best way to understand personality) was quite active. In 1946, Cattell defined the R-technique as an approach to studying personality that focused on correlating variables; hence, a variable-oriented approach to analyzing data. Alternatively, the Q-technique correlated persons. Some researchers (e.g., Eysenck, 1954) felt that the two approaches were the same; that is, they always led to similar results. Along with others (e.g., Cronbach, 1953; Stephenson, 1953), Block (1955) disputed this assertion, arguing that depending on the heterogeneity or homogeneity of the sample, the results would be different. In other words, in a heterogeneous sample, composed of more than one subgroup, it was possible that the covariance structures for specific variables would differ by subgroup. In homogeneous samples, the covariance structures would be the same. Writings throughout his career reflected Block’s concern that personality psychologists focused too much attention on variables and not enough on individuals. His strongest defense and advocacy of the person orientation can be found in Lives Through Time, Block’s (1971) seminal investigation of personality development. Many of the principles of the person orientation that he elaborated can be seen in the following quote from this book, when he wrote that psychologists should aspire to understand personality in the way that laypeople do, which is “[a] configuration and systematic connection of personality variables as these dynamically operate within a particular

816

Person-Oriented Approaches

person” (Block, 1971, p. 13),8 a point made by some current personality researchers, as well (e.g., Donnellan & Robins, 2010). There are a number of important components to this statement. First, laypeople naturally discuss people, not variables; whereas the opposite tends to be true of many psychologists. Block acknowledged that this problem was not unique to the study of personality. For example, he cited an article wherein the authors ask “where is the perceiver in perceptual theory?” The second important aspect of this statement is that personality consists of a configuration of characteristics. One characteristic cannot describe someone’s personality adequately. A person who is quiet in large social situations may be perceived as shy; however, when that same person is seen talking freely and intimately among his close friends, the observer may need to adjust his or her perception. Finally, the third component of Block’s statement speaks to the importance of configurations of personality variables operating in a dynamic fashion. He argued that personality theory put too much emphasis on examining stability (i.e., uniformity of development across people and time) at the expense of overlooking instances of change and what drives such change. For example, when personality psychologists find differences in individuals across time, they are likely to ascribe these to measurement error in the assessment instrument (in particular when the instrument was constructed based on classical test theory), random error, or the choice of an unimportant (i.e., nonuniversal) variable. Block disputed this bias by demonstrating in his longitudinal research that there were groups of individuals whom he labeled changers and nonchangers; importantly, these subgroups were better able to explain personality development compared to a priori groupings based on such variables as gender. Block recognized that there was strong pressure among personality researchers to adhere to the uniformity myth.9 Many researchers had argued that if uniformity across people did not exist, the result would be a lack of lawfulness; that is, the discovery of too many different types 8

Block was not arguing that laypeople should provide scientific definitions of personality for the researcher. In fact, one of his ongoing critiques of the FFM was that the list of adjectives used to define the traits had been derived from undergraduate students asked to list traits that best described themselves and their acquaintances. 9 The term uniformity myth, as discussed earlier, was also used by psychotherapy outcome researchers to suggest the folly in assuming that clients, therapists, and behavior were similar and that individual differences did not matter (e.g., Kiesler, 1966).

of personality would undermine the search for universal principles or types. Block argued against this perspective by decoupling lawfulness and universality. He wrote, when the two are conflated it means that “the seeking of one may preclude finding the second . . . [in fact, in not doing so] a greater lawfulness may be discerned as what is general and undifferentiated is partitioned into smaller but more homogeneous classes for study” (1971, p. 11). In fact, Block believed smaller groups were likely of great interest to the researcher, and, furthermore felt that random samples might not include certain types of infrequently occurring individuals. He felt so strongly about this that his recommendation for future longitudinal researchers was to not recruit a random sample, but rather to focus theory and research questions on a relevant, non-random, sample. As mentioned earlier, Block felt that theory and research were synergistic. In other words, it was inevitable that researchers holding variable-oriented theories would find support for these theories, as they employed variableoriented assessments and data analytic techniques that could not find support for person-oriented theories. This was a point he made continually throughout his life. For example, in his critique of the FFM, he eschewed the notion that five personality factors derived for adults could be readily transposed onto childhood personality (Block, 2010). In other words, he implied that researchers who believed in stability rather than change and who focused on variables, rather than people, might conclude (falsely) that adult and child personality traits were interchangeable. Thus, early in his career, to examine the person in personality, Block turned to a new assessment tool just gaining popularity. Q-Technique Although often attributed to Block, the Q-technique (also called Q-set or Q-sort) was not invented by him. Stephenson (1935) introduced this technique in an article aptly titled “Correlating Persons Instead of Tests.” Stephenson was concerned with a specific problem within psychology—that large numbers of participants were needed to collect the data that would allow psychologists to correlate tests or estimates and then conduct factor analyses, especially as related to Spearman’s g, c, and w factors. He went on to argue that laboratory experiments become almost “burdensome” because of the need to collect data from so many individuals. He proposed the Q-technique as an antidote to this problem: “Can we make factor studies on a few individuals, and bring the

Three Early Protagonists of the Person-Oriented Approach

methods of correlations and factor analysis into the laboratory?” (p. 18, italics in original; see also Nesselroade & Molenaar, 2010). His solution was to invert the typical process. Instead of obtaining a few tests on a large number of people, he proposed that psychologists obtain many tests on a few people. Stephenson’s demonstration of his new technique was a study on taste (i.e., aesthetics), what he considered an important, but understudied, facet of an individual’s personality. He took 60 color samples (no blacks, pure grays, light creams or whites were included) and gave them to 12 women and eight men with the instruction to grade each quickly and spontaneously from best to least liked. His small experiment found two factors; that is, two groups of individuals whose preferences correlated positively with each other. Factor 1 was composed of 12 individuals who preferred the subtle colors (which Stephenson labeled as the most tasteful); Factor 2 was composed of eight individuals who preferred the pure bright colors (which Stephenson labeled as vivid but crude). Interestingly, although Stephenson’s procedure had the participants sort the cards (self-report), Block’s initial Q-set was developed so that professionals could make appraisals of others (other-report). Also, although Stephenson developed this technique as a special type of factor analysis, the method became used as an assessment tool unrelated to factor analysis (e.g., Rogers, 1961). Block’s Q-Set Block’s (1961) original purpose in developing his California Q-set was to define more precisely the characteristics of personality widely used by professionals (e.g., authoritarian personality) as an aid to future research, including better communication and precision among scientists. Thus, as stated already, the Q-set was given to professionals to evaluate patients or research participants. Although it had earlier permutations, the Q-set Block presented in his 1961 monograph consisted of 100 statements, each of which described a specific personality trait (e.g., “is an interesting, arresting person; seeks reassurance from others; emphasizes being with others; gregarious”). Although the statements do not represent conceptualizations from any specific theoretical orientation, Block felt they reflected current informed opinion among professionals in personality psychology, clinical psychology, and psychiatry. Block conceived of the standard statements as a language and the sorting algorithm of the traits as a grammar. Each observer is given the 100 statements on separate cards; the task is to sort the statements into nine categories where one represents the category least representative of

817

the person and nine represents the category that is most representative. The number of statements that the observer places in each category differ and range from 5 to 18. Block saw the Q-set as offering both ipsative and normative data. That is, the information provided a distinct profile of characteristics specifically about the individual being assessed; however, across all the participant data, Block was able to find representative types. The normative data were particularly interesting over time; for example, certain of the 100 characteristics in the Q-set were rarely if ever used to rate individuals at specific stages of development. As noted earlier, Block was essentially concerned with issues of change and stability over time, as he felt that such knowledge would lead us to understand the developmental paths that lead to the same outcome (equifinality) or similar developmental paths that lead to different outcomes (multifinality). To study trends, Block developed profiles of personality variables based on their correspondence, salience level, and salience heterogeneity over time. This resulted in seven categories. For example, the category “sameness” described personality variables where the correspondence across time was high, and changes in the salience level and salience heterogeneity were minimal. The category “decreasing salience values, order maintained” described personality variables that correlated well over time but whose salience, as a group, declined over time. Using these categories of personality variables, Block examined continuity and change in specific groups (e.g., men, adolescents). The findings using this approach were deemed uninteresting by Block, “ . . . the plethora of slight to moderate across-time relationships supports only the by-now banal recognition that human beings do indeed remain somewhat true to themselves over considerable lengths of time” (p. 88). Block understood that there was an inherent flaw in summarizing information about variables across diverse people; he recognized that such an approach would mask subgroups. For example, in taking a different approach (i.e., factor analyzing the individual profiles), Block found that there were some subjects who did not fit the majority types he found. He labeled these individuals residuals and noted that they had one of two characteristics. The first were individuals who did not fit any of the types; the second were individuals who fit well within multiple types. It should be noted that in 1971, Block was not entirely convinced that factor analysis was the best data analytic strategy to find types or residuals. In fact, he wished to find natural groupings, but did not have the computational power to use techniques that could find them. In later years, some of his most pointed criticisms of the FFM were based on

818

Person-Oriented Approaches

researcher’s reliance on factor analysis (e.g., Block, 1995, 2001, 2010), including the assumption that mathematically forced orthogonality accurately represented psychological understanding (i.e., the five factors did not operate independently in an individual). It is obvious that Block’s theory of personality and research touched on each of the person-oriented tenets described earlier. One of the important components of the modern person-oriented approach is that context matters. Unfortunately, the Q-set was deficient in measuring context; a problem of which Block was aware. The Q-set questions tried to take into account individual behavior in diverse settings, but the questions did not specifically focus on this dimension of human behavior. When an individual’s behavior varied, Block often attributed this to context. In his longitudinal research, Block did not have the assessment tools or methods to explore the context of an individual’s behavior, so he simply averaged the various discrepancies. In summary, Block was considered a contrarian, and he focused many of his critical essays in his later years, and, even the year that he died, on arguing against the FFM, which had come to dominate personality research. He offered many specific criticisms; however, these were all embedded in his most fundamental belief, the personorientation. He promoted a focus on the person, rather than variables, with specific emphasis on examining profiles; he encouraged psychologists to develop intraindividual theories rather than theories of interindividual differences; and he promoted the notion of the individual as a dynamic system. Block (1995) wrote: Once the parameters that define the personality system of a generic individual have been conceptually posited and empirically identified, these parameters become the essential, nonarbitrary, overarching variables or concepts for a personality of interindividual differences. The differences between individuals would then be understandable as due to the different specific values these parameters take in different individual personality systems. It would be a sweet intellectual accomplishment if the theoretical constructs required for dynamically understanding within-individual functioning could also be used for understanding the differences between individuals. Personality psychologists might aspire to this goal. In that effort, our science—in its theoretical reach and empirical grasp—may better realize its deep aspirations: to provide a bio-social theory of intraindividual development, the moment-to-moment psychodynamic functioning within the individual, the life themes coherently reflected in the adaptive changes of individuals over time and context, and the personality differences among individuals. (p. 210)

STATISTICAL APPROACHES Methods of Person-Oriented Data Analysis In this section, we discuss and exemplify application of methods of person-oriented data analysis. With these aims in mind, we adopt the role of the data analyst. This poor soul is often handed a data set along with two requests. First, test the hypotheses that had been specified when the grant proposal was written. Second, let the data speak a little as well, and perform exploratory analyses, just to make sure anything interesting was not overlooked. In the current context, there may be a third request, which is that analyses be performed from a person-oriented perspective. For the third request to be realistic, data have to be collected so that person-oriented analyses can be performed. von Eye and Bogat (2007; von Eye, 2010) specified a set of conditions for person-oriented data collection and analysis. Two conditions are most important. First, the sample should be large enough for person-oriented analysis to become possible. When data are collected with the goal of performing person-oriented data analysis, the number of cases must correspond to the anticipated number of groups or populations the data are assumed to be collected from. Each of the subpopulations must be large enough for the analysis to be possible in each of the subpopulations, and for comparative analyses to be possible as well. This desiderate applies both when subpopulations are known prior to data collection and when the researchers just assume that more than one population might exist. In the first case, power analysis can be used to determine the required size of samples from each population. In the second case, the number of populations may be unknown, rendering the a priori determination of sample sizes complicated. The second condition for person-oriented data analysis is that dimensional identity prevails (see the earlier section on the tenets of person-oriented research). Only if variables, test instruments, or dimensions have the same meaning and characteristics in each of the subpopulations, comparative statements with these variables, instruments, or dimensions are possible. Person-oriented research implies that a specific set of methods of data analysis is employed; however, it also implies that theory, data collection, and data analysis complement each other. This is reflected in a statement by Sterba and Bauer (2010), according to which person-oriented approach = person-oriented theory + person-oriented methods

Statistical Approaches

The remainder of this section is structured as follows. We discuss three popular methods of analysis and ask whether and how they can be used in the context of person-oriented research. The methods are structural equation modeling, log-linear modeling, and cluster analysis. Configural frequency analysis (Lienert, & Krauth, 1975; von Eye, 2002; von Eye, & Gutierrez-Pena, 2004), among the chief methods used in person-oriented research, is presented in a separate chapter in this volume (von Eye & Mun, 2015). For each method, we highlight the elements of person-oriented research that are represented, and we give an empirical data example. We posit that each of these methods requires unique assumptions and allows researchers to analyze data in respects of importance for person-oriented research. Structural Equation Modeling Structural equation modeling (SEM; Bollen, 1989; Jöreskog & Sörbom, 2001; Kline, 2011; McArdle, 2012; Muthén & Muthén, 1998–2012) allows one to combine methods of factor analysis with methods of regression analysis. In SEM, factor analysis is enriched by providing the option of classifying variables into dependent and independent, and regression analysis is enriched by allowing latent variables to be regressed onto each other. It is also possible to perform exploratory or confirmatory factor analysis without distinguishing between dependent and independent variables, and to estimate regression or path models at the level of manifest variables. In Jöreskog and Sörbom’s (2001) LISREL notation, a structural model can be described as 𝜂 = −B𝜂 + Γ𝜉 + 𝜁 where • 𝜂 is a [m, 1] random vector of latent variables (factors) on the dependent variable side (the y-side of the model); • −B is an [m, m] matrix of regression coefficients that describe the structural relations among the 𝜂 factors; the diagonal elements of 𝛽 are zero; 𝛽 can be symmetric or asymmetric; • 𝜉 is a [n, 1] random vector of latent variables (factors) on the independent variable side (the x-side of the model); • Γ is an [m, n] matrix of regression coefficients that describe the structural relations among factors on the x- and the y-sides of the model (factors from the y-side are regressed onto factors on the x-side of the model); and • 𝜁 is a [m, 1] vector of residuals; this vector represents imperfections of the model in the sense that not all

819

of the variance of the 𝜂 factors can be explained by the model; these are random disturbances, also called equation errors. The measurement models for the manifest variables on the x- and the y-sides of the model describe the factorial structure of the dependent and the independent variables. The models are y = Λy𝜂 + 𝜀, and x = Λx𝜉 + 𝛿, where • x is a [p, 1] vector of observed (manifest) variables on the x-side of the model, with p indicating the number of manifest x-variables; • y is a [q, 1] vector of observed variables on the y-side of the model, with q indicating the number of y-variables; • Λ x is a [q, n] matrix of the loadings of the x-variables on their latent variables, 𝜉; • Λ y is a [p, m] matrix of the loadings of the y-variables on their latent variables, 𝜂; and • 𝜉 and 𝛿 are the residuals of the manifest y- and x-variables, respectively; 𝜀 and 𝛿 are convolutions of measurement errors and model imperfections. Later, in this section, we present an example of a structural model in which we only use manifest variables. One characteristic of the structural model that is important for the present discussion of methods of data analysis from a person-oriented perspective is that the input matrix for the estimation of the model is a variance-covariance matrix, that is, a matrix with the variances of the observed variables in its diagonal and all covariances in the off-diagonal cells. Here, it is not important that this matrix can be substituted, under certain conditions, by a matrix of, for example, correlations or tetrachoric coefficients. Important is that, by definition, information about individuals disappears when variances and covariances are calculated. The individual is no longer visible. There are, however, two interesting exceptions. The first involves estimating models at the level of individuals (see, e.g., Nesselroade & Molenaar, 1998; Molenaar & Newell, 2010). The second involves identifying groups by using, for example, mixture distribution models (Muthén & Muthén, 1998–2012) or defining groups, and then estimating multigroup models. In this chapter, we focus on multigroup models. Either of these approaches to modeling development is of great importance for person-oriented research. They

820

Person-Oriented Approaches

allow characteristics of individuals or groups of individuals to be modeled. This way, justice is done to unique characteristics of individuals or groups of individuals instead of averaging differences out and estimating models that describe the average individual, which, according to Walls and Schafer (2006), may not exist (see also the earlier section on Kurt Lewin). Multigroup models allow researchers to specify hypotheses concerning the parameters in which they do versus do not allow the comparison groups to differ. Specifically, let G groups be given, then, the model for group g involves the eight parameter matrices g g g g Λy , Λx , Bg , Φg , Ψg , Θ𝜀 , Θ𝛿 , where the superscript indexes the groups. Now, to give an example of the hypotheses that can be specified, suppose that, in a two-group comparison, the structural parameters in the B and the Γ matrices are supposed to be equal. This can be expressed as B1 = B2 and Γ1 = Γ2 . Under these hypotheses, the parameters in all other matrices can be different. In LISREL, the following options for parameter constraints and estimation are available. Matrices in the comparison groups can be constrained • to have the same patterns of fixed and freed elements, • to have the same patterns and starting values in the iteration process, and • to have invariant elements, that is, estimates are constrained to be numerically exactly the same for the comparison groups. It is interesting to note that the number of variables does not have to be equal in the comparison groups. In addition, residual variances and covariances can be estimated for individual groups only. Hierarchies of constraints have been proposed, but here, we will not use any of the hierarchical procedures. Data Example The following example uses data that were collected for a study on the development of aggression in adolescence. Finkelstein, von Eye, and Preece (1994) administered a questionnaire to 75 11-year-old girls and 39 same-age boys who indicated the degree to which they perceived in themselves aggressive impulses (AI), were verbally aggressive against adults (VAAA), and physically aggressive against peers (PAAP). The questionnaire was administered again when the adolescents were 13 and 15 years of age. We hypothesize that, at age 13, aggressive impulses are explanatory of physical aggression against peers and that this process is mediated by verbal aggression against

adults. To test this hypothesis, we first estimate a path model without the direct path from AI85 to PAAP85, that is, the direct effects model PAAP85 = 𝛼1 + 𝛽21 VAAA85 + 𝜁1 and VAAA85 = 𝛼2 + 𝛾11 AI85 + 𝜁2 If needed, we also estimate the model that does include the direct path PAAP85 = 𝛼3 + 𝛾21 AI85 + 𝜁3 The first model is, in a first run, estimated for the entire group of respondents. The model – data fit for this model is excellent. We obtain 𝜒 2 = 0.68 (df = 1; p = 0.41), RMSEA = 0.00, CFI = 1.00, and GFI = 1.00 and decide to retain the model. Because of this excellent fit, the model with the direct path does not need to be estimated. In a first person-oriented step of analyzing the relations among aggressive impulses, verbal aggression against adults, and physical aggression against peers, we ask whether any of the questionnaire variables can be predicted from gender. In other words, we ask whether responses are gender-specific. To answer this question, we include gender in the model. For the following model, we place gender on the x-side of the model, and AI85, VAAA85, and PAAP85 on the y-side. We predict PAAP85 from VAAA85, which, in turn is predicted from AI85. Each of these variables is predicted from gender. The model – data fit for this model is excellent as well. We obtain 𝜒 2 = 0.56 (df = 2; p = 0.75), RMSEA = 0.00, CFI = 1.00, and GFI = 1.00 and decide to retain the model. Figure 18.2 displays this model (standardized scores given). In the model in Figure 18.2, the path hypotheses are confirmed. AI85 predicts VAAA85, which predicts PAAP85. In contrast, only the PAAP responses are gender-specific (t = 4.89; p < 0.01). The paths from Gender to AI85 and to VAAA85 remain nonsignificant (t = 0.07, n.s.; t = 1.64, n.s.). From this result, we conclude that the two gender groups may differ in their responses to the aggression questionnaire, and we specify a two-group model. We adopt a descending strategy where the strictest model is estimated first. In this model, we set all parameters equal for the two groups. Specifically, we set 𝜃𝛿1 = 𝜃𝛿2 , B1 = B2 , Γ1 = Γ2 , and Θ1𝜀 = Θ2𝜀 . Not unexpectedly, this model fails to describe the data well. Based on the overall goodness-of-fit 𝜒 2 = 22.58 (df = 7; p = 0.002), RMSEA = 0.20, CFI = 0.63, and GFI = 0.89, none of

Statistical Approaches

Aggressive Impulses

.99 .39

.10

1.00

Gender

.13

.38 1.00 Verbal Aggression Against Adults

.82

Aggressive Impulses

Verbal Aggression Against Adults

821

.80

.35

Physical Aggression Against Peers

.39

.43 .27 Physical Aggression Against Peers

.62

Figure 18.3 Two-group model of aggressive impulses, verbal aggression against adults, and physical aggression against peers; standardized solution for female respondents given. Source: Data from 1985.

Figure 18.2 Path model of aggressive impulses, verbal aggression against adults and physical aggression against peers. All predicted from gender. Source: Data from 1985.

which is in support of the model, we reject it. We conclude already at this point that male and female respondents differ in their answers to Finkelstein et al.’s (1994) aggression questionnaire. We now relax the restrictions placed in the first model. Instead of navigating through one of the hierarchies of imposing or relaxing restrictions, we consider that Gender differences seem to exist in the responses concerning physical aggression against peers (PAAP85; Figure 18.2). In the present model, the parameters related to PAAP85 are located in matrix B (parameter 𝛽21 ) and in matrix Ψ (parameter 𝜓33 ). We relax the parameter constraints, and now specify that the parameters in the matrices B and Ψ exhibit the same pattern for the two comparison groups, but they are allowed to differ from each other in magnitude. The resulting model describes the data very well. Specifically, we obtain the overall goodness-of-fit 𝜒 2 = 1.22 (df = 4; p = 0.87), RMSEA = 0.0, CFI = 1.00, and GFI = 1.00. Each of these statistics is in support of the model, and we decide to retain it. This model is given in Figure 18.3 (standardized solution given for female respondents). Several elements of this model reflect group differences. First, we note that the model describes the male responses slightly better than the female responses. The female submodel contributes 79.02% to the overall 𝜒 2 , and the male submodel contributes only 20.98%. These differences are tiny, but they are interpreted as pointing in this direction. More important are possible differences between the parameters that were relaxed for this solution.

In Figure 18.3, we see that, for the female respondents, the parameter for the path from AI85 to VAAA85 is estimated to be 0.39. For male respondents, this parameter was fixed to be equal. In contrast, the parameter from VAAA85 to PAAP85 is 0.35 in the female group and 0.64 in the male group. The difference between these two estimates is significant when only the 𝛽 parameters are compared and the 𝜓 parameters are assumed to be invariant across the comparison groups (Δ𝜒 2 = 18.85; Δdf = 2; p < 0.01). We conclude that female and male adolescent respondents differ in the path from VAAA85 to PAAP85 in the sense that this path, although significant in both comparison groups, is stronger in males than in females. Discussion Not surprisingly, structural modeling of manifest variable path models allows one to identify the parameters in which comparison groups differ. In this respect, structural modeling is well capable to test hypotheses that are compatible with a person-oriented research perspective. SEM offers many options, including, for example, hierarchical modeling, modeling data of individuals, growth curve modeling in multigroup designs, or means modeling. Still, it is hard to label SEM a standard method for person-oriented research. We, therefore, have to identify the elements that make SEM suitable for such research. We discuss three options. First, as is well known, the input matrix for standard SEM is a covariance matrix. Therefore, the individual does not play a role or, even worse, the individual is not visible. However, most SEM programs are able to make individuals visible, in some form. For example, to the best of our knowledge, every SEM program is capable of creating and saving factor scores. These scores will, in person-oriented

822

Person-Oriented Approaches

research, not necessarily be used to estimate factor score regression models (for options, see Hoshino & Bentler, 2011). However, the individual’s relative position in a factor space can be ascertained. This way, outliers can be made out and factor scores from different factors can be related to each other, with the options of identifying outliers, or specifying individual profiles. In other words, every SEM program allows one to identify those individuals for which a factor solution is less than perfect in the sense that the individual becomes an outlier. From a person-oriented perspective, these are the individuals for which a model may not apply. (Note: When factor scores are used, the covariance matrix that is used as input information must contain all relevant information. That is, one must assume that higher order interactions do not exist or explain only negligible portions of the targeted variability, that is, the variability that allows one to distinguish between individuals.) The second option, and this was illustrated in the data example on aggression development, is that SEM is a very powerful methodology that can be used to distinguish among subpopulations that are known a priori. Differences in most any parameter can be ascertained. The third option involves SEM mixture modeling (Arminger & Stein, 1997; Dolan & van der Maas, 1997). Two approaches to SEM mixture modeling have been proposed (Vermunt & Magidson, 2004). The first involves specifying an SEM model and then modeling within-class covariance matrices under the assumption that there are multivariate normal mixtures. This approach is also known as latent class clustering or model-based clustering (Banfield & Raftery, 1993; Mun, von Eye, Bates, & Vaschillo, 2008). In the context of SEM mixture modeling, the class-specific mean and covariances are restricted to the specified SEM structure. This structure reflects, for instance, factors or growth curves. The second approach involves incorporating latent classes into a structural model, thus capturing unobserved heterogeneity. Here, the unobserved variable is class membership. SEM mixture models can be estimated by some SEM software packages (e.g., Mplus; Muthén & Muthén, 1998/2012) and is of importance in particular when researchers assume that data were collected from more than one population, but number and size of populations are unknown. These three options show that SEM is suitable for person-oriented research—when certain assumptions apply (Sterba & Bauer, 2010; von Eye, 2010). Most important, in the present context, is the assumption of dimensional identity that was previously discussed, in the context of the tenets of person-oriented research (von

Eye & Bergman, 2003). Dimensional identity implies that the variables that are used to describe individuals are the same in all parameters in the comparison groups. That is, they have the same meaning, the same factor and covariance structures, and they are commensurable, both cross sectionally and longitudinally. If this assumption holds, individuals and groups of individuals can be compared using these variables. If, however, this assumption must be rejected and there is, for example, differential item functioning, that is, item characteristics fail to be invariant across comparison groups, scores on the scales that include such items are no longer comparable. Not all is lost in such a case. The information that there is differential item functioning or that factorial structures vary across comparison groups may be of interest itself. However, group means cannot be compared, and neither can scale scores. Before testing multigroup structural models or estimating SEM mixture models, researchers, therefore, are well advised to make sure that dimensional identity prevails. There exists, however, an interesting exception. There are variables that have natural meaning. Examples of such variables are physiological measures, voxel counts for brain activity, response time, number of items recalled, and many other measures that are not based on instruments constructed using psychometric tools. When this kind of variable is employed, differences in means or covariance structures are most important for the person-oriented researcher because they are no longer indicative of flaws in instruments but of possibly meaningful differences between groups or individuals. Log-Linear Modeling Types of Log-Linear Models Log-linear models are special cases of generalized linear models in which categorical variables are analyzed (Agresti, 2012; von Eye & Mun, 2013). The generalized linear model (GLM) relates a vector, 𝜂, to one or more explanatory variables by way of a linear model, 𝜂 = X 𝛽, where X contains the predictors or the effects posited by a model, and 𝛽 is a parameter vector. In a linear regression model, 𝛽 contains the regression weights. Now, let 𝜂 be related to the expected value of the linear model by 𝜂 = g(𝜇), where 𝜇 is the expected value, and g(.) is a monotone function. This function is called the link function. For cases of the general linear model (e.g., analysis of variance or regression), the link function is the identity function. That is, 𝜂 = 𝜇. In log-linear modeling, one obtains log 𝜇 = X 𝛽

Statistical Approaches

and for the general linear model, one obtains 𝜇 = X𝛽 These equations illustrate that the general linear model and the log-linear model have the same form (see Agresti, 2012; McCullagh & Nelder, 1989; von Eye & Mun, 2013). Differences are evident in the link function and in the nature of the expected values. In regression and in analysis of variance, expected values are estimated for an observed variable, the outcome or dependent variable, typically in the units of this variable. In log-linear modeling, expected values typically are estimated for counts. By implication, main effects in regression or analysis of variance involve two variables, the independent and the dependent variables, and in log-linear modeling, main effects concern the distribution of the categories of just one variable. This applies accordingly to interactions. Three ways of specifying log-linear models have been discussed. The first involves hierarchical log-linear models, the second involves nonhierarchical models, and the third involves nonstandard models (Mair & von Eye, 2007; von Eye & Mun, 2013). Log-linear models are hierarchical if • all lower order relatives of terms included in a model are also part of the model; • the model contains only terms that can be expressed within the hierarchy of effects; and • no constraints are specified for parameters (except zero-sum constraints; see von Eye & Mun, 2013). Log-linear models are nonhierarchical if • they contain only a selection of the possible lower order relatives of the terms in the model; and • if the second and the third condition for hierarchical models are also fulfilled. Log-linear models are nonstandard if • they contain terms that cannot be specified in the hierarchy of models; or • if they include covariates, parameter constraints, or special effects that cannot be specified in the hierarchy of models. In the following sections, we first present examples of log-linear models, and we review parameter interpretation. This is followed by a data example and a discussion of log-linear modeling in the context of person-oriented research. For the examples, consider the four categorical

823

variables, A, B, C, and D. A first model that can be considered for these variables (and is not the null model of no effects) is that of variable independence. This model includes all main effects but proposes that there are no interactions. The model is log m = 𝜆 + 𝜆A + 𝜆B + 𝜆C + 𝜆D where m is the vector of model frequencies, and the superscripts indicate the variables involved in the main effects. A second model could propose that A is associated with B and C is associated with D. This model is log m = 𝜆 + 𝜆A + 𝜆B + 𝜆C + 𝜆D + 𝜆AB + 𝜆CD where the double superscripts indicate the associations (interactions) included in the model. When, for instance, the main effect of any variable is removed from this model, for example, when, after removal of the main effect of A log m = 𝜆 + 𝜆B + 𝜆C + 𝜆D + 𝜆AB + 𝜆CD researchers consider a nonhierarchical model. This model is nonhierarchical because the main effect of variable A that was removed is a lower order relative of the interaction between A and B. An example of a nonstandard model is log m = 𝜆 + 𝜆A + 𝜆B + 𝜆C + 𝜆D + 𝜆AB + 𝜆CD + 𝜆Covariate where the covariate is a term that cannot be expressed in terms of elements of a hierarchical model. Examples of covariates are continuous variables whose cell-specific averages are used as special effects. Parameter Interpretation The interpretation of log-linear model parameters is based on three arguments (von Eye & Mun, 2013). The first is statistical significance. To test the null hypothesis that an individual parameter is zero, researchers routinely employ t-tests. Second, the magnitude of parameters is of interest. Parameters can be transformed into odds ratios that have a natural interpretation. Third, and most important, is the meaning of a parameter. The meaning is expressed in terms of contrasts in the columns of the design matrix, X. For clear-cut parameter interpretation, that is, interpretation

824

Person-Oriented Approaches

that is independent of other parameters in the model, the 𝜆 parameters can be interpreted, based on X, as 𝜆 = (X ′ X )−1 X ′ log m where X is the design matrix that contains, in its column vectors, the effects included in a model. It is most important that the interpretation corresponds to the contrasts specified in the design matrix. To give an illustration of this type of parameter interpretation, we use the example given by von Eye and Mun (2013, p. 135). Consider a 2 × 2 table that results from crossing the variables A and B. For this matrix, we specify the saturated model, that is, the model with all possible parameters. The model is, in explicit form, ⎡log m11 ⎤ ⎡1 1 1 1⎤ ⎡ 𝜆 ⎤ ⎡𝜖21 ⎤ ⎥⎢ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ 1 −1 −1⎥ ⎢ 𝜆A ⎥ ⎢𝜖12 ⎥ ⎢log m12 ⎥ = ⎢1 + ⎢log m21 ⎥ ⎢1 −1 1 −1⎥ ⎢ 𝜆B ⎥ ⎢𝜖21 ⎥ ⎢log m ⎥ ⎢1 −1 −1 1⎥⎦ ⎢⎣𝜆AB ⎥⎦ ⎢⎣𝜖22 ⎥⎦ 22 ⎦ ⎣ ⎣ The first column vector of this model contains the logarithms of the observed cell frequencies. The first column vector in the design matrix represents the constant. The second, representing the main effect of Variable A, contrasts the first category of A with the second category. The third vector, representing the main effect of Variable B, contrasts the two categories of Variable B. The last column vector represents the A × B interaction. It results from element-wise multiplication of the main effect vectors of the interacting variables. After the design matrix follow the parameter and the residual vectors. One parameter is estimated for each effect, that is, each column in the design matrix. The inner products (and the correlations) of any two of the vectors of the design matrix in the previous equation are zero. Therefore, parameter interpretation is easy. Inserting into 𝜆 = (X ′ X )−1 X ′ log m yields, for the four parameters of this model, 𝜆 = 1∕4(log m11 + log m12 + log m21 + log m22 ),

equal weights of the logarithms of the cell frequencies and the signs of each term. These are the same as in the design matrix. For more examples, and for examples of design matrices that do not possess this characteristic, see von Eye and Mun (2013). Multiple approaches have been developed for log-linear analysis of longitudinal data. Models that come with a straightforward interpretation include those for which repeatedly observed variables are crossed (see, for example, von Eye & Mun, 2013). In the following data example, we present such a model. Data Example For the following example, we again use the data from the Finkelstein et al. (1994) study on the development of aggression. We ask whether the development of aggressive impulses over a span of 4 years (AI83 and AI87) is related to the development of physical aggression against peers (PAAP83 and PAAP85). For the following analyses, we dichotomize these variables at the grand median of aggressive impulses and the grand median of physical aggression against peers. This approach allows us to identify individuals who, overall, stay constant or display change relative to the average. We estimate two models. The first is the model of variable independence, log m = 𝜆 + 𝜆AI83 + 𝜆AI87 + 𝜆PAAP83 + 𝜆PAAP87 This model is estimated to have a reference for more complex models and to find out whether effects exist at all in the AI83 × AI85 × PAAP83 × PAAP87 cross-classification. For the model of variable independence, we obtain an overall goodness-of-fit maximum likelihood 𝜒 2 of 56.50, which, for df = 11 suggests that this model fails to describe the data well (p < 0.01). We need a more complex model. Skipping the model development steps in which we successively added terms to the reference model, we describe the final model. This model includes the associations between the repeated observations of AI and PAAP as well as the cross sectional associations between AI83 and PAAP83 and AI87 and PAAP87. In other words, this is the model

𝜆A = 1∕4(log m11 + log m12 − log m21 − log m22 ), 𝜆B = 1∕4(log m11 − log m21 + log m21 − log m22 ), 𝜆AB = 1∕4(log m11 − log m12 − log m21 + log m22 ). Each of these equations corresponds to the effects specified in the design matrix. This can be seen by the

log m = 𝜆 + 𝜆AI83 + 𝜆AI87 + 𝜆PAAP83 + 𝜆PAAP87 + 𝜆AI83,AI87 +ΛPAAP83,PAAP87 + 𝜆AI83,PAAP83 + 𝜆AI87,PAAP87 The new terms are presented in the second row of the equation. The overall goodness-of-fit maximum likelihood

Statistical Approaches TABLE 18.2 Observed and Expected Frequencies for Log-Linear Model of the Development of Aggressive Impulses and Physical Aggression Against Peers in Adolescents AI83 AI87 PAAP83 PAAP87

Observed frequencies

Expected frequencies

1111 1112 1121 1122 1211 1212 1221 1222 2111 2112 2121 2122 2211 2212 2221 2222

24 2 12 8 4 4 4 3 8 1 10 3 6 1 6 18

24.49 2.50 12.82 6.19 4.79 2.22 2.51 5.48 7.22 0.74 9.47 4.57 5.50 2.54 7.21 15.76

vectors of the design matrix (not shown here) correlate to zero, and each of the parameters can be interpreted as specified in the design matrix. We present one example, the parameter for the interaction between AI83 and AI87. This parameter has the interpretation 𝜆AB = 1∕16 log m1111 + log m1112 + log m1121 + log m1122 − log m1211 − log m1212 − log m1221 − log m1222 − log m2111 − log m2112 − log m2121 − log m2122 + log m2211 + log m2212 + log m2221 + log m2222

𝜒 2 of 6.64 suggests that, for df = 11, this model does describe the data well (p = 0.47). Table 18.2 displays the observed and the expected cell frequencies for this model. Table 18.2 shows that the observed and the expected cell frequencies are close to each other throughout. The largest standardized deviate is 1.20 (for cell 1212). This value is nonsignificant (p = 0.12), thus supporting the model. Table 18.3 displays the significance tests for the model parameters. The significance tests in Table 18.3 suggest that, of the parameters of interest (these are the interactions), only one is nonsignificant (AI83 × PAAP83). We conclude that the auto-associations AI83 × AI85 and PAAP 83 × PAAP87 and the cross sectional association AI87 × PAAP87 are sufficient to explain the data in Table 18.2. There is no need to include the cross-lagged associations AI83 × PAAP87 and PAAP83 × AI87. Before finally retaining this model, we ask whether the parameters of interest come with a clear-cut interpretation. This is the case. The column

We, therefore, retain the model. The substantive interpretation of this model is that self-perceived aggressive impulses and physical aggression against peers are stable over a span of four adolescent years and that, at age 15 (but not at age 11), aggressive impulses are significantly related to physical aggression against peers. Discussion The example given in the last section illustrates standard log-linear modeling. Many other and more complex models have been proposed (see, e.g., Agresti, 2012; von Eye & Mun, 2013). Examples of such models include logistic regression models (which can always equivalently be expressed in terms of log-linear models; see von Eye, Mair, & Bogat, 2005), latent class models, mixture distribution models, or path models. Even some IRT models can be expressed in terms of log-linear models (Meiser, 1996). Clearly, one of the greatest benefits of using log-linear models lies in their generality. We now ask whether log-linear models can be used for person-oriented research. This is possible in a number of designs. First, log-linear models allow researchers to naturally include classification or stratification variables. This is possible when sampling is multinomial, that is when every

TABLE 18.3 Parameters of the Model of the Development of Aggressive Impulses and Physical Aggression Against Peers in Adolescents Term tested

AI83 AI87 PAAP83 PAAP87 AI83*AI87 PAAP83*PAAP87 AI83*PAAP83 AI87*PAAP87

825

The model without the term

Removal of term from model

LOG(MLE)

Chi-square

df

p-value

Chi-square

df

p-value

−31.318 −31.635 −34.707 −36.575 −37.025 −37.662 −33.954 −37.940

6.795 7.429 13.574 17.311 18.210 19.484 12.068 20.040

8 8 8 8 8 8 8 8

0.559 0.491 0.094 0.027 0.020 0.012 0.148 0.010

0.153 0.787 6.932 10.668 11.568 12.842 5.425 13.398

1 1 1 1 1 1 1 1

0.696 0.375 0.008 0.001 0.001 0.000 0.020 0.000

826

Person-Oriented Approaches

new case can, in principle, be assigned to any cell of a cross-classification, but also when sampling is productmultinomial, that is, when the number of cases in a stratum or group of respondents is determined a priori. Parameter estimation is the same for both cases, but there are constraints concerning the models that can be estimated. In product-multinomial sampling designs, the estimated model must always reproduce the marginal probabilities of the variables for which these probabilities are fixed. In either case, log-linear models allow one to compare groups of individuals. The comparison can result in interactions that indicate the variable relations that are group-specific. Alternatively, one can estimate Mantel-Haenszel-type statistics that allow one to answer the question whether interactions differ across groups (for an overview, see Fidalgo, 2005). The important characteristic of the log-linear methods that allow one to compare groups of individuals, under certain conditions and over time, is that this comparison already represents an approach to personoriented research. These comparisons presuppose that researchers hypothesize that the comparison groups differ from each other, for example in a conditional pattern of interactions among variables. The only condition that must be met for this comparison is that of dimensional identity. Variables must have the same meaning for all comparison groups. Interestingly, if interaction patterns are defined to be constituting elements of dimensional identity, there should be no differential interaction patterns. If, however, a researcher asks whether such patterns vary across comparison groups under the condition that, otherwise, dimensional identity prevails, results of such a comparison would indicate where dimensional identity is violated (if groups differ). This result in itself amounts to a statement of interest to person-oriented researchers. If, in contrast, interaction patterns are not part of the definition of dimensional identity, a study on differential interaction patterns can be of interest to person-oriented, idiographic, and differential developmentalists alike. In brief, when dimensional identity can be made plausible, the comparison of a priori existing groups is of interest from a person-oriented research perspective. In this respect, log-linear models are comparable to analysis of variance (von Eye & Bogat, 2007) and structural modeling. In many instances, however, researchers just assume that multiple populations exist. Number and size of these populations are unknown. In these cases, groups need to be identified and group differences need to be established. Groups can be identified by using such methods as cluster analysis, latent class analysis, or mixture distribution decomposition.

It is important to realize that group differences can manifest in any parameter-(s). When metric variables are on the outcome side, mean differences can be examined (ANOVA). When variables are categorical, marginal distributions can be compared. In either case, groups can differ in variances, covariances, factor structures, parameters of structural models, trajectories, validity and reliability of variables, or any combination of these. Person-oriented researchers are interested in all of these. It is also important to realize that groups may not come in prespecified sizes. Groups can be as large as standard populations and as small as just one individual. Personoriented as well as idiographic developmental research recommend estimating parameters first at the level of the individual. Aggregation is done at the level of parameter estimates, if it is done at all. The development of methods of analysis of individual data and small group-data needs to make progress for this to be possible. Fortunately, methods for the analysis of single-subject data, in particular longitudinal data, are being developed at a rapid pace and integrated into the canon of methods for the analysis of longitudinal developmental data (e.g., Dolan, Molenaar, & Boomsma, 1989; Molenaar, 1997; Molenaar, Huizenga, & Nesselroade, 2003; Molenaar & Newell, 2010; Rovine, Molenaar, & Corneal, 1999). This development includes metric as well as categorical outcome variables (van Rijn, 2008; van Rijn & Molenaar, 2005). Cluster Analysis Taxometric methods, in particular cluster analysis, are the most frequently used methods for the identification of groupings that are unknown before analysis. Using these methods, one can extract groupings from a data set that contain members that • are maximally similar to each other, while being • maximally dissimilar to members of other groupings The number of taxometric methods is very large. The decision about which method to select is important, because results can differ dramatically depending on the method selected. von Eye, Mun, and Indurkhya (2004) present an example of an artificial data set that can lead to classifications that are completely unrelated to each other when different taxometric methods are used. Specifically, the cross-classification of the clusters created for this data set comes with a 𝜒 2 = 0.0. The most fundamental issue with taxometric methods concerns the measurement of the degree of similarity or

Statistical Approaches

dissimilarity between individuals. The choice of which variables to include in the profile to be analyzed and the type of similarity or dissimilarity measure to be used must be carefully considered and theoretically motivated. The scaling of the variables must be commensurable and, in most cases, continuous variables must be standardized before analysis (see Bergman, Magnusson, & El-Khouri, 2003, for exceptions). In the following section of this chapter, we first present a specification of the taxometry problem. We then review, discuss, and extend eight decisions to be made when selecting a taxometric procedure for person-oriented researchers by von Eye et al. (2004). This is followed by a data example. Taxometry: Specification of the Problem To illustrate the problem, we use a data example. We use Finkelstein et al.’s data on the development of aggression again. In this study, the variable self-perceived Aggressive Impulses (AI) was observed three times, in 1983, 1985, and 1987, in a sample of 114 adolescents (67 females). The trajectories AI are depicted in Figure 18.4, by Gender. Figure 18.4 shows that all possible patterns of ups and downs were observed, and that not a single adolescent gave the same responses at all three occasions. Now, proceeding as von Eye et al. (2004) with the artificial data set, we illustrate that two cluster solutions of the same data can result in quite different classifications. We apply two methods of clustering to the aggressive impulse data, thus creating clusters of trajectories of aggressive impulses during adolescence, and then examine the cross-classification of these two cluster solutions. We create a first cluster 35 30 25 20 15 10 5 0 AI83

TABLE 18.4 Cross-Classification of Two Cluster Solutions

1 2 3 4 Total

AI87

Figure 18.4 Developmental trajectories of self-perceived aggressive impulses in 114 adolescents. See footnote 1.

1

2

3

Total

21 20 24 3 68

9 9 15 12 45

0 0 1 0 1

30 29 40 15 114

solution by using the popular Ward method, based on Euclidean distances between the trajectories. We create the second solution with Ward’s method again, but, this time, based on the correlations between the trajectories. For the distance-based classification, we select the four cluster solution, and for the correlation-based classification, we select the three cluster solution. The cross-classification appears in Table 18.4. Table 18.4 shows that the clusters of the distance-based solution differ in size (row totals), but each contains at least 14 of the 114 cases of the sample, that is, more than 13% of the total. In contrast, the three clusters of the correlation-based solution differ dramatically in size. The smallest cluster contains a singleton, that is, just one individual. The largest contains 68 respondents, that is, almost 60% of the total. The cross-classification of the two cluster solutions created for the aggression data comes with a 𝜒 2 = 14.25. For df = 6, this value suggests that knowledge of one of the two cluster solutions also contains information about the second solution (p = 0.03). One might be tempted to interpret this result as indicating that the two solutions are similar. Fact is, however, that not one of the clusters from the distance-based solution remains intact when the two solutions are crossed. Each of the two larger clusters of the correlation-based solutions feeds from the four clusters of the distance-based solution. The singleton of the correlation-based solution is a peaceful member of the third cluster of the distance-based solution. We now proceed to the specification of the taxometry problem. In this section, we give a general definition of the clustering or classification problem (Bock, 1974, p. 22; Hartigan, 1975). Consider N objects (i.e., cases), O1 , . . . , ON . For these objects, a data matrix, xki , a similarity matrix,10 djk , or a relation, ≺, can be calculated, the elements xki , djk , or ≺ of which represent the similarity structure of the set of objects, S = {O1 , . . . , ON }. Searched for is a classification, 10

AI85

827

In later sections of this section we also discuss distance matrices. For the present purposes, we specify that the term similarity matrix subsumes distance matrices and other matrices that can be created to describe relations between objects.

828

Person-Oriented Approaches

A = (A1 , A2 , . . . ) of S. The subsets of this classification, that is, A1 , A2 , . . . are classes (groups or clusters) that reflect the similarity structure of the objects, and imply data reduction. These two requirements can be fulfilled if both the objects within a group are maximally similar to each other and cross-class dissimilarity is maximized. The first characteristic is called homogeneity; the second is called separation. In the following sections, we review and extend the decisions that von Eye et al. (2004) discussed for the selection of a method of clustering. Decisions for Clustering von Eye et al. (2004) considered the following eight cluster characteristics: clusters are (1) disjoint versus overlapping; (2) hierarchical versus nonhierarchical; (3) agglomerative versus divisive; (4) exhaustive versus selective; (5) stochastic versus deterministic; (6) based on correlation versus distance measure; (7) convex versus nonconvex; and (8) represent manifest versus latent variables. The authors also discussed methods of classification that are based on such a priori set criteria as cutoffs or ranges of admissible scores and natural classifications that result from cross-classifying categorical variables. Here, we focus on taxometric clustering and on natural classifications. Cross-classifications of categorical variables are discussed in the earlier section on log-linear modeling. Before discussing the eight criteria, it should be emphasized that there is no such thing as a true cluster solution. Instead, researchers use classification algorithms to structure data. The resulting clusters or groups reflect characteristics of the methods that were applied as well as data characteristics. Different methods are sensitive to different data characteristics. Therefore, different cluster solutions describe different data characteristics to the extent that different clustering methods are capable of depicting them. Criterion 1: Disjunct Versus Overlapping Clusters Disjunct classifications have the characteristic that each object Oj belongs to only one class, Ai , or cluster. In contrast, overlapping classifications allow each object Oj to belong to more than one class. Rarely, clusters naturally will be completely disjunct. Groups of cases will exist that are not easily assigned to a group. How can one deal with situations of this kind? We discuss two solutions. First, one can create nonoverlapping clusters by changing cluster parameters. This can be achieved by, for instance, reducing the diameters of clusters as long as necessary for them to no longer overlap. The number of cases located outside clusters will increase as the diameters shrink. This way,

the overlap can be reduced to zero. The cases not located inside the (imaginary) cluster hull can be (1) treated as a separate group, (2) considered members of no cluster, (3) assigned to clusters at random, (4) assigned to the cluster to which they most likely belong, or (4) allowed to form separate small groupings, even singletons. Second, one can consider whether overlapping cluster solutions may be natural or appropriate. There may be cases for which it is reasonable to assume that they belong to more than one group. For instance, an athlete may be both a soccer and a basketball player, or a student may be both a stellar performer and a political activist. Separation of clusters indicates the degree to which clusters are disjunct. Separation can be measured only if groupings are created that are allowed to overlap. If groupings are not naturally disjunct—groupings based on manifest categorical variables are naturally disjunct—the idea that clusters are nonoverlapping clusters is perceived, by some, as artificial. Therefore, application of methods that create nonoverlapping clusters by default may require justification. Based on this, it is surprising that a look at the literature suggests that, in empirical applications of clustering methods, researchers overwhelmingly use methods that result in nonoverlapping clusters. Similarly, most statistical software packages do not even include methods that allow researchers to create disjunct clusters. Two programs that provide this option include ADCLUS by Arabie and Caroll (1980) and Pyramid by Aude, Diaz Lazcoz, Codani, and Risler (1999; Everitt, Landau, & Leese, 2001). Criterion 2: Hierarchical Versus Nonhierarchical Clusters Classification algorithms can create solutions that are either hierarchical or nonhierarchical. A strength of a hierarchical cluster analysis is the simple relationship between solutions with different numbers of clusters. (For instance, a six-clusters solution is identical to a seven-clusters solution except that two clusters of the second solution are fused to constitute one cluster in the first solution.) This information is helpful for understanding the classification structure and in choosing the number of clusters. It also makes it feasible to present findings based on both a fine-grained classification with many clusters and a simpler classification with fewer clusters (since the relationship between the two classifications is straightforward and easily explained). Hierarchical clustering algorithms create series of cluster solutions. The series are either agglomerative or divisive. Agglomerative series result from grouping objects together

Statistical Approaches

into larger and larger groups. One of the chief characteristics of agglomerative hierarchical procedures is that smaller numbers of clusters result from merging clusters. Cases who were members of the same cluster will stay members of the same cluster. Divisive procedures are hierarchical if they create greater numbers of clusters by splitting larger clusters (the choice between agglomerative and divisive algorithms is discussed later under Criterion 3). A clustering method is hierarchical if cases that are members of the same cluster when clusters contain fewer cases are members of the same cluster when the cluster contains more cases. The results of hierarchical clustering are routinely displayed in the form of dendrograms (cluster trees). Figure 18.5 displays the dendrogram that results when Ward’s method, based on Euclidean distances, is used to cluster the aggressive impulses trajectories shown in Figure 18.4. Clustering algorithms are nonhierarchical when they create just one solution instead of a series of solutions. Examples of such procedures include methods that identify space density maxima and assign objects to the cluster that is defined by the nearest density center. Examples of such algorithms include the methods discussed by von Eye and Wirsing (1978, 1980) and the well-known k-means method. The number of clusters in nonhierarchical clustering is either user-determined as in k-means, or is determined by the number of space density maxima that were found based on a priori specified criteria. Hierarchical methods always create N − 1 solutions. A decision as to which solution to interpret is a key part of hierarchical cluster analysis. Cluster Tree

829

Several methods have been proposed to select from the hierarchy of cluster solutions. One involves examining a distance scale as the one used for the x-axis in Figure 18.5. Based on distance scales, the solution is selected that fulfills two conditions: 1. The number of clusters is near the desired number (see Bergman’s, 2000, postulate of targeting a number of 4–10 clusters) 2. A large increase in the distance scale value has occurred When these two conditions are fulfilled, one selects the solution right before the large increase on the distance scale. This large distance suggests that two heterogeneous clusters (clusters that are far apart from each other) had been fused into one. In the example in Figure 18.5, many researchers will select either the four- or the two-cluster solution. Because of weaknesses of selecting solutions based on distances, many other cluster evaluation methods have been discussed. Among these are methods that use various forms of the Rand Index (1971). This index measures the similarity between two cluster solutions, for example a random solution and the solution preferred by the researchers. Recent investigations have shown that the Hubert-Arabie-adjusted Rand index performs well in cluster solution evaluation (Steinley, 2004; Volkovich, Avros, & Golani, 2011). Interestingly, it has been shown that the Hubert-Arabie-adjusted Rand index is a twin of the well-known Cohen’s kappa (Cohen, 1960; Warrens, 2008). The choice between hierarchical or nonhierarchical methods can be based on two criteria. The first is whether the hierarchy of cluster solutions can be interpreted. When this is the case, hierarchical methods are selected. When, however, the series of cluster solutions is not interpreted or not interpretable, and when researchers have expectations concerning the number of clusters, nonhierarchical solutions may be selected. Only a very small number of solutions—maybe only one—needs to be calculated, and interpretation is typically straightforward. The second criterion is thus the extent to which researchers use prior knowledge when performing a cluster analysis. Criterion 3: Agglomerative Versus Divisive Clustering

0

50

100 Distance

150

200

Figure 18.5 Dendrogram of Ward’s clustering of aggressive impulses in adolescents. See footnote 1.

In this section, we discuss the choice between agglomerative and divisive clustering. Agglomerative clustering starts under the assumption that each object, Oj , constitutes a separate cluster. With the goals of reducing redundancy and creating a parsimonious solution, agglomerative methods join the clusters (at the beginning of the procedure, these are all individuals) that are the most similar to each

830

Person-Oriented Approaches

other, or, in more general terms, clusters are joined when some loss, measured, for instance, in units of decrease in information, increase on a distance scale, or difference in R2 , is minimal. This is repeated until either a stopping criterion has been reached or until the final two groupings are joined. Divisive clustering considers, before the first step is taken, all case members of one large cluster. Beginning with the second step, divisive clustering splits the existing cluster(s). This is done such that a benefit, measured in units similar to the ones used for agglomeration, is maximized. This step is repeated until either a stopping criterion, for instance, some R2 threshold, has been reached or until each individual constitutes its own cluster. Evidently, what is the starting point for the agglomerative procedure is the end point for the divisive procedure. The results of both agglomeration and division clustering can be depicted in the form of dendrograms (tree structures). Consider, again, Figure 18.5. If the dendrogram in this figure is the result of agglomeration, we read Figure 18.3 from left to right. At the left of the dendrogram, each individual is represented. In the first steps, the most similar objects are joined to form one cluster. In each of these steps a new cluster is created. Relatively soon, clusters are joined. In the last step, the clusters that resulted from the first n − 2 agglomeration steps are joined, and every case belongs to the one and only cluster. Researchers interpret the solution that best meets optimality criteria as the ones discussed previously, plausibility criteria, and criteria derived from theory. To illustrate the divisive procedure, we use Figure 18.5 again. This time, however, we read it from right to left. Before the first step (at Step 0), the procedure considers all objects members of one big cluster. Beginning with Step 1, this big cluster is subdivided into ever smaller clusters, until each object constitutes its own cluster. As for all hierarchical procedures, criteria of optimality, theory, and plausibility can be used to select a solution. When deciding which of the two procedures to select, agglomerative or the divisive clustering, researchers keep in mind that results from both methods are often the same. Therefore, method selection is largely guided by preferences and program availability. Most software packages offer exclusively agglomerative procedures. Examples of agglomerative procedures that are available in many software packages include Single Linkage (aka Nearest Neighbor), Complete Linkage (aka Furthest Neighbor), Average Linkage, Centroid Clustering, Median Clustering, Weighted-Average Linkage, Beta-Flexible, Density Linkage, and Ward’s method.

Criterion 4: Exhaustive Versus Nonexhaustive Clustering Classifications are exhaustive if every case, Oj , belongs to a cluster. Nonexhaustive classifications involve only some of the N objects. The cases assigned to clusters are sometimes called the most important objects. Cases not assigned to a cluster are often either ignored or discarded and thrown into a poubelle, that is, a garbage can. From a person-oriented perspective, the distinction between exhaustive and nonexhaustive classifications is not always conclusive. If a researcher assumes that populations can be as small as just one individual, nonclassified cases constitute classes also. In addition, Bergman (1988) suggested that it may be wise not to assign all cases to a grouping under all conditions. He called these unclassified cases a residue and argued that it can be important to remove them before cluster analysis for both theoretical reasons (true unique cases) and practical reasons (an outlier could be caused by errors of measurement that can distort the findings of a cluster analysis). Further, he did not regard them as a poubelle but as important cases to be analyzed separately and claimed that the proportion of residue cases in the sample can provide important information about the classificability of the data set. Still, in most applications, singletons (one-case clusters) are considered cases that are not classified. Even clusters with only two cases are often ignored in follow-up analyses or assigned to the poubelle. One way to illustrate nonexhaustive classifications involves configural frequency analysis (CFA; Lienert & Krauth, 1975; von Eye, 2002; more detail on CFA can be found in the chapter on CFA in this volume). CFA examines cross-classifications of categorical variables under the question whether the observed number of cases in a cell differs from the expected number. If the null hypothesis of no difference is rejected for a cell, CFA states that this cell constitutes a type if more cases than expected were found, and an antitype, if fewer cases than expected were found. It is a routine result of CFA that only a selection of cells constitutes types and antitypes. The frequencies in other cells do not deviate from expectancy. The designation as ‘belongs to a CFA type’ or ‘belongs to a CFA antitype’ is, therefore, nonexhaustive. Many methods of classification can create exhaustive as well as nonexhaustive solutions. In hierarchical classifications, a nonexhaustive solution would include one or more singletons or clusters that contain only two or three objects. In nonhierarchical classification, one can create solutions with relatively large numbers of clusters. Some of these may then contain very small numbers of cases. It is important to note that classification methods differ in their tendencies

Statistical Approaches

to yield small clusters. Methods that build clusters by letting objects gravitate toward a center of gravity (centroid) tend to allow larger clusters to swallow smaller clusters. Examples of such methods are the centroid method, complete linkage, average linkage, and Ward’s method. Pros and cons for exhaustive versus nonexhaustive classification are evident. On the pro side of exhaustive classification is that every object is assigned to a group. By implication, the sample under study is not reduced in size. At the expense of this strategy is, however, that, for some cases (e.g., outliers), it may be unnatural to belong to a group. Also on the con side of exhaustive clustering is the possibility that single-case clusters result. In follow-up analyses, researchers often experience problems with single-case clusters because individual cases can be hard to statistically compare with other clusters. For instance, multivariate analysis of variance (MANOVA) can be problematic for subsequent comparisons when one or more cells contain only one case. Criterion 5: Stochastic Versus Deterministic Clustering Stochastic models consider data points realizations of random variables. This seems appropriate under the following assumptions: (1) there is natural variation within each grouping; (2) the manifest variables that are used for classification are measured with error; and (3) the cases used for classification are a random sample. When these assumptions are met, stochastic models allow researchers to test hypotheses concerning possible group structures. In contrast, deterministic models consider data points fixed sets of objects. Variability is not considered random. Pros and cons of these two approaches are evident. Among the pros of stochastic models is that statistical decisions about hypotheses that are compatible with the existence of groupings (substructures) become possible. One downside is that a null hypothesis for the existence of a particular grouping structure is not always easily specified. Among the pros of deterministic models is that parametric assumptions do not need to be made. Another downside is that existing error structures are ignored. The question as to which of the two approaches, stochastic and deterministic clustering, is preferred by users, comes with a clear answer. Most applications use deterministic models. There certainly are multiple reasons for this preference. One reason may be that deterministic models seem easy to interpret. Another reason is that programs for stochastic clustering are part of only a few of the general purpose statistical software packages; S-plus, R, and SAS are among these.

831

An example of stochastic classification can be seen in a method introduced by von Eye and Gardiner (2004; von Eye & Bogat, 2005). The authors propose splitting the multivariate space of the variables to be used for classification into sectors. For each sector, the probability is estimated from the data under a hypothesis of data generation. Examples of data generation processes are those that generate multinormally distributed data, or uniformly distributed data. Based on the estimated probability, an expected frequency can be calculated, which then is compared with the frequency that was observed for this sector. Algorithmically, the following six steps are performed. 1. Splitting each of the d variables into cj segments, with cj > 1 and j = 1, . . . , d, where j indexes the variables; 2. Crossing the segmented variables; this results in a cross-classification with Πj cj sectors; 3. Estimating the probability of each sector. Typically, these probability do not exist a priori. Therefore, they are estimated from the data. Each of the Πj cj sectors has the boundaries c1i and c1i+1 on the first variable, c2j and c2j+1 on the second variable, . . . , and cdl and cdl+1 on the d-th variable, where the subscripts indicate the segments and the superscripts indicate the variables. The probability of being located in the sector with these boundaries is ) ( p z1i − z1i+1 , z2j − z2j+1 , . . . , zdl − zdl+1 z2i+1 z2j+1

=

∫ ∫ z1i z2j

zdl+1

...

∫

) ( Ψ z1 , z2 , . . . , zd dz1 dz2 . . . dzd

zdl

A numerical solution for this expression was proposed by Genz (1992). 4. Estimating expected sector frequencies as ei, j, . . . ,l = Npi, j, . . . ,l , where pi, j, . . . ,l is short for the expression given in the last equation. 5. Performing sector-specific tests. The null hypothesis for each of the tests is E[oi, j, . . . ,l ] = ei, j, . . . ,l , where oi,j, . . . ,l is the observed frequency for the sector with subscripts i, j, . . . , l. The tests can be performed using most of the residual tests from log-linear modeling. These include, for example, the Pearson X2 –component test, Xi, j, . . . , l = (oi, j, . . . , l − ei, j, . . . , l )2∕ei, j, . . . , l , where oi,j, . . . , l is the observed frequency for the sector with subscripts i, j, . . . , l, and ei, j, . . . ,l is the corresponding expected frequency. 𝛼-protection is advisable.

832

Person-Oriented Approaches

Based on the sector-specific test, an omnibus test can be devised by summing up the X2 –components. The resulting summary statistic can be used to test whether, overall, the cross-classification of segments follows a multinormal distribution. The test is approximately distributed as 𝜒 2 with df = (Πd j=1 cj ) − 2d −dcov − 1, where cj is the number of segments of the jth variable. The term dcov indicates the number of covariances taken into account, typically dcov = (d). That is, typically all covariances are taken into account. 6. Changing the number of segments and restarting at Step 1. This sequence of steps is repeated until the clusters (sectors) are detected that contain significantly more (or fewer) cases than expected. A data example follows after all of the eight criteria have been reviewed. Criterion 6: The Selection of Base Measures Many clustering methods start from an N × N similarity matrix, e.g., a covariance matrix, a correlation matrix, a distance matrix, or a matrix that contains coefficients that count [in percent] the number of incidences in which two objects’ characteristics match. The coefficients in such similarity matrices are termed base measures. The results of simulation studies suggest that “ . . . resulting clusters depend more on the underlying similarity criterion than on the physical process of cluster formation” (Wishart, 1970, p.1). The most popular base measures used in applications are Pearson’s correlation, r, and the Euclidean distance. Correlation measures classify objects as identical if their profiles are parallel. r is not sensitive to differences in standard deviation, mean, or both. Distance measures quantify spatial distance, regardless of whether profiles are parallel or not. Correlation and distance scores coincide only if the distance, d = 0. For an illustration of these relations, see von Eye, Mun, and Indurkhya (2004). When it comes to selecting from the many available base measures, it is important to keep in mind that, as we said earlier in this section, there is no true cluster solution. Clustering algorithms, in particular base measures, are sensitive to specific data characteristics, but not to others. Therefore, before selecting a base measure, researchers can do worse than answering the question concerning the data characteristics they wish to base their cluster solution on. For example, when spatial proximity is considered, the focus is on level or magnitude of behavior over time, or, if there exists a distinction that separates top from bottom clusters of developmental trajectories, a distance measure makes sense. It should be noted that variables must be

commensurable when Euclidean distances are selected. If variables are not commensurable, variables with larger scores and variances can dominate the cluster solution. When, however, the similarity of profiles is of importance, that is, the degree to which profiles are parallel, trend characteristics are more interesting than levels, when researchers are interested in shape, fluctuation, or curvature similarity, and that regardless of possible differences in location and spread, a correlation measure may be more appropriate. In some instances, a selection of a particular base measure may be hard to justify. In these instances, two strategies can be considered. First, one can create classifications using a selection of base measures. The inspection of the resulting classifications may reveal that some are more meaningful than others. A second strategy can involve using base measures that possess the characteristics of distance as well as correlational base measures. An example of such a coefficient is Cattell’s (1949) coefficient, rP . This coefficient is sensitive to profile shape, mean differences, and differences in standard deviations. To the best of our knowledge, however, this coefficient is not included in any of the general-purpose statistical software packages. Criterion 7: Convex Versus Nonconvex Clusters The shape of clusters can be of importance, for at least two reasons. First, when clusters take a shape that can be depicted, for example, in the form of an ellipsoid or a rectangle, verbal description of clusters will be straightforward as well. Second, when new objects are classified based on an existing cluster structure, one can ask whether these objects are located within the (imaginary) hull that represents a cluster. Many classification methods do create clusters that are convex in shape. Subsets (clusters) Ai are convex if any two points (cases), Oi and Oj , for i ≠ j, can be connected by a straight line that is entirely located within Ai . Examples of convex clusters include ellipses, squares, circles, rectangles, and other types of quadrics (von Eye & Wirsing, 1978, 1980). A number of classification methods, however, create nonconvex clusters. Examples include algorithms that create clusters that can take the shape of bananas. These clusters are not convex because connecting straight lines between data points can be located outside the cluster hull. Among the classification methods that yield nonconvex clusters is the single linkage algorithm. This algorithm agglomerates clusters based on the shortest distance of any cluster member with another element that did not belong to the cluster before. Thus, it tends to create chains rather than convex agglomerates. Hartigan (1975) noted

Statistical Approaches

that “single linkage clusters are famously strung out in long sausage shapes, in which objects far apart are linked together by a chain of close objects” (p. 200) (for graphical illustrations see Hartigan, 1975, pp. 201, 202). Usually, researchers opt for convex clusters because they can be described by using measures of central tendency or parameters that represent their hull. In particular instances, however, chain structures may be interpretable. This can be the case, for example, when events such as the next best step are grouped or social network structures. Criterion 8: Manifest Variable Versus Latent Variable Grouping The clustering algorithms discussed thus far in this chapter operate at the level of manifest, that is, observed variables. Cases are classified based on their relative distances or similarities. The resulting clusters are described also at the level of manifest variables. In contrast, there are methods that create latent, that is, unobserved variables, based on relations among manifest variables. As in structural equation modeling or factor analysis, these latent variables are specified to explain the covariation among the observed variables. Most popular among these methods are latent class analysis (LCA; Lazarsfeld & Henry, 1968; Rost & Langeheine, 1997), and latent class mixture models (LCMM; Muthén, 2001; Nagin, 1999). In this section, we briefly review these methods of latent variable grouping. LCMM is a variant of latent growth curve modeling that is used to depict heterogeneous latent classes of trajectories. In a first approach, Muthén (2001) used baseline data to cluster cases. Then, a latent variable structure is specified to model classes’ trajectories. The method is Bayesian and classes’ trajectories are conditional on class membership. For evaluation, this method uses posterior predictors of class membership. In a second approach, the semiparametric, group-based approach introduced by Nagin (1999) enables researchers to model data with reference to the zero-inflated Poisson, censored normal, and binary logit distributions. A unified approach that combines these methods with IRT and multilevel modeling was recently proposed by Hsieh and collaborators (Hsieh, von Eye, & Maier, 2010; Hiseh, von Eye, Maier, & Chen, 2013). Unobserved latent classes are defined to explain individual differences from the aggregated developmental trajectory in a population. Fixed-effect growth, that is, a fixed trajectory, is hypothesized for each class. That is, all members of one latent class are hypothesized to exhibit the same within-class developmental trajectory. The trajectories are class-specific, that is, trajectories vary across classes.

833

LCA of categorical variables is based on the concept of local independence. For each latent class, the joint probability of the observed frequencies is expressed as the product of the marginal probabilities. In other words, within each latent class, the observed variables are independent. Note, however, that local independence is a common characteristic only of the original latent class models. Mixed Markov models, in particular those for longitudinal data, do not create latent classes based on the restriction of local independence (Langeheine & van de Pol, 1993). When it comes to describing individuals, LCA models share the characteristic that all members of a latent class have the same response probabilities for the categories of the variables included in an analysis. In other words, individuals within the same latent class are treated as identical both in their developmental trajectories and their response probabilities. In addition, members of a latent class are different than individuals in other latent classes who are also considered identical to each other. This concept is more restrictive than the concepts used in cluster analysis where members of a cluster can differ from each other but are, on average, more similar to each other than to members of other clusters. Many researchers consider LCA the categorical variable analogue to factor analysis (Molenaar & von Eye, 1994). In the context of classification, the analogy to cluster analysis may be more interesting: LCA can also be considered a method of cluster analysis for categorical variables. Data Example For the following example, we reanalyze data used by von Eye and Bogat (2005). The data were collected for the first wave of a longitudinal study examining male violence toward their female partners and their children. When the women were pregnant, they were given a structured interview (the Working Model of the Child Interview; Zeanah, Benoit, Hirshberg, Barton, & Regan, 1994) to assess their perceptions and subjective experiences of their unborn children. The interviews were audio taped, transcribed, and coded on 5-point Likert scales. The three variables analyzed here—fear, joy, and anxiety—represent affective features of maternal representations (for more detail see Huth-Bocks, Levendosky, Theran, & Bogat, 2004). Each scale was coded so that low scores represent low levels of the mood represented by the scale. We now analyze these variables using three taxometric methods, complete linkage, Ward’s method, and the space-segmenting method proposed by von Eye and Gardiner (2004; von Eye & Bogat, 2005). The selection of these three methods can be explained as follows. The first

Person-Oriented Approaches 75 WARD 3 1 2 3

Count

50

25

0

0

1

2

3 FEAR

4

5

6

75 WARD 3 1 2 3

50 Count

two methods are both hierarchical. We use the Euclidean distance as base measure to make results comparable and because the space-segmenting method is sensitive to agglomerations in data space as well. Of the two hierarchical algorithms, Ward’s method is known for small groupings to be swallowed by larger groupings. Therefore, we expect larger clusters to result from this method than from complete linkage. The space-segmenting method can be based on the assumption of multinormality. It is, thus, a stochastic method. The other two are deterministic. All three of the methods used here are exhaustive although, theoretically, singletons could result from either hierarchical method and the space segmenting method could result in empty sectors and sectors with very small numbers of cases. Based on the distance scales in the dendrograms (not shown here), we select for both hierarchical methods the three-cluster solution. The density plots of the three clustering variables for the three clusters from Ward’s method are given in Figure 18.6. The corresponding plots for the three clusters from complete linkage are given in Figure 18.7. Cluster 1 from Ward’s method contains 48 cases. It is the smallest of the three clusters. Women in this group report average levels of fear, above average levels of joy, and slightly above average levels of anxiety. Cluster 2 is the largest cluster. It contains 103 women. These women report below average levels of fear, average levels of joy, and below average levels of anxiety. The 50 women in the third cluster report slightly above average levels of fear, clearly below average levels of joy, and elevated levels of anxiety. Cluster 1 from complete linkage is the same as Cluster 1 from Ward’s method. Cluster 2 contains 67 women who report below average levels of fear, a rather wide range with an average close to the scale midpoint of joy, and below average levels of anxiety. Cluster 3 contains 86 women who report slightly below average levels of fear, below average levels of joy, and elevated levels of anxiety. We do not see any problems interpreting any of these profiles. More important, however, for the present discussion is that the two hierarchical solutions overlap only in part. This is shown in the cross-tabulation of the two cluster solutions in Table 18.5. Table 18.5 shows that the first clusters from both methods contain the same respondents. The big second cluster from Ward’s method is split such that 67 of the 103 women are in the second cluster from complete linkage, and the remaining 36 women are now in the third cluster from complete linkage. The third cluster from Ward’s method goes entirely into the third cluster from complete linkage.

25

0

0

1

2

3 JOY

4

5

6

75 WARD 3 1 2 3

50 Count

834

25

0

0

1

2

3 4 ANXIETY

5

6

Figure 18.6 Density plots of fear, joy, and anxiety in three clusters (Ward’s method). See footnote 1.

From this pattern of results, we conclude that the women in Cluster 1 are different enough from the other respondents in their location in the Fear × Joy × Anxiety space that the different characteristics of the two clustering algorithms do not affect the composition of this group. Cluster 2 from Ward’s method is the largest cluster. It is possible that it resulted from swallowing small numbers of groups that, in complete linkage, stay together until they join Cluster 3. In support of this argument are (1) the agglomeration sequences that can be seen in the dendrograms (not shown

Statistical Approaches 140

1 2 3

100 Count

TABLE 18.5 Cross-Tabulation of the Three-Cluster Solutions from Ward’s Method (Rows) and Complete Linkage (Columns)

COMPLINK 3

120

835

1 2 3 Total

80 60

1

2

3

Total

48 0 0 48

0 67 0 67

0 36 50 86

48 103 50 201

40 20 0

0

1

2

3 FEAR

4

5

6

140 COMPLINK 3 1 2 3

120

Count

100 80 60 40 20 0

0

1

2

3 JOY

4

5

6

140 COMPLINK 3 1 2 3

120

Count

100 80 60 40 20 0

0

1

2

3 4 ANXIETY

5

6

Figure 18.7 Density plots of fear, joy, and anxiety in three clusters (complete linkages). See footnote 1.

here) and (2) the fact that, in all three variables, Cluster 2 in Ward’s solution has the largest variances (Figure 18.6). The third classification of the fear–joy–anxiety data was created with the space segmenting methodology. To keep the number of sectors in the data space manageable, we dichotomized the variables at their scale-specific mean. To compare the cross-classification of the three dichotomized variables with the solutions from the hierarchical methods, we crossed the 2 × 2 × 2 table of JoyD, FearD, and AnxietyD (where the D indicates that the variables

are dichotomized) with the classifications from Ward’s method and complete linkage. Table 18.6 displays the cross-classification with the solution from Ward’s method. Table 18.6 shows an interesting pattern. The first of Ward’s clusters appears unchanged in the crossclassification of the three dichotomized variables. These are the individuals with the variable profile 2 2 2. The third of Ward’s clusters remains intact also. These are the respondents with variable profile 2 2 1. In contrast, the second cluster from Ward’s method is distributed over six of the eight possible variable profiles. This supports, from a new perspective, the interpretation of this cluster as possibly heterogeneous. When the parallel cross-classification is created with the solution from complete linkage (table not shown here), a slightly different picture emerges. The first cluster—it is identical with the one from Ward’s method—remains intact again. The second cluster is spread over the first four of the eight possible variable profiles, and the third cluster is spread over three of the eight variable profiles. This difference is not surprising, considering that the two hierarchical solutions differed in the second and the third clusters. We estimated the cell frequencies of the 2 × 2 × 2 three-way cross-classification based on the assumption of symmetric distributions, we examined each of the eight cells by using the exact binomial test, and we protected the significance threshold 𝛼 by invoking Bonferroni’s procedure. Table 18.7 displays the results. Table 18.7 shows, in its left column, the eight patterns from low levels of fear, joy, and anxiety (Pattern 1 1 1) through high levels in all three variables (Pattern 2 2 2). In its second column, the table shows the observed frequencies, m, for each pattern. This is followed by the expected ̂ and the binomial tail probabilities, p. The frequencies, m, final column contains asterisks for those three patterns that were observed at different rates than expected under the assumption of independent symmetric distributions. The asterisks thus indicate clusters that deviate from the expected density in the sector that is described by the pattern in the first column. The frequencies in the remaining five patterns do not differ from expectancy.

836

Person-Oriented Approaches TABLE 18.6 Cross-Classification of JoyD, FearD, and AnxietyD with the Three-Cluster Solution Created with Ward’s Method WARD3

ANXIETYD

FEARD

JOYD

Frequency

Cumulative frequency

Percent

1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2

1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2

1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2

0 0 0 0 0 0 0 48 13 20 12 22 13 23 0 0 0 0 0 0 0 0 50 0

0 0 0 0 0 0 0 48 61 81 93 115 128 151 151 151 151 151 151 151 151 151 201 201

0.000 0.000 0.000 0.000 0.000 0.000 0.000 23.881 6.468 9.950 5.970 10.945 6.468 11.443 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 24.876 0.000

1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3

TABLE 18.7 Space-Segment Clusters of the Cross-Classification of the Dichotomized Variables Fear, Joy, and Anxiety Pattern

m

̂ m

p

111 112 121 122 211 212 221 222

13. 13. 20. 23. 12. 50. 22. 48.

25.942 16.435 32.740 17.065 16.267 31.565 17.254 27.492

.00241467 * .22963101 .00701107 .08851520 .16502180 .00051574 * .14311317 .00007017 *

The first cluster with a discrepant density is constituted by Pattern 1 1 1. Almost 26 respondents were expected to report below average levels of fear, joy, and anxiety, but only 13 did display this pattern. The second cluster with a discrepant density is constituted by Pattern 2 1 2. These are women who report above average levels of fear and anxiety and below average levels of joy. Fifty women exhibit this pattern, but only about 31 had been expected. The third cluster with a discrepant number of cases is constituted by Pattern 2 2 2. These are women who indicate above-average levels in all three emotions. About 27 had been expected to show this pattern, but 48 did. Comparison of Cluster Solutions The two hierarchical cluster solutions, created with Ward’s method and complete linkage, can only differ because of

Cumulative percent 0.000 0.000 0.000 0.000 0.000 0.000 0.000 23.881 30.348 40.299 46.269 57.214 63.682 75.124 75.124 75.124 75.124 75.124 75.124 75.124 75.124 75.124 100.000 100.000

differences in the agglomeration algorithms. As is well known, in Ward’s method smaller clusters tend to gravitate toward larger clusters, thus overriding, in part, distance characteristics of cases. This can be illustrated by the large variances of the big, second cluster of the emotion data in Figure 18.6. This cluster exhibits large variances in all three variables that were used for clustering. Ward’s method is, therefore, the method of choice when researchers assume that one large group of individuals exists as well as smaller groups that are remote in the data space from the large group. Ward’s method also tends to create convex and, by default, nonoverlapping clusters. So does complete linkage. However, with complete linkage, there is less of a tendency for larger clusters to absorb smaller ones. The space-segmenting method creates natural groupings by segmenting the data space in blocks. The solution of this method can be optimized by varying the number of segments (more segments result in smaller numbers of cases per segment) and by varying assumptions concerning the data-generating mechanism. For example, researchers can test the hypothesis that agglomerations near the centroid of the data space are excessive assuming the data are multinormally or, alternatively, uniformly distributed. Different numbers of segments and different distributional assumptions can result in different segments that contain more cases or segments that contain fewer cases than expected under these assumptions.

Conclusion and Future Directions

Discussion: Cluster Analysis and Person-Oriented Research With respect to the tenets of person-oriented research, cluster analytic methods, mixture distribution decomposition, latent class analysis or, in general, taxometric methods are most interesting. These methods allow one to create groups of individuals that are similar because of their profiles and not by definition. By implication, clusters are more homogeneous than populations. Developmental trajectories are, therefore, cluster-specific and do not describe everybody. They only describe those who belong to a particular cluster. If singletons are considered clusters, cluster analysis will allow researchers to also identify those cases that cannot meaningfully be assigned to a cluster. Reasons for these individuals to be different include unique profiles and, of course, measurement error. Another most important characteristic of cluster solution is that it is far more likely that dimensional identity holds for a cluster than for an entire population. The meaning and the test statistical characteristics of measurement instruments can be expected to be the same for everybody within a cluster (can be expected means should be ascertained). This is less important when variables are nominal level. It is important, however, when the variables used in a study are test scores. As was discussed by von Eye and Bergman (2003), test statistical characteristics of measurement scales can differ across populations (this is also discussed as differential validity or differential item functioning; von Eye, Bergman, & Hsieh, 2015). Whenever items or tests differ in their characteristics across groups, scores from these instruments cannot easily be compared when they describe individuals from different groups, even if they were given the same score on a scale. Individuals from the same group, however, can be compared. Because of these interesting characteristics, many studies in which researchers used taxometric methods have been published under the auspices of person-oriented research. There are downsides, however. The first of these is that, although taxometric methods are designed to capture unmeasured heterogeneity, it can be very hard to make sure the groups that appear as clusters do exist. To establish a cluster solution, researchers typically undertake two steps. In the first, the best solution is identified. This is done, when latent class analysis is performed, by comparing information criteria from solutions that differ in the number of latent classes. The most likely grouping can thus be identified. When hierarchical clustering algorithms are applied, researchers select from solutions that differ in their distance measure, R2 , or the Rand index (Steinley, 2004). The performance of these indexes is, however, still subject to discussion and research. This is why the second

837

step is equally important. In this step, cluster members are compared on measures that have not been used for clustering. This way, external validity of clusters can be established. There is another reason why these steps are very important. Many taxometric methods deliver solutions under virtually all conditions. Not all of these solutions, however, describe groups that naturally or meaningfully exist. Matters are complicated by the fact previously illustrated that multiple solutions can be created, and many of these may be interpretable and meaningful. The selection from multiple solutions that have these characteristics is no easy task. Still, taxometric methods are extremely useful options when researchers suspect that unmeasured heterogeneity exists that may be interesting and should be captured because it has the potential of fogging results that are created when this heterogeneity is ignored.

CONCLUSION AND FUTURE DIRECTIONS There is a striking continuity that connects the ideas of Stern, Lewin, and Block with those of present-day advocates of the person-oriented approach (e.g., Magnusson, Bergman, von Eye). We briefly summarize those continuities as well as suggest where, going forward, personoriented approaches might lead research and application in developmental psychopathology. The basic principles for psychological inquiry from a person-oriented approach have existed for more than 100 years. Understanding the individual is widely held to be central to the enterprise of psychological science. Concerns regarding a focus on lawfulness at the expense of understanding important aspects of individual human experience have been raised consistently from Stern to the present. Importantly, all theorists focus on the complexity of the process of development, indicating that there will be multiple determinants of the phenomena or process under investigation. The analogy of the individual as a dynamic system, whose component parts are integrated in a meaningful and coherent way also underpins much of the writing. Within this framework, concerns with change as well as stability are considered important aspects of human behavior. Researchers also emphasize the false dichotomy of the psychological and physiological aspects of an individual and the importance of understanding the individual and how he or she transacts with the environment. Although research methods have evolved dramatically in the past 100 years, basic concerns with the inadequacies

838

Person-Oriented Approaches

of extant statistical techniques as a means to elucidate individual behavior as well as attempts to develop alternative methods has been ongoing. Concerns with aggregation of data to explain individuals is a theme that continually emerges in person-oriented research. For example, as we noted previously, Block was interested in subgroups (i.e., heterogeneity within the aggregate) and concerned about the use of factor analysis in personality research, including his own. Lewin’s sentiments about the uniqueness of the individual child, quoted earlier, speak to the ecological fallacy decades before it received its moniker. Modern researchers, as we have noted, critique methods that aggregate individuals as well as those that make questionable and simplifying assumptions about variable relationships, data distributions, and random error. These researchers have also developed new, and revised existing, methods that circumvent some of these problems. Finally, following from the dynamic systems view, with integrated components together reflecting the system, person-oriented researchers have long realized the need to make, as far as possible, the whole person the object of study. In the early days of psychology, this goal was strived for by applying subjective methods, for example, when psychiatrists used observations and interviews to divide a sample of patients into different types. During the last decades, objective statistical methods of pattern analysis and classification have been increasingly applied; methods that are based on the pattern of variable values as the analytic unit. Hence, the old objections against a typological approach as being subjective and unscientific are no longer valid. What measures are used to assess individuals is another thread of continuity between the past and present. Early assessments of intelligence were highly individualized (e.g., Binet). Stern’s consternation with the overuse of intelligence tests and the intelligence quotient led him to develop methods that reflected person–world convergence; that is, test instruments should not compare the individual to others (as in classical test theory) but rather to himself or herself vis-à-vis the environment in which he or she interacts (Lamiell, 2010a). Loevinger’s focus on differential scaling predates von Eye et al.’s (2015) reformulation of IRT as a person-oriented tool. Block’s Q-set, described earlier, attempted to better reflect the complexity of the individual. von Eye (e.g., von Eye & Bergman, 2003; von Eye et al., 2015) has consistently raised concerns regarding dimensional identity of assessment instruments—items, measures, and constructs may have different meaning for different individuals at different times in their development. Going forward, issues of scaling and measurement are in need of more attention; researchers still, too often,

rely on the principles of classical test theory that do not align with the person-oriented approach. How should the field of developmental psychopathology integrate the person-oriented approach as a legitimate paradigm and not just the poor stepchild of the variable-oriented approach? One idea is that developmental psychopathology researchers explicitly identify whether they employ a variable- or person-oriented approach. At present, as far as we are aware, only research that employs a person-oriented approach identifies it as such. von Eye et al. (2015) remind us that the search for universals, which is best done via variable-oriented approaches, can be important in the early stages of a new field of inquiry. However, research conducted from a variable-oriented approach cannot normally make statements about the individual. Thus, variable-oriented research should refrain from making such statements in the hypotheses and in the interpretation of results. The language that researchers employ in this regard is important. Hypotheses and results must be formulated such that the person- versus variable-orientation of researchers becomes visible and, more important, defensible. For example, it is not valid to state that a person with long experience as a victim will be depressed, if the basis for such a conclusion is simply a correlation between duration of exposure and depression (that would amount to committing the sin of ecological fallacy). Person-oriented statements require person-oriented research and vice versa (but the first of these is more important). Discussion sections often include statements that the findings cannot be generalized to a population other than the one studied, but limitations inherent in the variable-oriented approach are not presented (e.g., cross sectional data do not elucidate process, universals can only be deemed such if validated by person-oriented research, assessment instruments may not have dimensional identity, aggregation of data limits the interpretation of results). Framing the argument and hypotheses as either variableor person-oriented and acknowledging the limitations of each approach in the discussion section would be more than a mere exercise; it would provide scaffolding for clearer thinking in theory and methods. The need for transparency about the underlying paradigm driving research is analogous in some ways to Stern’s (1911) delineation of four specific types of research in differential psychology (Lamiell, 2010a). More recently, in discussing the tension between person- and variable-oriented approaches, Cervone (2005) noted the need to acknowledge explicitly that these two approaches lead to two different types of information.

References

Focusing specifically on person-oriented research, what does the future hold as it applies to the field of developmental psychopathology? The NIMH Research Doman Criteria (RDoC, 2011), in an attempt to generate innovative and integrative research on brain and behavior, provides a starting point. The research domain criteria matrix suggested has units of analysis (from genes to paradigms) as the columns and domains/constructs (e.g., negative valence systems, cognitive systems) as the rows. The initiative emphasizes linking behavior with brain functioning and genetics. Unfortunately, developmental and environmental aspects of psychopathology are not in the matrix, although NIMH indicates that these are critical elements for researchers to consider. Research from a person-oriented approach would include environment and development in the matrix itself. Because the individual is a dynamic system, the person-oriented approach cannot explain the transaction of brain and behavior factors without taking environment and development into consideration. In addition, the patterning of these different factors within individuals or groups of individuals would be necessary for the research to be person oriented. Already we see the first signs of person-oriented research moving in this direction. The cortisol/behavior (Towe-Goodman et al., 2011) and cortisol/mental health/environment research (Bogat et al., 2012) cited earlier, research on ADHD that bases classifications on differential patterns of brain functioning (e.g., Fair, Bathula, Nikolas, & Nigg, 2012), and formulations of temperament calling for person-oriented methods that incorporate aspects of child behavior as well as central and peripheral nervous system factors (Nigg, 2006) all begin to integrate facets of the RDoC matrix. The NIMH RDoC is, of course, part of a wider movement within psychology and medicine toward translational research; that is, taking findings from basic research and applying them to clinical research with the goal of enhancing health or preventing problems from developing in the first place. Person-oriented research has a strong role to play in translational research. In medicine, cutting-edge cancer research eschews a one-size-fits-all treatment approach. Person-oriented approaches, whereby physicians and researchers analyze genetic and molecular abnormalities in an individual’s cancer, allow for more precise treatments to cure the cancer and cause the fewest side effects possible. Person-oriented approaches offer the same promise for psychological interventions. For example, batterer intervention programs (BIPs), often mandated by courts in cases where men perpetrate intimate partner violence, are generally ineffective. A meta-analysis by Feder

839

and Wilson (2005) found that the programs had no effect on victim’s reports of revictimization and only minimal effects on police reports of violence. An NIJ meeting convened to address this problem provided guidelines for practitioners and researchers going forward (National Institute of Justice, 2011); however, only one research recommendation was somewhat person oriented (“Study experiences in BIPs of men by ethnic group”). This is unfortunate. In the case of BIPs or other interventions, translational research, using person-oriented approaches to understand individual behavior (genetic, psychological, and environmental causes), holds the promise of individualized treatments to help the broadest number of individuals. In conclusion, a core goal of developmental psychopathology is to understand individual development. The direct and natural path to follow is then to study individual development of persons in real life, taking into consideration that we function as integrated organisms in interaction with our environment. This standpoint naturally leads to a person-oriented approach. Those who claim they can understand individual development by applying variable-oriented methods must carry the burden of proof that such an indirect approach is defensible.

REFERENCES Agresti, A. (2012). Categorical data analysis (3rd ed.). New York, NY: Wiley. Allport, G. W. (1937). Personality—a psychological interpretation. New York, NY: Holt. Ansara, D. L., & Hindin, M. J. (2010). Exploring gender differences in the patterns of intimate partner violence in Canada: a latent class approach. Journal of Epidemiology & Community Health, 64, 849–854. doi: 10.1136/jech.2009.095208 Arabie, P., & Caroll, J. D. (1980). MAPCLUS: A mathematical programming approach to fitting the ADCLUS model. Psychometrika, 45, 211–235.doi: 10.1007/BF02294077 Arminger, G., & Stein, P. (1997). Finite mixture of covariance structure models with regressors: loglikehood function, distance estimation, fit indices, and a complex example. Sociological Methods and Research, 26, 148–182. doi: 10.1177/0049124197026002002 Aude, J. C., Diaz Lazcoz, Y., Codani, J. J., & Risler, J. L. (1999). Applications of the pyramidal clustering method to biological objects. Computers and Chemistry, 23, 303–325. doi: 10.1016/S00978485(99)00006-6 Axen, I., Bodin, L., Bergström, G., Halasz, L., Lange, F., Lövgren, P. W., . . . Jensen, I. (2011). Clustering patients on the basis of their individual course of low back pain over a six month period. BMC Musculoskeletal Disorders, 12, 99–108. doi: 10.1186/1471-2474-12-99 Banfield, J. D., & Raftery, A. E. (1993). Model-based Gaussian and non-Gaussian clustering. Biometrics, 9, 803–821. doi: 10.2307/ 2532201 Bergman, L. R. (1988). You can’t classify all of the people all of the time. Multivariate Behavioral Research, 23, 425–441. doi: 10.1207/ s15327906mbr2304_1

840

Person-Oriented Approaches

Bergman, L. R. (2000). The application of a person-oriented approach: Types and clusters. In L. R. Bergman, R. B. Cairns, L. G. Nilsson, & L. Nystedt (Eds.), Developmental science and the holistic approach (pp. 137–154). Mahwah, NJ: Lawrence Erlbaum.

Borkenau, P., & Ostendorf, F. (1998). The Big Five as states: How useful is the five-factor model to describe intra-individual variations over time? Journal of Research in Personality, 32, 202–221. doi: 10.1006/jrpe. 1997.2206

Bergman, L. R., & Andersson, H. (2010). The person and the variable in developmental psychology. Journal of Psychology, 218, 155–165. doi: 10.1027/0044-3409/a000025

Boscardin, C. K., Muthén, B., Francis, D. J., & Baker, E. L. (2008). Early identification of reading difficulties using heterogeneous developmental trajectories. Journal of Educational Psychology, 100, 192–208. doi: 10.1037/0022-0663.100.1.192

Bergman, L. R., & Magnusson, D. (1991). Stability and change in patterns of extrinsic adjustment problems. In D. Magnusson, L. R. Bergman, & B. Törestad (Eds.), Problems and methods in longitudinal research: Stability and change (pp. 323–346). New York, NY: Cambridge University Press. doi: 10.1017/CBO9780511663260.017 Bergman, L. R., & Magnusson, D. (1997). A person-oriented approach in research on developmental psychopathology. Development and Psychopathology, 9, 291–319. doi: 10.1017/S095457949700206X Bergman, L. R., Magnusson, D., & El-Khouri, B. M. (2003). Studying individual development in an interinidividual context: A person-oriented approach. Mahwah, NJ: Erlbaum. Bergman, L. R., & Trost, K. (2006). The person-oriented versus the variable-oriented approach: Are they complementary, opposites, or exploring different worlds? Merrill-Palmer Quarterly, 52, 601–632. doi: 10.1353/mpq.2006.0023 Block, J. (1955). The difference between Q and R. Psychological Review, 62, 356–358. doi: 10.1037/h0042760 Block, J. (1961). The Q-sort method in personality assessment and psychiatric research. Springfield, IL: Charles C. Thomas. Block, J. (1971). Lives through time. Berkeley, CA: Bancroft Books. doi: 10.1037/13141-000 Block, J. (1995). A contrarian view of the five-factor approach to personality description. Psychological Bulletin, 117, 187–215. doi: 10.1037/ 0033-2909.117.2.187 Block, J. (2001). Millennial contrarianism: The five-factor approach to personality description 5 years later. Journal of Research in Personality, 35, 98–107. doi: 10.1006/jrpe.2000.2293 Block, J. (2010). The five-factor framing of personality and beyond: Some ruminations. Psychological Inquiry, 21, 2–25. doi: 10.1080/ 10478401003596626 Bock, H. H. (1974). Automatische Klassifikation [Automatic Classification]. Gottingen, Germany: Vandenhoeck & Ruprecht. Bogat, G. A. (2009). Is the person orientation necessary in community psychology? American Journal of Community Psychology, 43, 22–34. doi: 10.1007/s10464-008-9215-7 Bogat, G. A., Leahy, K. L., von Eye, A., Maxwell, C., Levendosky, A. A., & Davidson, W. S. (2005). The influence of community violence on the functioning of battered women. American Journal of Community Psychology, 36(1–2), 123–132. doi: 10.1007/s10464-005-6237-2 Bogat, G. A., Levendosky, A. A., DeJonghe, E., Davidson, W. S., & von Eye, A. (2004). Pathways of suffering: The temporal effects of domestic violence on women’s mental health. Maltrattamento e abuso all’infanzia, 6(2), 97–112. Bogat, G. A., Levendosky, A. A., Lonstein, J. S., Martinez-Torteya, C., Muzik, M., & Granger, D. A. (2012, October). Effects of pre and postpartum intimate partner violence (IPV) on maternal and HPA Axis functioning. Paper presented at International Society for Developmental Psychobiology, New Orleans, LA. Bollen, K. A. (1989). Structural equations with latent variables. New York, NY: Wiley.

Bronfenbrenner, U. (1977). Toward an experimental ecology of human development. American Psychologist, 32, 513–531. doi: 10.1037/0003066X.32.7.513 Bronfenbrenner, U. (1986a). Ecology of the family as a context for human development: Research perspectives. Developmental Psychology, 22, 723–742. doi: 10.1037/0012-1649.22.6.723 Bronfenbrenner, U. (1986b). Recent advances in research on the ecology of human development. In R. K. Silbereisen, K. Eyferth, & G. Rudinger (Eds.), Development as action in context: Problem behavior and normal youth development (pp. 287–309). Heidelberg, Germany: Springer-Verlag. doi: 10.1007/978-3-662-02475-1_15 Bronfenbrenner, U., & Morris, P. A. (1983). The evolution of environmental models in developmental research. In P. H. Mussen & W. Kessen (Eds.), Handbook of child psychology: History, theory, and methods (Vol. 1, pp. 357–414). New York, NY: Wiley. Cairns, R. B. (1983). The emergence of developmental psychology. In P. H. Mussen (Series Ed.) & W. Kessen (Vol. Ed.), Handbook of child psychology: History, theories, and methods, (4th ed., Vol. 1, pp. 41–101). New York, NY: Wiley. Cattell, R. B. (1946). Description and measurement of personality. Oxford, UK: World Book Company. Cattell, R. B. (1949). rp and other coefficients of pattern similarity. Psychometrika, 14, 279–298. doi: 10.1007/BF02289193 Cattell, R. B., & Tsujioka, B. (1964). The importance of factor-trueness and validity versus homogeneity and orthogonality in test scales. Educational and Psychological Measurement, 24, 3–30. doi: 10.1177/ 001316446402400101 Cauce, A. M. (2008). Parenting, culture, and context: Reflections on excavating culture. Applied Developmental Science, 12, 227–229. doi: 10.1080/10888690802388177 Cervone, D. (2005). Personality architecture: within-person structures and processes. Annual Review of Psychology, 56, 423–452. doi: 10.1146/annurev.psych.56.091103.070133 Cicchetti, D., & Scheider-Rosen, K. (1986). An organizational approach to childhood depression. In M. Rutter, C. Izard, & P. Read (Eds.), Depression in young people: Clinical and developmental perspectives (pp. 71–134). New York, NY: Guilford. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20, 37–46. doi: 10.1177/ 001316446002000104 Costa, P. T., Herbst, J. H., McCrae, R. R., Samuels, J., & Ozer, D. J. (2002). The replicability and utility of three personality types. European Journal of Personality, 16, S73–S87. doi: 10.1002/per.448 Cronbach, L. J. (1953). Correlations between persons as a research tool. In O. H. Mowrer (Ed.), Psychotherapy: Theory and research, (pp. 376–388). New York, NY: Ronald. doi: 10.1037/10572-014 Davies P. T., & Forman, E. M. (2002). Children’s patterns of preserving emotional security in the interparental subsystem. Child Development, 73, 1880–1903. doi: 10.1111/1467-8624.t01-1-00512 Davies, P. T., Sturge-Apple, M. L., Cicchetti, D., & Cummings, E. M. (2007). The role of child adrenocortical functioning in pathways between interparental conflict and child maladjustment. Developmental Psychology, 43, 918–930. doi: 10.1037/0012-1649.43.4.918

References

841

Del Giudice, M., Ellis, B. J., & Shirtcliff, E. A. (2011). The adaptive calibration model of stress responsivity. Neuroscience & Biobehavioral Reviews, 35, 1562–1592. doi: 10.1016/j.neubiorev.2010.11.007

Greenland, S. (2001). Ecologic versus individual-level sources of bias in ecologic estimates of contextual health effects. International Journal of Epidemiology, 30, 1343–1350. doi: 10.1093/ije/30.6.1343

Del Giudice, M., Hinnant, J. B., Ellis, B. J., & El-Sheikh, M. (2012). Adaptive patterns of stress responsivity: A preliminary investigation. Developmental Psychology, 48, 775–790. doi: 10.1037/a0026519

Gunnar, M. R., & Quevedo, K. M. (2007). Early care experiences and HPA axis regulation in children: A mechanism for later trauma vulnerability. Progress in Brain Research, 167, 137–149. doi: 10.1016/S0079-6123(07)67010-1

American Psychiatric Association (2013). Diagnostic and statistical manual of mental disorders, 5th edition. Arlington, VA: American Psychiatric Association Publishing. Dolan, C. V., Molenaar, P. C. M., & Boomsma, D. I. (1989). LISREL analysis of twin data with structured means. Behavior Genetics, 19(1), 51–62. doi: 10.1007/BF01065883 Dolan, C. V., & van der Maas, H. L. J. (1997). Fitting multivariate normal finite mixtures subject to structural equation modeling. Psychometrika, 63, 227–253. doi: 10.1007/BF02294853 Donnellan, M. B., & Robins, R. W. (2010). Resilient, undercontrolling, and overcontrolling personality types: Issues and controversies. Social and Personality Psychology Compass, 4, 1070–1083. doi: 10.1111/ j.1751-9004.2010.00313.x Ellis, B. J., Del Giudice, M., & Shirtcliff, E. A. (2013). Beyond allostatic load: The stress response system as a mechanism of conditional adaptation. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (2nd ed., pp. 251–284). New York, NY: Wiley. Everitt, B. S. (1998). The Cambridge dictionary of statistics. Cambridge, UK: Cambridge University Press. Everitt, B. S., Landau, S., & Leese, M. (2001). Cluster analysis (4th ed.). London, UK: Arnold. Eysenck, H. J. (1954). The science of personality: nomothetic! Psychological Review, 61, 339–342. doi: 10.1037/h0058333 Fair, D. A., Bathula, D., Nikolas, M. A., & Nigg, J. T. (2012) Distinct neuropsychological subgroups in typically developing youth inform heterogeneity in children with ADHD. Proceedings of the National Academy of Sciences, 109, 6769–6774. doi: 10.1073/pnas.1115365109 Feder, L., Wilson, D. B. (2005). A meta-analytic review of courtmandated batterer interventions programs: can courts affect abusers’ behavior? Journal of Experimental Criminology, 1, 239–62. doi: 10.1007/s11292-005-1179-0 Fidalgo, A. M. (2005). Mantel-Haenszel methods. In B. S. Everitt & D. C. Howell (Eds.), Encyclopedia of statistics in behavioral science (Vol. 3, pp. 1120–1126). Chichester, UK: Wiley. doi: 10.1002/ 0470013192.bsa364 Finkelstein, J. W., von Eye, A., & Preece, M. A. (1994). The relationship between aggressive behavior and puberty in normal adolescents: A longitudinal study. Journal of Adolescent Health, 15, 319–326. doi: 10.1016/1054-139X(94)90605-X Fogel, A., Garvey, A., Hsu, H., & West-Stroming, D. (2006). Change processes in relationships: Relational-historical research on a dynamic system of communication. Cambridge, UK: Cambridge University Press. doi: 10.1017/CBO9780511489686 Foti, R. J., Thompson, N. J., & Allgood, S. F. (2011). The pattern-oriented approach: A framework for the experience of work. Industrial and Organizational Psychology–Perspectives on Science and Practice, 4(1), 122–125. doi: 10.1111/j.1754-9434.2010.01309.x

Hartigan, J. A. (1975). Clustering algorithms. New York, NY: John Wiley. Hartshorne, H., & May, M. A. (1928). Studies in the nature of character. Vol. 1: Studies in deceit. New York, NY: Macmillan. Hartshorne, H., & May, M. A. (1929). Studies in the nature of character. Vol 2: Studies in service and self control. New York, NY: Macmillan. doi: 10.1037/11334-000 Hartshorne, H., May, M. A., & Shuttleworth, F. K. (1930). Studies in the nature of character. Vol. 3: Studies in the organization of character. New York, NY: Macmillan. Horney, J. D., Osgood, D. W., & Marshall, I. H. (1995). Criminal careers in the short-term: Intra-individual variability in crime and its relation to local life circumstances. American Sociological Review, 60, 655–673. doi: 10.2307/2096316 Hoshino, T., & Bentler, P. M. (2011). Bias in factor score regression and a simple solution. Department of Statistics, University of California–Los Angeles. Retrieved from http://www.escholarship.org/ uc/item/8w9514jw Hsieh, C. A., von Eye, A., & Maier, K. S. (2010). Using a multivariate multilevel Item Response Theory model to study parallel processes of change: The dynamic association between adolescents’ social isolation and engagement with delinquent peers in the National Youth Survey. Multivariate Behavioral Research, 45, 508–552. doi: 10.1080/00273171.2010.483387 Hsieh, C. A., von Eye, A., Maier, K. S., & Chen, S. H. (2013). A unified latent growth curve model. Structural Equation Modeling: A Multidisciplinary Journal, 20, 592–615. doi: 10.1080/10705511.2013.824778 Huth-Bocks, A. C., Levendosky, A. A., Theran, S. A., & Bogat, G. A. (2004). The impact of domestic violence on mothers’ prenatal representations of their infants. Infant Mental Health Journal, 25, 79–98. doi: 10.1002/imhj.10094 Jöreskog, K. G., & Sörbom, D. (2001). LISREL 8: User’s reference guide. Lincolnwood, IL: Scientific Software International. Keenan, K., Grace, D., & Gunthorpe, D. (2003). Examining stress reactivity in neonates: Relations between cortisol and behavior. Child Development, 74, 1930–1942. doi: 10.1046/j.1467-8624.2003.00647.x Kelso, J. A. S. (1995). Dynamic patterns: The self organization of brain and behavior. Boston, MA: MIT Press. Kelso, J. A. S. (2000). Principles of dynamic pattern formation and change for a science of human behavior. In L. R. Bergman, R. B. Cairns, L. G. Nilsson, & L. Nystedt (Eds.), Developmental science and the holistic approach (pp. 63–83). Mahwah, NJ: Erlbaum. Kiesler, D. J. (1966). Some myths of psychotherapy research and the search for a paradigm. Psychological Bulletin, 65, 110–136. doi: 10.1037/h0022911 Kim, D. (2008). Blues from the neighborhood? Neighborhood characteristics and depression. Epidemiologic Reviews, 30, 101–117. doi: 10.1093/epirev/mxn009

Genz, A. (1992). Numerical computation of multivariate normal probabilities. Journal of Computational and Graphical Statistics, 1, 141–149. doi: 10.1080/10618600.1992.10477010

Kim, J., & Fram, M. S. (2009). Profiles of choice: Parents’ patterns of priority in child care decision-making. Early Childhood Research Quarterly, 24(1), 77–91. doi: 10.1016/j.ecresq.2008.10.001

Gottfredson, M. R. (2005). Offender classifications and treatment effects in developmental criminology: A propensity/event consideration. The ANNALS of the American Academy of Political and Social Science, 602, 46–56. doi: 10.1177/0002716205279937

Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.). New York, NY: Guilford Press. Klump, K. L., Culbert, K. M., Slane, J. D., Burt, S. A., Sisk, C. L., & Nigg, J. T. (2012). The effects of puberty on genetic risk for disordered

842

Person-Oriented Approaches

eating: Evidence for a sex difference. Psychological Medicine, 42(3), 627–638. doi: 10.1017/S0033291711001541 Klump, K. L., Keel, P. K., Sisk, C. L., & Burt, S. A. (2010). Preliminary evidence that estradiol moderates genetic effects on disordered eating attitudes and behaviors during puberty. Psychological Medicine, 40(10), 1745–1754. doi: 10.1017/S0033291709992236 Klump, K. L., Perkins, P., Burt, S. A., McGue, M., & Iacono, W. G. (2007). Puberty moderates genetic influences on disordered eating. Psychological Medicine, 37, 627–634. doi: 10.1017/ S0033291707000189 Krauth, J., & Lienert, G. A. (1973). Die Konfigurationsfrekvenz-analyse (KFA) [Configural frequency analysis (CFA)]. Freiburg, Germany: Karl Alber. Kuhn, T. S. (1962/2012). The structure of scientific revolutions: Fiftieth anniversary edition. Chicago, IL: University of Chicago Press. Kunnen, S. E. (2012). A dynamic systems approach to adolescent development. New York, NY: Psychology Press. Laible, D., Carlo, G., Panfile, T., Eye, J., & Parker, J. (2010). Negative emotionality and emotion regulation: A person-centered approach to predicting socioemotional adjustment in young adolescents. Journal of Research in Personality, 44(5), 621–629. doi: 10.1016/j.jrp.2010.08.003 Lamiell, J. T. (2003). Beyond individual and group differences: Human individuality, scientific psychology, and William Stern’s critical personalism. Thousand Oaks, CA: Sage. Lamiell, J. T. (2009). Reviving person-centered inquiry in psychology: Why it’s erstwhile dormancy? In J. Valsiner, P. C. M. Molenaar, M. C. D. P. Lyra, & N. Chaudhary (Eds.), Dynamic process methodology in the social and developmental sciences (pp. 31–43). New York, NY: Springer. doi: 10.1007/978-0-387-95922-1_2 Lamiell, J. T. (2010a). Reflections on some neglected ideas about psychological measurement from the personalistic perspective of William Stern (1871–1938). In A. Toomela & J. Valsiner (Eds.), Methodological thinking in psychology: 60 years gone astray? (pp. 180–207). Charlotte, NC: Information Age Publishing, Inc. Lamiell, J. T. (2010b). Why was there no place for personalistic thinking in 20th century psychology? New Ideas in Psychology, 28, 135–142. doi: 10.1016/j.newideapsych.2009.02.002 Langeheine, R., & van de Pol, F. (1993). A unifying framework for Markov modeling in discrete space and discrete time. Sociological Methods & Research, 18, 416–441. doi: 10.1177/0049124190018004002 Laub, J. H., & Sampson, R. J. (2003). Shared beginnings, divergent lives: Delinquent boys to age 70. Cambridge, MA: Harvard University Press.

Lewin, K. (1931). The conflict between Aristotelian and Galileian modes of thought in contemporary psychology. Journal of General Psychology, 5, 141–177. doi: 10.1080/00221309.1931.9918387 Lewin, K. (1936). Principles of topological psychology. New York, NY: McGraw-Hill. doi: 10.1037/10019-000 Lewin, K. (1939). Field theory and experiment in social psychology: Concepts and methods. American Journal of Sociology, 44, 868–897. doi: 10.1086/218177 Lewin, K., Lippitt, R., & White, R. K. (1939). Patterns of aggressive behavior in experimentally created “social climates.” Journal of Social Psychology, 10, 271–299. doi: 10.1080/00224545.1939.9713366 Lienert, G. A., & Krauth, J. (1975). Configural frequency analysis as a statistical tool for defining types. Educational and Psychological Measurement, 35, 231–238. doi: 10.1177/001316447503500201 Loevinger, J., & Wessler, R. (1970). Measuring ego development: Vol. 1. Construction and use of a sentence completion test. San Francisco, CA: Jossey-Bass. Lundh, L. G., Saboonchi, F., & Wangby, M. (2008). The role of personal standards in clinically significant perfectionism. A person-oriented approach to the study of patterns of perfectionism. Cognitive Therapy and Research, 32(3), 333–350. doi: 10.1007/s10608-006-9109-7 Magnusson, D. (1985). Implications of an interactional paradigm for research on human development. International Journal of Behavioral Development, 8, 115–137. doi: 10.1177/016502548500800201 Magnusson, D. (1988). Individual development from an interactional perspective: A longitudinal study. Hillsdale, NJ: Erlbaum. Magnusson, D. (1995). Individual development: A holistic, integrated model. In P. Moen, G. H. Elder, & K. Luscher (Eds.), Examining lives in context (pp. 19–60). Washington, DC: American Psychological Association. doi: 10.1037/10176-001 Magnusson, D. (2000). The individual as the organizing principle in psychological inquiry: A holistic approach. In L. R. Bergman, R. B. Cairns, L.-G. Nilsson, & L. Nystedt (Eds.), Developmental science and the holistic approach (pp. 33–47). Mahwah, NJ: Erlbaum. Magnusson, D., & Törestad, B. (1993). A holistic view of personality: A model revisited. Annual Review of Psychology, 44, 427–452. doi: 10.1146/annurev.ps.44.020193.002235 Mair, P., & von Eye, A. (2007). Application scenarios for nonstandard log-linear models. Psychological Methods, 12, 139–156. doi: 10.1037/ 1082-989X.12.2.139

Lazarsfeld, P. F., & Henry, N. W. (1968). Latent structure analysis. Boston, MA: Houghton Mifflin.

Mandara, J. (2003). The typological approach in child and family psychology: A review of theory, methods, and research. Clinical Child and Family Psychology Review, 6, 129–146. doi: 10.1023/A:1023734627624

Lennon, M. C., McAllister, W., Kuang, L., & Herman, D. B. (2005). Capturing intervention effects over time: Reanalysis of a critical time intervention for homeless mentally ill men. American Journal of Public Health, 95(10), 1760–1766. doi: 10.2105/AJPH.2005.064402

Martinez-Torteya, C., Bogat, G. A., von Eye, A., & Levendosky, A. A. (2009). Resilience among children exposed to domestic violence: The role of protective and vulnerability factors. Child Development, 80, 562–577. doi: 10.1111/j.1467-8624.2009.01279.x

Lerner, R. M., & Busch-Rossnagel, N. (1981). Individuals as producers of their development: Conceptual and empirical bases. In R. M. Lerner & N. A. Busch-Rossnagel (Eds.), Individuals as producers of their development: A life-span perspective (pp. 1–36). New York, NY: Academic Press. doi: 10.1016/B978-0-12-444550-5.50008-0

Maughan, B. (2005). Developmental trajectory modeling: A view from developmental psychopathology. The ANNALS of the American Academy of Political and Social Science, 602, 118–130. doi: 10.1177/ 0002716205281067

Lerner, R. M., Taylor, C. S., & von Eye, A. (Eds.). (2002). Pathways to positive youth development among diverse youth. San Francisco, CA: Jossey-Bass.

McAdams, D. P., & Walden, K. (2010). Jack Block, the Big Five, and personality from the standpoints of actor, agent, and author. Psychological Inquiry, 21, 50–56. doi: 10.1080/10478401003648740

Lerner, R. M., von Eye, A., Lerner, J. V., & Lewin-Bizan, S. (2009). Exploring the foundations and functions of adolescent thriving within the 3-H study of positive youth development: A view of the issues. Journal of Applied Developmental Psychology, 30, 567–584. doi: 10.1016/j.appdev.2009.07.002

McArdle, J. J. (2012). Foundational issues in contemporary modeling. In B. Laursen, T. D. Little, & N. A. Card (Eds.), Handbook of developmental research methods (pp. 385–410). New York, NY: Guilford. McClelland, D. C. (1951). Personality. New York, NY: Holt. doi: 10.1037/10790-000

References McCrae, R. R., & Alik, J. (Eds.). (2002). The five-factor model of personality across cultures. New York, NY: Kluwer. doi: 10.1007/9781-4615-0763-5 McCullagh, P., & Nelder, J. A. (1989). Generalized linear models (2nd ed.). London, UK: Chapman and Hall. doi: 10.1007/978-1-4899-3242-6 Meehl, P. E. (1950). Configural scoring. Journal of Consulting Psychology, 14, 165–171. doi: 10.1037/h0058049 Meiser, T. (1996). Loglinear Rasch models for the analysis of stability and change. Psychometrika, 61, 629–645. doi: 10.1007/BF02294040 Miller, G. E., Chen, E., & Zhou, E. S. (2007). If it goes up, must it come down? Chronic stress and the hypothalamic-pituitary-adrenocortical axis in humans. Psychological Bulletin, 133, 25–45. doi: 10.1037/ 0033-2909.133.1.25 Molenaar, P. C. M. (1997). Time series analysis and its relationship with longitudinal analysis. International Journal of Sports Medicine, 18, 232–237. doi: 10.1055/s-2007-972720 Molenaar, P. C. M. (2004). A manifesto on psychology as idiographic science: Bringing the person back into scientific psychology, this time forever. Measurement: Interdisciplinary Research and Perspectives, 2, 201–218. doi: 10.1207/s15366359mea0204_1 Molenaar, P. C. M. (2007). On the implications of the classical ergodic theorems: Analysis of developmental processes has to focus on intra-individual variation. Developmental Psychobiology, 50, 60–69. doi: 10.1002/dev.20262 Molenaar, P. C. M. (2010). Testing all six person-oriented principles in dynamic factor analysis. Development and Psychopathology, 22, 255–259. doi: 10.1017/S0954579410000027 Molenaar, P. C. M., Huizenga, H. M., & Nesselroade, J. R. (2003). The relationship between the structure of interindividual and intraindividual variability: A theoretical and empirical vindication of developmental systems theory. In U. M. Staudinger & U. Lindenberger (Eds.), Understanding human development: Dialogues with lifespan psychology (pp. 339–360). Dordrecht: Kluwer Academic Publishers. doi: 10.1007/978-1-4615-0357-6_15 Molenaar, P. C. M., & Nesselroade, J. R. (1998). A comparison of pseudo-maximum likelihood and asymptotically distribution-free dynamic factor analysis parameter estimation in fitting covariance-structure models to block-Toeplitz matrices representing single-subject multivariate time-series. Multivariate Behavioral Research, 33, 313–342. doi: 10.1207/s15327906mbr3303_1 Molenaar, P. C. M., & Newell, K. M. (Eds.) (2010). Individual pathways of change: Statistical models for analyzing learning and development. Washington, DC: American Psychological Association. doi: 10.1037/12140-000 Molenaar, P. C. M., & Rovine, M. (2015). Intraindividual analysis. In D. Cicchetti (Ed.), Developmental psychopathology (3rd ed.). New York, NY: Wiley. Molenaar, P. C. M., Sinclair, K. O., Rovine, M. J., Ram, N., & Corneal, S. E. (2009). Analyzing developmental processes on an individual level using nonstationary time series modeling. Developmental Psychology, 45, 260–271. doi: 10.1037/a0014170 Molenaar, P. C. M., & von Eye, A. (1994). On the arbitrary nature of latent variables. In A. von Eye & C. C. Clogg (Eds.), Latent variables analysis —Applications for developmental research (pp. 226–242). Newbury Park, CA: Sage. Morey, L. C., & Hopwood, C. J. (2009). An IRT-based measure of alcohol trait severity and the role of traitedness in trait validity: A reanalysis of Project MATCH data. Drug and Alcohol Dependence, 105, 177–184. doi: 10.1016/j.drugalcdep.2009.06.025 Mun, E. Y., von Eye, A., Bates, M. E., & Vaschillo, E. (2008). Finding groups using model-based cluster analysis: Heterogenous emotional

843

self-regulatory processes and heavy alcohol use risk. Developmental Psychology, 44, 481–495. doi: 10.1037/0012-1649.44.2.481 Muthén, B. (2000). Latent variable mixture modeling. In G. A. Marcoulides & R. E. Schumacker (Eds.), Advanced structural equation modeling: New developments and techniques (pp. 1–33). Mahwah, NJ: Erlbaum. Muthén, B. (2001). Second-generation structural equation modeling with a combination of categorical and continuous latent variables: New opportunities for latent class-latent growth modeling. In L. M. Collins, & A. G. Sayer (Eds.), New methods for the analysis of change (pp. 291—322). Washington, DC: American Psychological Association. doi: 10.1037/10409-010 Muthén, B. O., & Muthén, L. M. (2000). Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes. Alcoholism: Clinical and Experimental Research, 24, 882–891. doi: 10.1111/j.1530-0277.2000.tb02070.x Muthén, L. M., & Muthén, B. O. (1998–2012). Mplus: User’s guide (7th ed.). Los Angeles, CA: Muthén & Muthén. Nagin, D. S. (1999). Analyzing developmental trajectories: A semiparametric, group-based approach. Psychological Methods, 4, 139–157. doi: 10.1037/1082-989X.4.2.139 National Institute of Justice. (2011). Batterer intervention programs often do not change offender behavior. Retrieved from http:// www.nij.gov/topics/crime/intimate-partner-violence/interventions/ Pages/batterer-intervention.aspx#noteReferrer2 National Institute of Mental Health. (2011). NIMH Research Domain Criteria (RDoC). http://www.nimh.nih.gov/research-funding/rdoc/ nimh-research-domain-criteria-rdoc.shtml Nigg, J. T. (2006). Temperament and development psychopathology. Journal of Child Psychology and Psychiatry, 47, 395–422. doi: 10.1111/j.1469-7610.2006.01612.x Nesselroade, J. R., & Molenaar, P. C. M. (2010). When persons should be few and occasions should be many—Modeling a different kind of longitudinal data. International Society for the Study of Behavioral Development Bulletin. Serial No. 57, 2–4. Oakes, J. M. (2009). Commentary: individual, ecological, and multilevel fallacies. International Journal of Epidemiology, 38, 361–368. doi: 10.1093/ije/dyn356 Perrine, M. W. B. (1995). Daily self-reported drinking: Longitudinal patterns. NIAAA grant AA09684.Vermont Alcohol Research Center, Essex, VT. Ram, N., & Nesselroade, J. R. (2007). Modeling intraindividual and intracontextual change: Rendering developmental contextualism operational. In T. D. Little, J. A. Bovaird, & N. A. Card (Eds.), Modeling contextual effects in longitudinal studies (pp. 325–342). Mahwah, NJ: Erlbaum. Ramsay, D., & Lewis, M. (2003). Reactivity and regulation in cortisol and behavioral responses to stress. Child Development, 74, 456–464. doi: 10.1111/1467-8624.7402009 Rand, W. M. (1971). Objective criteria for the evaluation of clustering methods. Journal of the American Statistical Association, 66, 846–850. doi: 10.1080/01621459.1971.10482356 Robins, R. W., John, O. P., Caspi, A., Moffitt, T. E., & Stouthamer-Loeber, M. (1996). Resilient, overcontrolled, and undercontrolled boys. Three replicable personality types. Journal of Personality and Social Psychology, 70, 157–171. doi: 10.1037/ 0022-3514.70.1.157 Robinson, W. S. (1950). Ecological correlations and the behavior of individuals. American Sociological Review, 15, 351–357. doi: 10.2307/2087176 Roeder, K., Lynch, K., & Nagin, D. (1999). Modeling uncertainty in latent class membership: A case study in criminology. Journal

844

Person-Oriented Approaches

of the American Statistical Association, 94, 766–776. doi: 10.1080/ 01621459.1999.10474179 Rogers, C. R. (1961). On becoming a person: A therapist’s view of psychotherapy. Boston, MA: Houghton Mifflin. Rost, J., & Langeheine, R. (Eds.) (1997). Applications of latent trait and latent class models in the social sciences. Münster, Germany: Waxmann. Rovine, M. J., Molenaar, P. C. M., & Corneal, S. E. (1999). Analysis of emotional response patterns for adolescent stepsons using P-technique factor analysis. In R. K. Silbereisen & A. von Eye (Eds.), Growing up in times of social change (pp. 261–285). Berlin, Germany: De Gruyter. doi: 10.1515/9783110810110.261 Sampson, R. J., & Laub, J. H. (2005). A life-course view of the development of crime. ANNALS of the American Academy of Political and Social Science, 602, 12–45. doi: 10.1177/0002716205280075 Sampson, R. J., Laub, J. H., & Wimer, C. (2006). Does marriage reduce crime? A counterfactual approach to within-individual causal effects. Criminology, 44, 465–508. doi: 10.1111/j.1745-9125.2006.00055.x Schmidt, H. D. (1977). Methodologische Probleme der entwicklungspsychologischen forschung. Probleme und Ergebnisse der Psychologie, 62, 5–27. Schmitz, B. (2000). Auf der Suche nach dem verlorenen Individuum: Vier Theoreme zur Aggregation von Prozessen [Searching for the lost individual: Four theorems on the aggregation of processes.] Psychologische Rundschau, 51, 83–92. doi: 10.1026//0033-3042.51.2.83 Selvin, H. C. (1958). Durkheim’s suicide and problems of empirical research. American Journal of Sociology, 63, 607–619. doi: 10.1086/ 222356 Sjoquist, E. S., Almqvist, L., Asenlof, P., Lampa, J., Opava, C. H., & Para Study, G. (2010). Physical-activity coaching and health status in rheumatoid arthritis: A person-oriented approach. Disability and Rehabilitation, 32(10), 816–825. doi: 10.3109/09638280903314069 Stam, H. J. (2010). The tradition of personalism and its relationship to contemporary indeterminate functionalism. New Ideas in Psychology, 28, 143–150. doi: 10.1016/j.newideapsych.2009.02.004 Steinley, D. (2004). Properties of the Hubert-Arabie adjusted Rand index. Psychological Methods, 9, 386–396. doi: 10.1037/1082-989X.9.3.386 Stephenson, W. (1935). Correlating persons instead of tests. Character and Personality, 6, 17–24. doi: 10.1111/j.1467-6494.1935.tb02022.x Stephenson, W. (1953). The study of behavior: Q technique and its methodology. Chicago, IL: University of Chicago Press. Sterba, S. K., & Bauer, D. J. (2010). Matching method with theory in person-oriented developmental psychopathology research. Development and Psychopathology, 22, 239–254. doi: 10.1017/ S0954579410000015 Stern, W. (1900). Über Psychologie der individuellen Differenzen (Ideen zu einer ‘differentiellen Psychologie’). Leipzig, Germany: J.A. Barth. Stern, W. (1911). Die differentielle Psychologie in ihren methodischen Grundlagen. Leipzig, Germany: J.A. Barth. Stern, W. (1917/2010). Psychology and personalism (Trans. J. T. Lamiell). New Ideas in Psychology, 28, 110–134. doi: 10.1016/ j.newideapsych.2009.02.005 Strupp, H. H., & Bergin, A. E. (1969). Some empirical and conceptual bases for coordinated research in psychotherapy. International Journal of Psychiatry, 7, 18–90. Subramanian, S. V., Jones, K., Kaddour, A., & Krieger, N. (2009). Revisiting Robinson: The perils of individualistic and ecologic fallacy. International Journal of Epidemiology, 38, 342–360. doi: 10.1093/ije/dyn359 Susser, M. (1994). The logic in the ecological. American Journal of Public Health, 84, 825–835. doi: 10.2105/AJPH.84.5.825

Toomela, A. (2009). How methodology became a toolbox—and how it escapes from that box. In J. Valsiner, P. C. M. Molenaar, M. C. D. P. Lyra, & N. Chaudhary (Eds.), Dynamic process methodology in the social and developmental sciences (pp. 45–66). New York, NY: Springer. doi: 10.1007/978-0-387-95922-1_3 Towe-Goodman, N. R., Stifter, C. A., Mills-Koonce, W. R., Granger, D. A., & The Family Life Project Key Investigators. (2012). Interparental aggression and infant patterns of adrenocortical and behavioral stress responses. Developmental Psychobiology, 54, 685–699. doi: 10.1002/dev.20618 Valsiner, J., Molenaar, P. C. M., Lyra, M. C. D. P., & Chaudhary, N. (Eds.), Dynamic process methodology in the social and developmental sciences (pp. 31–43). New York, NY: Springer. van Rijn, P. W. (2008). Categorical time series in psychological measurement. Dissertation at the Faculty of Social and Behavioural Sciences of the University of Amsterdam. Retrieved from http://dare.uva.nl/ document/104220 van Rijn, P. W., & Molenaar, P. C. M. (2005). Logistic models for single-subject time series. In L. A. van der Ark, M. A. Croon, & K. Sijtsma (Eds.), New developments in categorical data analysis for the social and behavioral sciences (pp. 125–146). London, UK: Erlbaum. Vassend, O., & Skrondal, A. (2011). The NEO personality inventory revised (NEO-PI-R): Exploring the measurement structure and variants of the five-factor model. Personality and Individual Differences, 50, 1300–1304. doi: 10.1016/j.paid.2011.03.002 Vedeler, D., & Garvey, A. P. (2009). Dynamic methodology in infancy research. In J. Valsiner, P. C. M. Molenaar, M. C. D. P. Lyra, & N. Chaudhary (Eds.), Dynamic process methodology in the social and developmental sciences (pp. 431–453). New York, NY: Springer. doi: 10.1007/978-0-387-95922-1_19 Vermunt, J. K., & Magidson, J. (2004). Structural equation models: Mixture models. Retrieved from http://statisticalinnovations.com/ technicalsupport/2005c.pdf Volkovich, Z., Avros, R., & Golani, M. (2011). A comparative approach to cluster validation. Journal of Pattern Recognition Research, 2, 230–243. doi: 10.13176/11.208 von Eye, A. (2002). Configural frequency analysis—Methods, models, and applications. Mahwah, NJ: Erlbaum. von Eye, A. (2010). Developing the person-oriented approach—Theory and methods of analysis. Development and Psychopathology, 22, 277–285. doi: 10.1017/S0954579410000052 von Eye, A., & Bergman, L. R. (2003). Research strategies in developmental psychopathology: Dimensional identity and the person-oriented approach. Development and Psychopathology, 15, 553–580. doi: 10.1017/S0954579403000294 von Eye, A., Bergman, L. R., & Hsieh, C. A. (2015). Person-oriented approaches. In W. Overton & P. C. M. Molenaar (Eds.), Handbook of child development: Theory and methods (Vol. 1, pp. 789–841). New York, NY: Wiley. von Eye, A., & Bogat, G. A. (2005). Identifying sectors of deviations from multinormality. In K. Anaya Izquierdo, A. Contreras Cristan, A. Dominguez Molina, & E. Estrada Barragan (Eds.), Memorias del XIX Foro Nacional de Estadistica (pp. 201–207). Aguascalientes, Mexico: Instituto Nacional de Estadistica, Geografia e Informatica. von Eye, A., & Bogat, G. A. (2006). Person-oriented and variable-oriented research: Concepts, results, and development. Merrill-Palmer Quarterly, 52, 390–420. doi: 10.1353/mpq.2006.0032 von Eye, A., & Bogat, G. A. (2007). Methods of data analysis in person-oriented research. The sample case of ANOVA. In A. Ittel, L. Stecher, H. Merkens, & J. Zinnecker (Eds.), Jahrbuch Jugendforschung 2006 (pp. 161–182). Wiesbaden: Verlag fur Sozialwissenschaften.

References von Eye, A., Bogat, G. A., & Rhodes, J. (2006). Variable-oriented and person-oriented perspectives of analysis: The example of alcohol consumption in adolescence. Journal of Adolescence, 29(6), 981–1004. doi: 10.1016/j.adolescence.2006.06.007 von Eye, A., & Gardiner, J. C. (2004). Locating deviations from multivariate normality. Understanding Statistics, 3, 313–331. von Eye, A., & Gutierrez-Pena, E. (2004). Configural frequency analysis: The search for extreme cells. Journal of Applied Statistics, 31, 981–997. doi: 10.1080/0266476042000270545 von Eye, A., Mair, P., & Bogat, G. A. (2005). Prediction models for configural frequency analysis. Psychology Science, 47, 342–355. von Eye, A., & Mun, E. Y. (2013). Log-linear modeling—Concepts, interpretation and applications. New York, NY: Wiley. von Eye, A., Mun, E. Y., & Indurkhya, A. (2004). Classifying developmental trajectories—A decision making perspective. Psychology Science, 46, 65–98. von Eye, A., & Wirsing, M. (1978). An attempt for a mathematical foundation and evaluation of MACS, a method for multidimensional automatical cluster detection. Biometrical Journal, 20, 655–666. von Eye, A., & Wirsing, M. (1980). Cluster search by enveloping space density maxima. In M. M. Barritt & D. Wishart (Eds.), Compstat

845

1980, Proceedings in Computational Statistics (pp. 447–453). Vienna, Austria: Physica. Walls, T. A., & Schafer, J. L. (Eds.). (2006). Models for intensive longitudinal data. Oxford, UK: Oxford University Press. doi: 10.1093/ acprof:oso/9780195173444.001.0001 Warrens, M. J. (2008). On the equivalence of Cohen’s Kappa and the Hubert-Arabie Adjusted Rand Index. Journal of Classification, 183, 177–183. doi: 10.1007/s00357-008-9023-7 Widaman, K. F. (2007). Intrauterine environment affects infant and child outcomes: Environment as direct effect. In T. D. Little, J. A. Bovaird, & N. A. Card (Eds.), Modeling contextual effects in longitudinal studies (pp. 387–436). Mahwah, NJ: Erlbaum. Wishart, D. (1970). The treatment of various similarity criteria in relation to Clustan 1A. University of St. Andrews, St. Andrews, Scotland. Witherington, D. C. (2007). The dynamic systems approach as metatheory for developmental psychology. Human Development, 50, 127–153. doi: 10.1159/000100943 Zeanah, C. H., Benoit, D., Hirshberg, L., Barton, M. L., & Regan, C. (1994). Mothers’ representations of their infants are concordant with infant attachment classifications. Developmental Issues in Psychiatry and Psychology, 1, 9–18.

CHAPTER 19

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology MICHAEL J. ROVINE and PETER C. M. MOLENAAR

INTRODUCTION 846 Definition of Intraindividual Variation 848 Contrasting Intraindividual Variation with Interindividual Variation 848 MODELING INTERINDIVIDUAL VARIATION THROUGH GROUP-BASED INDIVIDUAL DIFFERENCE MODELS 849 The Multilevel Growth Curve Model 850 The Linear Latent Growth Curve Model 851 The Multilevel Autoregressive Model 852 Lagged Regression Models 852 The Vector Autoregression Model 853

THE DATA BOX AND FACTOR ANALYSIS 854 Factor Models for Longitudinal Data 855 Person-Oriented Approaches 857 P-Technique Factor Models 857 The Dynamic Factor Model 857 DATA EXAMPLES 858 Data Example 1: The Borkenau and Ostendorf Data Set 858 Data Example 2: PANAS Data 858 DISCUSSION 861 FUTURE DIRECTIONS 862 REFERENCES 862

INTRODUCTION

which the focus is on person-specific methods (Molenaar, Lerner, & Newell, 2014; Molenaar & Newell, 2010). The medical sciences also are undergoing major methodological changes under the banner of personalized medicine (Hamburg & Collins, 2010). While the initial approaches to personalized medicine were heavily based on genotypic characterization of patients, this now has been generalized to include intensive longitudinal assessments of phenotypic features such as physiology, behavior, etc. (Schwarz & Collins, 2007). The availability of such intensive longitudinal assessments opens up the possibility to optimize therapeutic treatment by means of new methods derived from computational engineering. In these so-called dynamic treatment methods the sequence of therapeutic manipulations is increasingly optimized in real time by means of machine learning techniques conditional on the newly available longitudinal assessments (e.g., Chakraborty & Moodie, 2013). The trend toward personalized and person-specific methodology also is corroborated by a recent development in the foundations of psychometrics. Using general theorems of mathematical ergodic theory (Birkhoff, 1931), Molenaar (2004) proved that pooling across individuals,

Developmental psychopathology constitutes the confluence of two important research traditions that both are undergoing major methodological changes— developmental science and psychopathology. Developmental science is increasingly based on the metamodel of developmental systems theory (Ford & Lerner, 1992). Developmental systems theory conceives of development as the result of multiple coacting influences that (1) are context sensitive and contingent (e.g., Gottlieb, 2001), (2) affect constructive processes in which nonlinear epigenetic influences play central roles (e.g., Lickliter & Honeycutt, 2009), and (3) evolve at multiple time scales and at multiple levels (e.g., Smith & Thelen, 2003). Accordingly, the emphasis in developmental systems theory is on the inherently stochastic and person-specific nature of developmental processes. Applications of developmental systems theory therefore require methodological approaches in

1

Color versions of Figure 19.3 are available at http://onlinelibrary .wiley.com/book/10.1002/9781118963418 846

Introduction

which is the hallmark of standard statistical methods in the analysis of psychological data, only is warranted if strong so-called ergodicty criteria are met. The criteria that define ergodicity imply that (1) all individuals making up a population (i.e., domain of generalization) should obey exactly the same dynamic model and (2) the dynamic model should be stationary. The latter stationarity criterion implies that the process under consideration should lack any developmental features (e.g., constant mean, constant variance, constant sequential dependencies). In case one or both of these criteria are violated, the statistical analysis should be based on person-specific approaches involving intensive measurements of individual subjects. Developmental process almost by definition violates the stationarity criterion and therefore should be investigated using person-specific methods. In an important publication Sterba & Bauer (2010) scrutinized the match between person-oriented methods and person-oriented theory. They showed that person-specific methods like dynamic factor analysis (i.e., factor analysis of intensive longitudinal assessments obtained in single-subject or replicated time-series designs) provide the best match with person-oriented principles (see also Molenaar, 2010a). Accordingly, person-oriented principles should preferably be applied using person-specific methods. Because person-specific methodology is relatively new (in contrast to other fields of research where time-series analysis often is the standard approach to data analysis, such as for instance physics, engineering, and econometrics), we highlight in what follows differences between standard developmental methods based on analysis of interindividual variation with person-specific methods based on intraindividual variation. In particular, we will argue that the individual differences created within standard statistical developmental methods such as latent growth curve modeling violate in important respects the tenets of person-specific methodology. This implies that these standard statistical methods based on interindividual variation cannot be used to accommodate violations of ergodicity and also may have limited applicability in personalized and person-oriented approaches (Molenaar & Campbell, 2009; Voelkle, et al., 2014). Here, we suggest applying these approaches to the field of developmental psychopathology, In summarizing the field as part of the recognition of the fiftieth anniversary of the Journal of Child Psychology and Psychiatry, Cicchetti and Toth (2009) indicated the importance of an approach to research in developmental psychopathology that considered all methods that allowed the description of process that typically unfolds over time. They emphasized

847

the importance of multiple levels of analysis to achieve this goal. An important level of analysis that has received less attention in the developmental psychopathology literature is the person-specific approach based on single subject models. Given their notion that there are “multiple pathways to similar manifest outcomes and that there are different outcomes [resulting from] the same pathway,” (page 4) typical prediction models based on group characteristics may be lacking when the requirement would be a prediction that can have two different outcomes given identical conditions. Given that treatment may be suggested by the prediction, the person-specific single subject approach which would concentrate more on individual characteristics may be the better bet. Add to this that there has been an increasing tendency in the field toward the notion of person-centered care (Adams & Grieder, 2005). The idea behind this is simple. Treatments developed based on sample characteristics and applied to the individual may not appropriately address that individual’s needs. A treatment regimen tailored on the individual’s unique requirements may produce better results. This is akin to developing a person-specific model. Molenaar (1987; Molenaar, 2010b) has shown how this modeling approach can be successful in the therapeutic situation. There has also been a greater emphasis on shared decision making in the therapeutic setting (Adams & Drake, 2006; Deegan et al., 2008; Ptasznik, 2011). These tendencies suggest that the person-specific, single-subject modeling approach may be an important new tool for moving the field of developmental psychopathology forward. As Cicchetti and Toth (2009) described the type of developmental analysis we suggest, it “presupposes change and novelty, highlights the critical role of timing in the organization of behavior, underscores multiple determinants, and cautions against expecting invariant relations between causes and outcomes” (p. 1). We extend this idea of expected invariance to include individual models that can be used to achieve optimal therapeutic outcomes. In this chapter we would like to first demonstrate the way certain common analytic approaches such as longitudinal regression modeling, linear mixed modeling (e.g., growth curve and repeated measures one-way analysis of variance [ANOVA]), multilevel autoregressive modeling, factor modeling, and cluster analysis create individual difference measures based on a common underlying model. In this sense, they are modeling interindividual variation. After showing that these approaches depend on sufficient (summary) statistics and creating individual parameter predictions based on an underlying assumed probability model, we will then compare these methods

848

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

to more person-specific approaches. We will focus on analogous single subject time-series approaches (e.g., vector autoregressive approaches for cross-lagged regression, P-technique and dynamic factor models for cross sectional or longitudinal factor models based on relatively few occasions) and show how these person-specific methods create parameters based on models that can vary from individual to individual. We will demonstrate a true intraindividual approach using personality data collected by Borkenau and Ostendorf (1996) and emotional mood data from Corneal and Nesselroade (1994; see also Molenaar et al., 2009; Rovine, Molenaar, & Corneal, 1998). We first define the difference between the concepts of intraand interindividual differences. Definition of Intraindividual Variation In its most basic form intraindividual variation is variation in time of repeated measurements obtained with a single subject. We use the term subject for the single entity (case) that is measured repeatedly, but we understand that this entity also can be a single group of subjects like a particular mother–infant dyad, a single family, or a school. The important point to note is that here, subject is fixed; that is, it is not considered to be a randomly drawn from some population of subjects. Even though the designation for each subject as a system or dynamic system would be technically more appropriate, as psychologists we select subject. Intraindividual variation is understood to be random variation associated with a stochastic process which is thought to describe the evolution of some system over time. Here, the system is a single subject and the process is measured by a set of repeated measurements indexed by time. The general idea of a stochastic process was introduced by Kolmogorov (1931) and can be illustrated as follows: suppose the cognitive performance of a single subject on the same test is measured at T consecutive occasions. According to Kolmogorov’s definition of a stochastic process, the actual score observed on each occasion t, say t = 1, constitutes a random draw from a probability distribution of possible scores that could have been obtained at t = 1 under the same circumstances. This allows the consideration of the mean of each time series. Moreover, each pair of scores, say at t = 1 and t = 2, is considered to be a random draw from a bivariate probability distribution. This allows the consideration of the correlation or covariance between each pair of occasions. Together the means, variances, correlations, and higher order moments of this each series can be modeled using methods typically called time-series analysis. Time can be continuous or discrete. In this chapter we consider only discrete time processes. To simplify the

presentation we suggest that the time interval between consecutive occasions is constant, although this assumption can be dropped when discussing data obtained in, for example, ecological momentary assessment (EMA) designs. By emphasizing stochastic processes we are ruling out deterministic process; for example, those that follow a predetermined shape or trajectory caused such as growth processes. Such processes can be described by deterministic differential or difference equations and would typically be modeled using a different set of methods (Banks, 1983). Contrasting Intraindividual Variation with Interindividual Variation The converse of intraindividual variation is interindividual variation. These two contrasting types of variation can be distinguished in the following way. In analysis of intraindividual variation, we model each subject separately treating the subject as fixed while the estimation proceeds by pooling across the replicated measurements obtained with this subject. Because its measurements are time-dependent, the hallmark of analysis of intraindividual variation is representing by pooling across repeated occasions. The finite set of observations for each subject can be conceived of as a sample from all possible times. As a consequence, the results of a standard analysis of intraindividual variation can be generalized to all times previous to the observation interval (retrodiction), to times within the observation interval for which no observations are available (interpolation), and to future times (prediction). Hence, the domain of generalization of a pure analysis of intraindividual variation is the time domain, while no statements regarding the generalization across subjects can initially be made. In contrast, in analysis of interindividual variation the sample of subjects is conceived of as randomly drawn from a population of subjects. Estimation proceeds by pooling data across the sample of subjects. The hallmark of analysis of interindividual variation is pooling across subjects. As a consequence, the results obtained in an analysis of interindividual variation are generalized to the population of subjects. This is the typical analytical situation in longitudinal designs. Again, analysis of interindividual variation of these repeated measures proceeds by pooling across subjects. For instance, as we will show, the standard latent growth curve model is fitted by pooling across subjects. The result of this is that any longitudinal analysis is not analysis of intraindividual variation but instead is analysis of interindividual variation. The difference in the definitions of inter- and intraindividual variation suggests how the results of a particular analysis can be generalized (generalization across populations of subjects versus generalization across time). We note

Modeling Interindividual Variation Through Group-Based Individual Difference Models

that one of the dangers of forgetting this distinction is that inappropriate generalizations are often made based on the results of analyses based on interindividual variation. We will emphasize these below.

MODELING INTERINDIVIDUAL VARIATION THROUGH GROUP-BASED INDIVIDUAL DIFFERENCE MODELS Many popular analytical methods create values for what are described as individual difference parameters that are based on group models. For example, the linear multilevel (or latent) growth curve model (Bryk & Raudenbush, 1992; Goldstein, 1995; Grimm, Davoudzadeh, & Ram, 2014; McArdle & Epstein, 1987) can be used to generate individual intercepts and slopes in addition to the group intercept and slope that are typically the focus of the model. While the individual measures can be used to describe differences among participants in a study, they are created based on the group parameter estimates (e.g., regression weights, common covariance structure of the residuals). The individual variable values are then predicted using these group parameter values based on certain assumptions regarding the distribution of these variables (e.g., the individual values will be normally distributed with a zero mean). This approach represents the assumption that the same model holds for all study participants. With a relatively few number of occasions of measurement, this assumption cannot be tested.2 With more intensive time-series data, however, this assumption can be tested, and individual models can be determined. It may be the case that different individuals require different models. There may be unknown subgroups whose data may be described by a common model, but this is not known a priori. With more intensive time-series data, a separate model can be determined for each study participant. Individuals with common models can then be clustered to form subgroups (Gonzales & Ferrer, 2014; Nesselroade & Molenaar, 1999). One common model where this can become evident is the longitudinal lagged regression model based on a panel design. Here, we will compare how the group-based model relates to the person-specific vector autoregressive model. A second family of analytic procedures that we will consider encompasses methods based on the linear mixed model, which include multilevel growth curve modeling, repeated measures ANOVA, and multilevel autoregressive models. As an example, for the linear multilevel growth 2

Unless mixture modeling is used, but then the same assumptions reapply in each of the finite number of subpopulations.

849

curve model, individual slopes and intercepts are generated as part of the solution. These are often interpreted as individual difference measures. They are based on post hoc predictions of the model and are common to all participants in the sense that they represent variations in the parameters of a single underlying model under the assumption of a particular probability distribution for the residuals (Robinson, 1991). The common underlying model makes these individual difference measures much different than those that would result from a separate time-series model for each individual. Another model we consider is the longitudinal factor model. With a few occasions of measurement investigators can test whether a factor structure is invariant across occasions and the degree of invariance that can be assumed. But this factor model is the result of a subjects × variables data set where the variables for the different occasions are the variables of the model. As an alternative we will consider a factor analysis based on collecting multiple variables on multiple occasions on a single subject. Here, occasions essentially takes the place of subjects. A correlation (or covariance) matrix can be constructed which can be used to estimate individual factor models. The factor loadings and factor correlations of these different individual models can be compared both to each other and to results of the group-based subjects × variables as we show for the data presented by Borkenau and Ostendorf (1998) and Corneal and Nesselroade (1991; Molenaar et al., 2009; Rovine, Molenaar, & Corneal, 1998). Other modeling approaches attempt to account for heterogeneity of intraindividual variation in creating individual difference parameters. Latent class trajectory analysis (Nagin, 1999) and growth mixture modeling (Muthén & Sheden, 1999) create subgroups by implicitly clustering individuals with similar predicted parameter values. Although these methods allow heterogeneity in terms of differences in the predicted parameter values, there will still be strong distribution assumptions within subgroups (Sterba & Bauer, 2010). For example, a particular subgroup within a growth mixture model will have a specified mean and covariance structure that assumes homogeneity within that subgroup. These types of models still require a common group model of which the individual parameters represent instances of that model though with wider variation across subgroups. By requiring the common model, these approaches still describe interindividual variability. Multilevel autoregressive (AR) models use an autoregressive model as the Level 1 (time-related level) model. The Level 2 or individual-level model allows covariates that are invariant across occasions to account for variability in the Level 1 parameters. However, this results in a set

850

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

of fixed effect parameters that are added to the overall or group level equation again resulting in a common model for all individuals. As in the multilevel growth curve model, individual AR parameters are predicted post hoc under the strict assumptions that the same AR process (e.g., AR(1), AR(2)) holds for all individuals, and the individual parameter values deviate from the group parameter values by residual values that conform to a multivariate normal distribution. If the order of the process is not identical for individuals, or if the order of the process is the same, but the true distribution of the parameter does not follow a strict normal distribution, this approach represents a misspecification of the model. The person-oriented literature (Bergman, 2012; Bergman & Andersson, 2010) often refers to a cluster analytic approach to time-series data. Given the strong similarity between many clustering methods and standard factor analysis, such person-oriented approaches are often really describing interindividual differences. The Multilevel Growth Curve Model The linear growth curve model (Bryk & Raudenbush, 1992; Goldstein, 1995; McArdle & Epstein, 1987) is member of the linear mixed model family (Laird & Ware, 1982). A simple linear growth curve model with a timeinvariant covariate x can be written as

convention to refer to the total covariance matrix of the random effects as 𝜎 2 V , which defines V by factoring the residual variance out of the total covariance. This covariance matrix along with the fixed effects can be used to predict values of the individual random effects under the assumptions of this single common model under a multivariate normal distribution (Henderson, 1975). The equation for best linear unbiased predictions (BLUPs) was given by Robinson (1991) as ̂ 𝛾̂i = DZT V̂ −1 (yi − X 𝛽)

where 𝛾̂i is the vector of predicted random effects (intercept and slope) for subject i, X is the design matrix for the fixed effect parameters (group intercept and slope), and 𝛽 is the set of estimated fixed effect parameters. These random effects are often described as individual difference measures, and in the sense that they vary across individuals, they are. This interpretation is the result, in part, of the multilevel formulation of the linear mixed model as presented by Goldstein (1995), and included in the formatting of software such as HLM (Bryk & Raudenbush, 1992). For example, the multilevel model version of the linear growth curve shown above can be rewritten as Level 1:

yit = 𝛽0 + 𝛽1 TIME + 𝛽2 xi + 𝛽3 (x ∗ TIME) + 𝜐i0 + 𝜐i1 TIME + rit

(19.1)

where 𝛽0 , 𝛽1 , 𝛽2 , and 𝛽3 are fixed effects regression coefficients, and 𝜐i0 , 𝜐i1 , and rit are the random effects which are the deviation of the individual intercept around the group intercept, the deviation of the individual slope around the group slope, and the deviations of the individual observed variable values around their own regression line. As in all linear mixed models we estimate the expected values of the fixed effects and the covariance matrix of the random effects. These estimates are based on data pooled across individuals. The covariance matrix of the random effects is given by (19.2) 𝜎 2 V = 𝜎 2 (ZDZT + I) where 𝜎 2 is the pooled variance of the individual residuals around the individual’s regression (assumed equal across occasions for the typical multilevel linear growth curve model), Z is the set of design column vectors for the random intercept and slope parameters, and D is covariance matrix of the random intercepts and slopes. It is a

(19.3)

yit = 𝜋i0 + 𝜋i1 TIME + rit

Level 2: 𝜋i0 = 𝛽00 + 𝛽10 xi + 𝜐i0 𝜋i1 = 𝛽01 + 𝛽11 xi + 𝜐i1

(19.4)

where in Level 1, 𝜋i0 is the individual’s intercept, and 𝜋i1 is the individual’s linear slope. These random effects comprise the individual equation. In Level 2 we included a covariate, x, that varies from individual to individual but is otherwise time invariant. Here the individual intercept that has been passed on from Level 1 deviates around the group intercept, 𝛽00 , by an amount, 𝜐i0 , and the individual slope deviates from the group slope, 𝛽01 , by an amount, 𝜐i1 . In both Level 2 equations, we included a covariate, which is purported to account for variability in the individual slopes and intercepts. Here, the random effects are rit , 𝜐i0 , and 𝜐i1 . The fixed effects are the group intercept, 𝛽00 , the group slope, 𝛽01 , and the regression coefficient related to the covariate, 𝛽11 . For the linear mixed model, we estimate the expected values of the fixed effects, and the covariance matrix of the random effects.

Modeling Interindividual Variation Through Group-Based Individual Difference Models

Given the format of this multilevel model, it appears that a separate equation is estimated for each individual. This appearance is part of the reason why the model is mistaken for a true model of intraindividual differences. However, substituting Level 2 into Level 1, we get yit = (𝛽00 + 𝛽10 xi + 𝜐i0 ) + (𝛽01 + 𝛽11 xi + 𝜐i1 )TIME + rit yit = 𝛽00 + 𝛽01 TIME + 𝛽10 xi + 𝛽11 (x ∗ TIME) + (𝜐i0 + 𝜐i1 TIME) + rit

(19.5)

which is identical to the linear mixed model equation presented above in which the individual difference equation has disappeared and is replaced by a single equation with fixed and random effects. It is this equation which is estimated in programs like SAS PROC MIXED, SAS MIXED, and R LME. The random effects here and as estimated by programs such as HLM are the same regression residuals predicted under the same set of assumptions under a specific group-based regression model. The empirical Bayes estimates of HLM are identical to the solution of the Henderson linear model equations implemented by programs such as SAS PROC MIXED (Littell et al., 2006). This model assumes that the same trajectory shape describes each individual, and individuals only differ in, for example, the steepness of the slope, or the height of the intercept. When a common model does not hold, the growth curve approach may misrepresent how individuals differ resulting in an inference that gives a less than optimal description about the course of change (Liu et al., 2012). Given the short longitudinal data that are most typically available in growth curve analysis, the possibility of testing to see whether the model adequately describes the sample is often not available. The person-specific analog to the growth curve model would require fitting a separate polynomial equation to each individual. The proper order (i.e., the parameters that can describe the shape of the curve) would be empirically determined for each individual. With most studies, these data are not available. When sufficient data to estimate the individual curve would be available, the estimate of a relatively low order polynomial would only make sense for certain simple true growth processes. Since the growth curve model is often used for modeling some arbitrary change process that does not follow a known growth function, the fitting of a growth curve to time-series data may not even make much sense. For a relatively few number of repeated occasions of measurement, the growth curve approach often represents a necessary practical compromise.

851

The Linear Latent Growth Curve Model Latent curve models (Bollen & Curran, 2006; Meredith & Tisak, 1990) represent a very common approach to fitting curves through the means of repeatedly measured data. Hearkening back to the work of Tucker (1958) and Rao (1958), these approaches are primarily based on factor analytic methods. A very popular set of models based on a parameterization is presented by McArdle and Epstein (1987) and McArdle and Aber (1990). The latent growth curve model with either fixed or estimated basis vector coefficients (actually fixed or estimated factor loadings), is strongly related to the multilevel growth curve model. For example, the linear growth curve model uses a highly constrained factor model to estimate the same set of fixed effects parameters as the linear mixed model. Here the model is again used to define straight lines with different intercepts and different slopes. To vary things slightly we have included two groups in the model and no covariate. While we concentrate on the straight-line model, curves other than straight lines can be modeled either by using polynomial terms (powers, centered powers, or orthogonal polynomials) or by estimating factor loadings (McArdle & Hamagami, 1991). This model can be estimated as a structural equation measurement model. Using LISREL notation, the general matrix equation for the measurement model is y⃗ = Λ⃗ 𝜂 + 𝜀⃗

(19.6)

where y⃗ is the vector of observed variables, Λ is a partitioned matrix of fixed factor loadings corresponding to the fixed and random effects, 𝜂⃗ is a partitioned vector representing the fixed and random regression parameters, and 𝜀⃗ is a vector of regression residuals. For two groups with four repeated measures, the matrix form of the linear growth curve model is

⎡yi1 ⎤ ⎡1 ⎢ ⎥ ⎢ ⎢yi2 ⎥ = ⎢1 ⎢yi3 ⎥ ⎢1 ⎢ y ⎥ ⎢1 ⎣ i4 ⎦ ⎣

⎡yi1 ⎤ ⎡1 ⎢ ⎥ ⎢ ⎢yi2 ⎥ = ⎢1 ⎢yi3 ⎥ ⎢1 ⎢ y ⎥ ⎢1 ⎣ i4 ⎦ ⎣

0 1 2 3

0 1 2 3

0 0 0 0

1 1 1 1

0 0 0 0

0 1 2 3

1 1 1 1

⎡ 𝛽1 ⎤ 0⎤ ⎢ 𝛽2 ⎥ ⎡𝜀i1 ⎤ ⎥⎢ ⎥ ⎢ ⎥ 1⎥ ⎢ 𝛽3 ⎥ ⎢𝜀i2 ⎥ Group 1 + 2⎥ ⎢ 𝛽4 ⎥ ⎢𝜀i3 ⎥ 3⎥⎦ ⎢𝛾i1 ⎥ ⎢⎣𝜀i4 ⎥⎦ ⎢ ⎥ ⎣𝛾i2 ⎦

1 1 1 1

⎡ 𝛽1 ⎤ 0⎤ ⎢ 𝛽2 ⎥ ⎡𝜀i1 ⎤ ⎥⎢ ⎥ ⎢ ⎥ 1⎥ ⎢ 𝛽3 ⎥ ⎢𝜀i2 ⎥ Group 2 + 2⎥ ⎢ 𝛽4 ⎥ ⎢𝜀i3 ⎥ 3⎥⎦ ⎢𝛾i1 ⎥ ⎢⎣𝜀i4 ⎥⎦ ⎢ ⎥ ⎣𝛾i2 ⎦

(19.7)

852

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

where 𝛽1 and 𝛽2 are the respective intercept and slope for the first group, and 𝛽3 and 𝛽4 are the respective differences in intercept and slope between the two groups. Because columns 3 and 4 are 0s for Group 1, 𝛽3 and 𝛽4 are not added to the Group 1 equation. 𝛽3 and 𝛽4 are added to the model for Group 2. The 𝛽s are fixed for the sample. 𝛾1 and 𝛾2 are the random effects (in this case random slopes and intercepts). These correspond to columns 5 and 6. In theory the random effects for each individual provide the difference between the group intercept and the individual intercept (𝛾1 ) and the group slope and the individual slope (𝛾2 ). In this model, factor scores take the place of the random effects in the linear mixed model. In estimating the model, the covariance matrix of the random effects, along with the covariance matrix of the 𝜀t are estimated. Factor scores are then predicted post hoc to yield the individual random effects under the same assumptions as in the linear mixed model. This highly constrained factor model is identical to a two-group linear mixed model/multilevel growth model. As such, it is a model of interindividual differences. The Multilevel Autoregressive Model Given that it is now more common to collect more intensive longitudinal data, some researchers have begun to turn to time-series methods in an attempt to model these data. Some of these approaches look at estimating a separate model for each individual, but others have attempted to use a multilevel-type approach. One method that is gaining more popularity is the multilevel autoregressive (AR) model in which the Level 1 model is an autoregressive model of a particular order. These models are thought to be particularly appropriate for data that Nesselroade (1991) described as belonging to a stable process, one without a discernible trend. Having a Level 1 autoregressive model has resulted in what some authors refer to as dynamic multilevel modeling (Suls, Green, & Hillis, 1998). Using a first-order autoregressive model, they referred to the AR parameter as a measure of inertia in that in the presence of a perturbation (a movement away from equilibrium), the autoregressive parameter tended to return the series toward the equilibrium point at the mean of the series (Hamaker, 2012; Kuppens, Allen, & Sheeber, 2010). Consider an AR(1) process that is considered appropriate to describe all members of a sample, the multilevel AR model can be written as Level 1: yi,t+1 = 𝜋i0 + 𝜋i1 yi,t + ri,t+1

Level 2: 𝜋i0 = 𝛽00 + 𝛽10 xi + 𝜐i0 𝜋i1 = 𝛽01 + 𝛽11 xi + 𝜐i1

(19.8)

where an intercept, 𝜋i0 , and the AR coefficient, 𝜋i1 , are both conditioned on some covariate, x. As required by the stationarity condition, −1 > 𝜋i1 > 1. As before, we could substitute Level 2 into Level 1 to come up with a single linear mixed model equation to estimate the fixed effects of the model along with the covariance matrix of the random effects (Rovine & Walls, 2006). But instead we just wish to point out that the same assumptions of the multilevel model hold here; namely, that the AR(1) process holds equally well for all individuals, and that the individual AR(1) coefficients, the random effects, can be predicted post hoc under a common model assuming that the coefficients conform to a multivariate normal distribution. Different processes for different individuals cannot be accommodated by this model. It is therefore a model concerned with interindividual variation. One innovation in this approach has been the addition of a random innovation variance (Wang, Hamaker, & Bergeman, 2012), which determines each individual’s innovation variance in addition to their AR parameter value. Again, these individual error variances are predicted post hoc based on a common model and a specific assumed error distribution. That makes this model markedly distinct from a true model of intraindividual differences. Lagged Regression Models As an example of a lagged regression we consider the cross-lagged regression model with two variables, x and y, each measured on four occasions with the assumption that x at each occasion predicts y at the next occasion, and in turn, y at each occasion predicts x at the next occasion (Kenny, 2005). The model for such data is shown in Figure 19.1. Given any path model, the set of regression equations that result is given by the set of tracing rules (Wright, 1934). Simply stated, each variable in the model is a dependent variable in its own regression equation. Predictor variables for that equation are connected via one-headed arrows. Two-headed arrows represent correlations3 or covariances (depending on the scaling) between 3

From this point when we use the term correlation it will be implicit that the association coefficient could be a correlation, covariance, or sum of the cross-products depending on the requirements of scaling for the model.

Modeling Interindividual Variation Through Group-Based Individual Difference Models

y1

β21

y2

β32

y3

β43

y4

853

𝜂1 = y1 ; 𝜂2 = y2 ; 𝜂3 = y3 ; 𝜂4 = y4 ; 𝜂5 = x1 ; 𝜂6 = x2 ; 𝜂7 = x3 ; 𝜂8 = x4 . Placing the 𝜂s in equation (19.9) yields 𝜂1 = 𝜁1

β25 β61 x1

β65

x2

β76

β36

β47

𝜂2 = 𝛽21 𝜂1 + 𝛽25 𝜂5 + 𝜁2

β72

β83

𝜂3 = 𝛽32 𝜂2 + 𝛽36 𝜂6 + 𝜁3

x3

β87

x4

Figure 19.1 A cross-lagged regression model (regression coefficients only).

variables. Errors are linked to a particular dependent variable through one-headed arrows. To simplify the figures, we do not include errors in the models as shown.4 We also do not include any concurrent regression relationships. The set of regression equations resulting from the model in Figure 19.1 is y1 = 𝜁1 y2 = 𝛽21 y1 + 𝛽25 x1 + 𝜁2 y3 = 𝛽32 y2 + 𝛽36 x2 + 𝜁3 y4 = 𝛽43 y3 + 𝛽47 x3 + 𝜁4 x1 = 𝜁5

(19.9)

x2 = 𝛽61 y1 + 𝛽65 x1 + 𝜁6 x3 = 𝛽72 y2 + 𝛽76 x2 + 𝜁7 x4 = 𝛽83 y3 + 𝛽87 x3 + 𝜁8 where 𝛽 represents a regression coefficient, and 𝜁 represents an equation residual. As y1 and x1 are not predicted by other variables, they are considered exogenous, and all other variables (which are predicted) are considered endogenous. Exogenous variables are marked by the fact that their correlation equals the correlation between their equation residuals. To estimate the regression coefficients and residual variances of the model, we use a covariance structure modeling approach that expresses the associations between variables (correlations, covariances, or sums of squares and cross-products) as functions of the parameters to be estimated. Here we use the LISREL model (Jöreskog & Sorböm, 1996) for these relationships. To conform to the LISREL model, we rename the variables as follows: 4

We also indicate the observed variables as squares. In many SEM programs, regression relationships among latent variables (often shown as circles) are modeled. For observed variable regression, latent variables are set as equivalent to observed variables.

𝜂4 = 𝛽43 𝜂3 + 𝛽47 𝜂7 + 𝜁4 𝜂5 = 𝜁5

(19.10)

𝜂6 = 𝛽61 𝜂1 + 𝛽65 𝜂5 + 𝜁6 𝜂7 = 𝛽72 𝜂2 + 𝛽76 𝜂6 + 𝜁7 𝜂8 = 𝛽83 𝜂3 + 𝛽87 𝜂7 + 𝜁8 This can be summarized by the matrix equation 𝜂⃗ = B⃗ 𝜂 + 𝜁⃗

(19.11)

where 𝜂⃗ is an 8 × 1 vector of variables, 𝜁⃗ is an 8 × 1 vector of equation residuals, and 𝛽 is an 8 × 8 matrix of regression weights indexed according to the previous regression weights. This is a regression equation with all variables acting as possible dependent variables in their own regression equation. Which variables predict each of the dependent variables is determined by the path model. Equation 19.11 can be used to develop an equation for the expected associations among all of the variables. The difference between the observed variable covariance matrix, S, and the expected Σ̂ based on the model can be used to determine the fit of the model. In terms of estimating this model, the required input for generating the regression estimates is the covariance matrix. Individual data points are not required for the estimation of the parameters. As in ordinary regression, the equation residuals, 𝜁 , can be predicted post hoc using the parameter estimates. These residuals are model based in that the group model parameters must be known before the residuals can be calculated. The Vector Autoregression Model We now present the person-specific analog of the crosslagged regression model. Here we consider two time series, x and y, measure on a single subject. We hypothesize that xt predicts yt+1 and yt predicts xt+1 . This model appears in Figure 19.2. This is an example of a first-order vector autoregression model (Lutkepohl, 2006). Here, the order of the process represents which of the previous occasions are used to predict the next occasion. With just the previous occasion as a predictor, it is a first-order process. If we used the previous

854

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

y1

β11

y2

yn–1

β11

β12

β12

β21 x1

β22

yn

β21 x2

xn–1

β22

xn

Figure 19.2 A first-order cross-legged vector autoregression model (regression coefficients only).

two occasions to predict, it would be a second-order process. The first-order model presented is equivalent to the set of equations yt+1 = 𝛽11 yt + 𝛽12 xt + 𝜀1t+1 xt+1 = 𝛽21 yt + 𝛽22 xt + 𝜀2t+1

(19.12)

The matrix representation of these equations is [ ] [ ][ ] [ ] yt+1 𝛽 𝜀 𝛽12 yt = 11 + 1t+1 (19.13) xt+1 𝛽21 𝛽22 xt 𝜀2t+1 While this looks much like the cross-lagged regression model described above, here we estimate model [ a separate ] 𝛽11 𝛽12 for each individual. As a result the matrix is 𝛽21 𝛽22 person specific and not dependent on other individuals’ results. By changing the regression weight matrix, we can create the analog for [many different lagged regression models. ] 𝛽11 0 represents the analog to the For example, 0 𝛽22 stability only[ model (i.e., ] xt predicts xt+1 , yt predicts yt+1 ). 𝛽11 𝛽12 The matrix, , represents prediction from xt to 0 𝛽22 yt+1 . This latter situation represents the test of Granger causality (Granger, 1969). Here, x is a Granger cause of y. Granger causality has been used in neural network analysis (Seth, 2008), connectivity mapping of fMRI brain scan data (Roebroeck et al., 2005), and econometric modeling (Hatemi, 2012). Here we assume that the same order of autoregression and the same order of lagged prediction (i.e., x predicting y) would hold for each individual. However, in a true individual approach we would estimate distinct models, which would include determining the order of autoregression for each time series and the order of the lagged prediction. With the proper models, we could determine which individuals have the same models—for

example, whether the order of their models is the same, whether the regression coefficients are the same, or whether the variance of the errors is the same. Group models could be constructed for those individuals who have identical models. Nesselroade and Molenaar (1999) suggested methods for pooling the results of individual time series. This approach of creating samples based on the equivalence of individual models guarantees that ergodicity, as defined already, is satisfied and that a subgroup with an identical pooled model adequately describes each individual in the subgroup. Using such a strategy, it is most likely that different subgroups based on different (sometimes qualitatively different) models can emerge. When there are different models, the underlying process generating the series would be expected to be different for the subgroups, and this becomes an important and interesting research result. These models can, of course easily be extended to any number of variables. THE DATA BOX AND FACTOR ANALYSIS The conceptual framework that allows one to consider the difference between intra- and interindividual approaches in multivariate studies is often linked to Cattell’s data box. In his Handbook for Multivariate Experimental Psychology, Cattell (1952, 1966a) presented a classification scheme for determining the appropriate analysis for data collected in a multivariate study. The data box as shown in Figure 19.3 has three dimensions: variables, individuals, and occasions. Each two-dimensional slice or facet of the box represents a way of constructing an association matrix with each slice having a dimension of interest and a dimension across which the data will be pooled. Cattell gave each two-dimension facet a letter name. The most commonly used are the R and P facets. When interested in pooling across individuals to create a variable × variable correlation matrix, we are using the R facet resulting in R-technique. This results in the most common form of factor analysis of interindividual variation. Both common cross sectional and longitudinal factor analysis are examples of R-technique. In longitudinal factor analysis the variables collected on multiple occasions are treated along the variables dimension. The data are still pooled across individuals to form a correlation matrix among the variables. Since multilevel growth curve models, repeated measures ANOVA, and latent growth curve models are pooled across individuals, each can be thought of as belonging to the R-technique class of analyses. For person-specific models, the P facet takes data for a single subject and pools across occasions to create

The Data Box and Factor Analysis

855

Interindividual analysis (R-Technique) Variables

Interindividual analysis (P-Technique)

Persons/ Entities

Occasions

Figure 19.3 The Data Box. See footnote 1.

a variable × variable correlation matrix of intraindividual variation. Factor analysis of this type of data set is referred to as P-technique. The difference between R- and P-technique is thus the dimension across which one pools based on the way data are collected, where R-technique pools across persons and P-technique pools across occasions. Facets for both R- and P-technique appear in Figure 19.3.

matrix of the latent variables, and Θ𝜀 is the covariance matrix of the measurement errors or uniquenesses. In the exploratory common factor model with set number of factors, each observed variable loads on each factor. This means that with, for an example with three factors, each factor predicts each of the observed variables. With means added to the model 𝜇y = 𝜏 + Λ𝜅

Factor Models for Longitudinal Data Factor models for longitudinal data represent an extension of the common factor model (Gorsuch, 1983; Harman, 1976). The common factor model can be expressed as y⃗ = Λ⃗ 𝜂 + 𝜀⃗

(19.14)

where y⃗ represents the vector of observed variables, Λ represents the matrix of regression weights linking the latent variables to the observed variable, 𝜂⃗ represents the vector of latent variable factor scores, and 𝜀⃗ represents the vector of measurement errors. Given any factor model the expected covariance matrix of the observed variable based on the model is Σ̂ = ΛΦΛT + Θ𝜀 (19.15) where Σ̂ is the expected covariance matrix of the observed variables based on the model, Λ is the matrix of factor loadings (i.e., regression weights), Φ is the covariance

where 𝜇y is the vector of means, 𝜏 is the vector of intercepts, and 𝜅 is the vector of factor score means5 . To consider a longitudinal version of factor analysis for the situation in which there are relatively few occasions of measurement, we move to the confirmatory version of the factor model based on R-technique. The exploratory longitudinal factor model (for instance, a longitudinal factor model in which at each occasion the factors are uncorrelated, there only are autoregressions between the factor scores, and in which the factor loadings at each occasion only have minimal identifiability constraints) still becomes a confirmatory factor model when fitted to the longitudinal covariance matrix. The confirmatory factor model allows the researcher to specify certain hypotheses based on the parameters of the model. For example, the researcher could expect certain variables to load only on certain factors. 5

Additional constraints are required; otherwise, the means model would be overparameterized.

856

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

Consider the situation in which the researcher has collected the same three variables on three separate occasions assuming that the three variables are indicators of the same latent construct. A sample pattern based on a confirmatory model would have each factor represent the same latent variable at a particular occasion resulting in loadings of 0 for those variables on the other occasions. This would result in the following in the following Λ matrix:

y1T1 y2T1 y3T1 y1T2 y2T2 y3T2 y1T3 y2T3 y3T3

𝜂T1

𝜂T2

𝜂T3

⎡𝝀11 ⎢𝝀 ⎢ 21 ⎢𝝀31 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢ ⎢ 0 ⎢ 0 ⎢0 ⎣

0 0 0 𝝀42 𝝀52 𝝀62 0 0 0

0 ⎤ 0 ⎥⎥ 0 ⎥ 0 ⎥ ⎥ 0 ⎥ 0 ⎥ ⎥ 𝝀73 ⎥ 𝝀83 ⎥ 𝝀93 ⎥⎦

(19.16)

and each row represents a regression equation where the observed along margin represents the dependent variable that is predicted by the latent variables along the top with regression coefficients indicated by the factor loadings. The covariance matrix among the factors is ⎡𝜙11 Φ = ⎢𝜙21 ⎢ ⎣𝜙31

𝜙22 𝜙32

⎤ ⎥ ⎥ 𝜙33 ⎦

(19.17)

In exploratory factor models, the variance of the measurement errors, the uniquenesses, are estimated. In CFA under certain circumstances measurement errors can be correlated. In this model, we expect that the same variable measured on different occasions may result in correlated measurement errors. This would result in the following covariance matrix of measurement errors: ⎡𝜃11 ⎢0 ⎢ ⎢0 ⎢𝜃 ⎢ 41 Θ𝜀 = ⎢ 0 ⎢0 ⎢ ⎢𝜃71 ⎢0 ⎢0 ⎣

⎤ ⎥ ⎥ 𝜃33 ⎥ ⎥ 0 𝜃44 ⎥ 0 0 𝜃55 ⎥ ⎥ 𝜃63 0 0 𝜃66 ⎥ 0 𝜃74 0 0 𝜃77 ⎥ ⎥ 0 0 𝜃85 0 0 𝜃88 𝜃93 0 0 𝜃96 0 0 𝜃99 ⎥⎦ (19.18) In this model each factor is indicated by three observed variables. In addition to the regression relationships 𝜃22 0 0 𝜃52 0 0 𝜃82 0

between the latent and observed variables (the factor loadings), characteristics of the latent variables could be of interest. For example, a research question may involve whether the means of the latent variables change across time, or whether there is a distinct pattern of correlation among the latent variables across occasions of measurement. The set of variables could be, for example, indicators of problematic behavior (e.g., subscales of the Child Behavior Checklist, Achenbach & Edelbrock (1981)), measures of temperament, attachment, or any number of other constructs that are repeatedly measured. A natural extension would have more than one latent variable at each occasion, each one repeatedly measured. To be able to discuss group differences or characteristics of change in a repeatedly measured latent variable, a certain level of measurement invariance must exist (Meredith, 1993; Millsap & Meredith, 2004; Reise et al., 1993; Vandenberg & Lance, 2000; Widaman & Reise, 1997). The underlying notion of factorial invariance is if that assumption is violated, the same latent construct may not be measured across all groups or across all occasions. This result would strongly limit the ability to interpret any change in the construct. Suggestions for how to deal with differing degrees of factorial invariance have not always been consistent. Improved model fitting capabilities allow one to more easily test whether invariance has been violated without necessarily suggesting what to do in response to a particular violation (Bontempo et al., 2012; Millsap & Meredith, 2004). Millsap & Kwok (2004) gave a good set of recommendations for how to assess the effect of a particular violation. Millsap (2011) emphasized the multigroup case. For the previously shown factor pattern, the parameters to be estimated are Λ, the factor loadings; Φ, the factor covariances; and Θ𝜀 , the covariance matrix of the measurement errors. The model above is based on a correlation or covariance matrix. We could include information about the means as input, which would allow us to also estimate parameters related to the mean structure6 . 6

Means in longitudinal factor models should be modeled according to a method described by Jöreskog (1978) and Sörbom (1974) to ensure that the means model is appropriate for interval-scaled variables. This requires that at the first occasion the means are freely estimated in 𝜏 y (a saturated means model at the first occasion). Next 𝜏 y is constrained to be invariant across later occasions and only at these later occasions the changes in means are modeled by the factor means in 𝛼. If one models all means in terms of 𝛼 (including at the first occasion and leaving out 𝜏 y ) then it is required that the variables have ratio scale.

The Data Box and Factor Analysis

A typical sequence for testing factorial invariance may include the following levels: 1. Configural invariance—the same patterns of nonzero loadings hold across groups or occasions. 2. Metric or pattern invariance—the values of the loadings are fully invariant across groups or occasions. 3. Strong factorial invariance—the pattern matrices and the latent intercepts are fully invariant across groups or occasions. 4. Strict factorial invariance—the pattern matrices, intercepts, and unique variances are fully invariant across groups or occasions. The first two invariance criteria were introduced by Thurstone (1947) and the third and fourth by Meredith (1993). To test models of the latent variable means, strong factorial invariance is required. Like the cross sectional common factor model, the longitudinal factor model requires as input the correlation (or covariance) matrix of the input variables. Individual values of the variables are not needed to estimate the model. The factor loadings are group-based, not person-specific. Factor scores are generated as post hoc predictions using a similar procedure to that of the latent growth curve’s random effects. In fact, the multilevel growth curve model can estimated as a structural equations model (Rovine & Molenaar, 2000). In this form, the random effects are the factor scores of the model estimated using the fixed effects and the covariance matrix of the random effects under certain specified distributional assumptions. Since estimated models are based on input matrices pooled across individuals, they are models of interindividual differences. Despite the importance and utility of this model, it is not person specific.

857

P-Technique Factor Models As previously stated, the ergodicity hypothesis implies that a group-based result can apply to an individual only when that model holds for each individual. Without true time-series data that assumption cannot be tested. As Molenaar (2004) argued, the result of this is that cross sectionally or longitudinally derived factor models based on a covariance matrix pooled across individuals (R-technique) can be quite different from those derived separately for each individual. The individual approach, based on P-technique factor analysis, is a true intraindividual method. P-technique factor analysis has been used to study stepchildren’s emotional experiences (Corneal & Nesselroade, 1994), changes in mood (Bath, Daly, & Nesselroade, 1976), and changes in attachment security (Mitteness & Nesselroade, 1987) using essentially the standard (R-technique) factor analytic approach with occasions taking the place of individuals. For the basic P-technique model, the data for each subject are pooled across occasions resulting in a variable × variable association matrix. For the analogous three variable factor model shown above as a longitudinal factor model, the set of equations represented by this model is represented by the matrix equation ⎡𝜆1 ⎤ ⎡𝜀1 ⎤ ⎡y1 ⎤ ⎢y ⎥ = ⎢𝜆 ⎥ [𝜂]t + ⎢𝜀 ⎥ ⎢ 2⎥ ⎢ 2⎥ ⎢ 2⎥ ⎣𝜀3 ⎦t ⎣y3 ⎦t ⎣𝜆3 ⎦

(19.19)

where a single factor replaces the occasion factors of the longitudinal factor analysis and a single set of loadings represents the requirement of invariance across the time series. This model would be estimated separately for each subject.

Person-Oriented Approaches

The Dynamic Factor Model

Certain methods described as person-oriented approaches (Bergman & Magnusson, 1997) are typically described as a counterpoint to variable-oriented analysis (Bergman & Trost, 2006). These often involve cluster analyzing repeatedly measured data to identify subgroups of individuals who have similar sets of variable values across occasions of measurement. As can be shown, under certain circumstances, a cluster analysis can be transformed into a factor analysis. As such, this analytic approach may pool across individuals. Under these circumstances, cluster analysis generalizes across individuals, not occasions. This approach is thus an interindividual differences approach rather than a person-specific approach.

The P-technique makes certain assumptions regarding the nature of relationships across occasions of measurement. Most importantly, the model assumes that there are no lagged relationships among the occasion variables. The association matrix that is analyzed pairs variables at each occasion to construct the correlation. While this model has been shown to be more widely applicable than one might expect (Molenaar & Nesselroade, 2009), models that address the question of lagged relationships and determine the order of lags represent an important approach. Molenaar (1985) presented an extension of the P-technique model, the dynamic factor model that properly addresses the question of whether or not the relationship among

858

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

variables is lagged, and if so, what is the order. The dynamic factor model is yt =

s ∑ Λu 𝜂t−u + 𝜀t

(19.20)

u=0

where Λu contains the factor loading for lag, u. With this model the researcher can determine the degree to which latent variables and residuals are correlated over time. When s = 0, the model reduces to the state space model in which the evolution of factors is explained by a vector autoregressive process (VAR).7 These models and their variants have appeared under somewhat different names (Nesselroade et al., 2002). Wood and Brown (1994) showed a relatively straightforward method for implementing this model as a structural equations model. DATA EXAMPLES Data Example 1: The Borkenau and Ostendorf Data Set Here, we demonstrate an instance of that situation of a violation of ergodicity assumption for personality factors based on the big five using a data set provided by Borkenau and Ostendorf (1998). This data set included 22 individuals followed on a daily basis over 90 days with 30 personality items being repeatedly measured. The participants were psychology students from the University of Halle, Germany; course credit was given for participation. The sample consisted of 19 women and three men who were mostly 19-year-olds or in their early 20s, except for three females who were in their 30s or 40s. In the current example, an exploratory factor analysis is carried out at both the cross sectional group level (R-technique) and for several individuals at the person-specific time-series level (P-technique; Nesselroade & Jones, 1991). Here, we will demonstrate that the number of cross sectional factors gives a poor indication of the factor structure for each of the three individuals considered. In the exploratory approach to factor analysis, we initially want to determine the number of factors or latent variables that underlie the set of observed variables considered. For the Borkenau and Ostendorf data set we first determined the number of factors cross sectionally (R-technique). The scree plot for these data appears in Figure 19.4. The scree plot (Cattell, 1966b) is a commonly used visual guide for determining the number of 7

This VAR is implicit in the general model (20) because it is simplest to treat the evolution of factor scores as a white-noise process. Molenaar and Nesselroade (2001) showed that other choices are possible, including VARs. Only these choices have to be fixed.

factors in a solution. In the plot the points represent the results of an eigenvalue decomposition of the input correlation matrix. The eigenvalues are ordered from largest to smallest and can be thought of as incrementally explaining the set of correlations among the variables. Each eigenvalue corresponds to one possible factor, and some subset of these factors will be able to adequately reproduce the input correlation matrix. The question is to find this optimal number of factors. The scree plot does so by indicating the point at which the remaining factors are most likely trivial. This point is usually depicted by an elbow in the scree. For the cross sectional data, the elbow seems to appear at the fourth eigenvalue. For Subject A, the elbow appears at either eigenvalue 2 or eigenvalue 4; for Subject B at the third eigenvalue; and for Subject C at the second, third, or fourth eigenvalue. Because this perusal approach can be somewhat subjective, several numerical criteria have been developed. Here we employ the acceleration factor (AF; Raiche et al., 2006). The AF has been established as a good selection procedure for both R- and P-technique. The approach looks at the derivative (the slope) of the scree plot and determines where the slope changes most rapidly (accelerates). This typically represents the elbow of the curve. Using this approach, we determined that the elbow did indeed occur at the fourth eigenvalue for the cross sectional data indicating a three-factor solution. For three individuals, Subject A had a one-factor solution (elbow at the second eigenvalue), and Subjects B and C both had a two-factor solution. The solution from the cross sectional data described none of the three subjects adequately. This example demonstrates how the number of factors determined in a group-level R-technique factor analysis may be inconsistent with subsequent individual factor models obtained via individually applied P-technique analyses. Data Example 2: PANAS Data Using data taken from a study of stepchildren, Corneal and Nesselroade (1994; see also Corneal & Nesselroade, 1991; Molenaar et al., 2009; Rovine, Molenaar, & Corneal, 1998; Molenaar et al., 2009) showed a similar result based on time-series assessments of emotions. The data were measures designed to analyze the emotional experiences of stepsons as they interacted with their stepfathers over a 2-month period. After a specified type of interaction with their stepfather (including mealtimes, leisure activities, conversations, arguments, and disciplines), the stepson would complete the Positive and Negative Affect Scale (PANAS; Watson, Clark, & Tellegen, 1988) to assess the emotional content of the interaction.

Data Examples

10 0

5

Eigenvalue

10 5 0

Eigenvalue

15

Person A: P-Technique

15

Full Group: R-Technique

2

4

6

8

10

12

2

14

4

6

Factors

8 10 Factors

12

14

Person C: P-Technique

0

10 5 0

5

Eigenvalue

10

15

15

Person B: P-Technique

Eigenvalue

859

2

4

6

8

10

12

14

2

4

6

Factors

8

10

12

14

Factors

Figure 19.4 A comparison of scree plots of the Borkenau data.

To the 20 items of the PANAS, eight involvement items were added. Since the 20 items were developed for high internal consistency, the 28 items were expected to have a two-factor structure. The items representing this factor structure along the with-fit indices for Stepson 1 are shown in Table 19.1. While the two-factor solution typically shows up in the analysis of cross sectional data, the two-factor confirmatory solution showed a very poor fit. In fact the two-factor solution was a poor model for all stepsons included in this study. For each subject we first tested the two-factor model. After the model was rejected, we moved to an exploratory factor analysis to determine the proper number of factors. Using the salient loadings for each stepson, we then fit a confirmatory factor model. We include the results for three of the stepsons in Table 19.2.

TABLE 19.1 Positive and Negative Items for the Two-Factor Solution Negative

Positive

Involved

Distressed Satisfied Interested Upset Content Excited Humiliated Strong Enthusiastic Hostile Proud Alert Ashamed Liking Inspired Irritated Accepted Active Determined Scared Closed Afraid Loved Nervous Jittery Discouraged Not wanted Goodness-of-fit statistics Chi-square with 349 degrees of freedom = 1066.93 (P = 0.0) Nonnormed fit index (NNFI) 0.44

860

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

TABLE 19.2 Confirmatory Factor Analysis for Stepsons 1, 5, and 6

TABLE 19.2 (continued)

Stepson 1: LISREL ML estimates

Stepson 5: LISREL ML estimates Factors

Variables Interested Excited Enthusiastic Inspired Alert Attentive Satisfied Proud Distressed Upset Strong Hostile Irritable Discouraged Accepted Not wanted Determined Content Scared Nervous Jittery Afraid Humiliated Loved Active Liking Ashamed Close

Interest

Anger

0.65 0.67 0.72 0.42 0.66 0.74 0.62 0.66

Anxiety

Factors Affection

Closeness

Alert Attentive Proud Strong Discouraged Determined Satisfied Active Humiliated Accepted Ashamed

−0.38 −0.36 0.34 0.33 −0.59 −0.39 0.85 0.87 0.30 0.91 0.88 0.74 −0.63 0.97 0.34 −0.46

0.51

0.46 0.51 0.32

−0.23 0.23

0.33

0.29 −0.32 0.24

0.45 0.35

Anger Interest Strength Shame

Interest

Anger

Anxiety

Affection

1.00 0.36 −0.06 0.13 0.37

1.00 −0.08 −0.03 −0.36

1.00 0.11 0.14

1.00 0.23

Closeness

1.00

Stepson 5: LISREL ML estimates Factors Anger

Distressed Liking Upset Scared Hostile Irritable Not wanted Loved Nervous Close Jittery Afraid Content Interested Excited Enthusiastic Inspired

0.71 −0.68 0.79 0.65 0.87 0.75 0.68 −0.75 0.91 −0.72 0.55 0.33 −0.49

Interest

Strength

0.25 −0.38 0.33 −0.19

0.46

0.45 −0.22

0.35

0.56 0.55

−0.98 0.45

Anger

Interest

Strength

Shame

1.00 −0.64 −0.44 0.27

1.00 0.70 0.25

1.00 −0.01

1.00

Factors

Shame

Variables

Anger

Distressed Upset Humiliated Scared Hostile Irritable Discouraged Ashamed Not wanted Nervous Satisfied Jittery Afraid Content Interested Liking Excited Enthusiastic Accepted Inspired Loved Alert Attentive Close Strong Determined Proud Active

0.96 0.92 0.57 0.79 0.79 0.92 0.82 0.82 0.74 0.78 −0.58 0.81 0.85 −0.55 0.17 −0.16

−0.14 −0.22 0.20

Interest

Strength −0.21

−0.21

0.61 0.14 0.57 0.77 0.72 0.81 0.81 0.65 0.56 0.69 0.77 0.58 0.68

0.21 0.56 0.75 0.65 0.81

−0.60 Factor Intercorrelations

−0.30 0.26 0.65 0.70 0.79 0.60

Shame

−0.19 0.47 −0.26 0.81 0.47 0.57

−0.13

0.62 −0.41 0.73

Variables

Strength

0.59 0.53 0.60

Stepson 6: LISREL ML Estimates

Factor intercorrelations

Interest Anger Anxiety Affection Closeness

Interest

0.27 −0.49 0.43 0.37

0.43

Anger

Factor Intercorrelations

0.25 0.71 0.92 0.73 0.98

Variables

Anger Interest Strength

Anger

Interest

Strength

1.00 −0.10 0.42

1.00 0.65

1.00

Discussion

The results indicate the very intuitive result that different children respond to a situation by experiencing different combinations of emotions. For Stepson 1, a five-factor solution seemed the best fit. Looking at the items that load, the factors can be described as Interest, Anger, Anxiety, Affection, and Closeness. A number of variables loaded on more than one factor. Certain variables can be expected to help define more than one factor. For example, enthusiasm is an indicator of both Interest and Anger with a positive loading on Interest and a negative loading on Anger. The combination of variables that load on a particular define that factor for a specific stepson. While Anger may exist as a factor for each of the stepsons, its definition in terms of the variables that comprise the factor can vary. When Stepson 1 felt anger during interactions with his stepfather, he felt less accepted. For Stepsons 5 and 6, anger was accompanied by a feeling of anxiety. Stepson 5 also felt less loved when he felt anger. Stepson 5 differed from Stepson 1 in some important ways. For Stepson 4, a four-factor solution seemed most appropriate. Two factors that we labeled Strength and Shame appeared for Stepson 5 that had not occurred in Stepson 1. Stepson 6 had a three-factor solution as the best fitting model. Here, Anger and Anxiety items appeared on a single factor. Interest and Affection items also tended to appear on the same factor. A Strength factor similar to Stepson 5 also appeared. The results shown here indicate the importance of at least considering the person-specific approach in looking at factor structures. By pooling across subjects, the investigator is assuming that differences in individuals should be treated as essentially error. If the differences are systematic and represent important descriptive information, that information is lost when pooling to estimate models. The result of using the assumed cross sectional two-factor solution to describe each subject in this study would be that a wealth of information describing how these stepsons differ in their interactions with the stepfathers would be lost. DISCUSSION Clearly pooling across heterogeneous cases whether they be replications, subjects, or systems can be ill advised. Applying analysis of interindividual variation to heterogeneous data that involves pooling across subjects that obey subject-specific models can lead to results that poorly describe individuals and poorly describe the group. We argue, instead, that such data should be subjected to the analysis of intraindividual variation that is specifically

861

designed to detect the kind of heterogeneity that analysis of interindividual variables essentially ignores. However, carrying out separate single-subject analysis of the intraindividual variation of a sample of subjects may yield a bewildering collection of different subject-specific dynamic models. For this situation there is obviously an urgent need for methods to bring order to these separate models. This represents the need to arrive at nomothetic models for intraindividual variation. One promising approach to accomplish this is the so-called Idiographic Filter proposed by Nesselroade et al. (2007). The basic idea underlying the Idiographic Filter is the recognition that the way latent factors manifest themselves in observed variables can be subject-specific. For instance, in our second data example, for Stepson 5, anger is accompanied by being nervous and scared, which for Stepson 1, that is not the case. The Idiographic Filter thus allows for subject-specific factor loadings while locating homogeneity at the latent level. Molenaar (et al., 2012) has applied a generalized version of the Idiographic Filter (iFACE) to the analysis of multivariate time series obtained with a pair of genetically related subjects allowing for subject-specific heritabilities and environmental effects. A first application of the iFACE to multilead EEG obtained in an oddball task shows clear evidence for considerable subject-specificity of heritable and environmental effects. While the Idiographic Filter can accommodate heterogeneity due to subject-specific factor loadings, Molenaar recently developed an alternative approach called the Group Iterative Multiple Model Estimation (GIMME; Gates & Molenaar, 2012) which can accommodate arbitrary subject-specificity of both factor loadings and any other parameters in state space models using a variant of the structural equations model referred to as the unified SEM (Gates et al., 2010). GIMME has been used in a number of studies including the analysis of a large and multifaceted simulated data set considered in Smith et al. (2011), who tested 38 methods to estimate connectivity networks based on multivariate fMRI time series. The methods included Bayesian net approaches and vector autoregressive modeling. None of these methods did well in terms of recovering the presence of directed links among component series. This surprising result, obtained with realistic simulation of BOLD activities of brain regions of interest under a wide variety of conditions (varying signal-to-noise, varying dimensionality of the observed time series, varying number and strength of links, etc.), created quite a stir in the brain imaging community. When GIMME was applied

862

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

to these data, all directed links and the numerical values of their weights were recovered with close to or exactly 100% fidelity (see Gates & Molenaar, 2012, for a complete report of these results). GIMME has been most often applied to brain imaging data, but more recently it has been extended to the description of behavioral processes (Belz, et al., 2013). GIMME can be freely accessed at www.nitrc.org/projects/gimme/. We note that GIMME is unique in providing a valid way to determine common dynamic models for heterogeneous replicated time series, thus yielding generalizable (nomothetic) results about idiographic processes. We feel that the importance of the subject-specific approach to the modeling of intraindividual variation in the statistical analysis of developmental processes in particular is clear. All nonergodic processes such as developmental processes should initially be analyzed at the level of intraindividual variation, i.e., single-subject time-series analysis. From this starting point several courses of action can be followed in attempts to obtain nomothetic knowledge about intraindividual variation. One could apply the methodologies used in cognitive neuroscience and/or neuropsychology to generalize to nomothetic knowledge from single case analyses. Or one could use the Idiographic Filter or GIMME approaches for the same purpose. We hope and expect that in addition to these well-established approaches other new alternative methodologies will be found to obtain nomothetic knowledge about intraindividual variation, because this is an essential endeavor to move our science forward. FUTURE DIRECTIONS Cicchetti and Toth (2009) emphasized the importance of “translat[ing] basic research knowledge into real world contexts.” We believe that an important way to allow this clinical translation to occur is to ensure that appropriate models are used to describe a successful therapeutic intervention. The approaches to targeting interventions based on single subject, person specific modeling seem very appropriate to this task. In particular, the notion of optimal control, of moving the individual toward a specific goal or set of goals using a model that can assess whether the person in on target and suggest a change of direction when not on target seems to be an important tool to add to the field. This approach has been successful in many other areas and disciplines. To move the field of developmental psychopathology forward we believe that this approach should be added to the toolbox of researchers. Using these methods requires the collection

of more intensive data, but given the increasing popularity of data collection methods such as those based on Ecological Momentary Assessment (EMA), the single subject, person-specific modeling approach is becoming more feasible. It may take some ingenuity on the part of the researchers to design approaches that make use of these methods, but that alone represents an important challenge for clinicians and other investigators. In some areas of research, sufficient data already exist to make use of single subject methods. For example, work in the area of experience dependent brain development makes use of brain-scan data. The development of models requires individual models of developing connectivities. Methods such as the GIMME method described above are designed specifically for such a problem. As mentioned above, all nonergodic processes such as developmental processes should be analyzed as a single-subject time-series analysis. Methods such as GIMME and the idiographic filter can and should be applied to a wide range of problems in the behavioral sciences. We believe that researchers in the area of developmental psychopathology should look for opportunities to apply these methods, which will hopefully help to better describe the pathways to and out of abnormal development while describing conditions of resilience and improving the outcomes for those experiencing these problems.

REFERENCES Achenbach, T. M. & Edelbrock, C. S. (1981). Behavioral problems and competencies reported by parents of normal and disturbed children aged four through sixteen Monographs of the Society for Research in Child Development, 46(1),1–82. Adams, J., & Drake, R. (2006). Shared decision-making and evidencebased practice. Community Mental Health Journal, 42(1), 87–105. Adams, N., & Grieder, D. (2005). Treatment planning for person-centered care: The road to mental health and addiction recovery. New York, NY: Academic Press. Banks, R. B. (1994). Growth and diffusion phenomena: Mathematical frameworks and applications. Berlin, Germany: Springer-Verlag. Bath, K. E., Daly, D. L., & Nesselroade, J. R. (1976). Replicability of factors derived from individual P-technique analyses. Multivariate Behavioral Research, 1(2), 147–156. Belz, A. M., Beekman, C., Molenaar, P. C. M., & Buss, K. A. (2013). Mapping temporal dynamics in social interactions with unified structural equation modeling: A description and demonstration revealing time-dependent sex differences in play behavior. Applied Developmental Psychology, 17, 152–168. doi: 10.1080/10888691.2013.805953 Bergman, L. R. (2012). Advancing developmental science: Some challenges and obstacles. International Journal of Developmental Science, 6, 41–43. Bergman, L. R., & Andersson, H. (2010). The person and the variable in developmental psychology. Journal of Psychology, 218(3), 155–165. doi: 10.1027/0044-3409/a000025

References Bergman, L. R., & Magnusson, D. (1997). A person-oriented approach in research on developmental psychopathology. Development and Psychopathology, 9, 291–319. Bergman, L. R., & Trost, K. (2006). The person-oriented versus the variable-oriented approach: Are they complementary, opposites, or exploring different worlds? Merrill-Palmer Quarterly: Journal of Developmental Psychology, 52, 601–632.

863

Goldstein, H. (1995). Multilevel statistical models (2nd ed.). London, UK: Edward Arnold. Gonzales, J. E., & Ferrer, E. (2014). Individual pooling for group-based modeling under the assumption of ergodicity. Multivariate Behavioral Research, 49(3), 245–260. doi: 10.1080/00273171.2014.902298 Gorsuch, R. L. (1983). Factor analysis (2nd ed.). Hillsdale, NJ: Lawrence Erlbaum.

Birkhoff, G. (1931). Proof of the ergodic theorem. Proceedings of the National Academy of Science USA, 17, 656–660.

Gottlieb, G. (2001). Individual development and evolution: The genesis of novel behavior. Mahwah, NJ: Lawrence Erlbaum.

Bollen, K. A., & Curran, P. J. (2006). Latent curve models: A structural equation approach. New York, NY: Wiley.

Granger, C. W. J. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica, 37(3), 424–438.

Bontempo, D., Grouzet, F., & Hofer, S. (2012). Measurement issues in the analysis of within-person change. In J. Newson, R. Jones, & S. Hofer (Eds.), Longitudinal data analysis: A practical guide for researchers in aging (pp. 97–142). New York, NY: Routledge.

Grimm, K. J., Davoudzadeh, P., & Ram, N. (2015, in press). Longitudinal analysis. In N. Card (Ed.), Developmental methodology. Society for Research in Child Development Monograph (pp. 82–116). New York, NY: Wiley.

Borkenau, P., & Ostendorf, F. (1998). The big five as states: How useful is the five-factor model to describe intraindividual variations over time? Journal of Research in Personality, 32, 202–221. doi: 10.1006/ jrpe.1997.2206

Hamaker, E. L. (2012). Why researchers should think “within-person”: A paradigmatic rationale. In M. R. Mehl & T. S. Conner (Eds.), Handbook of methods for studying daily life (pp. 43–61). New York, NY: Guilford Publications.

Bryk, A., & Raudenbush, S. (1992). Hierarchical linear models: Applications and data analysis methods. Newbury Park, CA: Sage Press.

Hamburg, M. A., & Collins, F. S. (2010). The path to personalized medicine. New England Journal of Medicine, 363, 301–304. doi: 10.1056/NEJMp1006304

Cattell, R. B. (1952). The three basic factor-analytic designs—Their interrelations and derivatives. Psychological Bulletin, 49, 499–520. doi: 10.1037/h0054245

Harman, H. (1976). Modern factor analysis (3rd ed., revised). Chicago, IL: University Chicago Press.

Cattell, R. B. (1966a). The data box: Its ordering of total resources in terms of possible relational systems. In R. B. Cattell (Ed.), Handbook of multivariate experimental psychology (pp. 67–128). Chicago, IL: Rand-McNally.

Hatemi, J. A. (2012). Asymmetric causality tests with an application. Empirical Economics, 43(1), 447–456. doi: 10.1007/s00181-011-0484-x

Cattell. R. B. (1966b). The scree test for the number of factors. Multivariate Behavioral Research, 1, 245–276. doi: 10.1207/s15327906 mbr0102_10

Jöreskog, K. G. (1978). Structural analysis of covariance and correlation matrices. Psychometrika, 43(4), 443–477. doi: 10.1007/BF02293808

Chakraborty, E., & Moodie, E. E. M. (2013). Statistical methods for dynamic treatment regimes: Reinforcement learning, causal inference, and personalized medicine. New York, NY: Springer. Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychological Psychiatry, 50(1–2), 16–25. doi: 10.1111/j.1469-7610.2008.01979.x. Corneal, S. E., & Nesselroade, J. R. (1991). Multivariate, replicated, single-subject, repeated measures design: Studying change in the adolescent. In R. Lerner, A. Petersen, & J. Brooks-Gunn (Eds.), Encyclopedia of adolescence (pp. 681–687). New York, NY: Garland. Corneal, S. E., & Nesselroade, J. R. (1994). A stepchild’s emotional experience across two households: An investigation of response patterns by P-technique factor analysis. Multivariate Experimental Clinical Research, 10(3), 167–180. Deegan, P. E., Rapp, C., Holter, M., & Riefer, M. (2008). A program to support shared decision-making in an outpatient psychiatric medication clinic. Psychiatric Services, 59, 603–605. Ford, D. H., & Lerner, R. M. (1992). Developmental systems theory: An integrative approach. Newbury Park, CA: Sage. Gates, K. M., & Molenaar, P. C. M. (2012). Group search algorithm recovers effective connectivity maps for individuals in homogeneous and heterogeneous samples. NeuroImage, 63, 310–319. doi: 10.1016/ j.neuroimage.2012.06.026 Gates, K. M., Molenaar, P. C. M., Hillary, F. G., Ram, N., & Rovine, M. J. (2010). Automatic search for fMRI connectivity mapping: An alternative to Granger causality testing using formal equivalences among SEM path modeling, VAR and unified SEM. NeuroImage, 50, 1118–1125. doi: 10.1016/j.neuroimage.2009.12.117

Henderson, C. R. (1975). Best linear unbiased estimation and prediction under a selection model. Biometrics, 31, 423–447.

Jöreskog, K. G., & Sörbom, D. (1996). LISREL8: User’s reference guide. Chicago, IL: Scientific Software International. Kenny, D. A. (2005). Cross-lagged panel design. In B. S. Everitt & D. Howell (Eds.), Encyclopedia of statistics in behavioral science. New York, NY: Wiley. doi: 10.1002/9781118445112.stat06464 Kolmogorov, A. (1931). Über die analytischen Methoden in der Wahrscheinlichkeitsrechnung. Mathematische Annalen, 104, 415–458. Kuppens, P., Allen, N. B., & Sheeber, L.B. (2010). Emotional inertia and psychological maladjustment. Psychological Science, 21, 984–991. doi: 10.1177/0956797610372634 Lickliter, R., & Honeycutt, H. (2009). Rethinking epigenesis and evolution in light of developmental science. In M. Blumberg, J. Freeman, & S. Robinson (Eds.), Oxford handbook of developmental behavioral neuroscience (pp. 30–47). New York, NY: Oxford University Press. Littell, R. C., Milliken, G. A., Stroup, W. W., Wolfinger, R. D., & Schabenberger, O. (2006). SAS for mixed models. Cary, NC: SAS Institute. Liu, S., Rovine, M. J., & Molenaar, P. C. M. (2012). Selecting a linear model for longitudinal data: Repeated measures analysis of variance, covariance pattern models and growth curve approaches. Psychological Methods, 17(1), 15–30. doi: 10.1080/10705511.2012.726918 Lutkepohl, H. (2006). New introduction to multiple time series analysis. New York, NY: Springer. McArdle, J. J., & Aber, M. S. (1990). Patterns of change within latent variable structural equation modeling. In A. von Eye (Ed.), New statistical methods in developmental research (pp. 151–224). New York, NY: Academic Press. McArdle, J. J., & Epstein, D. (1987). Latent growth curves within developmental structural equation models. Child Development, 58, 110–133. doi: 10.2307/1130295

864

Person-Specific Approaches to the Modeling of Intraindividual Variation in Developmental Psychopathology

McArdle, J. J, & Hamagami, E. F. (1991). Modeling incomplete longitudinal and cross sectional data using latent growth structural models. In L. Collins & J. L. Horn (Eds.), Best methods for the analysis of change (pp. 276–304). Washington, DC: APA Press.

Nesselroade, J. R., & Jones, C. J. (1991). Multimodal selection effects in the study of adult development: A perspective on multivariate, replicate, single subject repeated measures. Experimental Aging Research, 11, 21–27.

Meredith, W. (1993). Measurement invariance, factor analysis, and factorial invariance. Psychometrika, 58, 525–542.

Nesselroade, J. R., Gerstorf, D., Hardy, S. A., & Ram, N. (2007). Idiographic filters for psychological constructs. Measurement: Interdisciplinary Research and Perspectives, 5, 217–235. doi: 10.1080/ 15366360701741807

Millsap, R. E. (2011). Statistical approaches to measurement invariance. New York, NY: Routledge. Millsap, R. E., & Kwok, O. (2004). Evaluating the impact of partial invariance on selection in two populations. Psychological Methods, 9(1), 93–115. Millsap, R. E., & Meredith, W. (2004). Factorial invariance: Historical trends and new developments. Paper presented at the Factor Analysis at 100 Conference, L.L Thurstone Psychometric Laboratory, University of North Carolina. Mitteness, L. S., & Nesselroade, J. R. (1987). Attachment in adulthood: Longitudinal investigation of mother–daughter affective interdependencies by P-technique factor analysis. Southern Psychologist, 3, 37–44. Molenaar, P. C. M. (1985). A dynamic factor model for the analysis of multivariate time series. Psychometrika, 50, 181–202. doi: 10.1007/ BF02294246 Molenaar, P. C. M. (1987). Dynamic assessment and adaptive optimisation of the therapeutic process. Behavioral Assessment, 9, 389–416. Molenaar, P. C. M. (2004). A manifesto on psychology as idiographic science: Bringing the person back into scientific psychology, this time forever. Measurement: Interdisciplinary Research and Perspectives, 2, 201–218. doi: 10.1207/s15366359mea0204_1

Nesselroade, J., McArdle, J., Aggen, S., & Meyers, J. (2002). Alternative dynamic factor models for multivariate time series. In D. M. Moscowitz & S. L. Hershberger (Eds.), Modeling intraindividual variability with repeated measures data: Advances and techniques (pp. 325–365). Mahwah, NJ: Lawrence Erlbaum Associates. Nesselroade, J. R., & Molenaar, P. C. M. (1999). Pooling lagged covariance structures based on short, multivariate time series for dynamic factor analysis. In R. H. Hoyle (Ed.), Statistical strategies for small sample research (pp. 223–250). Thousand Oaks, CA: Sage Publications, Inc. Ptasznik, A. (2011). The person in personal medicine: Moving consumers from periphery to centre of decision making. Journal of Addiction and Mental Health, 14(3), 14–22. Raiche, G., Riopel, M., & Blais, J. G. (2006). Non graphical solutions for the Cattell’s scree test. Paper presented at the International Annual Meeting of the Psychometric Society, Montreal, Canada. Rao, C. R. (1958). Some statistical methods for the comparison of growth curves. Biometrics, 14, 1–17. Reise, S. P., Widaman, K. F., & Pugh, R. H. (1993). Confirmatory factor analysis and item response theory: Two approaches for exploring measurement invariance. Psychological Bulletin, 114, 552–566.

Molenaar, P. C. M. (2010a). Testing all six person-oriented principles in dynamic factor analysis. Development and Psychopathology, 22, 255–259. doi: 10.1017/S0954579410000027

Robinson, G. K. (1991). The BLUP is a good thing: the estimation of random effects. Statistical Science, 6(1), 15–32. doi: 10.1214/ss/ 1177011926

Molenaar, P. C. M. (2010b). Note on optimization of individual psychotherapeutic processes. Journal of Mathematical Psychology, 54, 208–213. doi: 10.1016/j.jmp.2009.04.003

Roebroeck, A., Formisano, E., & Goebel, R. (2005). Mapping directed influence over the brain using Granger causality and fMRI. NeuroImage, 25, 230–242. doi: 10.1016/j.neuroimage.2004.11.017

Molenaar, P. C. M., & Campbell, C. G. (2009). The new person-specific paradigm in psychology. Current Directions in Psychological Science, 18, 112–117. doi: 10.1111/j.1467-8721.2009.01619.x

Rovine, M. J., Molenaar, P. C. M., & Corneal, S. E. (1999). Analysis of emotional response patterns for adolescent stepsons using P-technique factor analysis. In R. K. Silbereisen & A. von Eye (Eds.), Growing up in times of social change (pp. 261–285). Berlin, Germany: De Gruyter.

Molenaar, P. C. M., Lerner, R., & Newell, K. M. (Eds.). (2014). Handbook of developmental systems theory and methodology. New York, NY: Guilford Press. Molenaar, P. C. M., & Nesselroade, J. R. (2001). Rotation in the dynamic factor modeling of multivariate stationary time series. Psychometrika, 66(1), 99–107. doi: 10.1007/BF02295735 Molenaar, P. C. M., & Nesselroade, J. R. (2009). The recoverability of P-technique factor analysis. Multivariate Behavioral Research, 44, 130–141. Molenaar, P. C. M., & Newell, K. M. (Eds.) (2010). Individual pathways of change: Statistical models for analyzing learning and development. Washington, DC: American Psychological Association. Molenaar, P. C. M., Sinclair, K. O., Rovine, M. J., Ram, N., & Corneal, S. E. (2009). Analyzing developmental processes on an individual level using non-stationary time series modeling, Developmental Psychology, 45(1), 260–271. doi: 10.1037/a0014170 Muthén, B., & Shedden, K. (1999). Finite mixture modeling with mixture outcomes using the EM algorithm. Biometrics, 55, 463–469. doi: 10.1111/j.0006-341X.1999.00463.x Nagin, D. S. (1999). Analyzing developmental trajectories: A semiparametric group-based approach. Psychological Methods, 4, 139–157. doi: 10.1037/1082-989X.4.2.139

Rovine, M. J., & Molenaar, P. C. M. (2000). A structural modeling approach to the random coefficients model. Multivariate Behavioral Research, 35(1), 51–88. doi: 10.1207/s15327906mbr3501_3 Rovine, M. J., & Walls, T. A. (2006). A multilevel autoregressive model to describe interindividual differences in the stability of a process. In J. L. Schafer & T. A. Walls (Eds.), Models for intensive longitudinal data (pp. 124–147). New York, NY: Oxford. Schwartz, D., & Collins, F. S. (2007). Environmental biology and human disease. Science, 316, 695–696. Seth, A. K. (2008). Causal networks in simulated neural systems. Cognitive Neurodynamics, 2(1), 49–64. doi: 10.1007/s11571-007-9031-z Smith, L.B., & Thelen, E. (2003). Development as a dynamic system. Trends in Cognitive Sciences, 7, 343–348. Smith, S. M., Miller, K. L., Salimi-Korshidi, G., Webster, M., Beckman, C. F., Nichols, T. E., Ramsey, J. D., & Woolrich, M. V. V. (2011). Network modeling methods for fMRI. NeuroImage, 54(2), 875–891. doi: 10.1016/j.neuroimage.2010.08.063 Sörbom, D. (1974). A general method for studying differences in factor means and factor structures between groups. British Journal of Mathematical and Statistical Psychology, 27(2), 229–239. doi: 10.1111/ j.2044-8317.1974.tb00543.x

References

865

Sterba, S. K., & Bauer, D. J. (2010). Matching method with theory in person-oriented developmental psychopathology research. Development and Psychopathology, 22, 239–254. doi: 10.1017/ S0954579410000015 Suls, J., Green, P., & Hillis, S. (1998). Emotional reactivity to everyday problems, affective inertia, and neuroticism. Personality and Social Psychology Bulletin, 24, 127–136. doi: 10.1177/0146167298242002

Wang, L., Hamaker, E. L., & Bergeman, C. (2012). Investigating interindividual differences in short term intraindividual variability. Psychological Methods, 17(4), 567–581. doi: 10.1037/a0029317 Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief measures of positive and negative affect: The PANAS scales. Journal of Personality and Social Psychology, 54(6), 1063–1070. doi: 10.1037/0022-3514.54.6.1063

Thurstone, L. L. (1947). Multiple factor analysis. Berkeley, CA: University of California Press. Tucker, L. (1958). Determination of parameters of a functional relation by factor analysis. Psychometrika, 23, 19–23. doi: 10.1007/BF02288975 Vandenberg, R. J., & Lance, C. E. (2000). A review and synthesis of the measurement invariance literature: Suggestions, practices, and recommendations for organizational research. Organizational Research Methods, 3(1), 4–69. doi: 10.1177/1094428102005002001 Voelkle, M. C., Brose, A., Schmiedek, F., & Lindenberger, U. (2014). Toward a unified framework for the study of between-person and within-person structures: Building a bridge between two research paradigms. Multivariate Behavioral Research, 49(3), 193–213. doi: 10.1080/00273171.2014.889593

Widaman, K. F., & Reise, S. P. (1997). Exploring the measurement invariance of psychological instruments: Applications in the substance use domain. In K. J. Bryant, M. Windle, & S. G. West (Eds.), The science of prevention: Methodological advances from alcohol and substance abuse research (pp. 281–324). Washington, DC: American Psychological Association.

Walls, T. A., & Schafer, J. L. (2006). Models for intensive longitudinal data. New York, NY: Oxford University Press.

Wood, P., & Brown, D. (1994). The study of intraindividual differences by means of dynamic factor models: Rationale, implementation, and interpretation. Psychological Bulletin, 116, 166–186. doi: 10.1037/ 0033-2909.116.1.166 Wright, S. (1934). The method of path coefficients. Annals of Mathematical Statistics, 5, 161–215. Zevon, M. A., & Tellegen, A. (1982). The structure of mood change: An idiographic/nomothetic analysis. Journal of Personality and Social Psychology, 43, 111–122. doi: 10.1037/0022-3514.43.1.111

CHAPTER 20

Configural Frequency Analysis for Research on Developmental Processes ALEXANDER VON EYE and EUN-YOUNG MUN

INTRODUCTION 866 Introduction to Configural Frequency Analysis 868 CFA MODELS FOR DEVELOPMENTAL RESEARCH 877 CFA of Differences I 878 CFA of Differences II: Structural Zeros 881 Extensions 883 CFA of Trajectories: Typical and Atypical Development 884 Analyzing Series That Differ in Length 886 CFA of Pre–Post Designs—An Application of Confirmatory CFA 889 Treatment Effects in Designs with Control Group 892

Single-Subject Designs 895 CFA of Lags 900 CFA of Cross-Lagged Designs 905 Comparing Individuals’ Trajectories 908 Predicting Events 912 PREDICTING END POINTS OF DEVELOPMENT 912 Predicting a Trajectory 915 DISCUSSION 916 Unique Characteristics of CFA 917 CFA and Translational Research 917 REFERENCES 919

INTRODUCTION

By the same token, however, this situation causes worry. A disconnect between theories of development and methodological application has also emerged. Theorists of person-oriented research (see Bergman & Magnusson, 1997; von Eye & Bergman, 2003; von Eye, Bergman, & Hsieh, 2015; see also Ch. 18) as well as idiographic psychology (Molenaar, 2004; Molenaar & Campbell, 2009) have suggested and shown that the results of typical statistical analysis are not applicable at the level of individuals. This is because most statistical methods perform analysis on aggregated data (e.g., means and covariances). A prominent example is the analysis of covariance matrices in structural equation modeling. These matrices contain variances in their diagonal cells and covariances in their off-diagonal cells. By definition, these aggregate measures indicate variability that is averaged over the individual cases included in a study. The cases themselves are considered random and, thus, replaceable (by other random cases). Because of these characteristics, it is rare that any results from aggregated data do indeed apply to the individual. In longitudinal research, in particular dynamic modeling, very strict ergodicity conditions must be fulfilled for the results of aggregate-level, statistical inference to be applicable to the individual (see Loken, 2010; Molenaar, 2004; von Eye, Bergman, & Hsieh, 2015). Examples of such rigid conditions

Developmental processes, either typical or atypical, are inherently about changes that occur over time in response to internal and external stimulation at the individual level. Recent development of statistical methods of analysis has resulted in a situation that is quite interesting and encouraging for researchers of developmental processes because both type and number of questions that can be answered has increased dramatically. This increased capacity allows researchers to answer scholarly questions that seemed intractable just a few years ago. Just think of dynamic modeling, modeling of trajectories and mixtures, the new models and methods for intensive longitudinal data, and for categorical or nonnormal data, or the new approaches to the analysis of causal hypotheses (for overviews, see Everitt & Howell, 2005; Laursen, Little, & Card, 2012; von Eye & Mun, 2013; von Eye & Wiedermann, 2013). The doors for sophisticated data analysis and, more important, for analysis that exactly tests the hypotheses that researchers ask are wide open.

1

Color versions of Figures 20.1, 20.5–20.7 are available at http://onlinelibrary.wiley.com/book/10.1002/9781118963418 866

Introduction

include stationarity, that is, the data follow a stochastic process whose joint probability distribution is not sensitive to shifts in time and space. The parameters of such stochastic processes, for example, the mean and variance, will also be insensitive to shifts in time and space. Clearly, most developmental processes violate the stationarity condition. Note that we argue not that any data analytic techniques based on aggregate level data are not useful but that these characteristics necessitate that one needs to be aware of these assumptions and be cautious in interpreting results. To be less confined and be truer to theories, what is needed, therefore, are methods of statistical data analysis that have the following two characteristics. First, they allow one to make statements about individuals or groups of individuals. Second, they do not require that improbable conditions for the process under study be met. Configural frequency analysis (CFA; Lienert & Krauth, 1975; von Eye, 2002; von Eye & Gutiérrez-Peña, 2004; von Eye, Mair, & Mun, 2010) is such a method. In this chapter, we introduce readers to CFA and give examples of questions that can be answered using this method. These questions differ from those that can be answered using standard methods of analysis. Instead of targeting relationships among variables, they target individual or groups of configurations, that is, individual profiles. Before we discuss the technical aspects of CFA, its application and computation, we ask, however, where the disconnect between developmental theories and analytic practices has occurred in the field of developmental research. The answer to this question can be used to locate or develop statistical methods that do not suffer from the problems mentioned previously. To answer this question, we use the example of Eysenck’s (1958) approach to the development of a personality theory (von Eye & Bogat, 2006). As is illustrated in Figure 20.1, Eysenck’s approach involves proceeding through a hierarchy of steps. At the lowest level of the hierarchy, indicated by the boxes at the bottom of the figure, researchers focus on what Eysenck terms specific reactions. These are behaviors that are observed for an individual, at a particular moment. For example, a child may act in defiance of his/her mother’s

867

instruction to clean up after playing with toys. Specific reactions are just observed and recorded. No evaluation or comparison is involved or intended. The next higher level in the hierarchy is that of habitual reactions. For example, the child tends to exhibit hostile behaviors toward the mother whenever instructed to clean up after playing. At the level of habitual reactions, ipsative comparisons (e.g., the same individual’s behavior compared with that from a day or week before) but also comparisons with other individuals (e.g., different individuals at a given moment) can be made. In more general terms, at this level, statements are already made that go beyond specific observations. The third level is that of personality characteristics such as perseverance, rigidity, or irritability. At this level, statements are made that carry the researcher beyond a particular behavior observation, even beyond the individual that was observed. For example, one can talk about communication style preferences, regardless of the individuals who may or may not exhibit these preferences. At the highest level, personality types are defined. These types are based on grouping personality characteristics. In Eysenck’s theory, we find the type of extraverted versus introverted personality characteristics, the type of stable versus labile personality characteristics, and the type of psychoticism. Figure 20.1 shows where the disconnect occurs: between the levels of habitual reactions and personality characteristics (separated by the dotted line). Below the dotted line, descriptive or comparative statements are made about individuals. Above the dotted line, statements describe personality characteristics. These statements are abstract, and references to the individual are not necessarily implied. Standard statistical analysis relates personality characteristics to one another (cluster analysis of cases being among the few exceptions). In contrast, CFA analyzes individuals with particular profiles. Specifically, CFA creates all possible profiles for a set of categorical variables, and answers the question whether individual profiles were observed at different rates than expected from prior knowledge or theory. Analyses that focus on the information provided above the dotted line (variables) have been termed variable oriented. Analyses that focus on the information

Figure 20.1 Constructing models of personality. See footnote 1. Source: Eysenck (1958).

868

Configural Frequency Analysis for Research on Developmental Processes

provided below the dotted line (individuals) have been termed person oriented (see Bergman & Magnusson, 1997; von Eye, Bergman, & Hsieh, 2015; Ch. 18). CFA is regarded as among the prime methods of analysis for person-oriented research (Bergman, Magnusson, & El-Khouri, 2003). Although person-oriented research requires many other considerations, such as study design and assessment, this chapter focuses on CFA, a particular analytical approach that is useful in person-oriented research. CFA proceeds from the person-oriented assumption that development is best described by patterns of characteristics (instead of variable relations). This is most relevant for research in developmental psychopathology, where we find that the development of pathological processes can be very specific to particular groups of individuals, causes, timing, and circumstances. This applies accordingly to translational and intervention research. These topics are taken up again later in this chapter. This chapter is structured as follows. First, we introduce readers to CFA. We cover conceptual and technical aspects of CFA. This introduction will be selective. More detail can be found in von Eye (2002) and von Eye, Mair, and Mun (2010). After this introduction, we present models of CFA that are of importance for research in developmental psychopathology. Throughout, examples will be given with real data.

Introduction to Configural Frequency Analysis CFA was proposed by Lienert in 1968 as a method that is particularly suited for clinical research and has since been developed further. CFA allows researchers to answer the question whether, given explicit expectations, individuals with certain profiles were observed at higher or lower rates than expected. The answers to this question can be used to provide support for a hypothesis or a theory. In addition, CFA can test whether the definition or the operationalization of a concept is valid. To give an example, consider a certain mood that is associated with depression. Individuals diagnosed as depressed often feel sad, anxious, hopeless, worthless, guilty, or irritable. In contrast, individuals diagnosed as depressed rarely feel happy or proud. Therefore, the null hypothesis stipulating that sad and happy moods can coexist in time within a person will, in an empirical study, be rejected. The two variables sadness and happiness are related such that the individual mood profile [sad, happy] is expected with below chance probability. Interestingly, it can also be discussed whether the profiles [not sad, happy], [not sad, unhappy], and [sad, unhappy] will be

observed with above or below chance probability. In CFA, individual profiles are examined. CFA is used to analyze cross-classifications of categorical variables. In this respect, CFA is comparable with log-linear modeling (Bishof, Fienberg, & Holland, 1975; von Eye & Mun, 2013), logistic regression (see Agresti, 1998), or latent class analysis (see Goodman, 1974). CFA differs, however, from these methods in that it does not attempt to explain the relations among variables. It allows one to examine individual cells of cross-classifications (also called configurations, cells, patterns, or profiles) with the question whether their occurrence frequencies differ from expectancy. Groups of cells can also be examined. More generally, cross-classifications of categorical variables can be analyzed with specific aims in mind. Goodman (1984) listed the aims of examining 1. the joint distribution of the variables that span a crossclassification; 2. the association structure of these variables; and 3. the dependency structure of these variables. CFA adds a fourth aim. It asks whether individual cells or groups of cells deviate from the expectancy that is specified using a probability model. Later in this chapter, we describe in more detail how a CFA probability model is specified. Von Eye and Gutiérrez-Peña (2004) note that CFA is used to identify those cells of a cross-classification where the action is. In these cells, local associations (Havránek & Lienert, 1984) become apparent. That is, in particular cells, there is an association among the variables that span a cross-classification, as is manifested by above-chance or below-chance frequencies of observations. In other cells, observations are as frequent as expected with reference to the probability model (also called base model). Cells that contain more cases than expected are labeled as constituting a CFA type. Cells that contain fewer cases than expected are labeled as constituting a CFA antitype. To illustrate these concepts, we present a first data example. The data in this example were collected in the context of the Mother–Infant Study (MIS; Bogat, Levendosky, DeJonghe, Davidson, & von Eye, 2004; Huth-Bocks, Levendosky, & Bogat, 2002). The study first assessed women in their last trimester of pregnancy and again at 2 months postpartum and followed up the women and their children every year until the children were 14 years old. The women in the study reported a range of experiences of intimate partner violence, from none to severe. Intimate partner violence was defined as

Introduction

male violence toward a female partner. Research questions concerned factors that predict risk and resilience in women exposed to intimate partner violence and their children as well as aspects of the home environment (e.g., maternal mental health, parenting style); and individual child characteristics (e.g., temperament) that predict maladaptive socioemotional outcomes (e.g., aggression; for results, see, e.g., Bogat et al., 2004; Huth-Bocks, Levendosky, Bogat, & von Eye, 2004; www.msu.edu/∼mis/). For the first data example, we use the variables posttraumatic stress disorder (PTSD) symptoms (P), Income (I), and Depression (D) observed at the first occasion, that is, in the last trimester of pregnancy. The variables were binned close to their median. For each of the variables, 1 indicates below the cutoff and 2 above the cutoff. Table 20.1 displays the CFA results for the 2 × 2 × 2, P × I × D cross-classification. Without going into the technical details, which will be introduced later, we report that this table was analyzed under the base model of variable independence, standard normal z scores were used to evaluate the individual cells, and 𝛼 was protected using the Holland–Copenhaver procedure. The overall goodness-of-fit likelihood ratio X 2 for the base model of variable independence for the data in Table 20.1 is 45.60. For df = 4, this value suggests that the model is rejected (p < 0.01). This is the typical situation for a CFA that results in types and antitypes. In Table 20.1, we see two types and one antitype. The first type is constituted by configuration 1 2 1. These are women whose PTSD and Depression symptom scores were below the cutoff, and income was above the cutoff. Significantly more women showed this profile than expected with reference to the base model of variable independence. The second type, constituted by configuration 2 1 2, shows just the opposite profile. These are women with elevated PTSD and Depression symptom scores and Income below the cutoff. Again, significantly more women than expected showed this profile. TABLE 20.1 First-Order CFA of the Cross-Classification of PTSD (P), Income (I), and Depression (D) Configuration PID

m

̂ m

Statistic

p

111 112 121 122 211 212 221 222

28.00 12.00 53.00 4.00 37.00 37.00 28.00 9.00

37.317 15.847 30.770 13.067 42.703 18.134 35.211 14.953

−1.5251 −.9663 4.0076 −2.5082 −.8727 4.4303 −1.2152 −1.5394

.063612 .166938 .000031 .006067 .191425 .000005 .112144 .061855

Type Antitype

869

The sole antitype in Table 20.1 is constituted by configuration 1 2 2. The women with this profile exhibit belowthe-cutoff scores in PTSD, and above-the-cutoff scores in Income and Depression. This profile was observed four times, which is significantly below the expected frequency of about 13. The results shown in Table 20.1 are typical of CFA results in a number of aspects. First, in Table 20.1, the largest cell does constitute a type and the smallest cell does constitute an antitype. This is not always the case. The main reason for this characteristic of CFA results is that CFA focuses on discrepancies from expectation instead of sheer size of cells (the base model of zero-order CFA is the only exception; Lienert & von Eye, 1989; von Eye, 2002). Therefore, even relatively small cells can contain more cases than expected, and relatively large cells can contain fewer cases than expected. Later in this chapter, Table 20.12 shows an example in which one of the smallest cells constitutes a type. Second, CFA tables of types and antitypes are interpreted after the base model is rejected. However, we note that the rejection of a base model does not guarantee that types and antitypes exist. If a base model describes the data well, there will be no noteworthy discrepancies between observed and expected cell frequencies, and the search for types and antitypes that indicate the location of significant discrepancies is largely pointless. And finally, only a selection of cells (configurations) emerges as type or antitype constituting. The remaining cells do not statistically deviate from the base model. In the following sections, we describe technical elements of CFA (for more detail, see von Eye, 2002; von Eye & Gutiérrez-Peña, 2004; von Eye, Mair, & Mun, 2010). CFA: Technical Elements We begin the description of technical elements of CFA with a discussion of CFA base models. These models are of crucial importance because the interpretation of CFA types and antitypes depends on the effects that are (not) included in the base model. In general, the base model includes all effects that are not of interest to the researcher. If the base model is rejected, at least some of the effects the researcher is interested in are bound to exist. Naturally, different base models can lead to different expected cell frequencies (Mellenbergh, 1996) and, thus, different patterns of types and antitypes. Therefore, the selection of base models is not arbitrary.

Type

Definition of CFA Base Models. Most CFA base models are log-linear models (for base models that are not

870

Configural Frequency Analysis for Research on Developmental Processes

log-linear, see von Eye, 2002). Therefore, expected cell frequencies for most CFA models can be estimated using the log-frequency model ̂ = X𝜆 log m ̂ is the array of model frequencies, that is, frequenwhere m cies that conform to the model specifications. X is the design matrix (also called indicator matrix). Its vectors reflect the CFA base model or, in other contexts, the log-linear model under study. 𝜆 is the vector of model parameters. These parameters are not of primary interest in frequentist CFA. Instead, CFA focuses on the discrepancies between the expected and the observed cell frequencies. In contrast to log-linear modeling, CFA is not applied with the goal to identify a model that describes the data sufficiently and parsimoniously. Rather, CFA is applied with the goal of identifying the configurations that contradict the base model. These configurations reflect the effects of interest. CFA base models have three main characteristics (von Eye, 2004a): 1. They contain all effects that are not of interest to the researcher; 2. They reflect theoretical assumptions concerning the nature of the variables as either of equal status or grouped into predictors and criteria; and 3. Consider the sampling scheme under which the data were collected. In the following paragraphs, we discuss two of these three characteristics. We begin with the third. In practically all CFA applications, the sampling scheme is either multinomial or product-multinomial. When the scheme is multinomial, cases can be placed in any cell of a cross-classification, with no constraints. Marginal probabilities of the data collected under this sampling scheme are neither a priori given nor determined. In contrast, when the sampling scheme is product-multinomial, marginal probabilities of the data collected using this sampling scheme are given a priori. So, when a study focuses on the effects of smoking, researchers may wish to implement a balanced design such that equal numbers of smokers and nonsmokers go into the table. In this case, no additional smokers can enter the study when this subsample is filled. From a data analytic perspective, the sampling scheme has no implications for parameter estimation. However, the number and type of base models that can be estimated are restricted under the product-multinomial sampling scheme (for more detail, see von Eye et al., 2010).

The first of these characteristics is most important for the interpretation of types and antitypes. A CFA base model contains all effects that are not of interest to the researcher. Consider two examples. The base model used for the analysis reported in Table 20.1 was the model of variable independence. That is, the base model included only the main effects of the three variables PTSD, Income, and Depression. This model prevails only if the three variables are indeed independent of one another. If, however, one or more of the three possible two-way interactions, or the three-way interaction among all three variables exist, the model is rejected. In other words, from the fact that the base model of variable independence is rejected, we conclude that interactions exist. From the emergence of types and antitypes and the existence of configurations that do not contradict the base model, we conclude that local associations exist (Havránek, & Lienert, 1984). Types and antitypes that result from the bae model of variable independence thus indicate that interactions exist and where the interactions exist. For the second example, consider the base model of Prediction CFA (PCFA; Heilmann & Lienert, 1982; Heilmann, Lienert, & Maly, 1979; Lienert & Krauth, 1973a; von Eye, Mair, & Bogat, 2005; von Eye & Rovine, 1994). This base model includes two sets of effects. The first is defined by all possible effects on the predictor side. The model is thus saturated within the predictors. Types and antitypes cannot emerge just because predictors are related to one another. Second and in parallel, the base PCFA model is also saturated in the criterion variables. Therefore, types and antitypes cannot emerge just because criterion variables are related to one another. Third, the PCFA base model proposes independence between the two sets of variables: predictors and criterion variables. If this model is rejected, types and antitypes, by necessity, reflect relationships that link predictor variables with criterion variables. This characteristic is exemplified when we specify log-linear CFA base models, in the following paragraphs. For the purposes of this chapter and in correspondence with the literature on log-linear modeling and CFA, we use two ways to specify a CFA base model. The first is the log-linear notation. Its main characteristic is that every main effect, interaction, special effect, and covariate are represented in the equation. Consider the base model for the example in Table 20.1. This model contains only the main effects of P, I, and D. The general log-linear model ̂ = X 𝜆 is then, in this particular case, log m ̂ = 𝜆 + 𝜆 P + 𝜆I + 𝜆D log m

Introduction

̂ is the vector of expected cell frequencies (model where m frequencies), 𝜆 is the intercept, and the superscripted 𝜆s represent the main effects of the variables P, I, and D, respectively. In this particular case, each 𝜆 corresponds to one parameter. When effects require more than one parameter, which is the case when variables have more than two categories, subscripts are used to differentiate these categories. To illustrate the PCFA base model, consider the two predictors, P1 and P2, and the two criterion variables, C1 and C2. The log-linear base model for these four variables is ̂ = 𝜆 + 𝜆P1 + 𝜆P2 + 𝜆C1 + λC2 log m +𝜆P1,P2 + 𝜆C1,C2 The first line of this equation contains the main effects of all variables in the model. The second line contains the interaction between the two predictor variables and the interaction between the two criterion variables. All log-linear base models can be expressed using this notation. In addition to the log-linear notation, we use the bracket notation. This notation includes each effect in a pair of brackets. For main effects, a bracket pair will include just one variable. For two-way interactions, the bracket pair includes two variables, and so forth. The base model of variable independence used for Table 20.1 contains three main effects. In bracket notation, this model is [P], [I], [D]. The PCFA base model is, for the last example, [P1], [P2], [C1], [C2], [P1, P2], [C1, C2]. Many CFA base models are hierarchical log-linear models. In these hierarchical models, the lower order relatives of higher order terms are implied. Therefore, these lower order terms can be omitted in the model specification. They are included in the model, but, in the specification, they are redundant. The base model of variable independence cannot be simplified further, because the main effects are the lowest order effects in a log-linear model. In contrast, the PCFA base model can be simplified to [P1, P2], [C1, C2]. In this chapter, to be consistent with the CFA literature, and because some CFA base models are not hierarchical, we specify base models so that every effect under consideration is made explicit. We now ask which effects can be the causes for types and antitypes in the base model of variable independence. As we said before, these are effects that represent interactions. In the first data example, these are the effects [P, I], [P, D], [I, D], and [P, I, D], that is three two-way and one three-way interactions. If any of these effects exists, types and antitypes can emerge from the base model [P], [I], [D].

871

For the PCFA base model of the second example, the following effects may exist and cause types and antitypes to emerge: • Two-way interactions: [P1, C1], [P1, C2], [P2, C1], and [P2, C2]; • Three-way interactions: [P1, P2, C1], [P1, P2, C2], [P1, C1, C2], and [P2, C1, C2]; • Four-way interaction: [P1, P2, C1, C2]. Each of these interaction terms contains at least one predictor and at least one criterion variable. Types from PCFA indicate which predictor configurations allow one to predict the occurrence of which criterion configuration. PCFA antitypes allow one to predict which criterion configuration cannot be expected to occur after which predictor configuration. Given the right context, causal inference may be possible. Types could then be interpreted as predictor configurations that cause criterion configurations to occur. Antitypes could then be interpreted as predictor configurations that prevent criterion configurations from occurring. Classes of CFA Base Models Two classes of log-linear CFA base models have been discussed: global models and regional models (von Eye, 1988, 2002). Global models share the characteristic that all variables have the same status. There are no predictors, no criterion variables, no mediators or moderators. This situation is comparable to standard exploratory factor analysis in which all variables have the same status as well. There is a hierarchy of global CFA base models. Beginning at the bottom of the hierarchy, the first model is that of zero-order CFA (Lienert & von Eye, 1984, 1989), which is also called Configural Cluster Analysis (CCA). The base model of CCA is a null model. It assumes that no effects exist in the cross-classification of the variables under study. Types, which, here, are also called configural clusters, reflect agglomerations of cases in the data space. This characteristic makes zero-order CFA comparable to the cluster analytic methods that use spatial distance to cluster individual cases. Antitypes, which are also called configural anticlusters in the context of CCA, reflect sectors of the data space that contain surprisingly small numbers of cases or none. These sectors of the data space are typically ignored in cluster analysis. The log-linear base model ̂ = 𝜆. In other words, the base model for for CCA is log m CCA proposes a uniform distribution of cases. On the next higher level of the hierarchy of global CFA base models, we find first-order CFA. This is the original model used by Lienert (1968). First-order CFA includes all

872

Configural Frequency Analysis for Research on Developmental Processes

main effects but nothing else. It thus takes differences in marginal probabilities into account. Types and antitypes reflect local associations of first or higher order. To illustrate, consider the four variables A, B, C, and D. For these four variables, the first-order CFA base model is ̂ = 𝜆 + 𝜆 A + 𝜆 B + 𝜆C + 𝜆D log m The following interactions are not part of the first-order base model and can, therefore, cause types and antitypes to exist (in bracket notation): • Two-way interactions [A, B], [A, C], [A, D], [B, C], [B, D], and [C, D]; • Three-way interactions [A, B, C], [A, B, D], [A, C, D], and [B, C, D] • Four-way interaction [A, B, C, D]. The first data example in this chapter (Table 20.1) used first-order CFA as the base model. Most applications of CFA utilize first-order CFA as the base model, which takes only main effects into account. In many instances, firstorder CFA base models are used to determine whether there are any effects. In most of the examples presented in this chapter, the first-order CFA base model is also used for this purpose. On the next-higher level of the hierarchy of CFA base models, we find second-order CFA. This model assumes that two-way interactions exist. Types and antitypes can, therefore, emerge only if third- or higher order interactions exist. To illustrate second-order CFA, consider, again, the four variables A, B, C, and D. The second-order CFA base model for these variables is ̂ = 𝜆 + 𝜆A + 𝜆B + 𝜆C + 𝜆D log m +𝜆AB + 𝜆AC + 𝜆AD + 𝜆BC + 𝜆BD + 𝜆CD The second line of this equation contains the two-way interactions that, in first-order CFA, may cause types and antitypes. The possible interactions that are not part of the second-order CFA base model can, therefore, cause types and antitypes to exist (in bracket notation) are • Three-way interactions [A, B, C], [A, B, D], [A, C, D], and [B, C, D]; and the • Four-way interaction [A, B, C, D]. Although rarely applied, second-order CFA can be of interest for a number of reasons. The most important of these is that it uses first-order relations as a basis for analysis as do factor analysis, multidimensional scaling,

correspondence analysis, and some structural equation models. However, in contrast to these models, second-order CFA asks whether there exists systematic variability above and beyond what is explained by first-order interactions (i.e., all pairwise interactions). Factor analysis as well as the other methods on the list attempt to model the structure evident only in first-order relations. For an application of second-order CFA in psychopathological research, see von Eye and Lienert (1984). Base models higher than second-order have not been discussed in detail in the CFA literature. However, higher order base models can easily be specified in the context of hierarchical log-linear models. Using the four variables A, B, C, and D, the third-order CFA base model is ̂ = 𝜆 + 𝜆A + 𝜆B + 𝜆C + 𝜆D log m +𝜆AB + 𝜆AC + 𝜆AD + 𝜆BC + 𝜆BD + 𝜆CD +𝜆ABC + 𝜆ABD + 𝜆ACD + 𝜆BCD Using this model, only the four-way interaction [A, B, C, D] can be the cause of types and antitypes. In some of the paradoxes discussed in the statistics literature, for example, the Meehl’s paradox (Meehl, 1950; von Eye, 2002), the highest order interaction carries the entire association in a cross-classification. To explore whether this is the case for a concrete data set, a CFA with a ‘d-1’st-order base model can be performed, where d denotes the number of variables. If types and antitypes still emerge from this base model, one can proceed and test hypotheses that are compatible with Meehl’s paradox. This applies accordingly to Simpson’s paradox (but not necessarily to other paradoxes; Krauth, 2005), and some of the recently discussed configural moderator models (von Eye & Mun, 2013). In all global CFA base models, types and antitypes contradict the assumptions about associations that are made when estimating the expected cell frequencies. The resulting configurations stand out in the sense that they are observed at different rates from those that conform to the base model. An example of such a configuration is a pattern of symptoms that describes a nosological unit. These symptoms are not necessarily predictors of each other, and, conceptually, they cannot necessarily be grouped as dependent or independent variables, nor as mediators or moderators. For this reason, increasing numbers of CFA applications can be found in developmental psychopathology (e.g., Bogat, Levendosky, & von Eye, 2005; Martinez-Torteya, Bogat, von Eye, & Levendosky, 2009; von Eye & Bogat, 2006). In contrast, regional CFA base models do distinguish groups of variables. Predictors are distinguished from

Introduction

criterion variables, dependent variables from independent variables, covariates are treated differently than the variables that span the cross-classification, and mediator as well as moderator models can be tested. It is most important to note that most longitudinal CFA base models are regional models. A selection of these base models is explained later in this chapter. Therefore, we review only two examples in the present context. The first example is that of Prediction CFA (PCFA). As discussed in previous sections of this chapter, the PCFA base model includes two sets of effects or variables. The first set includes all possible effects on the predictor side, rendering the model saturated in the predictors. This characteristic has the effect that types and antitypes cannot emerge because there are associations within the predictors. The PCFA base model is also saturated in the criterion variables. Therefore, types and antitypes cannot emerge because the criterion variables are related to each other. The third characteristic of the PCFA base model is that it proposes independence between predictor and criterion variables. If this base model is rejected, types and antitypes, by necessity, reflect relationships that link predictor variables to criterion variables. The second example is that of Multigroup CFA. This CFA model also contains two groups of variables. The first group contains the variables that are used to compare the groups, that is, the discriminator variables. The Multigroup CFA base model is saturated in these variables. Therefore, types and antitypes (also called discrimination types) cannot emerge because the discriminator variables are related to each other. The second group of variables describes the groups. In most CFA applications, so far, this group contains only one variable, the grouping variable. For this variable, only the main effect is considered. This is done to take differences in group size into account. Interactions among the discriminator variables and the grouping variables are not part of the base model. Therefore, discrimination types and antitypes reflect such interactions, by necessity. Consider, for example, the four variables G, A, B, and C. G is the grouping variable and A, B, and C are used to discriminate the groups. The multigroup CFA base model for these variables is ̂ = 𝜆 + 𝜆G + 𝜆A + 𝜆B + 𝜆C + 𝜆AB + 𝜆AC + 𝜆BC + 𝜆ABC log m Considering that this model is hierarchical, it could also have been expressed as [G], [A, B, C]. More examples of regional CFA models follow when we discuss models in more detail that are of particular interest in developmental research. Before we embark on the discussion and

873

illustration of these models, we present, in the next section, a selection of significance tests that can be used to make type/antitype decisions. Significance Testing in CFA A large number of significance tests for CFA has been proposed. These tests differ in power, characteristics of the design under which they can be applied, their performance in samples of varying size, and whether they are exact or approximative. In this section, we present a selection of more popular tests. In addition, we present a test that can be applied when hypotheses are tested that involve groups of cells instead of individual cells. The CFA Null Hypothesis. Each of the tests to be discussed and applied in this chapter can be used to test the CFA null hypothesis H0 ∶ E[mr ] = m∗r where E[mr ] is the expectancy for cell r, where r goes over all cells of a cross-classification, and m∗r is the expected frequency for cell r under a base model. For the test of this null hypothesis, there are three possible outcomes: type, neither type nor antitype, and antitype. First, the cell constitutes a CFA type. In this case, the null hypothesis is rejected because BN,𝜋r (mr − 1) ≥ 1 − 𝛼 where N is the sample size, 𝜋r is the probability of cell r, and mr is the observed frequency of cell r. In other words, in the first case, the null hypothesis is rejected because cell r contains more cases than expected. In the second case, the null hypothesis prevails because cell r contains as many cases as expected. In the third case, the null hypothesis is rejected because BN,𝜋r (mr ) ≤ 𝛼 In other words, in the third case, the null hypothesis is rejected because cell r contains fewer cases than expected under a base model. It constitutes an antitype. CFA Tests for Individual Configurations. The first test to be presented here is the nonparametric binomial test. It can be employed under any sampling scheme (multinomial, product-multinomial sampling). Given mr , the exact tail probability of mr and more extreme frequencies is

Br (mr ) =

l ( ) ∑ N j N−j pq , J j=a

874

Configural Frequency Analysis for Research on Developmental Processes

where q = 1 – p, a = 0 if mr < m∗r and a = mr if mr ≥ m∗r , l = mr if a = 0, and l = N if a = mr . Usually, the value p is estimated from the sample. It is a priori known only in rare cases. As can be seen from the equation, the probability of each frequency is calculated and added to the total. There is no need to assume that a theoretical distribution is reasonably approximated. The binomial test is exact. This test is recommended when a sample is so small that the approximation properties of approximative tests are doubtful. More powerful than the binomial test is the z test, mr − Np z= √ Npq The z test can also be applied under any sampling scheme. Among the most powerful tests is Lehmacher’s (1981) normal approximation of the hypergeometric test, zL,r

m − m∗r = r 𝜎r

where 𝜎r is the square root of the exact variance, which can be calculated as σ2r = Npr (1 − pr − (N − 1)(pr − p̃r )),

A test that is similarly powerful was recently proposed by von Eye and Mair (2008a): the standardized Pearson residual, z′ r . The test statistic is m − m∗r z′ r = √ r m∗r (1 − hr ) where hr is the r-th diagonal element of the well-known hat matrix, H. This matrix is H = W 0.5 X (X ′ WX )−1 X ′ W 0.5 In this equation, W is a diagonal weight matrix that contains the expected frequencies, m∗r , in its diagonal. As the binomial test and the z test, the standardized Pearson residual can be used under any sampling scheme or base model. Testing Groups of Configurations The sum of the previous z statistics has an expectancy of E(Σzr ) = 0 and a variance of t, where t is the number of terms (cells) in the summation. Now, if a group of t configurations is subjected to a CFA null hypothesis that includes the whole group, Stouffer’s Z (Stouffer, Suchman, DeVinney, Star, & Williams, 1949) can be used, which is (

and when d = 3 variables are studied, the values of pr and p̃r are estimated as pr = and p̃r =

mi.. m.j. m..k N2

(mi.. − 1)(m.j. − 1)(m..k − 1) (N − 1)d

where d is the number of variables that span the crossclassification. For the present description of the values of pr and p̃r , we used the case of d = 3. Lehmacher’s test is powerful and tends to suggest types or antitypes even when they may not truly exist. To prevent this test from suggesting nonconservative type–antitype decisions, Küchenhoff (1986) recommended using a continuity correction for Lehmachers test, which involves reducing the difference in the numerator by 0.5. Lehmacher’s test can be used only under a product-multinomial sampling scheme. That is, the marginal probabilities must be fixed for the test to be valid. Another constraint is that Lehmacher’s test can be applied only under the base model of variable independence.

Z=

t ∑ zi

)/

√ t.

i=1

Data Example for Stouffer’s Z. The following data example illustrates the use of Stouffer’s Z in CFA. For the example, we reanalyze the data in Table 94 in von Eye (2002). In a study on experimental psychoses, Lienert (1964) administered LSD50 to 65 students. When the effects of the drug set in, three symptoms were observed: Narrowed Consciousness (C), Thought Disturbance (T), and Affective Disturbance (A). Each symptom was scaled as either present (1) or absent (2). Table 20.2 displays the 2 × 2 × 2 C × T × A cross-tabulation, along with the expected frequencies that were estimated under the base model of variable independence. We used the z test along with the Bonferroni-protected 𝛼 = 0.00625 (methods of 𝛼 protection are discussed in the following section). The overall goodness-of-fit likelihood ratio X 2 for the base model of variable independence for the data in Table 20.2 is 45.07. For df = 4, this value suggests that the model be rejected ( p < 0.01). Still, none of the eight configurations constitutes a type or an antitype. When looking at the patterns in Table 20.2, we notice that the

Introduction

variables Committed a Crime (C), Hyperactivity (H), and Adrenaline level (A) were observed. Crime was coded: 1 = no crime committed, 2 = crime committed before the age of 18, and 3 = persistent crime. Hyperactivity and Adrenaline levels were both coded as 1 = low and 2 = high. Cutoffs were the means. Table 20.3 displays the 3 × 2 × 2 C × H × A cross-classification, along with the expected cell frequencies and the tail probabilities from the binomial test, the z-test, and the standardized Pearson residual. 𝛼 was protected using the Holland–Copenhaver procedure (which is explained in the next section). The expected cell frequencies were estimated under the base model of variable independence. The overall goodness-of-fit likelihood ratio X 2 for the base model of variable independence for the data in Table 20.3 is 47.14. For df = 7, this value suggests that the model is rejected (p < 0.01), and we expect types and antitypes to emerge. We now interpret the results in Table 20.3 from two perspectives. First, we discuss the substantive meaning of the type and antitype in this table. Then, we compare the three CFA significance tests. As far as the analysis of crimes committed in adolescence is concerned, we note that of those 11 male adolescents who were grouped as persistent in their criminal behavior (C = 3), nine showed the pattern of high hyperactivity and low adrenaline. Pattern 3 2 1 constitutes a type. The antitype shows the same pattern (. 2 1), suggesting that with high levels of hyperactivity and low levels of adrenaline, it is particularly unlikely to be in the group of nonoffenders. In fact, not a single nonoffender showed this pattern, 1 2 1. We also ask whether including all pairwise interactions would allow us to fully explain the frequency distribution in Table 20.3. The corresponding model is hierarchical. It is [C, H], [C, A], [H, A]. The results suggest that this model

TABLE 20.2 First-Order CFA of the Cross-Classification of Narrowed Consciousness (C), Thought Disturbance (T), and Affective Disturbance (A) Configuration CTA

m

̂ m

z

p

111 112 121 122 211 212 221 222

20.00 1.00 4.00 12.00 3.00 10.00 15.00 .00

12.506 6.848 11.402 6.244 9.464 5.182 8.629 4.725

2.1193 −2.2348 −2.1921 2.3035 −2.1011 2.1162 2.1690 −2.1738

.017034 .012715 .014185 .010626 .017815 .017164 .015041 .014862

bi-symptomatic configurations are all observed at low rates. These are the configurations 1 1 2, 1 2 1, and 2 1 1. We, therefore, ask, whether these three configuration constitute a composite antitype (also called joint antitype). To test the corresponding null hypothesis that they do not constitute a composite antitype, we take the z scores from Table 20.2, insert them into the formula for Stouffer’s Z, and obtain Z=

−2.24 − 2.19 − 2.10 = −3.7689 √ 3

This value suggests that the null hypothesis can be rejected. We conclude that the three bisymptomatic configurations, 1 1 2, 1 2 1, and 2 1 1, constitute a joint antitype. It is particularly unlikely that just two of the symptoms are observed. Data Example for Comparison of CFA Tests. For the following example, we reanalyze data that were published by Magnusson in 1996. The data describe a sample of 70 male adolescents. To study the role played by physiological and psychological factors in criminal behavior, the three

TABLE 20.3 First-Order CFA of the Cross-Classification of Crime (C), Hyperactivity (H), and Adrenaline Level (A) Configuration CHA 111 112 121 122 211 212 221 222 311 312 321 322

875

m

̂ m

B

z

p(z)

z′

p(z′ )

25.00 18.00 .00 3.00 2.00 8.00 2.00 1.00 1.00 1.00 9.00 .00

20.1367 16.0061 5.4918 4.3653 5.6908 4.5235 1.5520 1.2337 4.8153 3.8276 1.3133 1.0439

.12564275 .32783713 .00328220 .35776569 .06902106 .08160703 .46147235 .64975053 .04204248 .09856945 .00000666 .34932965

1.0838 0.4984 −2.3435 −0.6535 −1.5472 1.6346 0.3596 −0.2104 −1.7387 −1.4453 6.7076 −1.0217

.139235 .309110 .009553 .256728 .060912 .051067 .359583 .416685 .041046 .074191 .000000 .153461

2.1456 0.9069 −3.2172 −0.8450 −2.2438 2.1978 0.4067 −0.2326 −2.4868 −1.9171 7.5137 −1.1204

0.016 0.182 0.001 0.199 0.012 0.014 0.342 0.408 0.006 0.028 0.000 0.131

Note. B = Binomial test; z = z-test; z′ = standardized Pearson residual test.

Antitype

Type

876

Configural Frequency Analysis for Research on Developmental Processes

must be rejected also (LR − X 2 = 11.05; df = 2; p = 0.004). Although this model is significantly better than the original base model of variable independence (ΔX 2 = 36.09; Δdf = 5; p < 0.01), it cannot stand by itself. As far as the comparison of the three CFA tests is concerned, we first see that, in this example, the z test is generally more powerful than the binomial test. As we said previously, we trust the binomial test more than the z test when a sample is small. Therefore, we interpreted the results from the binomial test. The standardized Pearson residual is much more powerful than the other two tests in this example. All configurations had smaller p values than under the other two tests. Protection of 𝜶 CFA involves multiple testing. There are mainly two reasons why multiple tests in CFA (and in other methods of analysis, for example, ANOVA) can become problematic. The first is known as capitalizing on chance, the second lies in the dependency of tests. Only the first of a series of tests on the same data keeps the a priori specified 𝛼 level. When more than one test is performed, the risk of making an 𝛼 error increases, even if the tests are independent. For example, if, in the context of testing for CFA types and antitypes, each of the 27 cells of a 3 × 3 × 3 table is examined under 𝛼 = 0.05, the probability of incorrectly rejecting two null hypotheses is 0.24. Performing multiple tests on one sample therefore includes a high risk of capitalizing on chance if no precautions are taken to protect 𝛼. This situation gets even worse when tests are dependent. von Weber, Lautsch, and von Eye (2003) showed that, when the expected cell frequencies in a 2 × 2 table are estimated under a log-linear main effect model, that is, the model of variable independence, the first CFA test indeed keeps the a priori specified 𝛼 level. However, the outcomes of the second, third, and fourth CFA tests in the 2 × 2 table are completely dependent upon the outcome of the first test. Similarly, Krauth (2003) showed that the number of possible patterns of types and antitypes increases as the size of a table increases, but that the tests in a CFA never become completely independent. A number of procedures exists for the protection of 𝛼. The most conservative of these, the Bonferroni adjustment, is the most popular procedure. This procedure specifies (1) that the sum of all adjusted significance levels not exceed the nominal 𝛼, or Σr 𝛼 r ≤ 𝛼, where r indexes the cells subjected to CFA tests; and (2) that the significance threshold be the same for each test, or 𝛼 r = 𝛼*, for all r, where 𝛼* is the adjusted threshold. The value of 𝛼* that meets both of these conditions is 𝛼* = 𝛼/R.

More efficient procedures have been proposed. Based on Holm’s (1979) procedure, Holland and Copenhaver (1987) proposed a procedure that does not require that Σr αr ≤ α. In this procedure, the tail probabilities of the cell-wise test statistics are rank ordered. The smallest probability is given Rank 1. The protected value of 𝛼* is then, for the i-th test, 1

𝛼i∗ = 1 − (1 − 𝛼) r−i+1 This procedure is clearly less restrictive than the Bonferroni protection. It is also (slightly) less restrictive than Holm’s (1979) procedure (for comparisons of procedures for the protection of 𝛼, see Olejnik, Li, Supattathum, & Huberty, 1979; von Eye, 2002; for procedures for special cases, see von Eye, 2002). The Five Steps of CFA When performing a CFA, researchers proceed in five steps (von Eye, 2002). 1. Selection of a base model and estimation of expected frequencies. In general, a CFA base model is a chance model that takes those effects into account that are NOT of interest to the researcher. The base model is used to estimate the expected cell frequencies in the table under study that are then compared with the observed cell frequencies, cell-by-cell. If a deviation between an expected and an observed cell frequency is significant, it reflects, by necessity, effects that are of interest to the researcher. 2. Selection of a concept of deviation from independence. Some of the measures of deviation from a base model take into account marginal frequencies, others do not. The corresponding measures are called marginaldependent and marginal-free (Goodman, 1991). Examples of marginal-dependent measure include the Φ-coefficient, which is a relative of Pearson’s X 2 . Φ quantifies the strength of the association between two dichotomous variables. The strength corresponds to the degree of deviation of the observed distribution from the base model of independence between these two variables. Examples of marginal-free measures include the odds ratio, 𝜃. Marginal-dependent and marginal-free measures can give different appraisals of deviation from a base model. Marginal-free measures have been proposed for use in two-group CFA (von Eye, Spiel, & Rovine, 1995). 3. Selection of a significance test. Significance tests of the null hypothesis that types or antitypes do not exist differ in that some are exact, whereas others are

CFA Models for Developmental Research

approximative. They also differ in statistical power and in the sampling schemes under which they can be employed. Simulation studies have shown that the tests that perform well under many conditions include Pearson’s X 2 , the z-test, and the exact binomial test. When sampling is product-multinomial, the best performing tests include Lehmacher’s (1981) exact and approximative hypergeometric tests. 4. Performing significance tests under protection of 𝛼. In both exploratory and confirmatory application of CFA, typically, a large number of tests is conducted. This number of tests is generally smaller in confirmatory CFA than in exploratory CFA. In either case, when more than one significance test is performed, the significance level, 𝛼, needs to be protected. The routine method for 𝛼 protection, the Bonferroni procedure, can suggest rather conservative decisions about the existence of types and antitypes. Therefore, Holm’s (1979) procedure and even less restrictive methods that have been derived from it have gained in popularity. 5. Interpretation of types and antitypes. The interpretation of types and antitypes utilizes the following five sources of information. 1. The substantive meaning of a configuration. This meaning is determined by the definition of the categories that constitute a configuration. 2. The base model. For example, when the base model distinguishes between predictor and criterion variables, the interpretation of types and antitypes differs from the one without this distinction (Mellenbergh, 1996). 3. The concept of deviation from expectation as marginal-dependent versus marginal free. 4. The sampling scheme (e.g., multinomial, productmultinomial). 5. External information that is used to discriminate among individuals in different cells in the cross-classification under study (from each other and from the configurations that constitute types or antitypes). The discrimination is not part of CFA itself. Instead, discrimination is performed in follow-up tests that are undertaken, for example, to establish external validity of CFA types and antitypes (see von Eye, Mair, & Mun, 2010). For discrimination, variables are used that were not part of CFA. In the remainder of this chapter, we discuss and illustrate CFA models that are of particular interest in developmental research. The models are selected based

877

on the developmental questions they allow the researcher to answer.

CFA MODELS FOR DEVELOPMENTAL RESEARCH The questions asked in developmental research can concern any individual parameter of the population under study, and any combination of parameters. In the first two data examples of this chapter, we asked questions concerning local associations in cross sectional data. In the following sections, we focus on questions for longitudinal data. In general, CFA allows one to answer questions concerning typical trajectories (longitudinal types) and atypical trajectories (longitudinal antitypes). The trajectories can be defined with respect to means, characteristics of shapes of curves, variability of scores, trends, intervention effects, end points of development, the predictability of trajectories, lags, effects of covariates, mediation, moderation, causal hypotheses, and many other aspects of longitudinal data. To answer some of the questions, ordinal data are needed, for other questions, data must be at the interval level. For many questions, nominal level data are sufficient. Viewed from a data characteristics perspective, given nominal, ordinal, or interval level data, some questions can be asked, but other questions cannot be asked. For example, for nominal level data, local associations and prediction patterns can be examined. However, questions concerning constancy and change in means cannot be asked using nominal-level data. In the following sections, we point to the properties that data must possess for certain questions to be tractable. This part of the present chapter is structured as follows. We begin with CFA of differences and related methods. For the simplest CFA models in this section, ordinal-level data are sufficient. For more advanced models, interval or ratio scale-level data are required. In the following section, we model shapes of trajectories. The first two sections are closely related to each other. The third section covers pre–post designs. We show how this type of design can be analyzed with CFA. The role of control groups is discussed. The following section covers single subject designs. In the contexts of person-oriented (Bergman & Magnusson, 1997; von Eye & Bergman, 2003; von Eye, Bergman, & Hsieh, 2015; Ch. 18) and idiographic research (Molenear, 2004; Molenaar & Campbell, 2009), long series of data are collected for individuals. CFA models for such data are presented. Prediction and autoassociation models for shorter series of data follow in the next sections.

878

Configural Frequency Analysis for Research on Developmental Processes

CFA of Differences I In this section, we ask how CFA can be used to analyze trajectories of repeatedly observed variables. Specifically, we ask whether group differences in such trajectories can be ascertained. Longitudinal, that is, intraindividual differences can be ascertained at all scale levels. At the nominal level, differences are described by switching from one category to another. At the ordinal level, longitudinal differences imply changing rank or moving up or down on an ordinal scale such as grade in school, or on a well-being scale. At the interval level, longitudinal differences imply that change can be measured in comparable, commensurate units. For example, an increase in intelligence can be by 5 IQ points. These points indicate the same amount of change, regardless where, on the IQ scale, the change is observed. At the ratio scale level, longitudinal differences can also be assessed in units of proportion. An improvement in processing speed from 4 to 2 seconds can be interpreted as an increase by 100%. In the following discussion of CFA methods for analyzing differences, we focus on differences on variables that are measured at least at the ordinal level. The simplest way of measuring differences involves subtracting the subsequent observation, yi+1 , from the previous, yi , that is, Δy = yi+1 − yi . When the number of ranks or interval-level scores is small, the number of resulting differences will also be small, and the differences can be crossed with other differences or variables right away. When the number is large, researchers often resort to binning, even dichotomizing. The resulting number of cells in a cross-classification will then be much smaller with binning than without binning. In the literature, there has been a discussion concerning the reliability of difference scores. Quite often, one finds the argument that the reliability of difference scores is deplorable. Unfortunately, this argument is not always well grounded. We present two counter arguments why. First, when scales are ratio scale level, differences can be highly reliable. Consider a car that travels at a speed of 50 mph. After braking to come to a standstill, the speed difference is 50 mph. This difference is perfectly reliable. This applies accordingly to other ratio scale measures used in developmental research, for example, blood pressure, alpha waves, reaction time, or number of words recalled, and to some interval-level measures and rankings. Second, Zimmerman (2009) showed that, even when test scores are used, the reliability of differences is not as low and anomalous when correlations among the true scores are used instead of

correlations among the observed scores. Therefore, while we are fully aware of the potential information loss that can come with binning variables (see MacCallum, Zhang, Preacher, & Rucker, 2002; von Eye & Mair, 2011), we believe that difference scores and binned difference scores can be useful. Most popular are first differences, that is, differences between time-adjacent measures, e.g., the difference of the second observation, y2 , from the first, y1 , Δy = y2 − y1 . First differences are used for a number of reasons. The two most important ones are to (1) simply find out about change, and (2) to de-trend series of measures. The second reason is of importance when methods of time series analysis are employed that require the stationarity assumption. The first reason is of importance when individual change scores or change patterns are interpreted. To calculate second differences, one subtracts adjacent first differences, Δ2 y = Δyi+1 − Δyi for i = 1, . . . , m − 1, with m being the number of first differences in the series. The superscript in this equation indicates the order of differences. Second differences are of interest (1) when changes in change, that is, acceleration or deceleration, are of interest, or (2) when the quadratic trend is to be taken out of a series of measures. Higher order differences such as third or fourth differences can be defined in an analogous way. Table 20.4 displays a scheme that shows how differences can be calculated. In Table 20.4, the Y scores are the (repeated) measures to be modeled. The X scores are the points on the time axis used in repeated measures studies. Difference scores are most easily comparatively interpreted when the points on X are equidistant or, at the least, the distances are the same for every case in a study.

TABLE 20.4 Calculation of First Through Fourth Differences Raw scores X

Y

x0

y0

x1

y1

Differences ΔY first differences Δ y0 Δ y1

x2

y2 Δ y2

x3

y3 Δ y3

x4

y4

Δ2 Y second differences

Δ2 y0 Δ2 y

1

Δ2 y2

Δ3 Y third differences

Δ3 y0 Δ3 y1

Δ4 Y fourth differences

Δ4 y0

CFA Models for Developmental Research

Differences possess important characteristics that make them interesting for longitudinal research. 1. Differences can be used to estimate the degree of a polynomial that describes a series of measures. Specifically, a. the n-th differences of a polynomial of the nth degree are constant; this implies that, as soon as a series of differences is constant, the polynomial of corresponding degree will describe the series perfectly; and b. the differences of an order higher than n disappear; in other words, calculating differences of a degree higher than n is pointless. 2. Difference variables can be used to span crossclassifications. As we said before, when the number of differences is small, they can be directly crossed with other differences, with time-varying, and with time-constant (measured only once) variables. a. The method of differences can be used to identify scores that are suspected to be faulty. Scores that are faulty result in the existence of higher order differences when most lower order differences have long converged to very small values, or even disappeared. In addition, the magnitude of the error can be concluded from the magnitude of the differences. To illustrate the method of differences, we first present an example with artificial data (von Eye, 2002), and then a real data example. For the artificial data example, we use the five consecutive x-values of 1 through 5 and select the five y-values {1, 3, 4, 3, 1}. For these scores, we calculate the first, second, and third differences. The three resulting curves of the differences are depicted in Figure 20.2, along with the curve of the raw scores. 5 4 3

Y-Value

2 1 0 Curve

–1

Raw Data 1st Differences 2nd Differences 3rd Differences

–2 –3 –4

0

1

2

3 X

4

Figure 20.2 Method of differences.

5

6

879

Figure 20.2 shows that the raw data can be described as an inversely U-shaped function. Its slope changes from positive to negative (first differences). The change in slope is negative throughout and more pronounced around the middle of the curve (second differences). The third differences are constant, suggesting that the raw data can be perfectly described by using a third-order polynomial. Data Example For the real-data example, we use data that Finkelstein, von Eye, and Preece (1994) collected on the development of aggressive behavior. The authors administered a questionnaire concerning aggressive behavior to 67 adolescent girls and 47 boys at three points in time. The time intervals were 2 years apart. The questionnaire addressed the four dimensions of aggression: Aggressive Impulse, Aggression-Inhibitory Response, Verbal Aggression against Adults, and Physical Aggression against Peers. In addition, physical pubertal development was assessed using Tanner scores. In the following analyses, we use the data on physical aggression against peers (PAAP). Three scores existed per person, PAAP83, PAAP85, and PAAP87, where the numbers indicate the years in which the data were collected. When the data were collected, the respondents were, on average, 11, 13, and 15 years of age, respectively. In the following example, we ask whether the developmental change in self-rated physical aggression against peers is different for boys and girls. We perform the following three steps. 1. Calculate the differences PADIF1 = PAAP85 – PAAP83 and PADIF2 = PAAP87 – PAAP83; 2. Dichotomize the differences; here, we dichotomize the differences at their grand mean; the resulting split will not lead to groups that are exactly the same in size (the distributions of the difference scores have slightly elevated skewness); however, information about the mean is conserved; 3. Perform 2-group CFA with Gender as the grouping variable. Descriptive statistics for the difference scores PADIF1 and PADIF2 appear in Table 20.5. After dichotomization, both variables were coded as 1 = below the mean and 2 = above the mean. Gender was coded as 1 = female and 2 = male. Crossed, the three variables span the 2 × 2 × 2 Gender (G) × PADIF1 (P1) × PADIF2 (P2) cross-classification. We analyze this table using two CFA base models. The first is the two-group base model, the second is the standard base model of first-order CFA.

880

Configural Frequency Analysis for Research on Developmental Processes

TABLE 20.5 Descriptive Statistics for the Variables PADIF1 = PAAP85 − PAAP83 and PADIF2 = PAAP87 − PAAP85 PAAPDIF1 N of Cases Minimum Maximum Arithmetic mean Standard deviation

114 1.000 2.000 1.561 0.498

TABLE 20.6 Two-Group CFA of the Gender (G) × PADIF1 (P1) × PADIF2 (P2) Cross-Classification, with Gender as the Grouping Variable

PAAPDIF2 114 1.000 2.000 1.482 0.502

Similar to the PCFA base model, the two-group model proposes that (1) the model is saturated in the variables used to discriminate between the two groups, and (2) the grouping variable is independent of the discriminator variables. In the present example, the two-group base model is ̂ = 𝜆 + 𝜆G + 𝜆P1 + 𝜆P2 + 𝜆P1,P2 log m This model can be rejected only if the grouping variable, Gender, is associated with the discriminator variables, P1 and P2. Specifically, this model can be rejected if one or more of the following interactions exist: [G, P1], [G, P2], and [G, P1, P2]. This base model has an interesting characteristic. It makes the association between the variables P1 and P2 part of the model. Considering that these two variables describe temporal elements of the development of physical aggression against peers, this association reflects the serial dependency of P1 and P2. Types and antitypes cannot, therefore, emerge just because P1 and P2 show an autoassociation (comparable to autocorrelation). The second base model is that of first-order CFA, ̂ = 𝜆 + 𝜆G + 𝜆P1 + 𝜆P2 log m This model contains only the main effects of the variables that span the cross-classification of G, P1, and P2. Therefore, the autoassociation of P1 and P2 can be among the reasons why this model is rejected (if it is rejected). For the type–antitype decisions under both base models, we use the normal approximation of the binomial test and the Holland–Copenhaver procedure of 𝛼 protection. Table 20.6 displays the 2 × 2 × 2 Gender (G) × PADIF1 (P1) × PADIF2 (P2) Cross-Classification along with the results of two-group CFA. The overall goodness-of-fit likelihood ratio X 2 for the base model of two-group CFA for the data in the table is 12.58. For df = 3, this value suggests that the model is rejected (p = 0.006), and we expect that discrimination types and antitypes may exist. The results in Table 20.6 suggest that the two gender groups differ in two profiles. Significantly more male than

Configuration P1 P2 G

m

z

p

111 112

8.00 16.00

−2.849

.002190

121 122

17.00 9.00

.780

.217801

211 212

19.00 16.00

−.648

.258591

221 222

23.00 6.00

2.602

.004632

Type? Discrimination type

Discrimination type

female respondents display the trajectory of individuals who consistently change to self-perceived ever lower levels of physical aggression against peers. The second discrimination type suggests that significantly more female than male respondents change to self-perceived, ever higher levels of physical aggression against peers. The remaining two comparisons—both concern up-and-down patterns of change—were observed about as often as expected under the base model. The two types also come with significant odds ratios. Specifically, males are 1.337 times more likely than females to show trajectory pattern 1 1 (z = −2.474; p = 0.007), and females are 1.273 times more likely than females to show trajectory pattern 2 2 (z = 2.520; p = 0.006). From a data analytic perspective, it is important to point to the fact that two-group CFA involves only half the number of tests as the corresponding standard CFA. Therefore, the protected 𝛼-level is less restrictive in two-group CFA than in standard one-group CFA. Consider the example in Table 20.6. In this table, four tests were performed instead of eight tests that are needed in standard, first-order CFA. One implication of a less restrictive, protected 𝛼 is that types and antitypes are more likely to be detected. Table 20.7 displays the results from first-order CFA that was performed using the same significance test and 𝛼 protection as for the two-group CFA in Table 20.6. The overall goodness-of-fit likelihood ratio X 2 for the base model of first-order CFA for the data in Table 20.7 is 13.09. For df = 4, this value suggests that the model is rejected (p = 0.011). This result leads to the expectation that types and antitypes may exist. However, as is evident from Table 20.7, none of the configurations deviates significantly from expectation, when the stricter protected 𝛼 is used. In an add-on step of the analysis of the data in Tables 20.6 and 20.7, we ask which log-linear model could

CFA Models for Developmental Research TABLE 20.7 First-Order CFA of the Gender (G) × PADIF1 (P1) × PADIF2 (P2) Cross-Classification Configuration P1 P2 G 111 112 121 122 211 212 221 222

m

̂ m

z

p(z)

8.00 16.00 17.00 9.00 19.00 16.00 23.00 6.00

15.209 10.669 14.11 9.945 19.467 13.656 18.147 12.730

−1.9856 1.7144 .8011 −.3138 −.1162 .6761 1.2424 −2.0013

.023538 .043226 .211539 .376848 .453743 .249482 .107053 .022679

be used to adequately describe the frequency distribution. ̂ = 𝜆 + 𝜆G + 𝜆P1 + 𝜆P2 + 𝜆G,P2 We find that the model log m explains the data very well (LR − X 2 = 4.398; df = 3; p = 0.222). This model suggests that there is an association between Gender and change in Physical Aggression Against Peers between the ages of 13 and 15. This result is plausible and easy to interpret. However, the specifics of the results, in particular, the findings that (1) male adolescents are more likely than female adolescents to exhibit a repeated reduction in self-perceived physical aggression and that (2) repeated increases in aggression are comparatively more likely in female adolescents can most easily be obtained using CFA. In sum, two-group CFA of trajectories shows, in the present example, a typical pattern of CFA results. A selection of configurations allows one to discriminate between the comparison groups. Other configurations are observed as often as expected, in both comparison groups. In addition, two-group CFA of differences has the interesting characteristic that the analyzed cross-classification is much smaller than the cross-classification of the original variables. Consider a study in which an ordinal variable with five ranks is observed three times. Crossing these observations results in a table with 53 = 125 cells. Creating first differences reduces the number of variables from 3 to 2. Binning the differences into three categories such as change toward higher ranks, lower ranks, and no change, and crossing the two resulting change variables results in a table with 3 × 3 = 9 cells, a savings of 92.8%. Naturally, this savings goes hand in hand with loss of information. However, it provides the researcher with the opportunity to capture interesting parts of the original information and to make them the focus of data analysis, or to include additional information such as covariates. In the next section, we discuss a priori probabilities in the analysis of differences (von Eye, 2002; von Eye, Mair, & Mun, 2010; von Eye & Mun, 2007, 2012).

881

CFA of Differences II: Structural Zeros So far in this chapter and, in most applications of CFA, researchers proceed from the assumption that the variables under study can be completely crossed. That is, each category of every variable can be observed under each category of every other variable. When difference variables as the ones discussed in the last section are crossed with covariates or difference variables that are based on other variables, this assumption is practically always met. However, there is one interesting exception. This exception concerns crossing difference variables of different order of the same original variable with each other. Consider, for example, a variable that is observed three times. For the resulting three measures, two measures of first differences can be calculated (Δ11 = Time 2 – Time 1, and Δ12 = Time 3 – Time 2) and one measure of second differences (Δ22 = Δ12 − Δ11 ). Now, let the difference measures be dichotomized at zero (this is also called sign-transformed). The resulting scores indicate whether the differences indicate an upswing or a downswing of differences. Let the original variable also be dichotomized. Crossed, these three indicators span a 2 × 2 × 2 contingency table, with cell indices 0 0 0, 0 0 1, 0 1 0, 0 1 1, 1 0 0, 1 0 1, 1 1 0, and 1 1 1. Of these eight cells, two are logically impossible and therefore, cannot contain any case at all. These are the cells with patterns 0 1 0 and 1 0 1, where the order of indices is original variable—first differences—second differences. The first of these patterns suggests a decrease that is followed by an increase and, overall, a negative quadratic trend. This pattern would reflect a ∪-shaped curve that is, simultaneously, ∩-shaped. The second pattern indicates an initial increase that is followed by a decrease in tandem with an overall positive quadratic trend. This pattern would reflect a ∩-shaped curve that is, simultaneously, ∪-shaped. Neither is logically possible. Because these shapes are contradictory, configurations 0 1 0 and 1 0 1 must be declared structural zeros. Structural zeros are different from observed zeros. When, in a sample, a pattern was not observed but, theoretically, maybe in a larger sample, there could be cases that exhibit this pattern, the zero is called a sampling zero. If, however, it is impossible to observe a particular pattern, mostly for logical and design reasons, the zero is structural. These two types of zeros are treated differently when it comes to estimating expected cell frequencies. Sampling zeros are treated just as the other frequencies. An expected value is estimated and the zero is compared with the estimate (exceptions exist when there are so many

882

Configural Frequency Analysis for Research on Developmental Processes

zeros that parameter estimation is impeded). In contrast, cells that contain structural zeros must be blanked out. For these cells, no expected frequency is estimated, and there is no comparison of observed with expected frequencies. Von Eye and Mun (2007) presented the following definition of impossible patterns in the analysis of categorized differences: Combinations of sign patterns are impossible if lower order and higher order signs suggest different orientations of a series of measures. This applies to patterns of difference signs of any order. The authors proposed a procedure with which all candidate patterns for contradictory descriptions of orientation can be identified. Candidates for a search are all patterns that indicate a change in orientation from up to down and vice versa. Patterns that show no change do not need to be considered. Now, let Δ1j and Δ1j+1 be the first differences between the adjacent measure pairs xj+1 and xj , and xj+2 and xj+1 , respectively, for j = 1, . . . , I − 1, and Δ2k the corresponding second difference, that is, Δ2k+1 − Δ2k . Let sgn Δ1j , sgn Δ1j+1 , and sgn Δ2k be the signs of these differences. The following two steps are performed to identify the structural zeros. Step 1: Compare each pair of sgn Δ1j and sgn Δ1j+1 with the corresponding sgn Δ2k . If both sgn Δ1j ≠ sgn Δ1j+1 and sgn Δ1j+1 ≠ sgn Δ2k , then the pattern is impossible. For example, if the sign pattern + − at a given difference level is combined with sign + at the next higher level, the sign pattern is impossible. The first differences would indicate a negative orientation, whereas the second difference would indicate a positive orientation. Accordingly, the sign pattern − + at a given level cannot be combined with the sign − at the next higher level. If a sign pair is classified as impossible, proceed to the next sign pattern. If a sign pattern is impossible, all other patterns that could be combined with the impossible pattern are impossible also. Step 2: Proceed to the next pair of signs, and start over, at Step 1. Continue until all patterns of signs are examined. This procedure can be applied to sign patterns of all levels. If a researcher decides to skip levels, the procedure can be adjusted. For example, pattern + − at level k cannot go with the sign – at level k + 2. To give an example of a possible pattern, let k indicate first differences and let k + 2 indicate third differences. In this situation, consider the sign pattern + − + for first differences and + for third differences. The first signs indicate an initial increase that is followed by a decrease, and then an increase. The third difference indicates an increase in acceleration. To give an example of an impossible pattern, consider the series 9, 15, 10, 6. For this sequence, the first differences sign pattern is + − −, the second differences sign

pattern is − +, and the corresponding third difference has the sign +, thus indicating a ∩-shaped trend. For this series, a negative sign for the third difference is impossible. A complete example with an artificial data set is given by von Eye and Mun (2007). A real-data example follows. Data Example In the data example in the following paragraphs, we use the data on adolescent development of aggression again that we used previously for the example in Table 20.6. For the earlier example, we had analyzed the first differences of Physical Aggression Against Peers. Here, we analyze first and second differences. Specifically, we create a 2 × 2 × 2 table by crossing PADIF1 = PAAP85 − PAAP83 (abbreviated by P1), PADIF2 = PAAP87 − PAAP85 (abbreviated by P2), and P2DIF = PADIF2 − PADIF1 (abbreviated by ΔP). The resulting cross-classification contains two structural zeros. The first is in cell 1 2 2. It only makes sense that nobody shows this pattern. The pattern would describe a ∪-shaped curve that is, simultaneously, ∩-shaped. The second structural zero is found in cell 2 1 2. This pattern describes a ∩-shaped curve that is, simultaneously, ∪-shaped. Neither is logically possible. The 2 × 2 × 2 P1 × P2 × ΔP cross-classification is analyzed using two CFA base models. The first is the log-linear main effect model of variable independence, ̂ = 𝜆 + 𝜆P1 + 𝜆P2 + 𝜆ΔP log m In this model, no provision is taken for the structural zeros. Therefore, the expected frequencies for Cells 1 2 2 and 2 1 2 will be greater than zero. The second base model declares these two cells structural zeros. The model is ̂ = 𝜆 + 𝜆P1 + 𝜆P2 + 𝜆ΔP + 𝜆s1 + 𝜆s2 , log m where the s1 and the s2 in the superscript represent the first and the second vectors that indicate structural zeros. Under this base model, the expected frequencies for Cells 1 2 2 and 2 1 2 will be zero. Table 20.8 displays the results of the first-order CFA under the first base model. Table 20.9 displays the results of first-order CFA under the second base model. For both CFA runs, we use the z test and protect 𝛼 using the Holland–Copenhaver procedure. The overall goodness-of-fit likelihood ratio X 2 for the base model of first-order CFA for the data in Table 20.8 is 85.72. For df = 4, this value suggests that the model is rejected (p < 0.01), and we anticipate that types and antitypes emerge. The table shows two types and two

CFA Models for Developmental Research TABLE 20.8 First-Order CFA of the P1 × P2 × ΔP Cross-Classification; Structural Zeros Not Taken into Account Configuration P1 P2 ΔP 111 112 121 122 211 212 221 222

m

̂ m

z

p(z)

10.00y 14.00 .00 26.00 35.00 .00 17.00 12.00

14.074 11.804 13.119 11.003 18.014 15.109 16.793 14.084

−1.0859 .6393 −3.6221 4.5210 4.0020 −3.8870 .0505 −.5554

.138771 .261317 .000146 .000003 .000031 .000051 .479843 .289314

Type/antitype?

Antitype Type Type Antitype

TABLE 20.9 First-Order CFA of the P1 × P2 × ΔP Cross-Classification; Structural Zeros Taken into Account Configuration P1 P2 ΔP 111 112 121 122 211 212 221 222

m

̂ m

z

p(z)

10.00 14.00 .00 26.00 35.00 .00 17.00 12.00

17.821 17.354 .000 14.825 23.825 .000 20.354 19.821

−1.8526 −.8051 – 2.9022 2.2894 – −.7434 −1.7567

.031969 .210380 – .001853 .011029 – .228617 .039488

Type/antitype?

Type

antitypes. The first type is constituted by cell 1 2 2. This pattern describes a curve that shows an initial decrease that is followed by an increase, with positive acceleration (positive quadratic trend). 11 cases were expected to exhibit this pattern, but 26 did exhibit it, a significant discrepancy. The second type is constituted by cell 2 1 1. It shows just the opposite pattern. Instead of the expected 18, 35 adolescents showed an initial increase in physical aggression that was followed by a decrease, with negative acceleration. For the present purposes, the two antitypes are most interesting. The first, constituted by cell 1 2 1, indicates that not a single adolescent showed a decrease that is followed by an increase which, at the same time, indicates a negative quadratic trend. The second, constituted by cell 2 1 2 indicates that no one showed an initial increase, followed by a decrease together with an overall positive quadratic trend. It only makes sense that nobody showed these patterns. The first antitype would be a ∪-shaped curve that is, simultaneously, ∩-shaped. The second antitype would be a ∩-shaped curve that is, simultaneously, ∪-shaped. Neither is logically possible. Because of this pattern of contradictory shapes, the Cells 1 2 1 and 2 1 2 must be declared structural zeros. The second CFA base model considered here takes the characteristic of Cells 1 2 1 and 2 1 2 as structural zeros into account. Results appear in Table 20.9.

883

The overall goodness-of-fit likelihood ratio X 2 for the base model of first-order CFA for the data in Table 20.9 is 20.40. For df = 2, this value suggests that the model is rejected (p < 0.01). This model is significantly better than the one that does not take the structural zeros into account (ΔX 2 = 65.32; Δdf = 2; p < 0.01), but it cannot stand by itself. Therefore, we anticipate that types and antitypes emerge. The table shows that only one of the two types from Table 20.8 reemerges. It is constituted by cell 1 2 2. We conclude that, in the present example, taking into account the structural zeros does not lead to a fitting model that is more parsimonious than the saturated model. Each of the structural zeros reduces the degrees of freedom by one. In the present example, there is one degree of freedom left to invest before we reach the saturated model (which is not more parsimonious than the data themselves). Including the interaction between the two first difference-variables results in a well-fitting model (likelihood ratio X 2 = 1.517; df = 1; p = 0.218), and the interaction parameter is significant. However, as before, the details of the interaction can be identified and interpreted in particular with CFA methods. Extensions Two extensions of the method of differences are particularly of note. The first concerns possible a priori probabilities. As shown by von Eye (2002), differences of any order come with a priori probabilities. Consider a series of three ordinal-level scores that describe the development of depression, with 1 indicating no depression, 2 indicating elevated depression, and 3 indicating clinical depression. For three observation points, the six sequences 1 2 3, 1 3 2, 2 1 3, 2 3 1, 3 1 2, and 3 2 1 are possible. These sequences are equally likely. Calculating, however, the first differences for the six sequences, one obtains the sign patterns + +, + −, − +, + −, − +, and − −. These patterns are not equally likely. Specifically, the probabilities are p(++) = 0.167, p(+ −) = 0.333, p(−+) = 0.333, and p(−−) = 0.167. Similarly, if differences are scored so that no change can be part of the resulting pattern, the possible patterns come with a priori probabilities that are not uniform. Uniformly distributed difference patterns can result only for continuous variables. This applies accordingly to higher order differences. Estimating expected frequencies without taking a priori probabilities into account can cause severely biased estimates for type and antitype patterns or, in log-linear modeling, severely biased parameter estimates. A second extension of CFA or log-linear models for difference scores concerns the inclusion of additional

884

Configural Frequency Analysis for Research on Developmental Processes

TABLE 20.10 First-Order CFA of the Gender × P1 × P2 × ΔP Cross-Classification; Structural Zeros Taken into Account Configuration G P1 P2 ΔP 1111 1112 1121 1122 1211 1212 1221 1222 2111 2112 2121 2122 2211 2212 2221 2222

m

̂ m

z

p(z)

4.00 4.00 .00 17.00 19.00 .00 14.00 9.00 6.00 10.00 .00 9.00 16.00 .00 3.00 3.00

10.474 10.199 – 8.713 14.003 – 11.962 11.649 7.347 7.155 – 6.112 9.823 – 8.392 8.172

−2.0003 −1.9411 – 2.8074 1.3355 – .5891 −.7761 −.4970 1.0637 – 1.1681 1.9710 – −1.8612 −1.8092

.022733 .026121 – .002497 .090860 – .277886 .218833 .309595 .143723 – .121391 .024364 – .031359 .035213

Type/antitype?

Type

variables or covariates. For example, given the results in Tables 20.6 and 20.9, one can ask whether the type/antitype pattern found in Table 20.9 is gender specific. Table 20.10 displays the results for this question, with structural zeros taken into account. To make these results comparable with the ones in Table 20.9, the z test was used again, along with the Holland–Copenhaver procedure for 𝛼 protection. The overall goodness-of-fit likelihood ratio X 2 for the base model of first-order CFA for the data in Table 20.10 is 33.55. For df = 7, this value suggests that the model is rejected (p < 0.01). Only one type results. It is constituted by Pattern 1 1 2 2, suggesting that females are significantly more likely than expected to initially show a decrease in aggression, that is followed by an increase, with an overall positively accelerated trend. For male adolescents, the same pattern was observed at a rate that did not differ from expectancy. Candidates for covariates include time-invariant variables such as gender, once-observed variables such as living conditions, and repeatedly observed variables such as diagnoses, or physiological measures of development. CFA of Trajectories: Typical and Atypical Development In person-oriented (Bergman & Magnusson, 1997; von Eye & Bergman, 2003; von Eye, Bergman, & Hsieh, 2015; Ch. 18) and idiographic research (Molenaar, 2004; Molenaar & Campbell, 2009), it has been shown that, to be able to replace individual series of repeated observations with cross sectional data from many individuals and to draw inference conclusions as if data came from repeated

observations, the data must be stationary. This condition is rarely reasonable and fulfilled in developmental processes such as learning or in temporal etiological processes. Therefore, it has been recommended that data not be aggregated across individuals. Instead, parameters are to be estimated first for longitudinal data, at the level of the individual. Aggregation can then be performed based on parameter profiles. In this section, we present a CFA method that does exactly this, in three steps, for interval-level or ratio scalelevel variables. In a first step, polynomials are estimated for each individual. These polynomials describe the trajectory of development. At least two consecutive interval- or ratio scale-level measures are needed. The method is more fruitful, however, when a series consists of three or more measures. In its second step, the polynomial parameters are binned, that is, reduced to a smaller number of binned parameters that are then crossed for CFA or log-linear analysis. In the third step, the parameter ranges are crossed with each other or with other variables and then subjected to CFA (von Eye, 2002; von Eye & Nesselroade, 1992). There are many forms of polynomials (Abramowitz & Stegun, 1972). Using an approach that is easily related to regression analysis, a series of scores yj can be approximated by the polynomial ̂ y = f (x) = PI (x) = b0 x0 + b1 x1 + b2 x2 + . . . + bI xI I ∑ = bi xi i=0

where the x are either the scores of a predictor variable that is observed at the same observation points as y, or scores that describe the observation points (1, 2, . . . ). The number I is called the degree of the polynomial or the order of the polynomial. In statistical data analysis, polynomials are well known and often used. For example, a standard regression analysis which fits a straight line to the data uses P1(x), that is, a first-order polynomial. Quadratic regression slopes are also often considered. In repeated measures analysis of variance with polynomial contrasts, polynomials up to order I can be estimated for I + 1 repeated measures. Polynomials are attractive tools for the description of series of measures, for a number of reasons. 1. Any series of scores can be perfectly described using a polynomial. Note, however, that estimating a polynomial that goes exactly through the observed data points is not advisable, because practically all observed

CFA Models for Developmental Research

data points come with measurement error and the polynomial parameter estimates must not be affected by this error. However, this characteristic of polynomials gives researchers the option of adjusting the degree of the estimated polynomial as needed. Typically, the degree of the estimated polynomial is much below the number of observation points. 2. The coefficients of orthogonal polynomials do not need to be calculated (in particular when the points on x are equidistant). They are either part of software already, or they can be found on the Internet (e.g., www .watpon.com/table/polynomial.pdf; for detail on the estimation of polynomial coefficients, see Abramowitz & Stegun, 1972). 3. When orthogonal polynomials are used, the polynomial parameters are independent of each other. Therefore, polynomial elements of any degree can be removed from an equation without affecting other parameters. The remaining parameters do not need to be reestimated. 4. Polynomial approximation of series of scores is attractive in particular when the series are relatively short, as is the case in most social and behavioral science applications. Among the main reasons for this are that (1) a polynomial can be devised so that it approximates the observed series as closely as desired, and that (2) polynomial parameters have the same interpretation as standard regression parameters. For example, the intercepts have the same interpretation and the parameter for the first-order polynomial is equivalent to the standard regression slope. On the downside, we find that 1. Although any series of scores can perfectly be approximated with polynomials, certain patterns are hard to approximate in the sense that polynomials of high orders may be needed for a satisfactory description. Among those patterns are cyclical series of scores (orthogonal trigonometric polynomials can be useful here) and series that approach an asymptote. 2. In many instances, the estimated polynomials cannot be used for purposes of extrapolation. Any polynomial approximates infinity, on both ends. Therefore, polynomials are valid only for the series used for estimation. In most applications, researchers use families of orthogonal polynomials instead of standard polynomials. In these

885

families, the element of order I is a polynomial of order I itself. Specifically, f (x) = a0 𝜙0 (x) + a1 𝜙1 (x) + a2 𝜙2 (x) + . . . + aI 𝜙I (x) =

I ∑ ai 𝜙i (x) i=0

where 𝜙i (x) is an i-th order polynomial. The coefficients ai are the estimated weights. They are comparable to the weights estimated in regression analysis (and can be estimated using least squares methods). This system of polynomials is orthogonal over the interval a < x < b, with respect to the weight function w(x) if b

∫a

w(x)Pn (x)Pm (x)dx = 0

where Pn (x) and Pm (x) are polynomials of degree n and m, respectively, and n ≠ m. At the level of individual polynomials, the two polynomials 𝜙i (x) and 𝜙j (x) are orthogonal to each other if n ∑ 𝜙i (xk )𝜙j (xk ) = 0 k=1

where n is the number of points on x, and i ≠ j. Data Example In the following example, we reanalyze data published by von Eye and Nesselroade (1992; von Eye, 2002). 42 college students were presented a state anxiety questionnaire four times (Nesselroade, Pruchno, & Jacobs, 1986). Each time, all respondents completed both parallel forms, A and B, of the questionnaire. This way, two series of four repeated measures resulted for each student. Von Eye and Nesselroade approximated each of these series using the second order orthogonal polynomial y = y + b1 t + b2 t2 where y is the arithmetic mean of the scores in the series, and t counts the equidistant observation points (t = 1, 2, 3, 4). For each student, one set of parameters was estimated for each of the two series. For example, the first student’s parameter estimates were y = 37.8, b1 = 1.0, and b2 = −0.8 for parallel form A, and y = 38.3, b1 = 0.95, and b2 = −0.6 for parallel form B. These estimates suggest that this student’s scores increase with time (linear trend, slope, b1 ). However, the negative quadratic trend b2 (also called

886

Configural Frequency Analysis for Research on Developmental Processes

curvature) suggests that toward the end this student’s scores increase less rapidly or even become lower, countering the linear trend. This applies to both parallel forms of the instrument. For the present re-analyses, we use the mean and the slope. The means were dichotomized at their medians, with 1 indicating above the median and 2 indicating below the median. The estimates for the slopes, b1 , were dichotomized at zero. Thus, 1 indicates a positive trend and 2 indicates a negative trend. Positive trends indicate increasing state anxiety, negative trends suggest decreasing state anxiety. Thus, zero is a natural cutoff point. The four variables A0 , A1 , B0 , and B1 resulted from dichotomizing the parameter estimates. A0 and B0 represent the means of the two parallel test forms, and A1 and B1 represent the two linear slopes. Crossed, these four variables form the 2 × 2 × 2 × 2 A0 × A1 × B0 × B1 contingency table. We use first-order CFA for the analysis of this cross-classification (Table 20.11). Because of the relatively small sample size, we use the binomial test and protect 𝛼 using the Holland–Copenhaver method. Table 20.11 displays the results. For the base model, we obtain an overall goodness-of-fit likelihood-ratio X 2 = 57.56. For df = 11, this value suggests that the base model must be rejected (p < 0.01), and we have reason to expect types and antitypes to emerge. First-order CFA suggests the existence of two types (note that von Eye, 2002, used Anscombe’s z test instead of the binomial test, and found an antitype in addition; it was constituted by cell 2 2 1 1). The types suggest that the parallel forms of the test perform as expected. The first

TABLE 20.11 First-order CFA of Means and Linear Trends of Two State Anxiety Questionnaire Parallel Forms Configuration A0 A1 B 0 B 1 1111 1112 1121 1122 1211 1212 1221 1222 2111 2112 2121 2122 2211 2212 2221 2222

m

̂ m

p(z)

10.00 .00 1.00 .00 5.00 4.00 .00 1.00 2.00 1.00 5.00 .00 .00 1.00 2.00 10.00

3.0967 2.1057 2.5581 1.7395 3.7486 2.5490 3.0967 2.1057 3.0967 2.1057 2.5581 1.7395 3.7486 2.5490 3.0967 2.1057

.00077668 .11528525 .26593006 .16921897 .31969519 .24977930 .04008495 .37085783 .39276833 .37085783 .11035198 .16921897 .01971179 .26776931 .39276833 .00003352

Type/antitype? Type

type, constituted by configuration 1 1 1 1, describes those students who scored above the median in both parallel test forms, and showed a linear increase in state anxiety over the four measurement occasions, also in both parallel test forms. Instead of the expected about 3 respondents, 10 displayed this pattern. The second type, constituted by configuration 2 2 2 2, contains individuals who scored below the median in both parallel test forms and exhibited a trend toward lower scores over the four occasions, in both parallel test forms. Only about 2 individuals had been expected to show this pattern, but 10 did show it. The two configurations 1 1 1 1 and 2 2 2 2, thus, describe typical (or prevailing) developmental patterns. In this example, there are no atypical developmental patterns, which would be indicated by antitypes. However, readers are encouraged to test whether patterns that indicate that the two parallel forms result in diverging statements constitute a composite antitype. As always, after performing a CFA, we ask whether there is a log-linear model that explains the frequency distribution in the cross-classification in Table 20.11. The model [A0 ], [A1 ], [B0 ], [B1 ], [A0 , B0 ], [A1 , B1 ] describes the data well (LR − X 2 = 12.79; df = 11; p = 0.172). The two interactions suggest that the means as well as the linear slopes are associated with the other. This model supports the interpretation that the two forms of the state anxiety instrument are parallel. CFA suggests that this result is carried in particular by patterns 1 1 1 1 and 2 2 2 2. Extensions of the method of polynomial configural analysis are straightforward. For example, one can ask whether groups of individuals differ in characteristics of their trajectories; one can ask whether covariates allow one to come to more differentiated conclusions; one can ask whether once-observed variables play a role; or one can ask whether characteristics of trajectories can be predicted, at the level of configurations, from other variables such as age or gender. Later in this chapter, we discuss CFA methods to predict trajectories. Analyzing Series That Differ in Length

Type

Thus far, the series of scores that we subjected to CFA shared the characteristic that they were all equal in length. However, there are many situations in which series differ in length. Apart from missing data which are dealt with using different tools, series can differ in length, for example, because clients need different numbers of therapy sessions, learners need different numbers of trials before they reach a criterion, workers are employed for different numbers of years before they retire, chess players need different

CFA Models for Developmental Research

numbers of moves before a problem is solved, or students take different numbers of semesters before they complete their degree. In all these and other cases, estimating and imputing missing data makes no sense, and neither does using such methods as full information maximum likelihood (FIML) estimation. These and other methods of missing data handling require the assumption that missing data could have been available to the analyst. In the sample cases, however, data do not and cannot exist because the process has come to an end (see the distinction between zeros, missing data, and nonexisting data; von Eye, 1989). Different methods are, therefore, needed to deal with the situation in which series of scores differ in length. Lienert and von Eye (1986) proposed methods to characterize the characteristics of series of different length so that they can be analyzed with CFA. The principle behind these methods is that every feature of a series can be scaled, including length itself. For example, consider a sample of series of measures, the shortest of which includes k scores. For this series, polynomials can be estimated up to an order of I = k − 1. Now, because all other series include k scores or more, polynomials of order I = k − 1 can also be estimated for these series. The series can, therefore, be compared based on these polynomials. Technically, this procedure is the same as the one described in the last section, and will not be repeated here. Similarly, when longer series are modeled using trigonometric functions, as is done in spectral analysis, spectral coefficients can be used for comparisons for the wave lengths that are applicable to the shortest of the comparison series. CFA methods for the analysis of coefficients from spectral analysis have not been discussed, so far. This possibility should be pursued in more detail in future work. Here, we review and exemplify the methods proposed by Lienert and von Eye (1986). The authors proposed describing series of ordinal or higher level scores, regardless of length, with respect to the following three characteristics: 1. Trend. Series can be coded based on whether trends exist. The original method proposed by Lienert and von Eye (1986) focused on linear trends. However, depending on scale level and length of the shortest series, not only change in the linear trend (quadratic, second order polynomial), but also change in curvature (cubic, third order polynomial) and higher order components can be considered as well. Naturally, the intercept or mean can also be made part of the CFA analysis. A special case is what is called the monotonic trend of a series of length k. For ordinal variables, trends

887

in changes in ordinal scores or ranks can be considered. A series is said to exhibit a monotonic trend if the inequality yi+1 ≥ yi holds for every i = 1, . . . , k − 1. This characteristic is known as weak monotonicity. This applies accordingly for yi+1 ≤ yi . If this inequality is violated just once, the series is said to not display a monotonic trend. Monotonicity is interesting because the trend it describes is not necessarily linear. Suppose the differences between time-adjacent measures fulfill the inequality but decrease over time, the curve approximates an asymptote but still is monotonic. Similarly, if the differences increase systematically with x, the curve is monotonic but can look similar to a quadratic curve. Finally, if the differences vary unsystematically, the curve zigzags its way to higher scores. The minimum number of scores needed for this criterion to be applicable is k = 2. 2. Early completion. In many studies, a series terminates when a particular criterion is reached. Examples of such criteria include recall rates in learning experiments and behavior thresholds in psychotherapy. Here, at least two options can be considered. First, one can specify a threshold and code a series asking whether the threshold was reached or not. Second, one can create an ordinal variable that provides a binning of the lengths of the series. The minimum number of scores needed for the early completion criterion to be applicable is 1 (e.g., a participant in a learning experiment reaches a 100% recall rate after the first trial). 3. Qualitative characteristics. In addition to the two criteria just discussed, any other characteristic of a curve can be considered. Examples of such characteristics include the number of errors in a problem solving task, the type of therapy used to cure a patient, the lengths of the intervals between therapy sessions, or the number of comorbid conditions that improved over the course of a treatment. For use in CFA, the scores that describe these and other characteristics are binned, if needed, and the resulting categorical variables are crossed. The cross-classification from this procedure can then be analyzed using, for example, log-linear modeling, latent class analysis, or CFA. Data Example For the following example, we use data from the MIS (Bogat et al., 2004; Huth-Bocks et al., 2002; see first data example). The study assessed a sample of 206 women in their last trimester of pregnancy and at 2 months postpartum, and the women and their children every year until

888

Configural Frequency Analysis for Research on Developmental Processes

the children were 14 years old. The women in the study reported a range of intimate partner violence experiences, from none to severe. For the following data example, we use the variables Depression (D), Anxiety (A), and PTSD (P). We use data from the first three observation points, that is, the first three years of observation, and ask whether trends as measured in terms of weak monotonicity of the three variables are related to each other. In other words, we ask whether the three variables develop in unison. Most important from the perspective of analyzing series that differ in length is that, after three years, a portion of the original participants had left the study already. Reasons for dropping out include death of child (three instances), loss of custody, and moving away. Still, we include data from the entire sample of 206. For each of the participants, weak monotonicity could be evaluated, based on the procedure outlined before. The three variables were coded as 1 if the criterion of weak monotonicity for a decline in scores was fulfilled and as 2 if the criterion was not fulfilled. In other words, the respondents were given a 1 if, for both comparisons of time-adjacent scores, the later score was below the earlier score. For respondents who dropped out after the second wave of data collection, only one comparison was used to determine the coding. This was done for all three variables. Using the coded complete data, we estimate two CFA base models. For both models, we use the normal approximation of the binomial test and the Holland–Copenhaver procedure for 𝛼 protection. The first base model is that of standard first-order CFA of variable independence. The second is that of two-group CFA. The first model is ̂ = 𝜆 + 𝜆 D + 𝜆A + 𝜆 P log m

TABLE 20.12 First-order CFA of the Cross-Classification of Monotonic Trend Information of Three Data Waves of Depression (D), Anxiety (A), and PTSD (P) Configuration DAP 111 112 121 122 211 212 221 222

m

̂ m

z

p(z)

4.01 5.01 4.01 33.01 .01 11.01 15.01 135.01

.494 3.951 4.617 36.938 1.728 13.827 16.160 129.284

5.0096 .5382 −.2858 −.7131 −1.3126 −.7843 −.2981 .8219

.000000 .295236 .387507 .237895 .094664 .216445 .382828 .205570

Type/antitype? Type

anticipate that types and antitypes may emerge. Table 20.12 displays results from first-order CFA. First-order CFA shows an interesting result. Only one type resulted. This type describes the women who show a decline in the scores of all three variables. Only four, but still significantly more women than expected showed this pattern. None of the other patterns deviates significantly from expectancy. The two-group CFA base model was estimated under the assumption that women who show monotonic decrease in PTSD may differ in levels of Depression and Anxiety from those women who do not show a monotonic decrease in PTSD. The base model for this two-group CFA can be rejected only if one or more of the following interactions exist: [D, P], [A, P], and [D, A, P]. Each of these interactions includes variable P, that is, PTSD. For this base model, we obtain the overall goodness-of-fit likelihood-ratio of X 2 = 9.178. For df = 3, this value leads to the rejection of the base model (p = 0.027), and we anticipate that discrimination types may emerge. Table 20.13 displays results from two-group CFA.

and the second model is ̂ =𝜆+𝜆 +𝜆 +𝜆 +𝜆 log m D

A

P

D,A

The second model was estimated to determine whether respondents with a monotonic decrease in PTSD symptoms differ from those with a nonmonotonic development of PTSD symptoms in the development of Depression and Anxiety. To prevent estimation problems that the zero frequency in cell 2 1 1 could cause for the second of these models, we invoke the delta option for both models with Δ = 0.01. For the first model, we calculate an overall goodness-offit likelihood-ratio of X 2 = 14.99. For df = 4, this value leads to the rejection of the base model (p < 0.01) and we

TABLE 20.13 Two-Group CFA of the Cross-Classification of Monotonic Trend Information of Three Data Waves of Depression (D), Anxiety (A), and PTSD (P), with PTSD as the Grouping Variable Configuration DAP

m

statistic

p

111 112

4.01 5.01

3.255

.000567

121 122

4.01 33.01

−.063

.474961

211 212

.01 11.01

−1.197

.115600

221 222

15.01 135.01

−.832

.202799

Type? Discrimination Type

CFA Models for Developmental Research

Two-group CFA presents one discrimination type. This type results from the comparison of cell 1 1 1 with cell 1 1 2. It suggests that significantly more women whose PTSD symptom scores systematically decrease show Pattern 1 1 than women whose PTSD scores do not decrease. In other words, women whose PTSD symptom scores show a monotonic decrease are significantly more likely to also show monotonic decreases in their depression and anxiety symptom scores than women whose PTSD symptom scores do not show a monotonic decrease. On a methodological note, this result may sound implausible, when one considers that Pattern 1 1 2 is exhibited by five women but Pattern 1 1 1 only by four. However, Tables 20.12 and 20.13 also show that, in general, far more women fail to fulfill the weak monotonicity criterion than fulfilling it. From this perspective, the number five is far fewer than one would expect based on the ratios found for the other three comparison patterns. A similar result is found when the odds ratio is calculated for cells 1 1 1 and 1 1 2 compared with all other cells combined. This odds ratio is 7.54 (log odds ratio = 2.019). This value indicates that the rate of women with Pattern 1 1 1 is more than twice as often found than women with Pattern 1 1 2 when the aggregate of all other patterns is used for a reference (p < 0.01). Extensions The approach to analyzing series that differ in length can be extended in several ways. First, one can include once-measured variables and time-invariant measures in the base model. Examples of such measures are grouping variables or information that remains unchanged such as the year of birth of respondents. Second, other characteristics of the series under study can be taken into account. For example, one can ask whether the score of a variable at the first occasion is predictive of trends, completion, or qualitative characteristics of a series. Later in this chapter, the predictive value of a score at the beginning of a series will be made part of a CFA prediction model. CFA of Pre–Post Designs—An Application of Confirmatory CFA In evaluation studies and in the study of effects of experiments or interventions, researchers typically follow a priori planned experimental designs. Hypotheses compatible with these designs are subjected to statistical testing, and detailed hypotheses are pursued, for example, by way of planned contrasts or between-group post hoc mean comparisons. In many cases, the dependent variables in such research

889

are metrical, that is, measured at the interval or ratio scale level. In other cases, however, the dependent variables are frequency variables such as the number of correctly reproduced syllables or the number of times a person diagnosed with obsessive-compulsive disorder washes his or her hands. Both the number of syllables a person is able to reproduce and the number of hand washes may depend on intervention procedures. In these and similar cases, CFA is the method of choice for data analysis. In the examples of CFA application used thus far, the method was employed in an exploratory context. When effects of experiments or intervention are examined, precise hypotheses often exist. In this situation, confirmatory application of CFA is in order. In confirmatory CFA, a priori specified hypotheses are tested. In other words, in confirmatory CFA, researchers determine before data collection which configurations are subjected to CFA tests. The number of these configurations is smaller than the number examined in exploratory CFA. Therefore, the procedures implemented to protect 𝛼 need to take into account only this number instead of the total number of cells in the cross-classification. This difference in the number of tests has an important implication for the testing procedure in CFA: the significance threshold 𝛼* will be less extreme. Let t′ be the number targeted in confirmatory CFA and t the total number of cells, with t′ < t. Then, we obtain, in confirmatory CFA, for the Bonferroni-protected α′∗ = α∕t′ and, in exploratory CFA, the Bonferroni-protected α∗ = α∕t, with α′∗ > α∗ . This indicates that investing information in the specification and selection of hypotheses results in a less extreme significance threshold. This effect is known from the differences in significance thresholds between one-sided and two-sided testing, where investing information in the derivation of one-sided tests also results in less extreme significance thresholds. Consider, for example, a cross-classification with 27 cells. The Bonferroni-protected α∗ for this table is, in exploratory application of CFA, α∗ = 0.05∕27 = 0.0019. If, in confirmatory research, only five cells are examined, the protected α′∗ is α′∗ = 0.05∕5 = 0.01. In this example, the difference between the protected 𝛼 in exploratory and in confirmatory research is a factor of 5.4. That is, the protected 𝛼 is, in exploratory research, 5.4 times more extreme than in confirmatory research. Types and antitypes are, therefore, more likely to emerge in confirmatory research than in exploratory research. The rest of the testing procedure is unchanged from standard, exploratory CFA testing. The cells that were selected for testing are either tested using a constant significance threshold (as is the case in the Bonferroni procedure)

890

Configural Frequency Analysis for Research on Developmental Processes

or the test statistics of these cells are ranked in descending order and tested until either the first discrepancy between the observed and the expected cell frequency fails to be significant or the number of a priori-specified cells is exhausted (Holm or Holland–Copenhaver procedures). Treatment Effects in Designs with No Control Group In the following sections, we present two data examples (Lienert & Straube, 1980; von Eye, 2002). The first example involves the examination of effects of neuroleptic drugs on inpatients diagnosed with schizophrenia, with no control group. The drugs were administered in a sample of 75 patients over a period of 2 weeks. At the beginning and end of this 2-week period, the patients were also administered the Brief Psychiatric Rating Scale (Overall & Gorham, 1962). Three of the 17 symptoms captured by this instrument are used in the following analyses: W = emotional withdrawal; T = thought disturbances; and H = hallucinations. Each of the symptoms was scaled as either present (+) or absent (−). Table 20.14 displays the data, along with the CFA results. We analyze these data in three different ways. The first is first-order, confirmatory CFA. For this CFA, we use the z test and the Holland–Copenhaver procedure for 𝛼 protection. The data in Table 20.14 are arranged in a somewhat unusual fashion. Instead of presenting all possible combinations of before–after treatment symptoms, we created an ordinal variable with the categories 3 symptoms observed (indicated by + + +), two symptoms observed (indicated by + + −, + − +, and − + +), one symptom observed (indicated by + − −, − + −, and − − +), and no symptom TABLE 20.14 Evaluation of Treatment of Schizophrenics with Neuroleptic Drugs in a Pre–Post Study Number of Configurations Number of symptoms symptoms after treatment before (expected number in italics) treatment WTH

3

+++

2

++− +−+ −++ +−− −+− −−+ −−−

1

0 Totals

3 2 1 +++ ++− +−− +−+ −+− −++ −−+

Totals

observed (− − −). This variable shows that, in this data example, we are interested in the number of symptoms instead of the individual symptoms. The overall goodness-of-fit likelihood-ratio for the base model of variable independence is X 2 = 17.79. For df = 9, this value leads to the rejection of the base model (p = 0.038), and we anticipate that CFA discrimination types may emerge. However, the tail probability, for none of the 16 tests, exceeds the threshold dictated by this exploratory analysis. Under standard conditions of exploratory application of CFA we would, therefore, conclude that the first and the second observations are independent of each other and that the neuoleptic drugs have no systematic effect. Having a hard time believing this, Lienert and Straube (1980) proposed using two additional methods of analysis for data as the ones presented in Table 20.14. The first involves using the diagonal-half sign test. This test can be used to compare the aggregate of the frequencies above the diagonal with the aggregate of the frequencies below the diagonal. For this test, the frequencies in the cells above the main diagonal (from upper left to lower right) in Table 20.14 are summed, and so are the frequencies below the main diagonal. Let the first sum be called b, and the second sum w. The patients in the upper triangle are those who moved, under the drug, from more to fewer symptoms. The patients in the lower triangle are those who moved from fewer to more symptoms. The diagonal-half sign test for the comparison of b with w is b−w z= √ b+w The test statistic z is approximately normally distributed. As an alternative, for small samples, the binomial test can be considered, which is h ∑ (N)pj qN−j B(h) = j=0

0 −−−

1 1.60

10 5.20

4 6.00

0 2.20

15

6 4.05

11 13.17

17 15.20

4 5.57

38

1 1.71 0 0.64 8

4 5.55 1 2.08 26

7 6.40 2 2.40 30

4 2.35 3 0.88 11

16 6 75

where, in the present case, p = q = 0.5, N = b + w, and h is the lesser of b and w. To illustrate the z test, we use the data in Table 20.14. For the upper triangle, we obtain b = 10 + 4 + . . . + 4 = 39. For the lower triangle, we obtain 6 + . . . + 2 = 14. Inserting into the formula for z results in 39 − 14 = 3.434 z= √ 39 + 14 The corresponding tail probability is 0.0003. From this result, we conclude that the neuroleptic drugs reduce the number of symptoms in schizophrenic inpatients.

CFA Models for Developmental Research

The second method of analysis proposed by Lienert and Straube (1980) allows one to test more detailed hypotheses than the one concerning general change in the number of symptoms. One can create a preintervention × postintervention cross-tabulation for each symptom and analyze the resulting I × I table using the Bowker test (1948; von Eye & Spiel, 1996), where I represents the number of categories, or the McNemar test (McNemar, 1947), when I = 2. The test statistic for the Bowker test is ∑∑ i

2

X =

j (mij

− mji )2

mij + mji

for i > j and i, j = 1, . . . , I. This test statistic is approximately distributed as 𝜒 2 with df = (I). When I = 2, one can use the McNemar test which has the same form but simplifies to X2 =

(m12 − m21 )2 m12 + m21

This statistic is approximately distributed as 𝜒 2 with df = 1. When cell frequencies are small, the continuitycorrected equation X2 =

(∣ m12 − m21 ∣ −0.5)2 m12 + m21

This pattern corresponds with the pattern of disagreement in which one rater systematically uses higher ratings than the other. Von Eye and von Eye (2005) proposed coefficients for this case. These coefficients are derivatives of Cohen’s (1960) 𝜅 (kappa) and can therefore be interpreted as proportionate reduction in error (PRE) coefficients. Focusing on the upper triangle of an agreement matrix, one can ask whether disagreement differs from expectation. To answer this question, we need to define a base model. A first option is that we define this model in a fashion similar to the base model for first-order CFA, that is, as a model of rater independence. The same model is used for Cohen’s 𝜅. A second option is to define the base model as the null model, as for Brennan and Prediger’s (1981) 𝜅 n . Using these base models, one can define coefficients of disagreement or, as in the present context, temporal shifts in symptoms, for the upper triangle of an agreement table. A coefficient of disagreement in the upper triangle of an agreement table can be defined in a fashion analogous to Cohen’s 𝜅. The probability of disagreement in the upper triangle of the cross-classification, that is, the probability that Rater B (columns) selects higher scores than Rater A (rows) or the probability that fewer symptoms are diagnosed at Time 2 than at Time 1 is 𝜃1t =

∑

pij = 1 −

∑ ∑ pii − pij

ij

where the superscript indicates that we consider the disagreement cells in the upper triangle. Under the base model of independence of the first from the second observation (or the base model of rater independence), the expected probability of finding cases in the upper triangle in the upper triangle is 𝜃2t =

∑ ∑ ∑ pi. p.j = 1 − pi. p.i − pi. p.j ij

Using 𝜃1t and 𝜃2t , one can define a 𝜅-like PRE coefficient as 𝜃 t − 𝜃2t 𝜅t = 1 1 − 𝜃2t This coefficient also has a range of −∞ < 𝜅 t ≤ 1 (for specifics on range and special cases, see von Eye & von Eye, 2005). To define a coefficient 𝜅nt that is analogous to Brennan and Prediger’s (1981) coefficient 𝜅n , we select the null model as the base model. This model is also used in CFA of zero

892

Configural Frequency Analysis for Research on Developmental Processes

order. For 𝜃1t , we use the same definition as above. For the expected probability, 𝜃2t , we obtain t 𝜃n,2

( ) I 1 I −1 = = 2 I2 2I

and, thus, 𝜅nt

=

t 𝜃1t − 𝜃n,2 t 1 − 𝜃n,2

This coefficient has a range of −∞ < 𝜅nt ≤ 1 (again, for specifics on range and special cases, see von Eye & von Eye, 2005). There are three major differences between the diagonalhalf sign test proposed by Lienert and Straube (1980) and the PRE coefficients proposed by von Eye and von Eye (2005). First, the PRE coefficients can always be interpreted in units of proportionate reduction in error. That is, the PRE coefficients tell the user the increase in the percentage of cases that can be found in the upper triangle in comparison to the expected percentage. Second, the user can select from a number of base models. These models are selected with the goals of CFA in mind. Third, the effects of covariates can be made part of the analysis by estimating the expected frequencies using the model ̂ = X 𝜆, where the ̂ = X 𝜆 + X c 𝜆c instead of log m log m extended model includes, in its second term, the covariate(s). The rest of the calculations remains unchanged from the case with no covariates. To illustrate the PRE coefficients, we reanalyze the data example in von Eye and von Eye (2005; Fennig, Craig, Tananberg-Karant, & Bromet, 1994; Schuster & Smith, 2002). The data describe psychiatric diagnoses of the same 223 patients, once made by practicing psychiatrists and once in a research context. The goal of the original analyses was to determine whether there is agreement between the diagnoses provided by the service facility and research teams. The patients were rated on an ordinal 4-point severity scale with the categories 1 = severe psychosis, 2 = average severity psychosis, 3 = mild psychosis, and 4 = no clinical diagnosis. Table 20.15 gives the cross-classification of the facility with the research diagnoses. The frequency distribution in Table 20.15 shows that the two diagnoses coincide in most cases. In each row of the cross-classification, the agreement cell (diagonal cell) contains the largest number of cases. Also in each row, the agreement cell contains more cases than expected under the base model of independent diagnoses. This is also indicated by the coefficients discussed by von Eye and von Eye (2005).

TABLE 20.15 Cross-Classification of Facility and Research Diagnoses for 223 Psychiatric Patients Facility diagnosis 1 2 3 4 Totals

Research diagnosis 1 40 4 4 17 65

2 18.95 10.20 10.49 25.36

6 25 2 13 46

3 13.41 7.22 7.43 17.95

4 1 21 12 38

4 11.08 5.96 6.14 14.83

15 5 9 45 74

Totals 21.57 11.61 11.95 28.87

65 35 36 87 223

Note: Expected cell frequencies in Italics (base model of first-order CFA). Agreement cells underlined. Source: Fennig et al. (Fennig et al., 1994)

Looking at the upper triangle, we obtain 𝜅 t = −0.24. This value suggests that the proportionate reduction in these error cells is 24.14%. In other words, the upper triangle contains 24.14% fewer cases than one would expect under the assumption of independence of the two diagnoses. Proceeding from the null model, that is, the model that does not take the differences in the marginal probabilities into account, we use the expected frequency of 223∕16 = 13.9375, for each cell. For the 𝜅 n equivalent, the value of 𝜅nt = −0.31 indicates that 31.30% discrepant diagnoses are made less than expected under the assumption of a uniform diagnoses distribution. The coefficient of raw disagreement for the upper triangle (this is just the proportion of cases in the upper triangle) of the disagreement cells is rat = 0.18. None of these values is significant. For example, rat = 0.18 comes with a z score of 0.10 and p = 0.46 (for this significance test, see von Eye, Schauerhuber, & Mair, 1997). For comparison, we also calculate the test discussed by Lienert and Straube (1980). We obtain w = 6 + 4 + . . . + 9 = 40, b = 52, and z = √40−52 = −1.25. The tail probability 40+52

for this value is 0.11, thus indicating that there is only an insignificant shift toward more severe ratings in the facility as compared to the research unit. Apart from indicating nonsignificant effects, this example shows that the various coefficients are sensitive to different characteristics in the data table. Lienert and Straube’s test responds to shifts toward to upper or lower triangle. Von Eye and von Eye’s 𝜅 equivalents compare the observed frequencies in sectors of the cross-classification with expectancies that reflect assumptions concerning marginal probabilities and, thus, effects in the table. Treatment Effects in Designs with Control Group Experimental designs without a control group are deemed less informative or more biased than designs with control group. However, designs without a control group can still

CFA Models for Developmental Research TABLE 20.16 2 × 2 × 2 Table for the Comparison of Pattern Shift in Two Groups Posttreatment patterns treatment group Pretreatment patterns A B Totals

A

B

Totals

m111 m121 m1.1

m112 m122 m1.2

m11. m12. m1..

Control group A

B

A B Totals

m211 m221 m2.1

m212 m222 m2.2

m21. m22. m2..

Sample Totals

m..1

m..2

N

provide information about whether or not treatment effects can be expected at all. If, in a design without a control group, results suggest that effects may exist, controlled studies can be undertaken (if they are not undertaken from the very start). Randomized clinical trials (RCTs) are necessary for a number of reasons. One is to compare treatments with each other. For example, a new form of psychotherapy is compared with an established one. Most important, in particular in psychotherapy research, RCTs are needed to distinguish spontaneous recovery, assessment reactivity, or expectancy effects from treatment effects. CFA can be used to compare experimental groups and control groups with two-group CFA, or, when there are more than two groups, with multigroup CFA. Here, we discuss methods that allow one to compare two groups in regard to the change from one configuration to another. To explain the scenario, consider one treatment group and one control group. For both groups, Patterns A and B are observed before and after the intervention administered to the treatment group. Table 20.16 schematizes the data situation. The goal of analysis is the comparison of the control with the treatment group in a particular change pattern. In the example in Table 20.16, there are only two change patterns, the change from m.11 to m.12 and from m.21 to m.22 , where, in the table as well as here, in the text, a dot indicates the variable aggregated across. The null hypothesis for the present comparison is that the comparison groups do not differ in their change rates. Specifically, we ask whether m112 differs from m212 , or whether m121 differs from m221 . The answer to these questions is given on the basis of 2 × 2 cross-classifications. Table 20.17 displays this cross-classification for the first comparison. In Table 20.17, the notation for the cell frequencies carries over from Table 20.16. Table 20.17 shows a 2 × 2

893

TABLE 20.17 2 × 2 Table for the Comparison of Pattern Shift in Two Groups Patterns

Comparison groups Control treatment

Totals

m112 versus m212 all others combined Totals

m112

m212

m212 + m212

m111 + m121 + m122

m211 + m221 + m222 N − m212 + m212

m1..

m2..

N = m1.. + m2..

cross-classification in which Frequency m112 is compared with m212 , in the context of the rest of the treatment and control group samples. Tables as the one in Table 20.17 can be analyzed with marginal-dependent measures such as Pearson’s X 2 or with marginal-free measures as the odds ratio (for the distinction between marginal-dependent and marginal-free measures, see Goodman, 1991, or von Eye, Spiel, & Rovine, 1995). Data Example In the following example, we reanalyze the data presented by Martinez-Torteya, Bogat, von Eye, Levendosky, and Davidson (Martinez-Torteya et al., 2009). The data were collected in the context of the MIS (//www.msu.edu/∼mis/). As was indicated before, the study first assessed a sample of 206 women in their last trimester of pregnancy and at 2 months postpartum, and the women and their children every year until the children were 14 years old. The women in the study reported a range of experiences of intimate partner violence, from none to severe. In the example, we focus on the 93 women who reported one or more episodes of partner violence during the second year of their child’s life. We ask whether frequency of violence (F) experienced during the second year of their child’s life is predictive of Depression (D) and PTSD (P) in the same year. Each of these three variables was dichotomized at the median, and coded such that the value of 1 indicates a score below the median and 2 indicates a score above the median. Table 20.18 displays the 2 × 2 × 2 D × P × F cross-classification along with CFA results. We first analyze these data using a standard first-order CFA with ̂ = 𝜆 + 𝜆D + 𝜆P + 𝜆F . Then, we ask the base model log m whether particular comparison patterns stand out. For the CFA tests, we use the binomial approximation of the normal distribution and the Holland–Copenhaver procedure for 𝛼 protection. The overall goodness-of-fit likelihood-ratio for the base model of variable independence is X 2 = 44.15. For df = 4, this value leads us to reject of the base model (p < 0.01), and we anticipate that types and antitypes may

894

Configural Frequency Analysis for Research on Developmental Processes

TABLE 20.18 Cross-Classification of Depression (D), PTSD (P), and Frequency of Violence (F)

TABLE 20.19 2 × 2 Table for the Comparison of Two Groups in a Two-Variable Pattern

Configuration

Violence Frequency

DPF

m

̂ m

z

p(z)

111 112 121 122 211 212 221 222

22.00 9.00 3.00 11.00 14.00 3.00 5.00 26.00

10.989 12.237 10.302 11.472 11.721 13.053 10.989 12.237

3.5374 −.9930 −2.4125 −.1490 .7120 −3.0011 −1.9238 4.2218

.000202 .160343 .007922 .440789 .238228 .001345 .027191 .000012

Type/antitype?

Patterns

Type

m111 versus m211 all others combined Totals

Antitype

Below Average Above Average Totals m111 = 22 22 44

n211 = 9 40 49

m111 + m211 = 31 62 N = 93

Antitype Type

Note: Expected cell frequencies are estimated under the base model of first-order CFA.

emerge. Indeed, Table 20.18 shows two types and two antitypes. The first type is constituted by configuration 1 1 1. These are women who report below-average frequencies of violence, along with below-average scores in Depression and PTSD. The second type is constituted by just the opposite pattern, 2 2 2. These are women who report above average frequencies of violence, along with above average Depression and PTSD. The two antitypes are constituted by Patterns 1 2 1 and 2 1 2, indicating that it is very unlikely that low frequencies of violence lead to high levels of depression but low levels of PTSD or high levels of depression but low levels of PTSD, in the presence of intimate partner violence. To illustrate the application of the scheme shown in Table 20.17, we now consider the two levels of violence the comparison groups. The variables Depression and PTSD are crossed to form the patterns in which the two treatment groups are compared. It is important to note that the analysis we are about to illustrate is not just an extension of first-order CFA. In terms of CFA base models, the following analysis represents a 2-group CFA that is saturated in the comparison variables. Specifically, the expected frequencies are estimated using the first-order base model ̂ = 𝜆 + 𝜆D + 𝜆P + 𝜆F + 𝜆DP log m The main reason for using this base model is that types and antitypes (or discrimination types) can no longer appear just because, maybe, the discrimination variables Depression and PTSD are associated. The two treatment groups can differ in a pattern only if one or both of the discrimination variables is related to the grouping variable. We insert, for instance, Pattern D = 1 and PTSD = 1 into a scheme as it is shown in Table 20.17 and obtain the cross-classification in Table 20.19.

The Pearson X 2 for the data in Table 20.19 is 10.44 (df = 1; p < 0.01). This result suggests that below-average frequency of violence is predictive of higher frequencies of low Depression and PTSD than above average violence. Two points concerning the method used for analysis must be discussed, in this context. First, it is important to note that this result is not just a reformulation of the description of the first type in Table 20.18. The type suggests that the pattern low frequency of Violence–low Depression–low PTSD is more likely than expected. Here, we can make a comparative statement in which we contrast the two violence groups. Second, the MIS study oversampled, on purpose, women who were victims of violence. Therefore, the marginal frequencies in Table 20.19 (as well as the univariate frequencies for tables that use the MIS data) may not reflect population characteristics. X 2 measures are known to be marginal-dependent. Thus, the result from the 2 × 2 cross-classification in Table 20.19 may be distorted or biased. A measure that is marginal-free is the odds ratio. Consider a 2 × 2 table with the cell frequencies m11 , m12 , m21 , and m22 . A cross-classification of this type (parallel to the one displayed in Table 20.17) is given in Table 20.20. Table 20.20 contains the information for four odds and one odds ratio (see von Eye & Mun, 2013): 1. Ω1 = p11 ∕p12 : the odds of a diagnosis of the undesirable outcome versus no such undesirable outcome under treatment; 2. Ω2 = p21 ∕p22 : the odds of a diagnosis of the undesirable outcome versus no such undesirable outcome without treatment;

TABLE 20.20 Cross-Classification of Treatment and Outcome Pattern and Cell Frequencies Diagnosis of undesirable outcome Treatment

Yes

No

Yes No

m11 m21

m12 m22

CFA Models for Developmental Research

3. Ω3 = p11 ∕p21 : the odds of a diagnosis of the undesirable outcome in the treatment versus the comparison groups; 4. Ω4 = p12 ∕p22 : the odds of a diagnosis of no undesirable outcome in the treatment versus the comparison groups. Each of these odds has an easy, natural interpretation, and so does the odds ratio that relates these odds to each other. Consider θT = Ω1 ∕Ω2 . This odds ratio can be used to test the hypothesis that respondents in the treatment group are more likely to be diagnosed with the undesirable outcome than respondents who were not exposed. It is interesting to note that there are more odds ratios that could be considered for this 2 × 2 table. However, because they all use the same information, they all have the same interpretation. Specifically, we obtain p11 p11 p p p12 p 𝜃 = p = p21 = 11 22 p12 p21 21 12 p22 p22 As far as the odds ratio is concerned, one can formulate the hypothesis that the odds ratio θT = Ω1 ∕ Ω2 is greater than one in the population, where the superscript T indicates that the odds ratio contrasts the treatment and the comparison groups. To test this hypothesis, one can divide the logarithm (base e) of the odds ratio by its standard error (se), which can be estimated by √ 1 1 1 1 + + + m11 m12 m21 m22

̂ = se(log 𝜃)

The test statistic for the logarithm of the odds ratio is z𝜃 =

log 𝜃̂ . ̂ se(log 𝜃)

This statistic is approximately normally distributed. Applied to the data example in Table 20.19, one obtains ̂ = 0.48; z = 3.13; and p < 𝜃̂ = 4.44, log 𝜃̂ = 1.49; se(log 𝜃) 0.01. We note that, in this example, the X 2 and the odds ratio analyses suggest the same conclusion. If the present X 2 result is biased, the distortion has no effect on the present result. Single-Subject Designs Why would one conduct research using single-subject designs when generalization of results is the main goal of scholarly activity? There are many answers to this question. Here, we focus on two answers that are of interest from

895

a person-oriented perspective. First, in person-oriented (Bergman & Magnusson, 1997; von Eye & Bergman, 2003; von Eye, Bergman, & Hsieh, 2015) and idiographic research (Molenaar, 2004; Molenaar & Campbell, 2009; von Eye, 2004b), it has been shown that results that are based on longitudinal aggregated raw data can be trusted only under rare and unrealistic conditions such as ergodicity. What needs to be done is (1) estimation of parameters at the level of individuals and (2), if needed, aggregation of cases based on their parameters. If this research strategy is followed, one needs longitudinal single-subject designs instead of cross sectional studies. Second, researchers can be interested in an individual in his or her own right, in a treatment study in which treatment can be extensive, or when random assignment can be ethically difficult. There can be a situation in which an individual can be different from other individuals to a degree that aggregation would bias the aggregation of parameter estimates beyond the tolerable. On the other extreme there may be an individual that is known to belong to a population and researchers are interested in how the individual responds to a new form of treatment. A number of single-subject designs has been developed for CFA (von Eye, 2002; von Eye et al., 2010). Here, we begin with single-subject CFA. This is followed by CFA of runs and cross-lagged CFA. Single-Subject CFA When a CFA is considered for multidimensional data that stem from a single individual, the resulting crossclassification that is being analyzed does not differ in form from the cross-classification of variables across multiple individuals. Single subject data are different however, in that they were collected over multiple occasions. Consider the relationships between the three variables A, B, and C, observed over n occasions for one individual. The A × B × C cross-classification has n entries. It allows one to test hypotheses concerning the association structure of the three variables, for that particular individual. It also allows one to search for types and antitypes that contradict a particular base model, also for the particular individual. Data Example To illustrate this approach, we use data from an intensive longitudinal study of self-identified alcoholics (Perrine, Mundt, Searles, & Lester, 1994). A sample of alcoholic adult men provided every morning information about their drinking the day before, as well as their subjective ratings of mood, health, and quality on the day of the interview. Here, we ask whether there are associations among mood (M) of

896

Configural Frequency Analysis for Research on Developmental Processes

Respondent 3000, his subjective reports of health (H) and beer drinking (B). Data assessed from 734 consecutive days were available. Mood was rated on a 10-point Likert scale with 1 indicating “terrible mood” and 10 indicating “just wonderful.” Respondent 3000 rarely endorsed the low and the high end points on this scale. Therefore, scores 4 and below were combined into a single category. Similarly, this respondent nearly never used scores 7 through 10. Therefore, scores 7–10 were also combined into a single category. The resulting scale thus ranged from 1 to 4 with 1 indicating average mood or below, and 4 indicating good or better mood. Similarly, the range of health ratings was condensed from a 10-point Likert scale to the values 1–4. The resulting scale values indicate 1 = perception of average health or below through 4 = perception of good to excellent health. The range of the numbers of beer consumed by Respondent 3000 was wide, with the extremely high numbers rarely reported. Therefore, this scale was also condensed to range from 1 = no beer consumed through 3 = three or more beers consumed in one day. For the analysis of the resulting 4 × 4 × 3 H × M × B cross-classification, we use the z-test and the Holland– Copenhaver procedure of 𝛼 protection. The first base model we use is that of first-order CFA, ̂ = 𝜆 + 𝜆 H + 𝜆M + 𝜆 B log m This model can be rejected when any combination of the interactions [B, M], [B, H], [M, H], [B, M, H] exists, that is, any interaction at all. For the base model of first-order CFA, we obtain a LR − X 2 = 89.60 (df = 39; p < 0.01). Therefore, we anticipate that types and antitypes emerge. Table 20.21 summarizes the CFA results. The results in Table 20.21 suggest that only one type exists and no antitype. This type indicates that comparatively low ratings of mood and health co-occur with low consumption of beer more often than expected. This result reflects an extreme case of a local association (Havránek & Lienert, 1984). The associations in this table of 48 cells are carried by just one cell. One could argue that cell 1 1 2 also makes a contribution. However, under the restrictions of 𝛼 protection, this contribution is nonsignificant if this cell is examined by itself. If one tests whether cells 1 1 1 and 1 1 2 constitute a composite type, one obtains a Stouffer Z = 6.93, which suggests a strong deviation from the independence model for this pair of cells. This composite type suggests that health and mood ratings are at their lowest when Respondent 3000 consumed two beers or less on the day before the interview.

TABLE 20.21 First-Order CFA of the Cross-Classification of Numbers of Subjective Health (H), Mood (M), and Beer Consumed (B), for Respondent 3000 Configuration HMB 111 112 113 121 122 123 131 132 133 141 142 143 211 212 213 221 222 223 231 232 233 241 242 243 311 312 313 321 322 323 331 332 333 341 342 343 411 412 413 421 422 423 431 432 433 441 442 443

m

̂ m

z

p(z)

11.00 21.00 38.00 7.00 38.00 110.00 .00 15.00 77.00 .00 5.00 19.00 .00 8.00 12.00 3.00 11.00 77.00 1.00 5.00 43.00 .00 1.00 7.00 .00 5.00 14.00 .00 16.00 47.00 .00 1.00 24.00 .00 1.00 10.00 .00 2.00 9.00 2.00 11.00 38.00 .00 4.00 30.00 .00 2.00 9.00

1.823 11.089 42.837 5.469 33.267 128.512 3.038 18.482 71.396 .820 4.990 19.277 .898 5.463 21.105 2.694 16.390 63.314 1.497 9.105 35.174 .404 2.458 9.497 .631 3.837 14.823 1.892 11.512 44.470 1.051 6.395 24.706 .284 1.727 6.671 .572 3.480 13.442 1.716 10.439 40.325 .953 5.799 22.403 .257 1.566 6.049

6.7972 2.9762 −.7391 .6549 .8205 −1.6330 −1.7430 −.8099 .6633 −.9057 .0044 −.0630 −.9477 1.0853 −1.9819 .1863 −1.3313 1.7200 −.4061 −1.3605 1.3195 −.6357 −.9302 −.8103 −.7942 .5936 −.2139 −1.3756 1.3228 .3793 −1.0253 −2.1335 −.1420 −.5328 −.5531 1.2891 −.7563 −.7932 −1.2115 .2168 .1737 −.3661 −.9764 −.7472 1.6051 −.5073 .3470 1.2000

.000000 .001459 .229927 .256278 .205954 .051236 .040665 .208996 .253577 .182548 .498231 .474865 .171650 .138893 .023747 .426105 .091542 .042715 .342351 .086831 .093503 .262482 .176141 .208888 .213533 .276405 .415319 .084468 .092951 .352224 .152603 .016442 .443541 .297093 .290106 .098681 .224736 .213836 .112856 .414167 .431041 .357140 .164439 .227484 .054234 .305958 .364301 .115073

Type/antitype? Type

We also ask whether the pairwise associations that involve mood, that is, [B, M], [M, H], allow us to explain the data. To answer this question, we estimate the log-linear model ̂ = 𝜆 + 𝜆H + 𝜆M + 𝜆B + 𝜆BM + 𝜆MH log m

CFA Models for Developmental Research

and a multinomial logit model. This model describes the data well. Specifically, we obtain a LR − X 2 = 33.87 (df = 24; p = 0.251). This model clearly stands by itself. It is also significantly better than the base model of first-order CFA (ΔLR − X 2 = 55.73; Δdf = 15; p < 0.01). Both interactions are significant. Similarly, a multinomial logit model comes with an X 2 of 53.54, which indicates that beer consumption and subjective health can be used as predictors of mood (df = 15; p < 0.01). This result illustrates again the difference between variable-oriented and person-oriented research. The former allows one to state that, in the present example, mood is associated with both beer consumption and subjective health ratings. The person-oriented results, however, suggest that (1) comparatively low beer consumption goes hand in hand with comparatively poor mood and low health ratings and that (2) otherwise the three variables are independent of one another. Although interesting, the approach to single subject CFA just described lacks a developmental perspective in the sense that possible change over time is aggregated over. Therefore, we present, in the next example an approach to CFA that takes the longitudinal nature of data into account. Data Example We use data from the study on the development of alcoholics again. Instead of Respondent 3000, we now use the answers from a different participant, Respondent 3004. He provided data on 742 consecutive days. To keep the example simple, we ask whether the relationship between beer consumption (B) and subjective health (H) varies from the first to the second half of the observation period (P). Each variable was coded so that small numbers indicate low levels and high numbers indicate high levels (or perceptions indicating good levels). This question can be answered using a base model that has two characteristics. First, the observation period is proposed to be independent of beer consumption and subjective health. Second, beer consumption is proposed to be associated with health rating. The base model is that of a Prediction CFA, ̂ = 𝜆 + 𝜆B + 𝜆H + 𝜆P + 𝜆BH log m This model can be rejected only if one or more of the following interactions exist: [B, P], [H, P], [B, H, P]. Each of these interactions contains the variable that indicates the first versus the second half of the observation period. If one or more of these interactions exist, the association

897

TABLE 20.22 Prediction CFA of the Cross-Classification of Numbers of Beer Consumed (B) and Subjective Health (H), over the First and the Second Halves of the Observation Period (P); for Respondent 3004 Configuration BHP

m

̂ m

z

111 112 121 122 131 132 211 212 221 222 231 232 311 312 321 322 331 332 411 412 421 422 431 432

1.000 0.000 0.000 1.000 3.000 131.000 1.000 0.000 0.000 1.000 0.000 29.000 2.000 1.000 1.000 1.000 3.000 28.000 19.000 22.000 3.000 1.000 331.000 155.000

0.496 0.504 0.496 0.504 66.452 67.548 0.496 0.504 0.496 0.504 14.381 14.619 1.488 1.512 0.992 1.008 15.373 15.627 20.332 20.668 1.984 2.016 241.014 244.986

0.716 −0.710 −0.704 0.698 −7.784 7.720 0.716 −0.710 −0.704 0.698 −3.792 3.761 0.420 −0.417 0.008 −0.008 −3.156 3.130 −0.295 0.293 0.722 −0.716 5.796 −5.749

Type/antitype?

Antitype Type

Antitype Type

Antitype Type

Type Antitype

between beer consumption and health varies over the two observation periods. For the following Prediction CFA, we use the z test and the Holland–Copenhaver procedure for 𝛼-protection. Table 20.22 displays results of CFA. The results in Table 20.22 suggest that four pairs of prediction type/antitype exist. Each of these pairs suggests that a pattern that stands out in the first half of the observation period also stands out in the second half, but for the opposite reason. Specifically, the first antitype, constituted by configuration 1 3 1, suggests that it is very unlikely that, for this respondent, very low levels of beer consumption together with subjective ratings of average to good health are reported during the first half of the observation period. More than 66 daily instances of this pattern are expected, but it was found only for 3 days. This pattern reverses dramatically in the second half of the period. More than 67 instances of the pattern of very low levels of beer consumption and average to good health are expected, but 131 instances were found. Configuration 1 3 2 thus constitutes a prediction type. Similarly, the second antitype, constituted by configuration 2 3 1, suggests that low beer consumption in combination with average or better subjective health is unlikely for the first half of the observation period. It constitutes a prediction antitype. In the second half, however,

898

Configural Frequency Analysis for Research on Developmental Processes

this pattern is particularly likely, and constitutes a prediction type. Similarly, Pattern 3 3 (average beer consumption in tandem with subjectively average or better health) is unlikely in the first half (three instances observed versus over 15 expected) but very likely in the second half of the observation period (28 instances observed versus over 15 expected). For Pattern 4 3, the relationship is reversed. This pattern indicates above average beer consumption in tandem with subjectively above average or better health, can be predicted to occur more often than expected for the first half of the study (331 observed versus 241 expected instances) but less often than expected in the second half (155 observed versus 244 expected instances). Pattern 4 3 is also the most frequent of all patterns for this respondent. If the temporal aspect of the observations is not taken into account, that is, after aggregating responses over time, the B × H cross-classification (aggregated over the two periods of time; not shown here) would have contained just one type. This type suggests that drinking small amounts of beer on any given day goes hand in hand with average or better self-reported health. This type can, evidently, not be generalized over time. In the second half of the observation period, it transmutes into an antitype. CFA of Runs Published for the first time in the context of CFA of intensive longitudinal data, CFA of runs focuses on particular characteristics of series of data. Using runs, researchers answer the question whether there is randomness in a series of data. For a definition of runs, consider a series of observations. A run is defined as a pattern in this series. In general, a run is (1) an uninterrupted series of bits that share one or more characteristics that is (2) bounded on both sides by bits of different characteristics. In the following paragraphs, we give a selection of examples (for the well-known runs tests, see Stevens, 1939; Swed & Eisenhart, 1943; Wald & Wolfowitz, 1940; for runs in the context of CFA and an algorithmic perspective, see von Eye et al., 2010). The definitions involve single scores, but they can also involve two or more scores. For single scores, a run can be defined as 1. The length, k > 1, of uninterrupted series of scores of the same value. This is the classical definition of a run. For example, the series 1 2 3 3 3 4 4 5 6 7 contains two runs. The first involves the three 3s, that is k1 = 3. The second involves the two 4s, that is k2 = 2, where the subscripts index the runs. This definition of runs requires no more than nominal level scales.

2. The length, k > 1, of an uninterrupted series of scores of increasing value. The series 8 9 9 3 4 5 6 contains 2 runs. The first involves the first two scores, and the second involves the last four scores. For runs of series of increasing scores, one needs scales that are at least ordinal. 3. The length, k > 1, of an uninterrupted series of scores of decreasing value. As for runs of decreasing scores, runs of decreasing scores require no more than ordinal scales. 4. The length, k > 1, of an uninterrupted series of scores within a pre-specified range. For this definition, at least ordinal information is needed. 5. For pairs of scores, a run can also be defined as 6. The length, k > 3, of an uninterrupted series of scores that go up-and-down or down-and-up (Wallis & Moore’s, 1941, runs up-and-down test). The series 1 2 5 4 6 7 6 9 contains two up-and-down runs. The first run contains the values 2, 5, and 4, and the second contains the values 6, 7, and 6. The series also contains two down-and-up runs. The first contains the values 5, 4, and 6. The second contains the values 7, 6, and 9. To describe the runs of up-and-down or down-and-up series of scores, one needs at least ordinal information. 7. The length, k > 1, of an uninterrupted series of pairs of scores in which one particular score follows the other particular score. Consider, for example, the series 4 1 0 1 0 1 0 6 7 8. This series contains one run of length k = 3 of 1s and 0s that follow each other. More complex definitions of runs are conceivable. As was illustrated for the case of up-and-down and downand-up runs, a series can contain runs that follow more than one definition. In most applications of runs tests, however, just one definition is used. For use in CFA, runs information from different variables can be crossed directly, if the number of different runs is limited. This is the case when the series under study are relatively short. When series are long, many runs can result that differ in length. Therefore, for long series, researchers either create classes of runs, e.g., ‘did not drink hard liquor for a week’ or ‘depression did not relapse for one month’, or they focus on particular runs such as the longest. The questions that CFA allows one to answer concerning runs information differ from the questions asked when runs tests are applied. With runs tests, one asks whether a series of scores or observations is random in nature. If the corresponding null hypothesis is rejected, one can entertain hypotheses concerning the processes that cause deviations from randomness. In CFA, one asks whether runs information for one variable is related to runs information

CFA Models for Developmental Research

from other variables or to other variables, in general. The answers to these questions are given in terms of local deviations from CFA base models, that is, in terms of types and antitypes. Data Example For the following example, we use data from the Finkelstein, et al. (1994) study on the development of aggressive behavior. The authors administered a questionnaire concerning aggressive behavior to 67 adolescent girls and 47 boys at three points in time. The time intervals were two years each. The questionnaire addressed the four measures of aggression: Aggressive Impulse, Aggression-Inhibitory Response, Verbal Aggression against Peers, and Physical Aggression against Peers. In the following analyses, we use the data on physical aggression against peers (PAAP), verbal aggression against adults (VAAA), and aggressive impulses (AI). For each of the variables, three scores are available, PAAP83, PAAP85, and PAAP87, VAAA83, VAAA85, and VAAA87, and AI83, AI85, and AI87, where the numbers indicate the years in which the data were collected. When the data were collected in those years, the respondents were 11, 13, and 15 years of age, respectively. In the following example, we look at uninterrupted series of scores of decreasing value. The question we ask concerns the decrease in aggression in adolescents. Specifically, we ask whether there are local associations among patterns of steady decrease in the three aggression variables: physical aggression against peers (PAAP), verbal aggression against adults (VAAA), and aggressive impulses (AI). For three observation points, there can only be increasing- or decreasing-score runs of length k = 2. In the present example, a run of k = 2 indicates a decrease in scores from 1983 to 1985 and again from 1985 to 1987. The run variables were coded as follows: • physical aggression against peers (PAAP): the run variable for PAAP, PR, was coded as • PR = 1 if there was no run of decreases in PAAP scores, and • PR = 2 if there was a run (two consecutive decreases); • verbal aggression against adults (VAAA): the run variable for VAAA, VR, was coded as • VR = 1 if there was no run of decreases in VAAA scores, and • VR = 2 if there was a run (two consecutive decreases); • aggressive impulses (AI): the run variable for AI, AR, was coded as • AR = 1 if there was no run of decreases in AI scores, and

899

• AR = 2 if there was a run (two consecutive decreases). The 2 × 2 × 2 PR × VR × AR cross-classification can be approached with two questions. First, we ask whether there are any local associations at all. This question can be answered with the base model of first-order CFA, ̂ = 𝜆 + 𝜆PR + 𝜆VR + 𝜆AR log m This base model can fail if any interaction exists. Second, we can also ask whether association patterns are specific to adolescents who differ in their patterns of change in aggressive impulses. This question can be answered using the following base model of 2-group CFA, ̂ = 𝜆 + 𝜆PR + 𝜆VR + 𝜆AR + 𝜆PR,VR . log m This base model can be rejected only if interactions exist that link the run variables for physical aggression against peers and verbal aggression against adults on one side with runs of aggressive impulses on the other. These are the interactions [PR, AR], [VR, AR], and [PR, VR, AR]. Table 20.23 presents results for the base model of firstorder CFA, that is, the model of variable independence. Table 20.24 presents results for two-group CFA. For both CFAs, we use the normal approximation of the binomial test and protect 𝛼 using the Holland–Copenhaver procedure. The overall goodness-of-fit likelihood-ratio for the base model of variable independence is X 2 = 13.93. For df = 4, this value leads us to reject the base model (p < 0.01), and we anticipate that types and antitypes may emerge. However, in this example, we encounter the rare case, in which no type or antitype emerges although the base model is rejected. Inspecting the frequency distribution in Table 20.23, we notice that, for three of the four cell pairs . . 1/. . 2, cell . . 1 contains (numerically) more TABLE 20.23 First-Order CFA of the PR × VR × AR Cross-Classification Configuration PR VR AR 111 112 121 122 211 212 221 222

m

̂ m

z

p(z)

45.00 6.00 18.00 3.00 15.00 13.00 9.00 5.00

38.078 11.817 16.870 5.235 22.212 6.893 9.841 3.054

1.3746 −1.7874 .2981 −1.0002 −1.7054 2.3996 −.2804 1.1288

.084620 .036937 .382811 .158601 .044063 .008207 .389594 .129500

Type/antitype?

900

Configural Frequency Analysis for Research on Developmental Processes

TABLE 20.24 2-Group CFA of the PR × VR × AR Cross-Classification, with AR as the Grouping Variable Configuration PR VR AR

m

Statistic

p

111 112

45.00 6.00

2.693

.003537

121 122

18.00 3.00

1.122

.131014

211 212

15.00 13.00

−3.259

.000559

221 222

9.00 5.00

−1.130

.129145

Type? Discrimination type

Discrimination type

cases than expected but cell . . 2 contains (numerically) fewer cases than expected. We, therefore, conduct the 2-group CFA and expect that discrimination types emerge. Table 20.24 displays results. The overall goodness-of-fit likelihood-ratio for the base model of 2-group CFA is X 2 = 13.72. For df = 3, this X 2 value leads to the rejection- of the base model (p < 0.01), and we anticipate that discrimination types may emerge. The fit of this model is only marginally better than the fit of the model of variable independence. We conclude that the interaction between the runs of physical aggression against peers and verbal aggression against adults is weak, and makes no contribution to the explanation of the frequency distribution in Tables 20.23 and 20.24. The results in Table 20.24 show that two-group CFA identified two discrimination types. The first is constituted by configuration pair 1 1 1/1 1 2. It suggests that those adolescents who show inconsistent increases/decreases in aggressive impulses (Pattern . . 1) are more likely than those adolescents who show consistent decreases in aggressive impulses (Pattern . . 2) to also show inconsistent patterns of decreases/increases in verbal aggression against adults and physical aggression against peers. The first group contains 45 respondents, the second group contains 6 respondents. This ratio is more extreme than the group ratio, which is 74 to 27. The corresponding log odds ratio is 1.32 (se = 0.51; z = 2.59; p < 0.01). The second discrimination type is constituted by configuration pair 2 1 1/2 1 2. This discrimination type suggests that those adolescents who show inconsistent increases/decreases in aggressive impulses (Pattern . . 1) are more likely than those adolescents who show consistent patterns of decreases in aggressive impulses (Pattern . . 2) to show consistent patterns of decreases only in verbal aggression against adults, but not in physical aggression against

peers. The first group contains 15 respondents, the second group contains 13 respondents. This ratio is less extreme than the group ratio, which is 74 to 27. The corresponding log odds ratio is −1.50 (se = 0.48; z = −3.12; p < 0.01). From the perspective of conducting research with CFA, we conclude: 1. Significant discrepancies between a base model and the observed frequency distribution are a necessary but not a sufficient condition for types and antitypes to emerge. Types and antitypes require, in addition, that the discrepancies between observed and expected cell frequencies are particularly strong in a few cells instead of being distributed more evenly over all cells in a table. Evidently, the distribution of discrepancies in Table 20.23 was not extreme enough. 2. Two-group CFA is able to find discrimination types in particular when the model–data discrepancies in corresponding cells go in opposite directions. This applies even when other models of CFA do not suggest that types or antitypes exist. As we said before, in this chapter, patterns of types and antitypes depend on the base model. Different base models can result in different patterns of types and antitypes (and come with different interpretations; Mellenbergh, 1996). 3. Depending on the number of groups to be compared, the number of tests performed in multigroup CFA is only a fraction of the number performed in other CFA models. Therefore, the protected 𝛼 is less extreme and, thus, less prohibitive for the detection of discrimination types. CFA of Lags In longitudinal research, one can distinguish between lagged values and current values. Lagged values are the values of a variable from a different period. Current values are from the last observation point in time. Lags are measured in units of time (e.g., days or years), or in experimental units (e.g., trials). In the simplest case, one variable is related to itself over time. For example, one can ask whether the number of technical problems in cars produced on a Monday are predictive of problems for cars produced the following Monday, or whether the number of beers an alcoholic individual consumes on a Friday is predictive of the number of beers consumed the following Friday. Creating data sets for the analysis of lags is relatively straightforward. All one needs to do is copy the series of scores and paste it back in with the lag of interest, shifting it by this lag. Table 20.25 displays the resulting arrangement for lags 1, 2, and 3.

CFA Models for Developmental Research TABLE 20.25 Series with Lags 1, 2, and 3 Time

Original scores

Scores with lag 1

Scores with lag 2

Scores with lag 3

1 2 3 4 5 . . . n−1 n

x1 x2 x3 x4 x5 . . . xn−1 xn

– x1 x2 x3 x4 . . . xn−2 xn−1

– – x1 x2 x2 . . . xn−3 xn−2

– – x1 x2 . . . xn−4 xn−3

TABLE 20.26 Cross-Classification of a Series of Scores with Itself, for a Lag of k Lag k series Original series

i=1

i=2

i=3

Totals

i=1 i=2 i=3 Totals

m11 m21 m31 m.1

m12 m22 m32 m.1

m13 m23 m33 m.1

m1. m2. m3. m.n = N − k

In Table 20.25, the subscripts of the x-scores indicate the observation point. The subscripts for the lagged series show that, with increasing lags, the number of scores that can directly be compared decreases by k, the size of the lag. When scores are categorical in nature, lagged series can be crossed with each other and subjected to methods of categorical data analysis, for example, log-linear modeling, logit modeling, or CFA. Table 20.26 displays a sample cross-classification for a variable with I = 3 categories. The frequencies in the diagonal cells in Table 20.26, which are the cells with subscripts ii, indicate the number of times a category of the original series is followed by an observation of the same category, for a lag of k. The frequencies in the off-diagonal cells, which are the cells with subscripts ij, with i ≠ j, indicate the frequency with which an observation of Category i is followed by the observation of Category j, for a lag of k. Extensions of the approach shown in Table 20.26 can be achieved, for example, by taking stratification variables or covariates into account (gender, age, ethnic background, etc.) or by increasing the number of lagged variables. Cross-classifications such as illustrated in Table 20.26 can be analyzed using well-known methods of categorical data analysis. It is interesting to note that rater agreement tables have the exact same form as the one given in Table 20.26. Therefore, the same methods can be used for analysis (see von Eye & Mun, 2005). These methods include the estimation of coefficients of agreement and the estimation of

901

manifest or latent variable models that reflect hypotheses concerning the pattern of agreement or disagreement (von Eye & von Eye, 2005). The best-known coefficient of rater agreement is Cohen’s 𝜅 (kappa), which is estimated from an I × I crossclassification as the one given in Table 20.26 by ∑ ∑ N i=j mij − i=j mi. m.j 𝜅= ∑ m2 − i=j mi. m.j The interpretation of 𝜅 can proceed as for all PRE (proportionate reduction in error) coefficients. Specifically, 𝜅 can be interpreted as proportion of agreement beyond chance. Data Example For the following data example, we use data from a study of self-identified alcoholic men again (Perrine et al., 1994). We use the data from Respondent 3053 who answered the telephone survey on 800 consecutive days. The following analyses focus on the numbers of beer (B) and liquor shots (L) consumed on a given day and on mood (M) on the following day. Mood was rated on a 10-point Likert scale with 1 indicating “terrible mood” and 10 indicating “just wonderful.” Respondent 3053 tended to not use the lower end points of the mood scale. Therefore, scores 4 and below were combined into a single category of 1. In addition, this respondent used scores greater than 6 rarely. Therefore, scores 6 through 10 were coded as 3. The resulting scale thus ranged from 1 to 3 with 1 indicating average mood or below, and 3 indicating good or better mood. Similarly, the range of shots of liquor consumed was condensed to include the values of 1 = no liquor consumed and 2 = 1 or more shots of liquor consumed. This respondent almost never had more than 2 shots of liquor on any given day. The range of the numbers of beer consumed by Respondent 3053 was narrow also, but he drank beer almost every day. This beer response was, therefore recoded to have values 1, 2, and 3, with 1 = one or no beer consumed through 3 = three or more beers consumed in one day. In the following paragraphs, we ask three questions that can be answered using the data thus created. 1. Can the number of beers consumed on consecutive days be predicted? 2. Can the number of shots of liquor consumed on consecutive days be predicted? 3. Is the mood on one day predictive of the mood on the following day? 4. Is the mood on one day predicted by the numbers of beer and shots of liquor the day before?

902

Configural Frequency Analysis for Research on Developmental Processes

TABLE 20.27 Cross-Classification of Beer Consumption for a Lag of One Day (Respondent 3053)

TABLE 20.28 Cross-Classification of Liquor Consumption for a Lag of One Day (Respondent 3053)

Beer consumption lag 1

Beer consumption

1 2 3 Total

1

2

3

Total

296 30 156 482

24 2 21 47

161 15 95 271

481 47 272 800

The first three questions are answered using CFA as well as methods otherwise employed in the analysis of rater agreement. Specifically, considering that the tables to be analyzed are all of the form I × I, we use first-order CFA for questions 1 and 3, Cohen’s 𝜅 for questions 1, 2, and 3, and the odds ratio for question 2. To answer question 4, we use CFA and log-linear modeling. Question 1: Beer consumption over a lag of one day. The cross-classification of the ordinal beer consumption scale for Respondent 3053 is given in Table 20.27. Table 20.27 shows an interesting marginal frequency pattern. Regardless of the number of beers the respondent consumed, the first observation day is often followed by a day during which the respondent consumed only a small number of beers or none. This becomes evident by comparing the frequencies in the first column with the frequencies in the second and third columns. This pattern implies that the number of beers consumed is not strongly related to the number of beers consumed on the following day. Accordingly, the Pearson X2 for this 3 × 3 table suggests a nonsignificant association (X 2 = 3.21; df = 4; p = 0.52). First-order CFA of the frequency distribution in Table 20.27 (not shown here) identified no types or antitypes. This is expected from the nonsignificant X2 . Considering that the diagonal contains the largest number of cases only in the first row/column, we also expect a small and nonsignificant 𝜅. Indeed, the 𝜅 = 0.02 indicates that the diagonal contains only 2% more cases than expected under the assumption of independence between beer consumption on one day and the next. This value is also nonsignificant (z = 0.66; p = 0.51).

Liquor consumption lag 1

Liquor consumption

2

Total

531 110 641

110 35 145

641 145 786

for this 2 × 2 table suggests a nonsignificant association (X 2 = 3.83; df = 1; p = 0.05). As for beer consumption, the diagonal contains the largest cell frequency only for one cell, here in Row 1/Column 1. Also as for Table 20.27, we therefore expect 𝜅 to be small and nonsignificant. Indeed, the 𝜅 = 0.07 indicates that the diagonal contains only 7% more cases than expected under the assumption of independence between liquor consumption on one day and the next. This value is nonsignificant (z = 1.82; p = 0.07). Similarly, the odds ratio of θ = 1.54 suggests that the association of liquor consumption on consecutive days is nonsignificant (log θ = 0.43; z = 1.95; p = 0.052). Excursus: on the dependency of types and antitypes in 2 × 2 tables. Interestingly, it does not make much sense to perform a first-order CFA on a 2 × 2 cross-classification as the one given in Table 20.28. Von Weber, Lautsch, and von Eye (2003) showed that the CFA tests in 2 × 2 tables are completely dependent, and therefore the result of the first CFA test fully determines the results of the remaining three tests (for a discussion of type/antitype patterns in larger tables, see Krauth, 2003). To illustrate the dependency of cell-wise CFA tests in 2 × 2 tables, we consider the cross-classification shown in Table 20.29. Suppose we employ the base model of variable indepen̂ 1. , m2. = m ̂ 2. , m.1 = dence. Under this base model, m1. = m ̂ .2 . Now suppose also that CFA identifies a ̂ .1 , and m.2 = m m ̂ 11 > 0. We set m11 − m ̂ 11 = type for cell 1 1, that is, m11 − m ̂ 11 + m ̂ 12 = m ̂ 1. , we obtain d. Because m11 + m12 = m1. = m ̂ 12 = −d. Accordingly, we find that m21 − m ̂ 21 = −d, m12 − m ̂ 22 = d. and m22 − m In words, in 2 × 2 tables for which the expected cell frequencies are estimated using the log-linear main effect TABLE 20.29 2 × 2 Cross-Classification with Observed Frequencies mij ̂ ij and Expected Frequencies m

Question 2: Liquor consumption over a lag of one day. The cross-classification of the ordinal liquor consumption scale for Respondent 3053 is given in Table 20.28. Liquor consumption of respondent 3035 is not associated with itself from one day to the next. The Pearson X 2

1 2 Total

1

Totals

Observed and expected cell frequencies

Totals

m11 ̂ 11 m m21 ̂ 21 m m.1 ̂ .1 m

m1. ̂ 1. m m2. ̂ 2. m m.. = N N

m12 ̂ 12 m m22 ̂ 22 m m.2 ̂ .2 m

CFA Models for Developmental Research

model, there is no space for independent CFA types and antitypes. Under the main effect model (which is also the model used for X2 testing), there is only one degree of freedom left. Therefore, if the difference between the observed and the expected cell frequency in cell 1 1 suggests the existence of a type, cells 1 2 and 2 1 will, by necessity, constitute antitypes, and cell 2 2 will, also by necessity, constitute a second type. Accordingly, if cell 1 1 constitutes an antitype, Cells 1 2 and 2 1 constitute types, and cell 2 2 constitutes a second antitype. In one word, if a type–antitype decision is made for any cell in a 2 × 2 cross-classification in first-order CFA, the type–antitype decisions in the remaining three cells are completely redundant. Therefore, we did not perform a first-order CFA on Table 20.28. Question 3: Mood over a lag of one day. The 3 × 3 crossclassification of the ordinal mood scale for Respondent 3053 is given in Table 20.30. Table 20.30 also shows an interesting pattern. The marginal frequencies of mood ratings are identical on consecutive days. Pearson’s X 2 for this 3 × 3 table suggests a significant association (X 2 = 86.18; df = 4; p < 0.01). Only one of the diagonal cells is the largest of its row/column. This is cell 3 3. Considering that this cell frequency is the largest of the entire table, by far, we anticipate a significant 𝜅 that is larger than those for the beer and liquor consumption data. Indeed, the 𝜅 = 0.22 indicates that the diagonal contains 22% more cases than expected under the assumption of independence between mood on one day and the next. This value is significant (z = 5.03; p < 0.01). Considering that the value of 𝜅, although significant, is small, and that the distribution in Table 20.30 is very uneven, we now perform a first-order CFA. The base model for this analysis is that of independence of the two mood ratings, M and M1. The model is ̂ = 𝜆 + 𝜆M + 𝜆M1 log m If this model is rejected, the two mood ratings must be related to each other. Types and antitypes tell us where TABLE 20.30 Cross-Classification of Mood Ratings for a Lag of One Day (Respondent 3053) Mood lag 1

Mood

1 2 3 Total

1

2

3

Total

3 6 11 20

8 11 36 55

9 38 663 710

20 55 710 785

903

TABLE 20.31 First-order CFA of Mood Ratings on Consecutive Days (Respondent 3053) Configuration M M1 11 12 13 21 22 23 31 32 33

m

̂ m

z

p

3.00 8.00 9.00 6.00 11.00 38.00 11.00 36.00 663.00

.510 1.401 18.089 1.401 3.854 49.745 18.089 49.745 642.166

3.4888 5.5744 −2.1371 3.8849 3.6405 −1.6653 −1.6668 −1.9488 .8222

.000243 .000000 .016297 .000051 .000136 .047929 .047776 .025657 .205492

Type/antitype? Type Type Type Type

local associations exist. Table 20.31 displays CFA results. We use the z test and the Holland–Copenhaver procedure of 𝛼 protection. The overall goodness-of-fit likelihood-ratio for the base model of independence of the mood ratings on consecutive days is X 2 = 54.11. For df = 4, this value leads us to reject the base model (p < 0.01), and we anticipate types and antitypes to emerge. Interestingly, only types emerge but no antitypes. The first type is constituted by cell 1 1. This cell contains the smallest number of ratings. Still, it constitutes a type. It indicates that, although Respondent 3053 rarely indicates average to low mood on any given day, on two consecutive days, he reports average to low mood more often than expected. The second type is constituted by cell 1 2, the third-smallest cell in the table. This type signifies that although an improvement in mood from average/low to slightly above average is rare, this occurs more often than expected. The same applies to the reverse change, from slightly above average mood down to average mood (configuration 2 1). This pattern is also rare but reported more often than expected. The fourth type, constituted by configuration 2 2, suggests that two consecutive days for which Respondent 3053 reports slightly above average mood occur more frequently than expected. The most frequent pattern (84.5%), for Respondent 3053, is 3 3. Respondent 3053 reports that his mood is above average on any two consecutive days for 84.5% of the times. We conclude that this respondent is remarkably stable and in a good mood. The observed frequency of 663 is only minimally greater than the frequency that was expected for this pattern (642.17). Therefore, configuration 3 3 does not constitute a type. Question 4: Is the mood on one day predicted by the numbers of beer and shots of liquor the day before? It is often assumed that alcohol consumption affects mood (First & Tasman, 2004). Here, we ask whether there

904

Configural Frequency Analysis for Research on Developmental Processes

are, for Respondent 3053, short-term predictive relationships of beer and liquor consumption to mood. To answer this question, we apply three methods and compare their results. First, we employ log-linear modeling. The results from this approach tell us whether there are interactions among the variables beer consumption (B), liquor consumption (L), and mood on the following day (M1), in general. Second, we estimate a log-linear model. The results from this model tell us whether these interactions are compatible with the hypothesis that beer and liquor consumption affects mood on the following day. Third, we perform a first-order and a prediction CFA. The results from these approaches tell us where local associations exist, if they exist. The B × L × M1 cross-classification is given in Table 20.32. Results from Log-Linear Modeling To create a base model that can be used for comparison, we first estimate the model of variable independence. This is the same model that will be used for first-order CFA. The overall goodness-of-fit likelihood-ratio for the base model of independence of alcohol consumption and mood ratings on the following day is X 2 = 33.34. For df = 12, this value leads us to the reject the base model (p < 0.01) (and we anticipate types and antitypes to emerge; see below). We then ask whether beer and liquor consumption are related to each other to the degree that this model–data discrepancy could be explained. The corresponding model is ̂ = 𝜆 + 𝜆B + 𝜆L + 𝜆M1 + 𝜆B,L log m TABLE 20.32 First-Order CFA of Beer and Liquor Consumption and Mood Ratings on the Following Day (Respondent 3053) Configuration B L M1 111 112 113 121 122 123 211 212 213 221 222 223 311 312 313 321 322 323

m

̂ m

z

p

6.00 26.00 338.00 3.00 6.00 94.00 1.00 4.00 26.00 2.00 1.00 13.00 7.00 15.00 222.00 1.00 3.00 22.00

9.777 26.886 349.521 2.198 6.044 78.574 .971 2.672 34.730 .218 .601 7.808 5.581 15.347 199.515 1.255 3.450 44.852

−1.2079 −.1709 −.6162 .5410 −.0180 1.7402 .0289 .8127 −1.4814 3.8123 .5154 1.8583 .6007 −.0887 1.5919 −.2273 −.2424 −3.4122

.113546 .432148 .268874 .294238 .492831 .040911 .488457 .208181 .069247 .000069 .303140 .031565 .274008 .464679 .055707 .410092 .404252 .000322

Type/antitype?

This is also the base model for the Prediction CFA that will be performed below. The overall goodness-of-fit likelihood ratio for this model of independence of alcohol consumption and the mood ratings on the following day, which incorporates the hypothesis that the two forms of alcohol consumption are related to each other, is X 2 = 7.71. For df = 10, this value leads us to the retain the model (p = 0.66) (and we anticipate that no prediction types and antitypes emerge). From the signs of the estimates of the interaction parameter, we conclude that the relationship between beer and liquor consumption is, for respondent 3053, negative, indicating that the more beer the respondent consumes, the less liquor he consumes, and vice versa. From this result, we conclude that, over a lag of one day, there is no relationship of beer and liquor consumption to mood on the following day. Based on this conclusion, we also expect that the log-linear model will suggest no predictive relationship. The overall goodness-of-fit likelihood-ratio for the log-linear model is X 2 = 28.19. For df = 6, this value leads us to reject the null hypothesis of no relationship between predictors (beer or liquor consumption) and criterion (mood on the following day) (p < 0.01). None of the parameters of the effects that link the predictors and the criterion is significant. The Nagelkerke R2 – equivalent is 0.01. This also indicates that close to nothing of the variability in the B × L × M1 cross-classification is explained by the model. A log-linear model that includes the interactions between beer and mood on the following day and liquor and mood on the following day, that is, the model ̂ = 𝜆 + 𝜆B + 𝜆L + 𝜆M1 + 𝜆B,M1 + 𝜆L,M1 log m also fails to describe the data well. The B × L interaction is needed. Evidently, the two-way interaction between beer and liquor consumption is not sufficient to explain the variability of the data that exists in Table 20.32 beyond the main effects. Results from CFA

Type

Antitype

The results from log-linear modeling make us expect that first-order CFA will identify types and antitypes. However, the same results make us also expect that these types and antitypes only indicate where liquor and beer consumption are locally associated. As soon as the interaction between beer and liquor consumption is included in the base model, all types and antitypes can be expected to disappear. This would, again, suggest that there is no relationship between alcohol consumption and mood on the following day.

CFA Models for Developmental Research

Table 20.32 displays the results from first-order CFA (the base model for this CFA was given previously). For both this run and the following Prediction CFA, we use the z-test and the Holland–Copenhaver procedure of 𝛼 protection. The overall goodness-of-fit for the base model of firstorder CFA is the same as for the log-linear model of variable independence discussed previously. The base model is rejected. However, the results in Table 20.32 are, from a CFA perspective, not overly rich. We find only one type, for a cell that contains only 2 ratings, and one antitype. When the corresponding expected frequency is small, we should interpret the findings of types and antitypes quite cautiously. Specifically, as soon as an expected cell frequency is below 0.5, we do not interpret a type. Then what is left in this table is the sole antitype that is constituted by configuration 3 2 3. It indicates that, significantly less often than expected, Respondent 3053 consumed above average quantities of beer together with liquor and was in a positive mood on the next day. From the log-linear modeling results, we already know that Respondent 3053 tended to drink predominantly either beer or liquor on drinking occasions but not both, and that his drinking had no effect on his mood on the next day. Therefore, we expect the type and the antitype to disappear when the beer × liquor interaction is taken into account. This is indeed the case (table of results not shown here). To provide closure on the substantive part of this example, we now ask whether alcohol consumption has effects on the mood as rated for the same day. We find almost the same results. The log-linear model of variable independence is rejected (LR − X 2 = 30.25; df = 12; p < 0.01), and the same antitype emerges. The model that takes the beer × liquor consumption into account describes the data very well (LR − X 2 = 3.77; df = 10; p = 0.96), and no type or antitype emerges. We conclude that, for this respondent, there is no short-term relationship between alcohol consumption and mood. CFA of Cross-Lagged Designs Cross-lagged designs allow researchers to test hypotheses about the temporal relationships between two or more repeatedly observed variables. Consider the two variables, X and Y, each observed twice, that is, the four realizations X1, X2, Y1, and Y2. Figure 20.3 displays the relationships that are usually estimated. All possible two-way relationships can be arranged in the form of a correlation matrix, as in Table 20.33.

X1

X2

Y1

Y2

905

Figure 20.3 Cross-lagged relationship. TABLE 20.33 Correlation Matrix for X1, X2, Y1, and Y2 Variables Variables

X1

X2

Y1

X2 Y1 Y2

rX2 X1 rY1 X1 rY2 X1

rY1 X2 rY2 X2

rY2 Y1

Five of the six relationships listed in Table 20.33 are included in Figure 20.3. The only one missing is the one between X2 and Y2. This relationship was omitted to illustrate that cross-lagged models do not necessarily involve all possible two-way relationships. In standard, aggregate-level, variable-oriented data analysis, the paths in Figure 20.3 are estimated as regression coefficients. The paths between X1 and X2 as well as between Y1 and Y2 are called autoregressions, and the paths between X1 and Y2 as well as between Y1 and X2 are cross-lagged regressions. The relationships between X1 and Y1 as well as between X2 and Y2 are typically estimated as correlations (they could be regressions also). These paths and correlations allow one to test hypotheses concerning 1. The contemporaneous associations between X1 and Y1 as well as between X2 and Y2; and 2. The lagged associations between X1 and Y2 as well as between Y1 and X2. For both manifest and latent variable models, structural modeling is most suitable for estimation. Occasionally, regression analysis is still applied for manifest variable models. One characteristic of such models that is of importance here is that all relationships are expressed at the level of aggregated raw data. That is, all relationships are expressed within a framework of variable-oriented research. In the present context, we take a person-oriented perspective for the analysis of categorical variables. That is, we do not assume that all categories of all variables carry a relationship. Instead, using the concept of local associations (Havránek & Lienert, 1984), we assume that, in many

906

Configural Frequency Analysis for Research on Developmental Processes

cases, only some of the categories of a variable carry an association. Under this assumption, CFA is the method of choice. Therefore, we present here, for the first time, a CFA approach to cross-lagged analysis. This approach is an extension of the CFA approach to lagged relationships that was discussed and illustrated in the last section (von Eye et al., 2010). To develop the base model for cross-lagged CFA of categorical data, we use the general definition of CFA base models. This definition states that a CFA base model contains all relationships that are not of interest to the researchers. If this model is rejected and types and antitypes emerge, they necessarily reflect the cross-lagged relationship under study. To illustrate, consider the model in Figure 20.3. This model contains five of the six possible bivariate relationships. These must, therefore, not be included in the base model. In contrast, the remaining two-way relationship (the one between X2 and Y2) must be included. None of the four possible three-way interactions and the four-way interaction are part of the tested hypotheses. Therefore, they are included in the base model. In log-linear modeling notation, the nonhierarchical base model for cross-lagged CFA is ̂ = 𝜆 + 𝜆X 1 + 𝜆X 2 + 𝜆Y 1 + 𝜆Y 2 + 𝜆X 2,Y 2 log m +𝜆

X 1,X 2,Y 1

+𝜆

X 1,X 2,Y 2

+𝜆

X 1,Y 1,Y 2

+𝜆

X 2,Y 1,Y 2

+ 𝜆X 1,X 2,Y 1,Y 2 As with all other CFA base models, if any of the interactions not included in the model exist, the model is rejected and resulting types and antitypes indicate the configurations that carry the relationship. Also, as was illustrated before, the rejection of a model is a necessary condition but no guarantee that types and antitypes emerge. It is interesting to note that the rejection of the base model of cross-lagged CFA gives no indication concerning which of the hypothesized relationships exist. Therefore, we recommend considering one of two strategies. The first involves estimating a categorical variable path model, in particular when there are more than two observation points. This can be achieved with Mplus (Muthén & Muthén, 2010) or with a specialized program for categorical data analysis such as Lem (Vermunt, 1997). The methods of functional CFA can also be used (von Eye & Mair, 2008a, 2008b, 2008c), which allow one to identify the configurations that carry the two-way interactions of the cross-lagged model.

Beer

Beer lag 1

Mood

Mood lag 1

Figure 20.4 Cross-lagged relationship between beer consumption and mood.

Data Example In the following example, we continue the analysis of the one-lag data of Respondent 3053 from the study of self-identified alcoholic men (Perrine et al., 1994). We ask whether, from one day to the next, beer consumption and mood are related to each other so that it can be described by a cross-lagged model such as the one in Figure 20.3. In this example, the model would be as shown in Figure 20.4. We approach this model first from the variable-oriented perspective. We estimate a log-linear model that includes all the variable relations indicated in the graph. That is, we estimate the model ̂ = 𝜆 + 𝜆B + 𝜆B1 + 𝜆M + 𝜆M1 log m + 𝜆B,M + 𝜆B,B1 + 𝜆B,M1 + 𝜆M,B1 + 𝜆M,M1 . In this section, we use the same abbreviations as in the last. That is, B = beer consumption, B1 = beer consumption on the following day, M = self-rated mood, M1 = self-rated mood on the following day. This is not the CFA base model. Therefore, we hope that this model describes the data well. If this is the case, we perform the CFA of the cross-lagged model described previously, which includes the terms that are not of interest. The base model for this CFA is ̂ = 𝜆 + 𝜆B + 𝜆B1 + 𝜆M + 𝜆M1 + 𝜆B1,M1 log m + 𝜆B,B1,M + 𝜆B,B1,M1 + 𝜆B,M,M1 + 𝜆B1,M,M1 + 𝜆B,B1,M,M1 If this cross-lagged base model is rejected, resulting types and antitypes indicate the interactions that are not specified in this model, which is nonhierarchical. We will know which of the interactions exist because, in this example, we estimate the log-linear model before we perform CFA. The types and antitypes will tell us where, in the cross-classification, these effects become manifest. In the example in the last section, we used the threecategory variables B and B1, and M and M1. Crossed,

CFA Models for Developmental Research

these variables span a 3 × 3 × 3 × 3 cross-classification with 81 cells. For about 800 observations, this table is not too large. However, due to the very uneven frequency distribution, many cells were empty. Sparseness was so extreme that the programs we use were unable to estimate the models. Therefore, we further combined categories of the variables. Specifically, we now use for B and B1 the categories 1 = two or fewer beers consumed and 2 = three or more beers consumed in one day. For M and M1, we use the categories 1 = good mood or below, and 2 indicating better mood. For the log-linear model, we obtain an overall goodnessof-fit likelihood-ratio of X 2 = 7.11. For df = 6, this value leads us to retain the model (p = 0.31), and we conclude that the higher order interactions are not needed to explain the frequency distribution shown in Table 20.34. We ask which of the five two-way interactions are significant. Table 20.35 displays the parameter estimates. Table 20.34 shows an interesting pattern of results. Although the model, overall, describes the data well, the hypothesis that the relationships among the variables that span the cross-classification in Table 20.35 can be described by a cross-lagged model must be rejected. Of the five bivariate relationships included in this model, only one is significant. It is the one between mood on a given day and mood on the following day. We conclude that, for Respondent 3053, mood is predictive of itself over a lag of one day. In contrast, the quantities of beer consumed are hard to predict, even from one day to the next, and there seems to be no relationship between mood and beer consumption. This applies both to the contemporaneous and the lagged beer–mood and mood–beer TABLE 20.34 Cross-Lagged CFA of the B × M × B1 × M1 Cross-Classification Configuration B M B1 M1 1111 1112 1121 1122 1211 1212 1221 1222 2111 2112 2121 2122 2211 2212 2221 2222

m

̂ m

14.00 16.00 7.00 11.00 21.00 296.00 7.00 148.00 3.00 15.00 4.00 5.00 12.00 145.00 7.00 79.00

4.690 19.927 1.348 17.586 25.154 292.911 12.814 146.571 1.571 20.814 0.911 9.154 18.586 139.348 9.927 69.690

z 4.299 −.673 4.867 −1.570 −0.828 0.181 −1.624 0.118 1.140 −1.274 3.238 −1.373 −1.528 0.479 −0.929 1.115

p 0, and the condition of Step 4 would also not be met, and a researcher applying the steps approach would erroneously conclude that no mediation was present. Another apparent difficulty of strictly applying the steps approach is that at Step 4, any difference in the appropriate direction could be deemed evidence of mediation, but given ever-present sampling variability, a stricter criterion requiring a significant difference seems more appropriate. An approximate standard error can be derived via first-order Taylor series expansion or the multivariate delta method (MacKinnon, 2008): sab =

√ b2 s2a + a2 s2b

(21.8)

where s2a and s2b are the sampling variances of the regression estimates of a and b. These can be obtained by squaring the standard errors of a and b produced in all standard regression output or by examining the diagonal elements of the co/variance matrix of the estimates which can generally be requested as additional output in statistical software for regression estimation. This particular form of the standard error of the mediated effect is often referred to as the Sobel standard error (Sobel, 1982, 1986). Baron and Kenny (1986) suggested a slightly more complex approximate standard error, which can be derived via second order Taylor series expansion and represents the exact variance under the condition of independence: sab =

√ b2 s2a + a2 s2b + s2a s2b

(21.9)

A statistical test of the mediated effect can be conducted by dividing either the ab or the equivalent c – c′ estimate of

Page 929

Trim Size: 8.5in x 11in

930

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

the mediated effect by the selected approximate standard error. This should be distributed as a z in large samples, and thus values more extreme than ±1.96 would be considered significantly different from zero at p < .05. However, this large sample z test assumes that the product ab is normally distributed, and the ab estimate of mediation generally is not (MacKinnon et al., 2007). This test has been found to be overly conservative (i.e., to have low power) in simulation studies (MacKinnon et al., 2002). This may or may not be problematic, depending on the particular circumstances in which the test is employed. If a researcher uses this test in an existing data set and finds a significant mediated effect, the fact that the test is conservative and therefore somewhat underpowered does not diminish the significance of the finding in any way, although a confidence interval formed around the significant point estimate of the mediated effect will still be too wide and indicate more uncertainty in the value than is truly warranted. In contrast, a nonsignificant result might indicate a null effect or might simply be the result of an underpowered test. In addition, using this test to determine the necessary number of subjects for a future study in an a priori power analysis would result in an unnecessarily large sample size recommendation. Because of this distributional difficulty, several alternative approaches have been suggested. One such alternative involves looking not at the distribution of ab, but at the product za zb , where za and zb are z scores formed as the ratios of the coefficient estimates to their standard errors. The za zb product is known to have a complex nonnormal distribution, but values obtained in such an analysis can be compared to tabled values (Meeker, Cornwell, & Aroian, 1981) and asymmetric confidence limits can be obtained from software written especially for this purpose (MacKinnon et al., 2007; Tofighi & MacKinnon, 2011). An alternate approach involves a resampling technique known as bootstrapping. Bootstrapping involves repeated resampling with replacement (Efron & Tibishirani, 1993). For example, if a researcher is bootstrapping cases from a sample of 100, many (typically over 1000) new samples of 100 cases are drawn from the original sample. These samples may differ from the original in that a given case may appear more than once in each bootstrap sample. Then the statistic of interest may be calculated for each bootstrap sample. These are ordered and values of the statistic associated with the desired percentiles (often the 2.5th and 97.5th for a 95% confidence interval) define the lower and upper limits of a confidence interval. This procedure may be refined using a procedure known as bias correction to adjust for the fact that the statistic obtained in the original analysis may not fall at the 50th

percentile of the distribution of bootstrap values (Efron & Tibishirani, 1993). In the case of a mediated effect, the cases themselves are resampled and ab estimates produced from regressions involving X, Y, and M in each of the new bootstrap samples are obtained (MacKinnon et al., 2004; Preacher & Hayes, 2004). The simple single-mediator model in Figure 21.1 can be expanded to multivariate situations that include multiple predictors and multiple mediators. With multiple predictor and mediator variables, it is common to run an initial set of analyses examining mediated effects of each predictor–mediator pairing individually, then to run a combined model to examine the unique effects of each variable over and above the others. For example, if we were to expand Figure 21.1 to include three mediators, equation 21.5 examining the effect of the single predictor on the single outcome would remain unchanged, producing a single estimate of the total effect c. The X to M associations examined with equation 21.6 would require three separate regression equations, each predicting a single mediator from the predictor X, producing three unique a coefficients, a1 , a2, and a3 . Equation 21.7 would need to include all three mediators simultaneously along with the predictor to produce three unique b coefficients b1 , b2, and b3 and a single estimate of the direct effect c′ . The c – c′ point estimate of mediation in this case represents the total mediated effect and the a1 b1 , a2 b2 , and a3 b3 point estimates represent specific mediated effects unique to each of the three mediators. A set of cross sectional mediation analyses was conducted in the ECLS-K data. These analyses tested whether the effect of ADHD symptoms on depression was mediated by academic performance and interpersonal skill at the initial kindergarten fall assessment. Specifically, teacher ratings of approaches to learning and externalizing behavior were treated as initial predictor variables, theta scores on math and reading and teacher ratings of interpersonal skill served as mediators, and teacher ratings of internalizing behavior were the outcome. Regression equations similar to those in equations 21.5 to 21.7 were first fitted to each predictor–mediator pair separately, then all predictors and mediators were entered simultaneously into a multiple mediator model. The results of these analyses are summarized in Table 21.2. The third and fourth columns of Table 21.2 present a and b coefficient estimates, with their standard errors below in parentheses. The results of these analyses show that both approaches to learning and externalizing behaviors significantly predict each of the three mediators, and each of the three mediators significantly affects the internalizing

Page 930

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Combining Basic Mediation and Moderation

12:38pm

931

TABLE 21.2 Coefficients, Standard Errors, Tests of Mediation, and Bootstrap Confidence Intervals for Single- and Multiple-Mediator Models Single-mediator models X

M

a

LEARN

INTERP

LEARN

READ

LEARN

MATH

EXTERN

INTERP

EXTERN

READ

EXTERN

MATH

b

0.6564*** (0.0053) 0.3115*** (0.0055) 0.3113*** (0.0049) −0.5545*** (0.0064) −0.1197*** (0.0062) −0.1227*** (0.0056)

Med(Z test)

−0.1515*** (0.0083) −0.0234** (0.0081) −0.0422*** (0.0091) −0.2348*** (0.0073) −0.1315*** (0.0076) −0.1699*** (0.0084)

Multiple-mediator models 99% bootstrap CI: LL, UL

−0.0994*** (0.0055) −0.0073** (0.0025) −0.0131*** (0.0028) 0.1302*** (0.0043) 0.0157*** (0.0012) 0.0208*** (0.0014)

outcome. The ab point estimates and the Sobel standard errors of each mediated effect appear in the fifth column of the table. When examined in separate analyses, each of these mediated effects is significant. In this particular case, the bootstrap 99% confidence intervals in the sixth column of the table are very close to those that would be computed using the normal theory and Sobel standard error provided in the previous column. The fact that zero is not included in any of these confidence intervals indicates significance at p < .01 and corresponds to the z test results. The remaining columns of the table present the results of a multiple predictor–multiple mediator analysis in which all predictors and mediators were analyzed simultaneously. A path diagram for this particular model is shown in Figure 21.2. The two initial predictor variables, approaches to learning and externalizing behavior, are highly negatively correlated (r = –.50, p < .0001). The two academic mediators are strongly positively correlated (.77, p < .0001), and these two measures are moderately positively correlated with interpersonal skill (.27 for math

a1

MATH

LEARN

−0.1282*** (0.0088) 0.0042 (0.0115) −0.0626*** (0.0130) −0.1282*** (0.0088) 0.0042 (0.0115) −0.0626*** (0.0130)

Med (Z test) −0.0668*** (0.0046) 0.0014 (0.0039) −0.0213*** (0.0044) 0.0353*** (0.0025) 0.0003 (0.0007) −0.0037*** (0.0009)

99% bootstrap CI: LL, UL −0.0808, −0.0550 −0.0085, 0.0112 −0.0334, −0.0094 0.0280, 0.0433 −0.0016, 0.0023 −0.0063, −0.0017

and .25 for reading, both ps < .0001). The a path estimates representing the unique effects of each of these predictors (over and above the others) on each of the three mediators appear in column 7 of Table 21.2, and all six of these remain significant, though the effects of externalizing behavior seem somewhat attenuated (when compared with the separate analyses reported in column 3 in the table), and the effects of externalizing on reading and math have, in fact, changed sign. The b path estimates representing unique effects of a particular mediator are provided in column 8. When all variables are in the model simultaneously, the path representing the unique effect of reading on internalizing is not significant (over and above the effects of math, interpersonal skill, approaches to learning, and externalizing behavior). The b paths from interpersonal skill and math to internalizing have also reversed sign in the analyses that simultaneously include the externalizing predictor. The z tests of point estimates of the mediated effects in column 9 and the bootstrap confidence intervals in column 10 indicate that the effects of approaches to learning and externalizing behavior on internalizing behavior are significantly—and uniquely—mediated by math and interpersonal skill but not reading.

COMBINING BASIC MEDIATION AND MODERATION

c´1

a3 READ

EXTERN

0.5213*** (0.0058) 0.3427*** (0.0064) 0.3405*** (0.0056) −0.2755*** (0.0060) 0.0638*** (0.0067) 0.0596*** (0.0059)

b

b1

a2

a4

−0.1145, −0.0842 −0.0147, −0.0014 −0.0212, −0.0072 0.1183, 0.1411 0.0137, 0.0199 0.0165, 0.0236

a

a5

b2 c´2

INTERN

b3

a6 INTERP

Figure 21.2 Path diagram for the two-predictor, three-mediator in the ECLS-K data.

Several researchers have described approaches that combine moderation and mediation into a single framework to examine more complex associations among sets of variables (e.g., Edwards & Lambert, 2007; Fairchild & MacKinnon, 2009; Hayes, 2013; MacKinnon, 2008; Muller, Judd, & Yzerbyt, 2005; Preacher, Rucker, & Hayes, 2007). Such approaches typically consider situations involving both

Page 931

Trim Size: 8.5in x 11in

932

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

moderated mediation and mediated moderation. Moderated mediation, or moderation of an indirect effect (Fairchild & MacKinnon, 2009), is moderation of the ab estimate of a mediated effect by a fourth variable, Z. Z could exert its moderating influence on the a path from X to M, on the b path from M to Y, or both (though it should be noted that circumstances exist when moderation of both paths does not necessarily imply moderation of the ab product). Mediated moderation, or the mediation of a moderator effect (Fairchild & MacKinnon, 2009), involves an exploration of the mechanism through which an interaction between two predictors X and Z (i.e., the XZ cross-product) exerts its effect on Y. The framework described in Fairchild and MacKinnon (2009) attempts to unify the various proposed approaches by providing a single comprehensive presentation that subsumes the specific differing approaches as special cases. Figure 21.3 presents the path diagram for this combined framework. The regression equations necessary to produce the various path coefficients are straightforward generalizations of each of the three regression equations necessary for a simple mediation analysis. For example, equation 21.5 in which the outcome is regressed on the focal predictor becomes Y = b0 + c1 X + c2 Z + c3 XZ + e

(21.10)

XMZ j XM h MZ

b1 Y

c´1 X

Z

XZ

a1

c´2 c´3

b2

a2

a3 M

Figure 21.3 Path diagram of the Fairchild and MacKinnon (2009) approach to moderated mediation and mediated moderation.

The c coefficients in this equation thus provide total effect analogs for each of two predictors, X and Z, as well as their interaction XZ. The coefficient c1 is the effect of X on Y when Z = 0, the coefficient c2 is the effect of Z on Y when X = 0, and the coefficient c3 is the interaction of X and Z on the outcome. Typically, all predictor variables in such an analysis would be centered around their means prior to estimation, which would make the c1 and c2 coefficients the average effects of X and Z on the outcome, respectively. The generalization of equation 21.6 is M = b0 + a1 X + a2 Z + a3 XZ + e

(21.11)

This equation provides a path coefficient estimates for X, Z, and their interaction. The coefficient a1 is the average effect of X on the mediator M, the coefficient a2 is the average effect of Z on the mediator, and the coefficient a3 is the interaction of X and Z on the mediator variable. The generalization of equation 21.7 is Y = b0 + c′1 X + c′2 Z + c′3 XZ + b1 M + b2 MZ + hXM + jXMZ + e

(21.12)

The c′ coefficients in this equation provides direct effect analogs for X, Z, and their interaction XZ. The coefficients c′1 and c′2 are the average direct effects of X and Z on the outcome, and the coefficient c′3 is the direct effect of the XZ interaction on the outcome. The b1 coefficient is the average effect of M on Y, and the remaining paths b2 , h, and j estimate possible moderation of this effect by Z, X, and the XZ interaction, respectively. Moderation of an indirect effect, in this framework, would involve the a3 or b2 coefficients. Fairchild and MacKinnon (2009) suggested that a joint test of these two coefficients may be acceptable to identify moderated mediation (i.e., mediation is deemed significant if both the a and b paths are significant). When the moderator is dichotomous, this effect can be more rigorously tested by calculating the ab point estimate of the mediated effect within each group and testing the significance of the difference between them using a pooled standard error. Details of this procedure as well as that for a continuous moderator are discussed in Fairchild and MacKinnon (2009). Mediation of a moderator effect could be estimated as the a3 b1 product. Table 21.3 presents the results of six analyses examining whether (1) the mediated effects of ADHD symptoms on internalizing behavior through academic performance and

Page 932

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Longitudinal Data

12:38pm

933

TABLE 21.3 Coefficients, Standard Errors, and Tests of Moderated Mediation and Mediated Moderation X

M

LEARN

INTERP

LEARN

READ

LEARN

MATH

EXTERN

INTERP

EXTERN

READ

EXTERN

MATH

a3

0.0018 (0.0102) 0.0324** (0.0109) −0.0042 (0.0097) −0.0545*** (0.0128) −0.0031 (0.0127) −0.0009 (0.0116)

b2

−0.0239 (0.0084) −0.0365* (0.0165) −0.0260 (0.0176) −0.0326* (0.0142) −0.0394* (0.0155) −0.0384* (0.0164)

b1

−0.1551*** (0.0081) −0.0281*** (0.0082) −0.0407*** (0.0088) −0.2399*** (0.0071) −0.1401*** (0.0078) −0.1719*** (0.0082)

interpersonal skill might be moderated by gender, and (2) the moderation effect of gender on the association between ADHD symptoms and internalizing might be mediated by academic performance and/or interpersonal skill. The a3 and b2 coefficients that suggest moderated mediation appear in columns 3 and 4, and the statistical test appears in column 8. The joint test would suggest moderation of the mediated effect by gender for the effect of LEARN on INTERN through READ and the effect of EXTERN on INTERN through INTERP. The test of the difference between the ab estimates of mediation for boys and girls indicates that the latter effect is significant at p < .001, but the former is only marginally significant. The a3 and b1 coefficients that comprise the a3 b1 estimate of mediated moderation appear in columns 3 and 5 of Table 21.3, and the test of mediated moderation is provided in column 9. These results indicate that the interactive effect of LEARN and GENDER on INTERN is mediated by reading performance, and the interactive effect of EXTERN and GENDER on INTERP is mediated by interpersonal skill.

LONGITUDINAL DATA The discussion of methods to detect mediation and moderation above are specific to cross sectional data. The introduction of longitudinal data, with repeated measurements of the same variables over time, provides several benefits that may help to elucidate the associations of interest to a researcher. Many of these are discussed in MacKinnon (2008), with attention to how these are relevant to mediated effects in particular. First, longitudinal data provide information about the temporal precedence of the variables involved in the association. This is especially salient when the researcher wishes to establish causal

MedMale

MedFemale

Moderated mediation test

Mediated moderation test

−0.0948*** (0.0078) −0.0033 (0.0034) −0.0086* (.0040) 0.1164*** (0.0058) 0.0127*** (0.0015) 0.0191*** (0.0019)

−0.1051*** (0.0078) −0.0122** (0.0038) −0.0171*** (0.0044) 0.1450*** (0.0064) 0.0171*** (0.0020) 0.0246*** (0.0025)

0.0103 (0.0110) 0.0089† (0.0051) 0.0086 (0.0060) −0.0286*** (0.0086) −0.0044† (0.0025) −0.0055† (0.0030)

−0.0003 (0.0016) −0.0009* (0.0004) 0.0002 (0.0004) 0.0131*** (.0.0031) 0.0004 (0.0018) 0.0002 (0.0020)

pathways in a mediation model. Causes must precede their effects, and mediation models necessarily involve either implicit or explicit causal claims. In cross sectional data, if the initial predictor is randomly assigned (e.g., randomization to treatment and control groups), then logic dictates that X precedes M and Y in the model. For nonrandomly assigned variables and for the M to Y relation, however, it is often difficult to establish the ordering of the variables, and thus it is common to acknowledge the possibility of reverse or reciprocal causation in such situations. With repeated measurements of the various constructs in longitudinal data, however, the researcher can explore lagged associations and changes in variables over time. If X at Time 1 predicts M at Time 2, which in turn predicts Y at Time 3, it is more likely that the earlier measures cause the later measures than vice versa (though it is still possible that some third variable causes both the earlier and later measures). Testing moderation does not necessarily involve ordering variables as in a mediation model, but here still it can be useful to explore lagged associations and change over time. If an interaction between two variables at Time 1 predicts an outcome at a later time point or change in the outcome over time, this makes a stronger argument for possible causality than examination of all variables cross sectionally. A second benefit of the repeated measurement of constructs in longitudinal data is that it allows the researcher to disentangle processes occurring within and between individuals. In cross sectional data, the only differences that can be detected are those between different individuals at the single time point. With two or more observations, however, it is possible to calculate intraindividual change, as well as differences between individuals. It may well be that the predictors of individual change are quite different from the

Page 933

Trim Size: 8.5in x 11in

934

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

predictors of differences between individuals, hence disentangling the within-person and between-person effects will provide a more fully accurate model of how real-world processes operate to produce their effects on behavior. A third benefit of longitudinal data is that having repeated measurements of variables allows the researcher to rule out some possible alternative explanations for the effect of one variable on another. In cross sectional data, it is possible to argue that virtually any association between two variables is, in fact, due to an unmeasured and omitted third variable that affects them both. With longitudinal data, however, a researcher can covary previous measures of an outcome construct when predicting the current measure of that construct. This sort of autoregressive effect essentially removes the influence of any stable person characteristics that might influence the current outcome measure, as each individual serves as his or her own control. While one could still argue that some characteristic of the individual influences both the Time 1 and Time 2 measurement of a construct, it would be difficult to assert that this stable characteristic influences change in the construct over time. However, it is still possible that some time-varying confound provides a reasonable alternative account of a significant effect, so though use of a longitudinal design guards against some kinds of third variable explanations, it is not a universal panacea. Another benefit of longitudinal data was noted by Gollob and Reichardt (1991), who point out that the effect that one variable has on another may not be instantaneous. It may take some time for certain types of variables to exert their effects, which cross sectional data may not be able to capture. Examining lagged associations in longitudinal data or changes over time may be more illuminating in such cases. The concepts of stability, stationarity, and equilibrium are helpful in describing longitudinal data, longitudinal relationships, and the models used to examine them. Stability refers to the extent to which some feature of the data stays the same across time. That feature may be a variable mean, a rank ordering of variable values, a trend over time, or a periodic process. Stationarity refers to the extent to which relations between variables stay the same across time. Equilibrium refers to the extent to which the pattern of variances and covariances among variables stays the same over time. Equilibrium is related to both stability and stationarity, but a stable and stationary model is not necessarily at equilibrium (Cole & Maxwell, 2003). A cross sectional model can accurately assess relations among variables only after equilibrium has been reached (Maxwell & Cole, 2007). In longitudinal data, stability can be assessed

by measuring the dependency of an observation at a given time point on its value at the previous time point; stationarity can be evaluated by assessing invariance of relations among variables across time points; and some information about equilibrium can be obtained by examining the variance–covariance matrix among the variables across time points (MacKinnon, 2008). The analysis of longitudinal data requires an analytic framework that can accommodate the unique features of such a data structure, most importantly the dependence that arises from the within-person stability of repeated measurements of an outcome measure assessed at multiple time points. Such dependence violates assumptions regarding independent errors that are necessary for many traditional statistical model techniques (e.g., GLM), resulting in downwardly biased standard errors, inflated test statistics, and inflation of the Type I error rate when such methods are improperly applied to longitudinal data. Two alternative analytic frameworks—multilevel modeling and structural equation modeling—are able to appropriately model the within-person similarity and thus are frequently used to analyze data from developmental studies. Below we briefly discuss these two frameworks, how cross sectional moderation and mediation are incorporated into each, and what types of models are commonly used to analyze longitudinal data with each approach.

MULTILEVEL MODELING When data are nested or clustered, application of standard regression analysis techniques can be problematic. For example, when individual subjects are nested within groups, it is quite possible that outcome measures for individuals within the same group may be more similar than a random grouping of individuals would be. This similarity could arise from a number of different sources, including shared group experiences (e.g., students within the same elementary school classroom all share the same teacher), interaction among group members (e.g., students within the classroom talk to each other, which might result in more similar attitudes), or nonrandomly distributed background variables (e.g., students within a school may all live in the surrounding neighborhood, which may be relatively homogeneous in terms of socioeconomic status). Such data would have a positive intraclass correlation (ICC). Such within-group similarity (or between-group mean differences) would not be modeled within a single-level regression analysis, resulting in downwardly biased standard errors and overly large

Page 934

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Multilevel Modeling

test statistics, ultimately increasing the probability of making a Type I error (alpha inflation) and concluding that a significant effect exists more often than warranted (Barcikowski, 1981). Multilevel models—also known as mixed models (Laird & Ware, 1982), random coefficient models (Longford, 1993), or hierarchical linear models (Raudenbush & Bryk, 2002)—are explicitly designed to accommodate such nested data structures. A multilevel model can be conceptualized as a series of models at the various levels of the data. For students nested within schools, the lower level (L1) model for students would be specified as L1:

Yij = 𝛽0j + 𝛽1j Xij + rij

(21.13)

where Yij , Xij , and rij are the outcome score, predictor score, and residual, respectively, for individual i within school j. 𝛽0j and 𝛽1j are intercept and slope values for school j. Thus, one can think of this part of the model as a series of regressions within each of the j groups. Each of the coefficients of the L1 model is treated as a criterion measure in higher or upper level (L2) models. In the students within schools example, these would be specified at the school level: L2:

𝛽0j = 𝛾00 + u0j 𝛽1j = 𝛾10 + u1j

(21.14)

The 𝛽0j intercept for school j is a function of the overall intercept for the sample (𝛾00 ) plus a group-specific deviation u0j . Similarly the 𝛽1j slope of Y on X is a function of the overall slope for the sample (𝛾10 ) plus a group specific deviation u1j . Including error terms at both L1 (rij ) and L2 (the u’s) appropriately models the within-group similarity, thus providing accurate tests of significance of predictor effects. The multiple equations at L1 and L2 may be algebraically combined into a single equation: Yij = 𝛾00 + 𝛾10 Xij + u0j + u1j Xij + eij

(21.15)

The single-equation form of the model is typically estimated using an iterative procedure that alternates between generalized least squares estimates of fixed effects (𝛾00 and 𝛾10 ) and maximum likelihood estimates of the variance components. The L2 u error terms are assumed to be distributed ∼ N(𝟎, 𝝉), and thus the variance components include 𝜏00 (the variance of u0j ), 𝜏11 (the variance of u1j ), 𝜏01 (the covariance between u0j and u1j ), as well

12:38pm

935

as 𝜎 2 (the variance of rij ). Estimating the combined equation rather than performing the two-step procedure implied by the multiple equation form (known as slopes and intercepts-as-outcomes; see Burstein & Miller, 1980) allows for more appropriate weighting of large versus small groups in the analysis and permits estimation even when some groups are rank deficient (i.e., when some groups contain too few individuals to run the within-group regression analysis). This particular analysis, also known as a random coefficient regression model (Raudenbush & Bryk, 2002), essentially describes the distribution of intercepts and slopes across groups, estimating an average intercept 𝛾00 , the variability of intercepts 𝜏00 , an average slope 𝛾10 , and the variability of slopes 𝜏11 . These averages are not the simple arithmetic averages that one could obtain from the two-step procedure but rather an Empirical Bayesian weighting that shrinks less reliable estimates (i.e., those from smaller groups) toward common group estimates to obtain a more representative measure of central tendency. Another common application of multilevel modeling involves repeated measures data where the same constructs are assessed on multiple occasions for a set of individuals. In this scenario, L1 is the repeated measures (within-person) level, and L2 is the individual (between-person) level. The multiple outcome measurements provided by a given individual are more likely to be similar than a random grouping of measurements, and thus we would expect a positive ICC value. When the repeated measurements are considered simple replications and no systematic trend across time is expected, the multilevel framework discussed above can be readily adapted. Such analyses are common, for example, in daily diary studies (see Bolger & Laurenceau, 2013) where days are nested within individuals. Often no trends are necessarily expected across the relatively brief spans of such studies; the multiple days are included simply to obtain many measures of predictor and outcome variables and to allow the researcher to distinguish between within- and between-person effects. The data for a multilevel analysis of repeated measures is typically organized as a long file with one observation for each of the assessments. The multiple-equation form of the model would be L1∶

Yti = 𝜋0i + 𝜋1i Xti + eti

L2∶

𝜋0i = 𝛽00 + r0i 𝜋1i = 𝛽10 + r1i

(21.16)

Page 935

Trim Size: 8.5in x 11in

936

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

One might think of the first equation as a regression across timepoints t for each individual i, providing individual-specific intercept and slope coefficients that serve as outcomes in the individual-level L2 equations. The single-equation form of the model is Yti = 𝛽00 + 𝛽10 Xti + r0i + r1i Xti + eti

(21.17)

This model provides estimates of the overall expected value of the outcome measure when the X predictor is equal to zero (𝛽00 ), the variance in this value across individuals (𝜏00 = var(r0i )) the overall association between X and Y (𝛽10 ), the variance in this association across individuals (𝜏11 = var(r1i )), as well as the residual variance among time points within an individual (𝜎 2 = var(eij )). Moderation in Multilevel Models The aforementioned multilevel equations could be expanded to accommodate a higher level predictor of both intercepts and slopes. This would be a group-level predictor in the scenario with individuals nested within groups and an individual-level predictor in the repeated measures scenario. Going forward with the repeated measures case, the lower level equation would remain unchanged: L1:

Yti = 𝜋0i + 𝜋1i Xti + eti

(21.18)

and each of the L2 equations would be expanded to incorporate the individual-level predictor: L2:

𝜋0i = 𝛽00 + 𝛽01 Wi + r0i 𝜋1i = 𝛽10 + 𝛽11 Wi + r1i

(21.19)

The single-equation form of the model becomes Yti = 𝛽00 + 𝛽01 Wi + 𝛽10 Xti + 𝛽11 Wi Xti + r0i + r1i Xti + eti (21.20) This model estimates four fixed effects. The 𝛽00 estimate is the expected value of the outcome when both the L1 and L2 predictors are equal to 0. The 𝛽 01 estimate is the effect of the individual-level predictor W on the outcome when the repeated measures predictor X is equal to 0 (i.e., the simple effect of W). The 𝛽10 estimate is the overall effect of the repeated measures predictor X on the outcome when the individual-level predictor W is equal to 0 (i.e., the simple effect of X). The 𝛽 11 estimate is the cross-level interaction between the repeated measures predictor X

and the individual level predictor W. Thus this estimate represents (1) the extent to which the association between X and Y changes for each unit increase in W, and/or (2) the extent to which the association between W and Y changes for each unit increase in X. Cross-level interactions are not entirely unique to multilevel models, but they arise most naturally within the multiple equation form within this framework. The t test of the 𝛽 coefficient associated with the cross-product (i.e., 𝛽11 ) is thus often an appropriate test of moderation when predictor and moderator variables are assessed at different levels of a hierarchical data structure. In addition to cross-level interactions, multilevel models can accommodate single-level interactions at either the lower or the higher level of the data structure. The techniques involved are essentially identical to the cross-product methods used in single-level regression models. Moderation of the effect of a repeated measures variable by another repeated measures variable would require expanding the L1 equation to include the second variable and the L1 cross-product vector: Yti = 𝜋0i + 𝜋1i Xti + 𝜋2i Zti + 𝜋3i Xti Zti + eti (21.21) and additional L2 equations would need to be specified for the two new L1 coefficients: L1:

L2:

𝜋2i = 𝛽20 + r2j 𝜋3i = 𝛽30 + r3j

(21.22)

The 𝛽20 coefficient represents the simple effect of the Z predictor when X equals 0, and the test of the 𝛽30 coefficient tests the overall moderated effect. Note that if an L2 predictor was added to the L3 equation for 𝜋3i , the associated coefficient would represent the three-way interaction among L1 X, L1 Z, and the L2 predictor. The fact that this is a three-way interaction would be apparent in the single-equation form of the model. Single-level moderation at L2 would involve expanding (at least) the first of the L2 equations (which predicts the L1 intercept) to include the second individual-level predictor and the cross-product vector: 𝜋0i = 𝛽00 + 𝛽01 Wi + 𝛽02 Zi + 𝛽03 Wi Zi + r0i (21.23) The 𝛽 02 coefficient represents the simple effect of Z when W equals 0, and the test of the 𝛽03 coefficient tests the moderated effect. Note that if the individual-level Z predictor and the WZ cross-product are added to other L2 equations L2:

Page 936

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Multilevel Modeling

(which predict L1 slope coefficients associated with particular L1 variables) the associated coefficients would represent three-way interactions among L2 W, L2 Z, and the particular L1 variable whose L1 parameter estimate serves as the outcome for the L2 equation. Calculation, testing, and plotting simple slopes for interactions in multilevel models is carried out in exactly the same way as in single-level regression models. The issue of centering, which features prominently in discussions of interactions within the single-level regression framework, is no less important when considering multilevel models. Centering serves the same two purposes, providing meaningful interpretations of lower order coefficients and increasing the stability of model estimation. In fact, the latter probably provides greater benefit in multilevel models than in single-level regression analyses, as estimation is much more difficult in the multilevel case. When considering an interaction between two L2 predictor variables, centering is essentially carried out just as in single-level regression: subtracting the mean of each L2 predictor from each observation of that predictor and creating the cross-product from these centered variables reduces nonessential collinearity and allows interpretation of lower order effects of one variable at the mean of the other. One slight wrinkle does emerge: the analyst must decide whether to calculate the mean of the L2 variable to use in the centering as a simple unweighted function, counting each L2 score equally, or as a weighted function which adjusts for the number of L1 units within each L2 unit (i.e., individuals within groups or observations within individuals). Of course, there will be no difference between the two when the number of L1 units per L2 unit is identical across all groups/individuals. When one or both of the variables involved in the interaction is at L1, however, centering becomes a slightly more complex issue. Consider first the situation with individuals nested within groups. Kreft, de Leeuw, and Aiken (1995) discussed two possible means to use in centering L1 predictors. Centering around the grand mean (CGM) simply involves subtracting the single grand mean value of the L1 predictor from each predictor observation. Centering within contexts (CWC) involves subtracting the mean of each group from the predictor observations within each group. Thus while GGM is a simple linear transformation of the raw variable score, CWC is not. Both provide meaningful (though different) zero points at which to interpret lower order coefficients, and both confer some computational advantage, though this is typically greater for CWC, which orthogonalizes effects at L1 and L2. Additional discussion of centering issues for multilevel models with individuals nested

12:38pm

937

within groups can be found in Hofmann and Gavin (1998) and Enders and Tofighi (2007). Repeated measurements of time-varying predictors can be centered with CGM or CWC analogs that subtract the mean value of the L1 predictor for the entire sample or a person-specific mean from each observed score. Hoffman and Stawski (2009) explain how such centered L1 variables can be entered into multilevel models that simultaneously include the means as L2 predictors to disentangle between-person and within-person effects in repeated measures data. Mediation in Multilevel Models Mediation in multilevel models is tested in essentially the same matter as for single-level regression models. Each of the three regression models necessary for producing the coefficients for a mediation analysis may be recast as a multilevel model, including an L2 intercept error term to appropriately model clustering of individuals within groups or repeated measures within individuals. Details are provided in Krull and MacKinnon (1999, 2001). The multiple equation forms of such models would differ depending on the level of measurement of X and M. Krull and MacKinnon (2001) adopted a model naming convention related to the sequential ordering of variables in a mediation model, where X → M → Y . In a 1 → 1 → 1 model, the X, M, and Y variables would each be measured at the lower level of the data hierarchy. So the initial variable, the mediator, and the outcome variable would all be characteristics of individuals in multilevel data with individuals nested within groups or time-varying characteristics measured on multiple occasions over a period of time in multilevel data with repeated measures nested within individuals. In a 2 → 1 → 1 model the initial predictor would be a L2 variable, while the mediator and outcome would be L1 constructs. In a 2 → 2 → 1 model, both the initial predictor and mediator are measured at L2, while the outcome is at L1. Models including a link from a lower level to a higher level are not estimable within a multilevel modeling framework, but these could potentially be handled within a multilevel structural equation model (Preacher, Zyphur, & Zhang, 2010). Fitting a 2 → 2 → 1 mediation model would involve running a regression model on L2 data to obtain an estimate of the a path linking predictor to mediator and a multilevel model of the form L1:

Yij = 𝛽0j + eij

L2:

𝛽0j = 𝛾00 + c′ Xj + b Mj + u0j

(21.24)

Page 937

Trim Size: 8.5in x 11in

938

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

to obtain an estimate of the b coefficient. These would be multiplied to produce the ab estimate of the mediated effect, and this can be tested using the Sobel standard error described earlier for single-level mediation models or more precisely by developing bootstrap confidence intervals (Pituch, Stapleton, & Kang, 2006). For a 2 → 1 → 1 model, the multilevel model equations for obtaining the a path are L1: L2:

Mij = 𝛽0j + eij 𝛽0j = 𝛾00 + aXj + u0j

(21.25)

And the equations for the b path are L1:

Yij = 𝛽0j + 𝛽1j Mij + eij

L2:

𝛽0j = 𝛾00 + c′Xj + u0j 𝛽1j = b

(21.26)

Note the similarity in form for the b path in the 2 → 2 → 1 model and the a path in the 2 → 1 → 1 model, both of which involve 2 → 1 links. For a 1 → 1 → 1 model, the multilevel equations for the a path are L1:

Yij = 𝛽0j + 𝛽1j Xij + eij

L2:

𝛽0j = 𝛾00 + u0j 𝛽1j = a

(21.27)

And the multilevel equations for the b path are L1:

Yij = 𝛽0j + 𝛽1j Xij + 𝛽2j Mij + eij

L2:

𝛽0j = 𝛾00 + u0j 𝛽2j = b

(21.28)

Again, note the similarity between the multilevel equations for these paths and those for the b path in the 2 → 1 → 1 model, which also involves a 1 → 1 link. Equations for producing c path estimates are not provided here but can be found in Krull and MacKinnon (2001). Two complications arise when working with multilevel rather than single-level equations for estimating the various paths in a mediation model. The first is that the two alternate point estimates of the mediated effect, ab and c − c′ , are generally not identical in the multilevel case.

The reasons for this are discussed in Krull and MacKinnon (1999). The differences are usually quite small and become smaller with increasing sample size. The second complication is unique to the 1 → 1 → 1 model. The previously specified equations allow for random intercepts to accommodate within group similarity, but slopes of all other L1 coefficients are specified as fixed. It may, in fact, be that the effects of the initial predictor on the mediator (i.e., the a path) or the effect of the mediator on the outcome (i.e., the b path) vary across L2 units. This would mean that different groups have different a and/or b paths in multilevel data with individuals nested within groups or that different individuals have different a and/or b paths in multilevel data with repeated measures nested within individuals. To appropriately model such data, it will be necessary to include random errors on the L2 equations predicting the L1 slope coefficients. If it should be the case that both the a path and the b path in a 1 → 1 → 1 model randomly vary across L2 units, it is possible that these a and b coefficients might covary. In this case, the point estimate of the mediated effects is not equal to ab but rather Med = ab + 𝜎ab

(21.29)

which involves the covariance between the a and b estimates (Kenny, Korchmaros, & Bolger, 2003). Moreover, the approximate standard error is not equal to equation 21.8 but rather sMed =

√ 2 b2 𝜎a2 + a2 𝜎b2 + 𝜎a2 𝜎b2 + 2ab𝜎ab + 𝜎ab

(21.30)

which involves the covariance between the a and b estimates as well as the sampling variance of the covariance estimate (Kenny et al., 2003). Finding an estimate of the covariance between the a and b estimates is challenging because these are typically produced in two completely separate analyses. Bauer, Preacher, and Gil (2006) presented a method that fits both the a and b paths in a single analysis, which allows the covariance and its corresponding sampling variance to be estimated. Their approach involves stacking the data so that each L1 unit has two observations with identical predictors and creating a generic outcome variable. The outcome variable is set equal to the M variable on the first observation for each L1 unit, and the outcome variable is set equal to the Y variable on the second observation. Then two selection variables (Sm and Sy ) are created to distinguish between the first (Sm = 1, Sy = 0) and second

Page 938

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Multilevel Modeling Approaches for Longitudinal Data

(Sm = 0, Sy = 1) observations for each case. A multilevel model with a L1 equation of the form: L1: Zti = Smti (1 + ai Xti ) + Syti (1 + bi Mti + c′i Xti ) + eZti = Smti + ai Smti Xti + Syti + bi Syti Mti + c′ti Syti Xti + eZti

(21.31)

is then fit to the stacked data, and each of the five L1 coefficients is allowed to vary randomly at L2. This model essentially fits multilevel models for the a path and the b and c′ paths simultaneously. To model the error variance appropriately, it will be necessary to allow the estimation of two separate values, corresponding to the observations with different values on one of the two indicator variables. Further details on how to implement the Bauer et al. (2006) method in a number of statistical software programs can be obtained from the website of the article’s first author. More information on implementation of the model in the Mplus programming language is also available (MacKinnon, 2008; Preacher et al., 2010). The multilevel mediation model is summarized in MacKinnon and Valente (2014) and a Bayesian formulation for the multilevel mediation model is in Yuan and MacKinnon (2009).

MULTILEVEL MODELING APPROACHES FOR LONGITUDINAL DATA The dominant multilevel model for the analysis of longitudinal data is the multilevel growth model (Hox, 2010; Raubenbush & Bryk, 2002; Singer & Willett, 2003). Growth in this sense refers to any sort of change over time, including increasing values, decreasing values, or trends that change direction. For longitudinal data that involve an intervention delivered at a particular point in time, an exponential decay model might provide a useful alternative that is especially well-suited for examining the change that follows the delivery of the intervention. Both approaches are discussed in detail below.

Multilevel Growth Modeling A multilevel growth model is very similar to the two-level model for repeated measurements nested within individuals shown in equations 21.16 and 21.17. The primary difference is that the growth model includes at least one variable representing time at L1. If a single time variable is linearly

12:38pm

939

coded and no covariates are included in the model, the multiple equation specification for the model would be L1:

Yti = 𝜋0i + 𝜋1i timeti + eti

L2:

𝜋0i = 𝛽00 + r0i 𝜋1i = 𝛽10 + r1i

(21.32)

where timeti is the time relevant variable. Each individualspecific growth parameter 𝜋 is a function of the overall estimate of the growth parameter 𝛽 in the sample and individual specific deviation r from that value. The single-equation form of the model is Yti = 𝛽00 + 𝛽10 timeti + r0i + r1i timeti + eti

(21.33)

𝛽00 estimates is the overall intercept, which is the expected value of the outcome variable when time is at its zero point. 𝛽10 estimates the overall linear slope estimate. The 𝜏 matrix contains the variances and covariances of the L2 errors r, and thus 𝜏00 is the variance of individual-specific intercepts, 𝜏11 is the variance of individual-specific slopes, and 𝜏 01 is the covariance between them. This particular parameter estimate is more often interpreted in growth models than in other multilevel models that do not involve time, because it can indicate whether individuals are becoming more similar (compensatory growth) or different (fan-spread growth) in their outcome variable scores. More complex forms of growth can be modeled by adding additional variables coding time information to the L1 model. For example, higher order polynomial trends can be fitted by adding powered functions of the linear time variable. Piecewise models can also be fitted by adding additional coded vectors to produce models that estimate either (1) the slopes of the different time segments, or (2) one slope and the difference between slopes (Raudenbush & Bryk, 2002). Both individual and time-varying covariates can be added to the model. Individual covariates are added to the L2 specification as predictors of intercepts and/or slopes. For example, a gender variable could be added to the L2 equations in 21.32 to test whether intercepts and slopes of the outcome variable differed for males and females. A time-varying covariate could be added to the L1 specification either to adjust for its effect when estimating growth or to examine its effect over and above growth. A decision would need to be made about whether the effect of this variable would vary across individuals (in which case the L2 equation associated with this variable

Page 939

Trim Size: 8.5in x 11in

940

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

would include an r error term) or would be assumed fixed (no L2 error term). Most information relevant to multilevel models in general applies equally to multilevel growth models, with just a couple of notable exceptions. First, the time variable in a growth model is typically “centered” on a meaningful value—often the initial time point of the study—rather than on the mean per se. Centering time at its mean would simply put the zero point in the middle of the study period, which may not be the most substantively meaningful point at which to evaluate effects. Second, typical independence and homoscedasticity assumptions about L1 errors may not be realistic in longitudinal data. L1 error variances may increase or decrease over time, and values measured at later time points may depend on previous values measured at earlier time points. Most multilevel modeling software includes options for specifying heteroscedastic and autoregressive error structures that might be better suited for growth models. Multilevel Exponential Decay Model In considering models for longitudinal data, a primary issue is the shape of the trajectory across time. As discussed by Fritz (2014), often when researchers encounter data that exhibit any nonlinearity, a quadratic growth model is chosen without regard for the implications to the underlying variables or theory. For longitudinal psychological data, especially following an intervention or treatment, Fritz suggests that an exponential decay model has several advantages over a quadratic model. First, the exponential decay function is more realistic for many variables in psychology because it approaches an asymptote. One can imagine many scenarios where an individual is given a treatment, causing a change in the individual’s underlying construct, which slowly returns to the individual’s pretreatment level over time. At the same time, the individuals in the control group remain unchanged, a scenario illustrated in Figure 21.4. If the treatment had a lasting effect on the individual, the individual’s score would still likely decrease across time, but not back to his or her pretreatment level (i.e., the construct would asymptote at a different value than the pretreatment level). While a quadratic growth model can approximate the exponential decay present in the figure, a quadratic function at some point must reach a minimum and then begin to grow towards infinity, an unlikely scenario given most variables in psychology. For this reason, many researchers have already used an exponential decay function (e.g., Blozis, 2004, 2007; Blozis, Conger, & Harring, 2007; Browne, 1993; Cudeck, 1996).

Control Group Treatment Group

Treatment

Time

Figure 21.4 Illustration of a treatment group decaying exponentially across time while the control group remains constant.

Another reason for choosing an exponential decay model is that exponential decay is a parsimonious function requiring the estimation of three parameters: Yi = 𝛽2 + (𝛽0 − 𝛽2 ) exp[𝛽1 t] + ri

(21.34)

where t is time, 𝛽0 is the initial state value at t = 0, 𝛽2 is the asymptote value as time goes to infinity, and 𝛽1 is the rate of change value, which is constrained to be less than zero to ensure that the asymptote exists (Pinheiro & Bates, 2000). While a quadratic model also contains three parameters, the interpretations of the coefficients from the exponential decay model are likely to be more useful to a researcher than those from a quadratic model. Analyzing longitudinal data using an exponential decay model requires the use of a nonlinear multilevel model. Using notation from Raudenbush and Bryk (2002; see also Davidian & Giltinan, 1995), the exponential decay model is L1:

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + rti

L2:

𝛽0i = 𝛾00 + u0i 𝛽1i = 𝛾10 + u1i 𝛽2i = 𝛾20 + u2i

(21.35)

where timei , 𝛽0i , 𝛽1i , and 𝛽2i have the same interpretation as in equation 21.34, rti are the Level 1 residuals, and the uji ’s are random effects distributed ∼ N(𝟎, 𝝉). As in linear multilevel models, time-varying predictors can be added at L1, while time-invariant predictors are added at L2. The inclusion of L2 predictors may be especially interesting for exponential decay because they test whether variance in the initial value, rate of change, and asymptote values can be predicted by another variable. Two of the most widely used software packages for estimating nonlinear multilevel

Page 940

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Structural Equation Modeling

models in the social sciences are NLME (R Development Core Team, 2013) and SAS PROC NLMIXED (SAS Institute, 2014), although many other statistical packages for linear multilevel models also allow the estimation of nonlinear multilevel models.

12:38pm

941

M a

X

b

c´

Y

STRUCTURAL EQUATION MODELING Structural equation modeling is a multivariate analytic technique for exploring patterns of association among variables. A typical structural equation model (SEM) has two components: (1) a measurement model that specifies how observed indicator variables are related to unobserved latent constructs; and (2) a structural model that specifies how the various observed and latent variables are related to one another. Structural equation modeling employs a matrix representation to specify a number of simultaneous equations that denote the associations among observed and latent variables, and the model can be fitted using estimation methods that maximize the fit between the observed covariance matrix of the variables and the covariance matrix implied by the model. In depth treatment of this topic can be found in a number of texts, including Bollen (1989) and Kline (2011). Diagrams are often used to depict the variables and relations specified in an SEM. By convention, observed variables are typically represented as squares or rectangles, and latent variables are typically represented as circles or ovals. Single-headed arrows represent directional causal relations between variables and double-headed arrows represent nondirectional correlational associations. The estimates produced in an SEM typically include factor loadings relating observed indicators to latent factors, path coefficients relating variables and factors to one another, variances and covariances of errors/disturbance terms. It is also possible to estimate means of latent variables, although these parameters are not typically part of standard SEM estimation. Mean structure can be indicated in SEM diagrams by arrows from a triangle representing a constant value of 1, though often such effects are not noted in the diagrams. In addition to parameter estimates, SEM programs provide a number of absolute and relative fit indices that allow researchers to examine how well the proposed model fits the data. These include a chi-square test of overall model fit, the root mean square error of approximation (RMSEA; Steiger, 1990), and the comparative fit index (CFI; Bentler, 1990). Particular aspects of the model can be tested by examining the size of the estimated coefficients relative to their standard errors or

Figure 21.5 variables.

SEM mediation model for manifest X, M, and Y

by comparing the fit of a series of models in which specific parameters, originally freely estimated, are systematically constrained. Mediation in SEM Specification and estimation of mediation in SEM is straightforward. Figure 21.5 depicts a simple SEM for a single mediator model. All three variables in this model are observed, and thus this particular SEM is a simple path analysis and does not include a measurement model. Fitting the model produces estimates of the a and b path coefficients and their standard errors. These can be used to test mediation in much the same way as with regression analysis. Some SEM programs offer specific options for testing mediated (or indirect) effects, using either delta standard error method tests or bootstrapping methods. If variables are measured with multiple indicators, SEM can be used to examine mediation among the underlying latent constructs. A measurement model can be specified to define the initial, mediator, and outcome factors from their various observed indicator variables (as shown in Figure 21.6), and the structural model specifies the appropriate paths between the latent constructs to test the mediated effect. The latent constructs are theoretically error-free, and the path coefficients will not suffer from the attenuation that occurs in estimated associations between observed measurements. Thus the mediated effects among latent constructs should be stronger and more easily detected than mediation among observed variables. Moderation in SEM There are a number of different approaches for examining moderation within an SEM. If the moderating variable is observed and categorical, the typical method involves fitting a multiple group model (Jöreskog, 1971). Such a model essentially fits a separate SEM within each level of the categorical moderator. For example, to examine

Page 941

Trim Size: 8.5in x 11in

942

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

M1

M2

X

M3

Z

Y

ηM b

a

XZ ηY

ηX

Figure 21.7 variables.

c´

X1

X2

Figure 21.6

X3

Y1

Y2

Y3

Mediation model with latent X, M, and Y variables.

whether the association between impulsive/hyperactive symptoms and interpersonal skill is moderated by gender, an SEM that allows different coefficients for the EXTERN → INTERP path would be fitted. The chi-square value for this model would be compared with that from a nested model that constrained the coefficients in the two groups to be equal. A significant chi-square difference would indicate that model fit improves when the two coefficients differ, supporting the hypothesis that the association is moderated by gender. If the two variables involved in the moderated effect are observed, but neither is categorical, a multiplicative cross-product variable can be created and entered as an additional observed variable in the SEM. This cross-product variable then (1) is allowed to covary with the two component variables from which it was produced, and (2) predicts the outcome variable as shown in Figure 21.7. A significant path coefficient from the cross-product variable to the outcome measure would indicate moderation. A moderated effect involving latent variables is somewhat more difficult to model. The most common approach to representing a latent variable interaction involves specifying an additional latent variable—essentially the cross-product of the two latent variables of interest—with cross-products of the original variable indicators serving as indicators of the interaction construct. This latent cross-product variable represents the interactive effect on the outcome measure when fitted with the two original latent variables simultaneously in the model. Specification

An SEM with an interaction between manifest

and fitting of this model can be difficult, as it may involve complex nonlinear constraints, include many highly correlated variables, and violate normality assumptions. A number of different methods have been proposed to address these issues, but no clear consensus as to the best method for analyzing latent variable interactions has emerged. The various issues and proposed solutions are summarized in Kline (2011, Ch. 12).

STRUCTURAL EQUATION MODELS FOR LONGITUDINAL DATA When longitudinal data are to be used in an SEM, the data set is typically organized in wide format (as opposed to the long format necessary for multilevel modeling), with one line of data per individual subject and multiple uniquely named variables representing the repeated measures of each construct. A number of different types of SEMs can be used to model longitudinal data, including autoregressive panel models, latent growth models, and latent change models. It is also possible to fit an exponential decay model within the SEM framework. Autoregressive Model In an autoregressive panel model (Duncan, 1969; Humphreys, 1960; Jöreskog, 1979; Wiley & Wiley, 1970), a measure of a given construct at a particular time point is presumed to be a function of value of that construct at the previous time point. The top series in Figure 21.8 illustrates an autoregressive model for an X variable, and the lower series illustrates a similar model for a Y variable. The arrows from time t to t + 1 within each series (i.e., X1 to X2 and X2 to X3 ; Y1 to Y2 and Y2 to Y3 ) are the autoregressive coefficients representing stability

Page 942

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Structural Equation Models for Longitudinal Data

X1

X2

X3

L1 Predictor1

L1 Predictor2

L1 Predictor3

Y1

Y2

Y3

X1

X2

X3

Figure 21.8 Autoregressive model with cross-lagged paths.

over time. Such models can be estimated with a series of OLS regression models or in a single simultaneous SEM. Autoregressive models for particular variables can be combined to form more complex models. For example, Figure 21.8 includes two types of cross-lagged paths: (1) paths from earlier measures of X to later measures of Y, and (2) paths from earlier measures of Y to later measures of X. Testing such a cross-lagged model—usually by fitting a series of SEMs that constrain particular sets of paths to be equal to zero or equal to each other (see Martens & Haase, 2006)—can allow a researcher to determine whether one set of effects leads the other or if there is reciprocal causality between the two constructs.

1

1

1

1

12:38pm

943

2

Linear Slope

Intercept

L2 Predictor

Figure 21.9 SEM latent growth curve model with L1 and L2 predictions.

Latent Growth Curve Model Another common SEM for longitudinal data is the latent growth curve (LGC) model (Duncan, Duncan, & Strycker, 2006). In such a model, multiple measurements of a given construct over time serve as indicator variables for latent growth factors. While typical SEM measurement models estimate the factor loadings that relate the latent variable to the observed indicators, the loadings in a latent growth model are fixed to define the shape of the growth trajectories. For example, Figure 21.9 presents a linear growth trajectory modeled based on three measurement time points. The factor loadings on the intercept factor are set to 1. The factor loadings on the linear slope factor are set to represent the amount of time elapsed since the first measurement occasion while setting the loading for the time point represented by the intercept factor to 0. Therefore, in the ECLS-K data, if we model linear growth from the kindergarten fall to fifth grade spring with kindergarten fall as the intercept, the first loading would be set equal to zero (essentially erasing the arrow) for the kindergarten fall assessment and the second loading would be set equal to .5, indicating that, on average, half of a year has elapsed

at kindergarten spring. The other loadings would be set equal to 1.5, 3.5, and 5.5 to represent elapsed time at the first, third, and fifth grade spring assessments. If we model a quadratic growth trajectory, another growth factor, a quadratic growth factor, is added to the linear growth model in Figure 21.9, with the factor loadings equal to the squared linear values: 0, .25, 2.25, 12.25, and 30.25. Estimating an SEM for a latent growth model produces means and variances for each of the latent growth factors. The means describe the normative growth trajectory. In a linear growth model, means of the intercept and linear growth factors would represent the expected value of the construct at the initial time point and the linear growth rate, respectively. The estimated variances of the growth factors represent differences between individuals in intercepts and linear rates of change. The covariance between intercepts and linear slopes is also estimated, along with an estimate of the error variance associated with the measurement of the outcome variable at each time point. Both time-varying and individual (invariant over time) covariates can be added to a latent growth model.

Page 943

Trim Size: 8.5in x 11in

944

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

Individual covariates are added as predictors of the latent growth factors, as shown in the lower portion of Figure 21.9. A time-varying covariate measured at a given time point can be added as a predictor of the outcome construct at that time point, as shown in the upper portion of Figure 21.9. Alternately, a researcher could let a time-varying covariate measured at an earlier time point predict the outcome construct measured at later time point, introducing an aspect of temporal precedence and a prospective lagged effect. SEM and multilevel models can produce equivalent estimates of growth model parameters (Curran, 2003). However, the two modeling frameworks differ somewhat in their defaults for model specification and estimation, and thus results from the two approaches may initially appear to differ. Specifically, multilevel models estimate a single L1 variance component (assuming homogenous error variances across time points), while SEMs usually estimate time point-specific error variances. Also, multilevel programs will usually constrain negative variance estimates to zero, while SEM programs will allow the negative estimate. However, if program defaults are appropriately changed to reflect common settings, results produced with multilevel models and SEMs will be essentially identical.

Latent Change Model Another type of latent variable model for investigating longitudinal data is the latent change score (LCS) model, also referred to as the latent difference score model (Ferrer & McArdle, 2003; McArdle, 2001; McArdle & Hamagami, 2001; McArdle & Nesselroade, 2003). An example of a four-wave LCS model is shown in Figure 21.10, where the 𝜂 variables represent the error-free true score for each measurement wave, the Δ variables represent the latent change scores, and the 𝜉1 variable is a second-order slope variable that represents the change in the latent change scores over time or, in other words, the change in the rate of change in score over time. Because each 𝜂 variable is represented by only one observed score, the paths between all of the observed Y variables and the 𝜂 variables are constrained to be equal to 1. A measurement model can be added to the LCS model to incorporate multiple indicators; Figure 21.11 shows an example of a four-wave multiple-indicator LCS model. McArdle and Hamagami (2001) described the Δ variables as being a rate of change for a given time lag such that for a time lag of 1, Δt = 𝜂t − 𝜂t−1

Y1

Y2

Y3

Y4

η1

η2

η3

η4

j1

j2

Δ1

Δ0

j3

Δ2

k1

k2

Δ3

k3

1

ξ1

Figure 21.10 A four-wave latent change score model.

1

(21.36)

Page 944

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Structural Equation Models for Longitudinal Data

Y11

Y21

Y31

Y12

η1

Y22

Y32

Y13

η2

j1

Δ0

Y23

Y33

η3

Δ1

Y34

j3

Δ2

k1

Y24

945

η4

j2

1

Y14

12:38pm

k2

ξ1

Δ3

k3

1

Figure 21.11 Multiple-indicator latent change score model.

except for Δ0 . An individual latent change score is then equal to (21.37) Δt = kt 𝜉1 + jt 𝜂t−1 where jt is the loading between the latent change score at time t to the latent construct score at time t – 1, and kt is the loading of the latent change score on the 𝜉 1 variable. Assuming equally spaced measurement, the values of kt are usually constrained to be equal across waves, as are the values of jt , but unequal spacing of measurements can be taken into account by adjusting these constraints to reflect the actual spacing. In addition, Ferrer and McArdle (2003) described a cross-lagged version of the LCS model where the change score at time t predicts the observed score at time t + 1. Equation 21.37 shows that the latent change score Δt is proportional to the score at the previous time point, jt 𝜂t−1 , and the latent slope, kt 𝜉1 . That is, an individual’s change score at time t is a function of the overall change across time and the individual’s score at the previous time point, which is why McArdle and Hamagami (2001) called this model a dual change score model. Constraining either the j paths, k paths or both to be equal to 0 changes the interpretation of the model. For instance, if the k paths are set to 0, the model becomes a proportional change score model because an individual’s change score at time t only depends on their score at the previous time point, which will not be the same for each individual or time point. Alternatively

if the j paths are set to 0, the model becomes a constant change score model because an individual’s score at time t only depends on the latent slope, which is the same for all individuals across all time points. If the k and j paths are both set to 0, then the latent change score is not proportional to the previous score or related to the latent slope. When this occurs, there is no change in the construct scores across time, so the model becomes a no-change score model. Much like a linear LGC model, the mean of Δ0 can be considered as a latent intercept because it is the average score across individuals at baseline and the mean of 𝜉 1 can be considered a latent slope because it is the average change in change scores across time. Overall model fit can be determined by typical SEM fit indices such as the chi-square goodness of fit test, the RMSEA, and the CFI. Stationarity in the LCS model can be investigated by comparing models that constrain the autoregressive paths, j paths, and k paths of the model to be equal across waves with models where the constraint is lifted, using a series of nested model chi-square change tests.

Exponential Decay Structural Equation Model A latent variable version of the exponential decay nonlinear mixed effects model can be specified using the general method for nonlinear latent variable growth curve models described by Browne (1993; see also Blozis, 2004, 2007;

Page 945

Trim Size: 8.5in x 11in

946

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

Blozis et al., 2007). The nonlinear LGC model for exponential decay is similar to the LGC model and is of the form X = 𝝉 x + 𝚲x 𝝃 + 𝜹

(21.38)

where the vector 𝝃 contains the latent growth parameters, the paths in the loading matrix 𝚲x are constrained to specific values, and the 𝝉 x vector contains the means of the observed variables. Also present is the vector 𝜿, which contains the means of the latent variables, in this case the growth parameters. The difference between a linear LGC model and one for exponential decay is in the interpretation of the means of the latent variables and in the values of the constrained path loadings in the 𝚲x matrix. To construct an exponential decay LGC model using Browne’s (1993) method, illustrated in Figure 21.12, the loadings in the 𝚲x matrix must be set equal to the first partial derivatives of the exponential decay function 𝜆0t =

𝜆1t =

d f (t, 𝜽) = Exp[t𝛽0 ] d𝛽0

(21.39)

d f (t, 𝜽) = (𝛽0 − 𝛽2 )tExp[t𝛽1 ] d𝛽1

(21.40)

d f (t, 𝜽) = 1 − Exp [t𝛽2 ] b𝛽2

(21.41)

𝜆2t = where

f (t, 𝜽) = 𝛽2 − (𝛽0 − 𝛽2 )Exp [t𝛽1 ]

1 0

ξβ1

ξβ0

X1

Figure 21.12 model.

X2

MODELS FOR MODERATION AND MEDIATION IN LONGITUDINAL DATA The next section of the chapter describes several models that incorporate mediation and moderation into the analysis of longitudinal data. Most employ the multilevel and SEM frameworks for model specification and estimation. Some of these approaches (e.g., the lag as moderator model, time-varying covariate interactions within multilevel models, and the mediation over time model) make use of the repeated measurements to explore how certain effects and associations change over time. Other approaches, especially those examining mediation using the SEM framework, focus more on the inherently time-ordered aspect of longitudinal data to establish the temporal precedence of effects necessary to make a causal argument about how predictor variables exert their effects on the outcomes of interest. Lag as Moderator Model

ξβ 2

X3

This defines the vector of latent variables to be 𝝃 = [𝜉𝛽0 𝜉𝛽1 𝜉𝛽2 ]′ and the vector of means for the latent variables to be 𝜿 = [𝛽0 𝛽1 𝛽2 ]′ where 𝛽0 , 𝛽1 , and 𝛽2 have the same interpretations as in equation 21.34. The latent exponential decay model created using the Browne (1993) method has two problems that arise during estimation. The first is that the mean of one of the latent variables must be constrained to be equal to zero. The second problem is that, as shown in equations 21.39 to 21.41, to constrain the loadings in the 𝚲x matrix correctly, the values of the means of the latent variables must already be known prior to estimating the model. Both of these problems may be circumvented, however, by setting 𝛽1 equal to 0 and then simultaneously estimating the path loadings and the means of the latent variables using nonlinear constraints, for example using the MODEL CONSTRAINT command in Mplus.

X4

A four-wave exponential decay latent growth curve

Selig, Preacher, and Little (2012) developed a moderator model unique to longitudinal data. Specifically, they propose the lag as moderator (LAM) approach to examine time-dependent associations. It is possible that the strength of the association between two variables assessed at different occasions will depend on the length of time elapsed between the occasions. Their presentation assumes just a single measurement of two variables for each subject, but variability of the time lag between the two measurements across subjects.

Page 946

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

The general approach involves using the length of the lag between measurement occasions as a predictor in a regression model. If the focal variables are X (predictor) and Y (criterion), the model would include X, the lag, and the interaction between X and the lag as predictors: ̂ = b0 + b1 X + b2 Lag + b3 Y

(21.42)

In this equation, the b1 coefficient represents the association between X and Y when Lag = 0 (i.e., simultaneous estimation), the b2 coefficient represents the expected change in Y for a 1 unit increase in the length of the lag, and b3 represents the linear change in the association between X and Y that accompanies a 1 unit increase in the length of the lag. Strategic centering of the lag variable can change the meaning of the b1 coefficient to test the association between X and Y if a particular lag is believed to be particularly important. Selig et al. pointed out that in many cases, change in the X–Y association as a function of the lag will not be linear. This nonlinearity might be captured with a straightforward polynomial specification (here, a quadratic form): ̂ = b0 + b1 X + b2 Lag + b3 Lag2 + b4 XLag + b5 XLag2 Y (21.43) or with a more complex nonlinear parameterization: [ ̂ = b0 + 𝛼Y ⋅X Y

( ) − 𝛼Y ⋅X − 𝛼0

(

)2 ] Lag X (21.44) −1 𝛼Lag

Though the latter expression requires nonlinear regression techniques for the estimation of its parameters, it provides estimates with more meaningful interpretations than those of the quadratic model (Cudeck & du Toit, 2002). 𝛼0 is the expected association between X and Y when Lag = 0, 𝛼Lag is the value of the lag at which the maximal X-Y association is obtained, and 𝛼X.Y is that maximum association. Alternately, an exponential model might be specified: ̂ = b0 + (aebLag )X Y

(21.45)

to represent a functional relationship that approaches an asymptote rather than symmetric form (i.e., any increase to a maximum must necessarily be followed by a decrease) implied by a quadratic specification. Selig et al. provide an empirical example illustrating the LAM approach using variables from the Early Head Start Research and Evaluation study (Department of Health

12:38pm

947

and Human Services, Administration for Children and Families, 1996–2001; available from the Inter-university Consortium for Political and Social Research website: www.icpsr.org/). As is often the case, the original intention was to collect the observational data at a single fixed lag; however, practical difficulties resulted in variation in time elapsed between measurement occasions across subjects. The LAM analysis examined effect of lag length on the association between the quality of stimulation provided to a child in his or her home at approximately 14 months and a measure of cognitive, language, and social development at approximately 24 months of age. Though the nominal lag between the 14-month and the 24-month observations was 10 months (and, in fact, the average lag was 10.08 months), there was considerable variability in the actual length of the lag, which ranged from 2.95 to 16.98 months (SD = 1.74 months). Initial examination of the data consisted of a series of scatterplots windowing different lag lengths and estimation of a regression slope within each window. A plot of these slope values against the average lag for each window showed a clear decrease in the strength of association with longer lags. To capture this change, Selig et al. fit a linear and an exponential LAM model, and both showed significant lag moderation. The two models seemed to fit the data equally well for the range of lags actually present in the data, but the authors pointed out that for even longer lags the linear model predicts a negative association between the two variables while the exponential model predictions approach a lower asymptote of zero. They cautioned, however, that their example data did not involve random assignment of the lag length, and thus interpretations must be circumspect. We used the LAM model to examine the effect of the lag between the measurement of ADHD symptoms at kindergarten fall and measurement of interpersonal skills, academic performance, and internalizing behavior at kindergarten spring. The lag between the two assessment occasions ranged from 3.8 to 9.3 months, with an average of 6.1 months. Linear and quadratic effects of lag length were tested. Only two interactions between predictor values and lag length emerged, both linear. The negative association between kindergarten fall EXTERN and kindergarten spring INTERP was stronger for shorter than for longer lags (b3 = 0.3340, SE = 0.1177, p < .01), and the positive association between kindergarten fall EXTERN and kindergarten spring INTERN was stronger for longer than for shorter lags (b3 = 0.2268, SE = 0.1063, p < .05). Selig et al.’s LAM approach was designed for a very specific data structure, with only two observations of

Page 947

Trim Size: 8.5in x 11in

948

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

measurement per subject, each variable assessed at only one of the measurement occasions, and variation in the lag between measurement occasions across subjects either produced intentionally (e.g., through random assignment of lag lengths) or as the result of vagaries in the implementation of a fixed-lag design. The single-level regression estimation that they recommend, whether linear or nonlinear, requires a single outcome observation for each subject. It should be possible, however, to adapt their framework for estimation as a multilevel model, which will widen the range of data structures to which the LAM approach might be applied. Moreover, such applications would convert what Selig et al. acknowledge as an interindividual approach to what is often an intraindividual question into an appropriately matched intraindividual analysis. When the predictor variable is time-invariant but the criterion is assessed at multiple later time points, each subject shows intraindividual variation in the predictor-criterion lag. The single equation form of a multilevel model would closely parallel those specified by Selig et al., simply adding L2 error terms to model the potential unmeasured similarity among the repeated criterion observations obtained from a given individual. At L1 (the repeated measures level), the equation would be Yti = 𝛽0i + 𝛽1i Lagti + eti

(21.46)

and at L2 (the individual level), we would have 𝛽0i = 𝛾00 + 𝛾01 Xi + u0i 𝛽1i = 𝛾10 + 𝛾11 Xi + u1i

(21.47)

The 𝛾11 coefficient captures the lag moderation effect in this model. Results of fitting such models in the ECLS-K data are provided in Table 21.4. These models predicted kindergarten spring through fifth grade spring assessments of interpersonal skills, academic performance, and internalizing behavior from kindergarten fall measures of ADHD symptoms. Initial models included interactions between ADHD symptoms and both linear and quadratic lag variables. If the interaction between the ADHD symptoms variable and the quadratic lag was not significant, the model was refitted with only the interaction involving the linear lag. Most of the final models included significant interactions with the quadratic lag function; however, the model predicting READ from EXTERN included only an interaction with linear lag. All of the models indicated that

TABLE 21.4 Coefficients and Standard Errors for Multilevel LAM Models Predictor

Outcome

Simple effect at Lag = 0

Predictor x Lag

Predictor x Lag2

LEARN

INTERP

LEARN

READ

LEARN

MATH

LEARN

INTERN

EXTERN

INTERP

EXTERN

READ MATH

EXTERN

INTERN

−0.1982*** (0.0060) −0.0451*** (0.0030) −0.0484*** (0.0025) 0.0665*** (0.0054) 0.0866*** (0.0066) 0.0146*** (0.0013) 0.0146*** (0.0027) −0.0508*** (0.0059)

0.0261*** (0.0011) 0.0029*** (0.0005) 0.0059*** (0.0004) −0.0103*** (0.0010) −0.0105*** (0.0012) NS

EXTERN

0.6204*** (0.0050) 0.3262*** (0.0052) 0.3231*** (0.0046) −0.2690*** (0.0052) −0.5390*** (0.0060) −0.1469*** (0.0061) −0.1242*** (0.0054) 0.2084*** (0.0057)

−0.0021*** (0.0005) 0.0078*** (0.0011)

the strongest effects of the kindergarten fall ADHD symptoms variables were the contemporaneous ones on the outcome variables at kindergarten fall with Lag = 0. The magnitude of the association between kindergarten fall EXTERN and READ declined linearly through the fifth grade spring assessment. For the other predictor/outcome combinations, the magnitude of the association declined quickly at the kindergarten spring and first grade spring assessments and gradually leveled off. Some of the models, in fact, indicate a reversal of the original trend near the fifth grade assessment, but this seems more likely an artifact of the necessary symmetry in quadratic models than a real pattern in the data, suggesting that exponential models might be a more appropriate choice for modeling some of these particular associations. Multilevel Growth Models with Time-Varying Covariate Interactions The multilevel growth model presents an interesting opportunity for examining a particular kind of moderation in longitudinal data. This is moderation of the effect of a L1 predictor variable on a L1 outcome variable by time. That is, one can test whether the association between repeatedly measured predictor and outcome constructs changes systematically over the course of a longitudinal study. Many theories would predict change in the strength of association between two variables during the course of development or when transitioning from one developmental stage to another. Judicious coding of time trends and modeling interactions between predictors and the time-related variables will allow such effects to be directly tested.

Page 948

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

For example, consider a simple linear multilevel growth model L1:

Yti = 𝛽0i + 𝛽1i timeti + eti

L2:

𝛽0i = 𝛾00 + u0i

(21.48)

𝛽1i = 𝛾10 + u1i Combined: Yti = 𝛾00 + 𝛾10 timeti + u0i + u1i timeti + eti This model estimates the linear trend in the outcome variable over time. If a researcher is interested in the effect of a time point-specific predictor variable (Xti , a time-varying covariate), it can be added to the L1 specification L1:

Yti = 𝛽0i + 𝛽i1 timeti + 𝛽2i Xti + eti

(21.49)

with an associated L2 equation L2:

𝛽2i = 𝛾02

(21.50)

Combined: Yti = 𝛾00 + 𝛾10 timeti + 𝛾02 Xti + u0i + u1i timeti + eti (21.51) With this specification, the time trend (𝛾 01 ) is adjusted for variation in the time-varying predictor and 𝛾 02 represents the effect of the predictor over and above the time trend. Now consider adding the interaction between the L1 predictor and the linear time variable: L1:

Yti = 𝛽0i + 𝛽i1 timeti + 𝛽2i Xti + 𝛽3i Xti timeti + eti

L2:

𝛽3i = 𝛾30

(21.52)

Combined: Yti = 𝛾00 + 𝛾10 timeti + 𝛾02 Xti + 𝛾03 Xti timeti + u0i + u1i timeti + eti As Singer and Willett (2003, p. 171) noted, the coefficient associated with the interaction between a time-varying predictor and the time variable (here, 𝛾 03 ) can be interpreted in two ways. One interpretation is that the rate of change

12:38pm

949

over time in the outcome variable depends on the level of the time-varying predictor. The second interpretation, and the one that we will focus on here, is that the effect of the time-varying predictor on the outcome changes over time. The equation above specifies linear change in the association between the time-varying predictor and the outcome, but (as discussed in the original Baron and Kenny (1986) paper), a more (e.g., quadratic) or less (e.g., step) complex function could be used to model a hypothesized pattern of change. Hedeker (2004) discussed such an effect in the context of a longitudinal psychiatric study originally described in Reisby et al. (1977). This study examined the association between imipramine and desipramine plasma levels (the time-varying predictors) and clinical response (the outcome variable) over time in a sample of depressed inpatients. Because the effectiveness of antidepressants is not immediate but develops over time, an exploration of such moderated effects seems warranted. The model included a linear time trend, main effects of the imipramine and desipramine plasma levels, and interactions between linear time and the plasma levels. The model including the two interactions fit better than the model without the interactions. Examination of specific coefficients showed that the imipramine level by time interaction was not significant in this sample, but the desipramine interaction was significant, indicating that, in fact, the effect of this particular antidepressant on clinical response became more pronounced over the course of the 5-week study. Table 21.5 reports the results of analyses in the ECLS-K data that included interactions between time-varying predictors and time. In particular, we explored whether the effects of the two ADHD symptoms variables on interpersonal skill, academic performance, and internalizing behavior change between fall of kindergarten and spring of fifth grade. Multilevel growth models incorporating an interactions between an ADHD symptoms variable and functions of the time variable were fitted for each possible pairing of LEARN and EXTERN with INTERP, MATH, READ, and INTERN. In each model, the base growth curve for the outcome construct was allowed to be quadratic with random effects for all growth parameters, and any nonsignificant variances were trimmed. As Table 21.5 shows, the interaction between LEARN and linear time on INTERP is significant. This indicates that the contemporaneous positive association between the approaches to learning score and interpersonal skill ratings become stronger over time. For example, as indicated in column 3 of Table 21.5, the expected association between these two variables at kindergarten fall (i.e., when

Page 949

Trim Size: 8.5in x 11in

950

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

TABLE 21.5 Coefficients and Standard Errors for Multilevel Models Including Time-Varying Predictor by Time Interactions Predictor

Outcome

Simple effect of predictor

Linear time interaction

LEARN

INTERP

LEARN

READ

LEARN

MATH

LEARN

INTERN

EXTERN

INTERP

EXTERN

READ

EXTERN

MATH

EXTERN

INTERN

0.6365*** (0.0035) 0.1602*** (0.0035) 0.1312*** (.0029) −0.2632*** (0.0037) −0.5120*** (0.0052) −.0234*** (0.0039) −0.0315*** (.0037) 0.2242*** (0.0049)

0.0070*** (0.0013) −0.0293*** (0.0030) −0.0436*** (0.0026) −0.0114*** (0.0014) −0.0787*** (0.0056) −0.0247*** (0.0033) −.0035** (.0012) 0.0369*** (0.0054)

Quadratic time interaction NS 0.0029*** (0.0005) 0.0064*** (0.0004) NS 0.0093*** (.0010) 0.0034*** (.0005) NS −0.0040*** (.0010)

Time = 0) is .6365 and can be interpreted in much the same way as an unstandardized regression coefficient: for each one unit increase in LEARN at kindergarten fall, the expected value of interpersonal skill at that time point increases by .6365 points. For each additional year that passes, the contemporaneous association is expected to increase by .0070 points (column 4). So at fifth grade spring, the expected value of interpersonal skills increases by .6750 (i.e., .6365 + .0070 ∗ 5.5) for each one unit increase in LEARN. Similarly, Table 21.5 shows that the negative effects of LEARN on INTERN and the negative effects of EXTERN on MATH become significantly more pronounced over time. The predicted associations between the ADHD predictor variables and the outcome variables at each time point correspond fairly well—but not perfectly—to coefficients produced via OLS regressions conducted on data from each time point separately. Although this model may not precisely reproduce the regression coefficients at each time point, it does allow for a single parameter test of the change in such coefficients across time that is not available from analyses conducted on data from each measurement occasion separately. Five of the eight analyses presented in Table 21.5 involve a significant interaction between the time-varying ADHD symptoms predictor and the quadratic time function, and for all of these, the interactions involving the linear and quadratic time functions are of opposite sign. This indicates that the initial changes in the magnitude of the ADHD symptoms variable and the outcome will eventually change direction. For example, the coefficients for the effect of LEARN on MATH in columns

3 and 4 of Table 21.5 indicate a positive association at kindergarten fall that initially decreases in magnitude over time. However, because of the positive sign of the quadratic interaction in column 5, this trend reaches a minimum around third grade spring, and the magnitude of the association increases thereafter. The maxima or minima for the other four associations with quadratic time interactions, however, reach their minima or maxima much closer to the end of the grade range included in the data. Care should be taken to ensure that these reversals are a true feature of the data and not just an artifact of the necessarily symmetric time function used in this particular example. The model specification in equation 21.50 as well as those in Hedeker (2004) and the previous ECLS-K analyses assume that the effects of the time-varying covariates are fixed. That is, the association between the time-varying covariate and the outcome measure is the same for all subjects. Moreover, the extent to which the covariate’s association with the outcome changes over time is also assumed to be the same for all subjects. However, with a sufficiently large sample, it may be possible to explore whether, in fact, these effects differ across individuals. This can be accomplished by adding error terms u2i and u3i to the L2 equations predicting the 𝛽2i and 𝛽3i coefficients, respectively. Then the variance of u2i estimates (𝜏22 ) indicates the extent to which individuals vary in the association between their predictor and outcome values when time is at its zero point, and the variance of the u3i estimates (𝜏33 ) indicates the extent to which individuals vary in how much this association changes over time. Mediation Change over Time Model The mediation change over time (M-COT) model builds upon the idea of creating interactions between predictors and the explicit time variable in a multilevel growth model to examine possible change in both the a and b paths of a mediation model over time. In contrast to other longitudinal mediation models that will be presented below, the focus of this model is not on establishing the temporal precedence of effects necessary for causal pathways. Rather, the M-COT model is largely descriptive and explores how the various paths in the basic model of Figure 21.1, specifically the a path from X to M and the b path from M to Y, change over the course of a longitudinal study. The X → M and M → Y relations within the model might be cross sectional, or X might be assessed prior to M, and M might be assessed prior to Y. In the former case, the M-COT model would simply describe how the contemporaneous associations between variables

Page 950

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

change over time. In the latter case, the M-COT model would describe how the lagged relations among variables change as time progresses. This discussion will concentrate on the M-COT model for exploring cross sectional mediation in longitudinal data. This model is appropriate for data with repeated measures of the outcome variable Y. The initial variable X and the mediator variable M could also be measured repeatedly or could be time-invariant individual variables. The following model assumes a 1 → 1 → 1 model with all three variables measured repeatedly at the same time points. The M-COT model can be specified as a two-level random coefficient (i.e., multilevel) model with time points nested within individuals (Arruda & Krull, 2013). For a simple model that examines linear change over time in the a and b paths, the multiple equation form of the model to estimate a paths would be: L1:

Mti = 𝛽0i + 𝛽1i timeti + 𝛽2i Xti 𝛽0i = 𝛾00 + u0i

𝛽2i = 𝛾20

𝛽1i = 𝛾10 + u1i

𝛽3i = 𝛾30

(21.53)

Mti = 𝛾00 + 𝛾10 timeti + 𝛾20 Xti + 𝛾30 timeti Xti (21.54)

In this particular model, the 𝛾20 coefficient estimates the association between X and M when time is at its zero point. The 𝛾30 coefficient estimates how much this association changes for each additional unit of time. The multilevel equation to generate 𝛽 path estimates would be: L1:

Yti = 𝛽0i + 𝛽1i timeti + 𝛽2i Xti + 𝛽3i timeti Xti + 𝛽4i Mti + 𝛽5i timeti Mti + eti

L2:

𝛽0i = 𝛾00 + u0i

𝛽2i = 𝛾20

𝛽4i = 𝛾40

𝛽1i = 𝛾10 + u1i

𝛽3i = 𝛾30

𝛽5i = 𝛾50

in the single equation form. The 𝛾 40 coefficient in this particular model estimates the association between M and Y when time is at its zero point. The 𝛾 50 coefficient estimates how much this association changes for each additional unit of time. Combining the appropriate coefficients from the two models provides an expression for how the mediated effect ab changes over time (Arruda & Krull, 2014a). The additional subscripts M and Y in the following equations indicate whether a given coefficient came from the first multilevel equation (M) regressing M on X or the second multilevel equation (Y) regressing Y on X and M simultaneously. The combined expression is simply the product of the terms relevant to the a effect in equation 21.54 and those relevant to the 𝛽 effect in equation 21.6: abti = (𝛾20M + 𝛾30M timeti )(𝛾40Y + 𝛾50Y timeti )

(21.57)

abti = 𝛾20M 𝛾40Y + (𝛾20M 𝛾50Y + 𝛾30M 𝛾40Y )timeti

and the single-equation form would be

+ u0i + u1i timeti + eti

951

Manipulating and rearranging this expression gives

+ 𝛽3i timeti Xti + eti L2:

12:38pm

(21.55)

in the multiple equation form, and

+ 𝛾30M 𝛾50Y timeti 2

(21.58)

The first term in this expression, 𝛾20M 𝛾40Y , estimates the mediated effect (i.e., the product of the a and b paths) when time is at its zero point. The second term, 𝛾20M 𝛾50Y + 𝛾30M 𝛾40Y , estimates the instantaneous linear change in the mediated effect when time is at its zero point. The third term, 𝛾30M 𝛾50Y , estimates the quadratic change in the mediated effect over time and indicates how the instantaneous linear change in the mediated effect itself changes over time. Approximate standard errors for the three terms in equation 21.58 can be obtained via the multivariate delta method (Arruda & Krull, 2014a). Specifically, the standard errors are as follows: SE(𝛾20M 𝛾40Y ) = sqrt(𝛾20M 2 𝜎 2 𝛾 40Y + 𝛾40Y 2 𝜎 2 𝛾 20M ) (21.59) SE(𝛾20M 𝛾50Y + 𝛾30M 𝛾40Y ) = sqrt(𝛾20M 2 𝜎 2 𝛾 50Y + 𝛾50Y 2 𝜎 2 𝛾 20M + 𝛾40Y 2 𝜎 2 𝛾 30M + 𝛾30M 2 𝜎 2 𝛾 40Y ) + (2 ∗ (𝛾20M 𝛾30M 𝜎𝛾 40Y 𝛾50Y ) + (𝛾40Y 𝛾50Y 𝜎𝛾 20M𝛾30M ))

Yti = 𝛾00 + 𝛾10 timeti + 𝛾20 Xti + 𝛾30 timeti Xti + 𝛾40 Mti + 𝛾50 timeti Mti + u0i + u1i timeti + eti (21.56)

(21.60) SE(𝛾30M 𝛾50Y ) = sqrt(𝛾30M 2 𝜎 2 𝛾 50Y + 𝛾50Y 2 𝜎 2 𝛾 30M ) (21.61)

Page 951

Trim Size: 8.5in x 11in

952

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

These standard errors can be used to conduct hypothesis tests or develop confidence intervals for each of the terms in equation 21.58. Thus we can explicitly test the trend in ab over time. One useful way to think about this model is to consider the ultimate form of change in the ab estimate of the mediated effect that will result from combining different trends in a and b. As currently specified, three types of trends are possible for the a path: the a path could be zero (both the 𝛾20M and 𝛾30M coefficients are nonsignificant), a nonzero constant over time (the 𝛾20M coefficient is significant but the 𝛾30M coefficient is nonsignificant), or a linearly increasing function of time (the 𝛾30M coefficient is significant, regardless of the significance of the 𝛾20M coefficient). The same three trends are also possible for the b path (i.e., zero—both the 𝛾 40Y and the 𝛾50Y coefficients are nonsignificant, constant—the 𝛾40Y coefficient is significant but the 𝛾50Y coefficient is nonsignificant, or linearly changing—the 𝛾50Y coefficient is significant, regardless of the significance of the 𝛾40Y coefficient). Then there are nine possibilities for the multiplicative combination, as shown in Table 21.6. Of course, whenever the nature of either of the two paths is zero, the multiplicative function ab will also be zero. Thus five of the nine combinations ultimately produce no mediated effect. Only four combinations produce nonzero values for the ab function over time, and these correspond to the four terms in equation 21.58. With this limited set of possibilities (i.e., allowing only zero, constant, and linear trends for a and b), the only way that the ab product could indicate a constant mediated effect over time is if both a and b are nonzero constants. There are two ways that ab could show a pattern of linear change over time, specifically (1) if a is a nonzero constant and b changes linearly over time, or (2) if b is a nonzero constant and a changes linearly over time. The product ab will change quadratically over time only if both a and b change linearly. By substituting each of the unique time values into the equation, it is possible to calculate the estimated mediated

effect ab at each time point. Standard errors of these time point-specific estimates can be calculated using the multivariate delta method. Simulation studies (Arruda & Krull, 2013, 2014b) comparing the time point-specific estimates of ab from the M-COT model with those generated by fitting a series of OLS regression models to the data from each time point found that both approaches generated unbiased estimates of ab at each time point. When the data generating process for a and b was truly linear over time (i.e., when the analysis model form matched the true model), however, time point-specific estimates produced by the M-COT model had smaller standard errors than those generated using standard regression techniques to estimate mediated effects at each time point separately. Thus the M-COT approach is more highly powered than the regression approach. Across the range of simulated conditions, the standard errors of mediated effects at specific time points were 14-47% smaller in the M-COT model, resulting in 25-89% more power to detect nonzero effects than using standard regression techniques (Arruda & Krull, 2014b). The M-COT model was used to assess changes in mediated effects over time in the ECLS-K data. Results of separate analyses in which the ADHD symptoms variables predicted interpersonal skill and academic performance mediators and an internalizing behavior outcome are presented in Table 21.7. The negative effect of LEARN on INTERN through INTERP has a negative linear trend and no significant quadratic trend, so this effect become increasingly pronounced over time. Similar patterns hold for the effects on LEARN on INTERN through READ and MATH, but these include a significant positive quadratic component, so the increasingly negative mediated effects level off at later grades for these academic performance mediators. The effects of EXTERN on INTERN through the three mediators all show similar patterns, with positive mediated effects at kindergarten fall and significant linear and quadratic trends. This indicates that the mediated effects become stronger over time, and the rate of increase speeds up at later grades.

TABLE 21.6 Combinations of Significant and Nonsignificant Coefficients That Produce Various Trends of the Mediated Effect over Time 𝛾 20M

𝛾 30M

a path trend

𝛾 40Y

𝛾 50Y

b path trend

ab trend

NS NS NS Sig Sig Sig NS or sig NS or sig NS or sig

NS NS NS NS NS NS Sig Sig Sig

Zero Zero Zero Constant Constant Constant Linear Linear Linear

NS Sig NS or sig NS Sig NS or sig NS Sig NS or sig

NS NS Sig NS NS Sig NS NS Sig

Zero Constant Linear Zero Constant Linear Zero Constant Linear

Zero Zero Zero Zero Constant Linear Zero Linear Quadratic

Page 952

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

12:38pm

953

TABLE 21.7 Coefficients, Standard Errors, and Mediated Effect Trends for the M-COT Model X

M

a at K fall

a linear trend

b at K fall

b linear trend

Med at K fall

LEARN

INTERP

LEARN

READ

LEARN

MATH

EXTERN

INTERP

EXTERN

READ

EXTERN

MATH

0.6365*** (0.0035) 0.2457*** (0.0042) 0.1915*** (0.0033) −0.5352*** (0.0043) −0.0396*** (0.0047) −0.0315*** (0.0037)

0.0070*** (0.0013) −0.0213*** (0.0012) −0.0174*** (0.0010) −0.0297*** (0.0016) −0.0058*** (0.0013) −0.0035** (0.0012)

−0.1333*** (0.0053) −0.0010 (0.0036) −0.0125** (0.0041) −0.1798*** (0.0046) −0.0681*** (0.0036) −0.0892*** (0.0040)

−0.0027 (0.0020) −0.0283*** (0.0019) −0.0247*** (0.0015) −0.0066*** (0.0018) −0.0451*** (0.0019) −0.0329*** (0.0015)

−0.0849*** (0.0034) −0.0003 (0.0009) −0.0024** (0.0008) 0.0962*** (0.0026) 0.0027*** (0.0004) 0.0081*** (0.0004)

The M-COT model is unique among techniques designed to assess mediation in longitudinal data because it assesses mediation cross sectionally, but the cross sectional associations are not necessarily stable over time. It is assumed that there is some short-term process that has reached equilibrium at each measurement occasion, but a longer-term process that is not stationary. This model might prove useful in explaining conflicting results in developmental studies that have examined mediation at different ages or across different timespans. The model specified here is a 1 → 1 → 1 mediation model, but the general framework could be easily adapted to accommodate individual-level predictor and mediator variables at L2. The model was also limited to linear trends over time for both the a and b paths. A simulation study is currently underway to examine more complex trends for a and b as well as ideal methods for determining the form of the a and b trends in real world data where the true data generating process is not known. The current model allows intercepts and time trends to vary across individuals but assumes that the a and b paths are fixed. If the data warrant, these assumptions could be removed by incorporating methods similar to those of Bauer et al. (2006). Finally, critiques about assessing mediation cross sectionally could be addressed by estimating lagged relations among variables (e.g., X 1 → M2 → Y3; X 2 → M3 → Y4; X 3 → M4 → Y5) rather than the cross sectional associations as shown here.

Med linear trend −0.0027* (0.0013) −0.0069*** (0.0005) −0.0045*** (0.0003) 0.0089*** (0.0011) 0.0022*** (0.0003) 0.0014*** (0.0002)

−1.89E –05 (1.44E –05 ) 0.0006*** (5.25E –05 ) 0.0004*** (3.59E –05 ) 0.0002*** (5.45E –05 ) 0.0003*** (5.97E –05 ) 0.0001** (3.98E –05 )

is contemporaneous. Gollob and Reichardt (1991) argued that it takes time for a cause to have an effect and thus, at least k + 2 waves of data are needed to estimate indirect effects in a model with a chain of k mediators. Thus, with a single mediator, data should ideally be collected at least across three waves or time points, so X can affect M and M can affect Y in time. Otherwise, some parameter constraints or assumptions are needed, based on theory and prior knowledge (Gollob & Reichardt, 1991), such as assumptions of stationarity and knowledge of test-retest reliability. Figure 21.13 presents an example of the autoregressive mediation model with two waves of data, where the effect of X on M and the effect of M on Y are specified as lag-1 effects in an SEM. The autoregressive models described in this section are based on the observed variables, ignoring the measurement errors. With inclusion of multiple

X1

X2

sx

a1

M1

b1

Y1

M2

sm

Longitudinal Mediation in Autoregressive Panel Models When X, M, and Y are measured repeatedly over time, autoregressive panel data models can be extended to the three variable case to investigate longitudinal mediation. When the autoregressive mediation model involves two waves of data, fitting the autoregressive models to estimate mediation effect is somewhat limited, in that either the relation between X and M or the relation between M and Y

Med Quad Trend

sy

c´1

Y2

Figure 21.13 An autoregressive model with longitudinal mediation across two waves.

Page 953

Trim Size: 8.5in x 11in

Cicchetti

954

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

indicators, the repeated measures of X, M, and Y can be modeled as latent variables and the indirect effects can be estimated based on the lagged effects between the latent variables. The model in Figure 21.13 can be described by equations 21.62 to 21.64.

X1

X2

sx

a1

X2 = ix2 + sx X1 + ex2

(21.62)

M2 = im2 + a1 X1 + sm M1 + em2

(21.63)

Y2 = iy 2 + b1 M1 + c′1 X1 + sy Y1 + ey 2

(21.64)

M1

where the variables measured at time 1 are treated as predetermined as follows: X1 = ix1 + ex1

(21.65)

M1 = im1 + em1

(21.66)

Y 1 = iy 1 + e y 1

(21.67)

where ix1 , im1 , and iy1 are the means of X1 , M1 , and Y1 , respectively, and ex1 , em1 , and ey1 are deviations from the means. Typically, X1 , M1, and Y1 are specified to be correlated. As described previously, im2 and iy2 in equations 21.63 and 21.64 are regression intercepts, and em2 and ey2 are residuals. The coefficients a1 and b1 represent the effect of X1 on M2 and the effect of M1 on Y2 , respectively, while the coefficient c′1 denotes the direct effect of X1 on Y2 . The product of the estimates of the coefficients a1 and b1 , i.e., a1 b1 , is the estimated longitudinal indirect effect. Although the coefficient b1 represents the relation between M and Y before X affects M, the a1 b1 estimator would be a reasonable estimate of the longitudinal indirect, because under the stationarity assumption the relation between M1 and Y2 (b1 ) would be equal to the relation between M2 and Y3 that would have been measured after the second wave (Cole & Maxwell, 2003). The standard error of the estimated indirect effect can be obtained using the formulas described earlier based on the relevant path coefficients and their standard errors to conduct the significance test and compute confidence intervals. Another example of the autoregressive mediation model with two wave data is the model in Figure 21.14 (Mackinnon, 2008). The relations relevant to the mediational pathways can be expressed by the following equations 21.68 and 21.69: M2 = im2 + a1 X1 + a2 X2 + sm M1 + em2

(21.68)

Y2 = iy 2 + b1 M1 + b2 M2 + c′1 X1 + c′ 2 X2 + sy Y1 + ey 2 (21.69)

M2

sm

b1

Y1

a2

sy

c´1

b2

c´2

Y2

Figure 21.14 An autoregressive mediation model with longitudinal and contemporaneous mediation across two waves.

With the addition of the contemporaneous relations (a2 , b2 , and c′2 ) to the previous model in Figure 21.13, the coefficient a1 in the above model represents the effect of X1 on M2 after controlling for the effects of X2 and M1 , while the coefficient a2 represents the effect of X2 on M2 after controlling for the effects of X1 and M1 . Similarly, the coefficients b1 and b2 indicate the effect of M1 on Y2 after controlling for the effects of X1 , X2 , M2 , and Y1 and the effect of M2 on Y2 after controlling for the effects of X1 , X2 , M1 , and Y1 , respectively. There are several ways to estimate mediated effects in the model in Figure 21.14: a1 b1 , a2 b2 , and a1 b2 . As described earlier, the indirect effect a1 b1 is based on the across-time or lag-1 relations, while the indirect effect a2 b2 is estimated based on the contemporaneous relations among the variables measured at time 2. The indirect effect a1 b2 is based on the across-time relation between X1 and M2 and the contemporaneous relation between M2 and Y2 . Cole and Maxwell (2003) recommend the a1 b1 estimator, because the estimates a2 b2 and a1 b2 involve cross sectional relations and are likely to be biased when time is needed for the cause to exert its effect on the outcome (Gollob & Reichardt, 1987). In the two-wave autoregressive mediation models, information on stability, stationarity, and equilibrium is quite limited, in that we cannot actually test whether these properties hold across time. However, estimating autoregressive effects is beneficial for avoiding omitted variable bias. An autoregressive effect is defined as the effect that a variable has on itself at a later time (Gollob & Reichardt, 1987) and

Page 954

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

is denoted by the coefficients sx , sm , and sy in the above models. These coefficients also estimate the stability of a variable across time (Kenny, 1979). When the dependent variable at time t is predicted by the predictors at time t – 1, the levels of the dependent variable at time t − 1 may represent unmeasured variables that are correlated with the predictors and estimating autoregressive relations can control for such unmeasured variables (Gollob & Reichardt, 1991). Without controlling for those omitted variables or confounders, the causal effects will be biased. Thus, the estimates of a1 and a2 in Figures 21.13 and 21.14 will be less biased when estimating sm and so will b1 and b2 , when estimating sy . These less biased coefficients in the mediational pathways will result in the indirect effect that is closer to the true effect size. When X, M, and Y are measured repeatedly over three or more occasions, the parameterizations in the autoregressive models become more flexible. For example, the effect of M2 on Y3 that relied on the stationarity assumption in the two wave data (b1 in Figures 21.13 and 21.14) can now be estimated. With three or more waves of data, estimating lag effects according to the causal order of X to M to Y is possible, although timing of measurements is a remaining issue, because it is difficult to predict how long it takes for one variable to have its optimal effect on another. Having more waves of data also provides opportunities to examine when stability, stationarity, and equilibrium are reached. It may take a long time for a system of X, M, and Y to achieve stability, stationarity, or equilibrium and sometimes it may be achieved beyond the time span the data cover. While

X1

X2

sx1

b1

Y1

sy1

(21.70)

X3 = ix3 + sx2 X2 + ex3

(21.71)

X4 = ix4 + sx3 X3 + ex4

(21.72)

M2 = im2 + a1 X1 + sm1 M1 + em2

(21.73)

M3 = im3 + a2 X2 + sm2 M2 + em3

(21.74)

M4 = im4 + a3 X3 + sm3 M3 + em4

(21.75)

X3

sx2

X2 = ix2 + sx1 X1 + ex2

M2

c´1

Y2

a3

M3

sm2

b2

sy2

X4

sx3

a2

sm1

955

these characteristics can only be assumed in the cross sectional or two wave designs, researchers can check these assumptions with more waves of data by testing equivalence of parameter estimates across waves. Finally, time specific indirect effects can be estimated across multiple waves. The magnitude of effect, either the autoregressive effects or the lagged effects, may vary across time (Cole & Maxwell, 2003). The effect of X1 on M2 may be larger or smaller than the effect of X2 on M3 and the resulting indirect effects may differ depending on which time points are involved (Cole & Maxwell, 2003; Gollob & Reichardt, 1987, 1991). It might be an interesting research task to detect when the significant indirect effect arises, holds, and desists. Figure 21.15 shows an example of the autoregressive mediation model with four waves of data, in which the relations among X, M, and Y are specified as lag-1 effects. The model in Figure 21.15 can be described in the following equations:

a1

M1

12:38pm

c´2

b3

Y3

M4

sm3

sy3

Figure 21.15 An autoregressive mediation model with longitudinal mediation across four waves.

c´3

Y4

Page 955

Trim Size: 8.5in x 11in

956

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

Y2 = iy 2 + b1 M1 + c′1 X1 + sy 1 Y1 + ey 2

(21.76)

Y3 = iy 3 + b2 M2 + c′2 X2 + sy 2 Y2 + ey 3

(21.77)

Y4 = iy 4 + b3 M3 + c′3 X3 + sy 3 Y3 + ey 4

(21.78)

Again, the variables measured at time 1 are predetermined in the same way as in equations 21.65 to 21.67. With more than two waves of data, the effect of X on M and the effect of M on Y are represented by more than one path coefficient. In Figure 21.15 and equations 21.70 to 21.78, the coefficients a1 , a2 , and a3 all represent the time-specific effect of X on M. Similarly, the coefficients b1 , b2 , and b3 represent the time-specific effect of M on Y controlling for X, although the coefficients b2 and b3 are the effects of M that were affected by X, while b1 is the effect of M before X exerts its effect on M. The coefficients c′ 1 , c′ 2 , and c′ 3 reflect the time-specific direct effects of X on Y. The indirect effects for the model in Figure 21.15 can be estimated in several ways. As we did for the two wave data, one can estimate the indirect effect for each lag such as a1 b1 , a2 b2 , and a3 b3 . Another option is to estimate the indirect effect reflecting the temporal ordering of X, M, and Y, such as a1 b2 and a2 b3 . The indirect effect and its standard error can be obtained following the formulas described earlier and can be used for significance testing and construction of confidence limits. If the stationarity assumption holds across 4 waves, all the a paths will be equivalent and so will the b paths. In such case, the estimated indirect effects (i.e., a1 b1 , a2 b2 , a3 b3 , a1 b2 , a2 b3 ) will be the same. On the other hand, if the path coefficients vary across waves, the indirect effects will be different depending on the time points involved and specific indirect effects can be examined. For example, if four measurement occasions span across different developmental phases, the indirect effect observed at the earlier phase may differ from the indirect effect at the later phase. Differences in the indirect effects across different phases may be an interesting research question. If the autoregressive mediation model is estimated in an SEM framework, the equivalence of path coefficients across time can be tested and the parameters can be constrained accordingly before obtaining the indirect effect. In addition to the indirect effects listed already that are based on the two longitudinal relations between Xt → Mt+1 and Mt → Yt+1 (e.g., a1 b1 , a2 b2 , a3 b3 , a1 b2 , a2 b3 ), there are other time-specific indirect effects in the model in Figure 21.15, such as sx1 a2 b3 and a1 b2 sy3 . These individual indirect effects are time-specific indirect effects, as they represent causal mechanisms from X to M to Y

for specific time intervals. Gollob and Reichardt (1991) argued that these individual time-specific indirect effects may be of less interest to researchers because mediation may evolve gradually over the course of time, rather than occurring at a specific time point. The overall indirect effect may then provide a better estimate for the indirect effect. The overall indirect effect is defined as the sum of all the individual time-specific indirect effects, i.e., all the indirect effects from X at time 1 to Y at the last time point mediated by M at any time between the first and the last time points (Gollob & Reichardt, 1991). In the model in Figure 21.15, there are three such individual time-specific indirect effects: (1) X1 → X2 → M3 → Y4 (sx1 a2 b3 ), (2) X1 → M2 → M3 → Y4 (a1 sm2 b3 ), and (3) X1 → M2 → Y3 → Y4 (a1 b2 sy3 ). Computing the total indirect effect and its standard error, as well as the total direct effect and its standard error, is available in most SEM software programs, including EQS (Bentler & Wu, 2006), LISREL (Jöreskog & Sörbom, 2006), Proc CALIS in SAS (SAS Institute, 2014), and Mplus (Muthén & Muthén, 2014). It is worth noting that the accuracy of the total indirect effect would rely on the specified mediation model, however. For example, if the mediation model in Figure 21.15 is a true model but the measurement at time 2 is omitted, resulting in the model with data measured at time 1, 3, and 4, the estimated total indirect effect will not equal the true overall indirect effect. See Gollob and Reichardt (1991) for illustration. We fit an autoregressive panel SEM with longitudinal mediation paths to first, third, and fifth grade ECLS-K data. Though it is possible to fit autoregressive models with unequal lags between time points (Voelkle, Oud, Davidov, & Schmidt, 2012), for the sake of simplicity, only the first, third, and fifth grade measures with constant 2-year lags are used in this example. Table 21.8 presents

TABLE 21.8 Coefficients and Standard Errors for Tests of Longitudinal Mediation in an Autoregressive Panel Model X1

M3

Y5

LEARN

INTERP INTERN

LEARN

READ

INTERN

LEARN

MATH

INTERN

EXTERN INTERP INTERN EXTERN READ

INTERN

EXTERN MATH

INTERN

a 0.1985*** (0.0131) 0.0514*** (0.0035) 0.0511*** (0.0041) −0.3229*** (0.0135) −0.0268*** (0.0037) −0.0157*** (0.0043)

b −0.0063 (0.0112) −0.1674*** (0.0204) −0.1270*** (0.0156) −0.0959*** (0.0091) −0.2468*** (0.0203) −0.1829*** (0.0155)

Med −0.0013 (0.0022) −0.0086*** (0.0012) −0.0065*** (0.0010) 0.0310*** (0.0032) 0.0066*** (0.0011) 0.0029*** (0.0008)

Page 956

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

coefficients and standard errors for a, b, and the longitudinal mediated effect from ADHD symptoms at first grade through interpersonal skills and academic performance at third grade on internalizing behavior at fifth grade. The results indicate significant longitudinal mediated effects of LEARN on INTERN through both READ and MATH (but not INTERP) and longitudinal effects of EXTERN on INTERN through all three mediators. However, CFI and RMSEA values indicated that these models did not fit the data well. Another type of autoregressive mediation model is shown in Figure 21.16 with the mediational pathways described in equations 21.79 to 21.84: M2 = im2 + a1 X1 + a4 X2 + sm1 M1 + em2

(21.79)

M3 = im3 + a2 X2 + a5 X3 + sm2 M2 + em3

(21.80)

M4 = im4 + a3 X3 + a6 X4 + sm3 M3 + em4

(21.81)

Y2 = iy 2 + b1 M1 + b4 M2 + c′1 X1 + c′4 X2 + sy 1 Y1 + ey 2 (21.82) Y3 = iy 3 + b2 M2 + b5 M3 + c′2 X2 + c′5 X3 + sy 2 Y2 + ey 3 (21.83) Y4 = iy 4 + b3 M3 + b6 M4 + c′3 X3 + c′6 X4 + sy 3 Y3 + ey 4 (21.84) In addition to the longitudinal relations between X, M, and Y specified in Figure 21.15, the model in Figure 21.16

X1

X2

sx1

a1

M1

M2

X3

sx2

a2

M3

b2

sy1

Y2

sy2

a6

M4

sm3

c´3 b5

c´4 Y1

a3

c´2 b4

X4

sx3

a5

sm2

c´1 b1

957

includes the contemporaneous relations among X, M, and Y at each wave. These relations are estimated by the coefficients a4 , a5 , and a6 for the relations between X and M at Time 2, Time 3, and Time 4, respectively, and the coefficients b4 , b5 , and b6 for the relations between M and Y. The product of path coefficients for the contemporaneous relations, such as a4 b4 , a5 b5 , and a6 b6 , provide the estimates of the contemporaneous indirect effect. When X and M have immediate effects, the longitudinal relations may miss the real effects due to the time lag between measurements. In such cases, the contemporaneous indirect effects may be closer to the true indirect effects (MacKinnon, 2008). We fit models including both longitudinal and contemporaneous mediation pathways to the ECLS-K first through fifth grade data. The results are summarized in Table 21.9. These models fit the data better than the previous set that included only the longitudinal pathways (all CFIs > .95), but RMSEA values were in the .08–.10 range, which is still larger than desired for optimal fit. Longitudinal mediated paths (X1 → M3 → Y 5) involving the INTERP and READ (but not MATH) mediators were significant or marginally significant and in the expected direction. Several mediated effects involving contemporaneous a or b paths were also significant in models including the INTERP mediator and READ mediators. Although having more than two waves for the autoregressive models may allow for the flexibility in model specification and the opportunities to obtain more detailed information about the indirect effects, it may introduce

a4

sm1

12:38pm

b3

b6

c´5 Y3

sy3

c´6 Y4

Figure 21.16 An autoregressive mediation model with longitudinal and contemporaneous mediation across four waves.

Page 957

Trim Size: 8.5in x 11in

958

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

TABLE 21.9 Coefficients and Standard Errors for Tests of Longitudinal and Contemporaneous Mediation in an Autoregressive Panel Model X = LEARN X → INTERP → INTERN X1 → M2 → Y2 X1 → M2 →Y3 X2 → M2 → Y2 X2 → M2 → Y3 X2 → M3 → Y3 X3 → M3 → Y3 X → READ → INTERN X1 → M2 → Y2 X1 → M2 → Y3 X2 → M2 → Y2 X2 → M2 → Y3 X2 → M3 → Y3 X3 → M3 → Y3 X → MATH → INTERN X1 → M2 → Y2 X1 → M2 → Y3 X2 → M2 → Y2 X2 → M2 → Y3 X2 → M3 → Y3 X3 → M3 → Y3

a

b

X = EXTERN Med

−0.1535***

−0.1326***

0.0204***

(0.0108) −0.1535*** (0.0108) 0.6912*** (0.0087) 0.6912*** (0.0087) −0.1445*** (0.0113) 0.6698*** (0.0086)

(0.0118) 0.0400*** (0.0122) −0.1326*** (0.0118) 0.0400*** (0.0122) −0.1435*** (0.0124) −0.1435*** (0.0124)

0.0232*** (0.0040) 0.0232*** (0.0040) 0.0618*** (0.0038) 0.0618*** (0.0038) 0.0144*** (0.0031) 0.0302*** (0.0030) 0.0246*** (0.0047) 0.0246*** (0.0047) 0.0569*** (0.0045) 0.0569*** (0.0045) 0.0160*** (0.0039) 0.0364*** (0.0038)

a

b

Med

−0.0308*

−0.2076***

(0.0023) −0.0061** (0.0019) −0.0917*** (0.0082) 0.0276** (0.0084) 0.0207*** (0.0024) −0.0961*** (0.0084)

(0.0128) −0.0308* (0.0128) −0.6232*** (0.0110) −0.6232*** (0.0110) −0.0057 (0.0135) −0.5954*** (0.0115)

(0.0109) −0.0255* (0.0113) −0.2076*** (0.0109) −0.0255* (0.0113) −0.2161*** (0.0110) −0.2161*** (0.0110)

0.0064* (0.0027) 0.0008† (0.0005) 0.1294*** (0.0072) 0.0159* (0.0070) 0.0012 (0.0029) 0.1287*** (0.0070)

−0.0172 (0.0273) −0.1066*** (0.0340) −0.0172 (0.0273) −0.1066*** (0.0340) 0.0472 (0.0356) 0.0472 (0.0356)

−0.0004 (0.0006) −0.0025* (0.0009) −0.0011 (0.0017) −0.0066** (0.0021) 0.0007 (0.0005) 0.0014 (0.0011)

−0.0083 (0.0044) −0.0083 (0.0044) −0.0365*** (0.0045) −0.0365*** (0.0045) −0.0022 (0.0035) −0.0198*** (0.0036)

−0.1257*** (0.0274) −0.1675*** (0.0350) −0.1257*** (0.0274) −0.1675*** (0.0350) −0.0428 (0.0362) −0.0428 (0.0362)

0.0010† (0.0006) 0.0014† (0.0008) −0.0046*** (0.0011) 0.0061*** (0.0015) 0.0001 (0.0002) 0.0008 (0.0007)

−0.0373 (0.0228) −0.0461 (0.0289) −0.0373 (0.0228) −0.0461 (0.0289) −0.0190 (0.0278) −0.0190 (0.0278)

−0.0009 (0.0006) −0.0011 (0.0007) −0.0021 (0.0013) −0.0026 (0.0005) −0.0003 (0.0005) −0.0007 (0.0010)

0.0048 (0.0051) 0.0048 (0.0051) −0.0400*** (0.0052) −0.0400*** (0.0052) −0.0083 (0.0044) −0.0222*** (0.0045)

−0.1095*** (0.0229) −0.0781** (0.0297) −0.1095*** (0.0229) −0.0781** (0.0297) −0.0894** (0.0282) −0.0894** (0.0282)

−0.0005 (0.0006) −0.0004 (0.0004) 0.0044*** (0.0011) 0.0031* (0.0013) 0.0007 (0.0005) 0.0020** (0.0007)

complexity of parameterization. For example, a variant of the model in Figure 21.16 may include cross-lagged effects among X, M, and Y by adding the direct effects from Mt−1 to Xt and Yt−1 to Mt based on the temporal sequence in which the variables are measured. Specifying these cross-lagged relations may reflect data more reasonably than the models in Figure 21.15 and 21.16, because X, M, and Y are related to each other. However, this model violates the causal sequence of X to M to Y specified by the mediation model because of the reversed paths from M to X and Y to M. As expected, choosing a correct model can be a tricky task with more parameters to be estimated. When the models are estimated in the SEM framework, it would be appropriate to estimate a series of nested models and conduct a series of chi-square difference tests to choose a model that better reflects the data, as well as

the theory. In addition, equality of path coefficients, such as paths for the effects of X on M (i.e., a1 , a2 , and a3 ) can be also tested by comparing the models with and without the equality constraints. Latent Growth Curve Mediation Models The latent growth curve (LGC) modeling (Meredith & Tisak, 1990; Muthén & Curran, 1997; Willett & Sayer, 1994) is another approach that can be used to evaluate longitudinal mediation models when X, M, and Y are repeatedly measured across three or more time points (Cheong, 2011; Cheong, MacKinnon, & Khoo, 2003; Mackinnon, 2008). Unlike the autoregressive mediation models, where the indirect effect is based on the differences among the study participants, the indirect effect in

Page 958

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

the LGC mediation model is conceptualized in terms of within-individual changes. In the autoregressive mediation models, the indirect effect is interpreted as the expected effect of X on Y at time t through M at time t – 1, if there was one unit change in X prior to time t – 1. Here, the changes in X, M, and Y are assumed equivalent to differences among participants in those variables. In LGC mediation models, individuals’ trajectories of X, M, and Y are estimated using the repeated measures across multiple measurement time points and the indirect effect is determined based on the relations between the within-individual changes in X, M, and Y. The capability of modeling mediation in terms of within-individual changes is important in developmental research, where the core research questions center on development, which is typically defined as within-individual change, and individual differences in development (Nesselroade, 1991). Using difference scores in a two-wave design is the simplest way of measuring individual change; however, difference scores contain minimal information about individual change (Rogosa, 1988; Rogosa, Brandt, & Zimowski, 1982; Willett & Sayer, 1994). The change patterns may follow nonlinear (e.g., quadratic or exponential) patterns or fluctuate in level between the two measurement time points. Growth curve modeling is considered a better method for investigating longitudinal within-individual changes and individual differences in those changes (Rogosa, 1988; Rogosa et al., 1982). Although growth curve models can be formulated both in the multilevel modeling framework (Raudenbush & Bryk, 2002; Goldstein, 1995) and in the SEM framework (Meredith & Tisak, 1990; Muthén & Curran, 1997; Willett & Sayer, 1994), we focus on the SEM approach in this section. In the LGC approach to mediation, the mediation process is typically modeled using a parallel process model (Muthén & Curran, 1997; Stoolmiller, 1994). The repeated measures of X, M, and Y are modeled as distinctive growth processes and the mediation is specified in the relations among the growth factors. As it is recommended that researchers investigate the longitudinal pattern of change in each variable before examining the longitudinal relations between the variables (Raudenbush, 2001; Rogosa, 1988), the estimation of growth trajectories that best describe the changes in X, M, and Y should precede modeling mediation based on the relations among the latent growth factors. In this approach, the growth trajectories of X, M, and Y are modeled in the measurement model, where the repeated measures relate to the latent growth factors according to the specified time scale. Once the growth

12:38pm

959

trajectory shape is appropriately modeled for each X, M, and Y, the full parallel process model is specified to model the mediation process in the structural model among the latent growth factors. Figure 21.17 shows an example of the LGC mediation model, where the growth trajectories of X, M, and Y are modeled as linear, based on repeated measures obtained with equal intervals across three waves. According to the specified time scales, the intercept factors (i.e., 𝜂1 , 𝜂3 , and 𝜂5 ) represent the level of each variable at time 1, i.e., the initial status. The slope factors (i.e., 𝜂2 , 𝜂4 , and 𝜂6 ) represent the change or growth rate of each variable per time unit. The mean and the variance of the intercept factor estimate the average level of initial status of each variable and the individual differences in the initial status. The mean and the variance of the latent slope factor estimate the average growth rate of each variable and the individual differences in the growth rates. Details of model specifications for LGC models can be found elsewhere (e.g., Muthén & Khoo, 1998; Willett & Sayer, 1994). The mediation process is specified in the relations among the latent factors shown in the middle of the model in Figure 21.17. More specifically, the individual differences in the initial status and the growth rates of M and Y are modeled as functions of the individual differences in the initial status and the growth rates of X and M. The relations among the latent factors in Figure 21.17 are described in the following equations: 𝜂3i = 𝜋0m + d1 𝜂1i + e3i

(21.85)

𝜂4i = 𝜋1m + a𝜂2i + d2 𝜂1i + e4i

(21.86)

𝜂5i = 𝜋0y + d3 𝜂1i + d5 𝜂3i + e5i

(21.87)

𝜂6i = 𝜋1y + b𝜂4i + c′ 𝜂2i + d4 𝜂1i + d6 𝜂3i + e6i

(21.88)

where 𝜂1i , 𝜂3i , 𝜂5i are the intercept factors and 𝜂2i , 𝜂4i , 𝜂6i are slope factors for the trajectory of X, M, and Y, respectively, varying across individuals indicated by the subscript i. In the previous equations, 𝜋0m , 𝜋0y , 𝜋1m , and 𝜋1y are the regression intercepts of the growth factors and e3i , e4i , e5i , and e6i are residuals of the growth factors. Equations 21.85 and 21.87 describe the intercept factors of M and Y, respectively, as functions of the intercept factors of X and M. These directional paths in Figure 21.17 are based on the causal theory of X to M to Y. If the first measures of X, M, and Y are obtained simultaneously, they can be specified as covariances. Equation 21.86 models the slope factor of M as a function of the intercept and the slope factors of

Page 959

Trim Size: 8.5in x 11in

960

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

M1

M2

1

1

M3 1 1

η3

2

η4

d1 d2

b a

d6 d5

η1

d4

1

1

η6

c´

η2

η5

d3

1 1

1 X1

X2

1

2

2 1

1

X3 Y1

Y2

Y3

Figure 21.17 A latent growth curve mediation model across three waves.

X, while equation 21.88 specifies the slope factor of Y predicted by the intercept and the slope factors of X and M. Although there are several possible indirect effects that can be of interest in this model, the main focus is on the relations among the slope factors, as the benefit of the LGC approach is the capability of modeling mediation in terms of within-individual changes in X, M, and Y. The coefficient a represents the effect of the growth rate of X on the growth rate of M. The coefficient b represents the effect of the growth rate of M on the growth rate on Y. The product of the two coefficients (i.e., ab) provides the estimate of the indirect effect that is based on the within-individual changes, determined by the extent to which the change in X affects the change in M (a) and the extent to which the change in M affects the change in Y (b). The standard error of the indirect effect and its standard error can be obtained in the same way described earlier and used for the significance test of the indirect effect. We fit a series of simple LGC mediation models to ECLS-K data from all 5 waves. Though preliminary models suggested that growth models did not fit the ADHD symptoms, interpersonal skill, and internalizing behavior data well, we nonetheless proceeded with the mediation models for the sake of illustration, specifying linear trajectories for the X, M, and Y variables. Unsurprisingly, we encountered

frequent convergence problems; estimation of the variance of the slope of the mediator trajectory seemed particularly problematic. This variance was estimated as negative in the LEARN → INTERP → INTERN model and was very close to zero in the LEARN → READ → INTERN and EXTERN → MATH → INTERN models. It was necessary to constrain this parameter (and its covariance with the intercept of the mediator trajectory) to zero to achieve convergence in the LEARN → MATH → INTERN and EXTERN → INTERP → INTERN models. The EXTERN → READ → INTERN model did not converge, even with this constraint. Both the CFI and RMSEA indicate poor fit for all models. Tests of the mediated effects involving predictor, mediator, and outcome slopes are reported in Table 21.10. For all models that converged, the estimate of the a path relating the slope of the predictor to the slope of the mediator was significant, though the relation between the slopes of LEARN and MATH was counterintuitively negative. Only two b path estimates were significant: the slope of READ was negatively related to the slope of INTERN, and the slope of INTERP was counterintuitively positively related to the slope of INTERN. Tests of the ab estimates of mediated effects show significant mediation in the LEARN → READ → INTERN and EXTERN → INTERP → INTERN models, but the

Page 960

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

12:38pm

961

TABLE 21.10 Coefficients and Standard Errors in an LGM Mediation Model X

M

Y

LEARN

INTERP

INTERN

LEARN

READ

INTERN

LEARN

MATH

INTERN

EXTERN

INTERP

INTERN

EXTERN EXTERN

READ MATH

INTERN INTERN

mediated effect in the latter model is in a counterintuitive direction. Though LGC mediation models can be very useful in explicating mediation effects in longitudinal data, this particular technique does not seem appropriate for this particular data, as a growth framework is not well suited to capture change over time in most of the variables. As mentioned earlier, it is required to inspect the pattern of change in the study variables carefully before putting the trajectories together in a comprehensive LGC mediation model. When the measures of X, M, and Y across waves are concurrent, the relations between the slope factors in the mediation model may be interpreted as correlational because the growth rates are estimated based on the measures across the same time span. However, this is less problematic than purely cross sectional data, because the growth rates are estimated based on the entire time span that the data cover and the effects of the changes in the predictors at earlier waves would be reflected in the changes in the outcome variables at later waves. An alternative method is to estimate growth trajectories in two-stage piecewise growth curve models (Raudenbush & Bryk, 2002; Khoo, 2001), where distinct growth trajectories are specified for early and later phases and the time ordering of the growth trajectories of X, M, and Y can be modeled accordingly. The indirect effects then can be estimated for different phases, such as the indirect effect via the earlier growth of M on the earlier growth of Y, the indirect effect via the earlier growth of M on the later growth of Y, and the indirect effect via the later growth of M on the later growth of Y.

Estimating LGC Mediation as a Multilevel Model The LGC mediation models discussed in the previous section were estimated using structural equation modeling techniques. The SEM framework is ideally suited for such analyses, as its ability to estimate a number of simultaneous equations allows for the estimation of

a 1.0538*** (0.0229) 0.1491*** (0.0087) −0.1135*** (0.0143) −1.0272*** (0.0242) Did not converge. −0.1056*** (0.0113)

b −0.0501 (0.0335) −0.1429*** (0.0307) −.01165 (0.1490) 3.7495*** (0.0507) 0.0796 (0.0409)

Med −0.0528 (0.0353) −0.0213*** (0.0047) 0.0013 (0.0169) −3.851*** (0.1046) −0.0084 (0.0044)

predictive and outcome effects of the same variable within a single analysis. For example, in a mediational analysis, the mediator serves as an outcome variable when examining the a path and as a predictor variable when examining the b path. It is relatively simple within this framework to have one arrow pointing to a variable (i.e., that variable is an outcome) and another pointing from that variable (i.e., that variable is a predictor of another variable) in an SEM diagram and to create the appropriate equations to estimate both effects simultaneously. The mediated effect involving latent slope parameters in Figure 21.17 could potentially also be tested within a multilevel modeling framework. The difficulty here is that latent growth in X is a predictor of latent growth in M, and that latent growth in M is a predictor of latent growth in Y. In a single multilevel model, it is simple to allow a variable to predict the growth parameter—just include that predictor in the L2 equation predicting that parameter—but there is no obvious way to expand the model to allow the growth parameter itself to predict another variable. However, after running a multilevel model, a researcher can obtain Empirical Bayesian (EB) estimates of growth parameters for each individual. These estimates combine individual-specific information from a single subject’s own data and the normative trajectory for the entire sample to produce stable and reliable estimates of person-specific effects (Candel & Winkens, 2003). The EB estimates of intercepts and slopes of the X variable can be entered as L2 predictors of the growth parameters in a multilevel growth analysis of the mediator variable to produce estimates of the d1 , d2 , and a parameters in Figure 21.17. Similarly, EB estimates of intercepts and slopes can be obtained when running the growth analysis for the mediator variable, and these (along with the EB estimates of intercepts and slopes from the first analysis) can be entered as L2 predictors of the intercept and slope parameters in the multilevel growth analysis of the outcome variable to produce the

Page 961

Trim Size: 8.5in x 11in

Cicchetti

962

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

remaining coefficients in Figure 21.17, including the b coefficient that relates the mediator slope to the outcome slope. Conducting multiple multilevel analyses seems a bit less elegant than the single comprehensive SEM, but equivalent estimates are produced, and thus longitudinal mediation can be assessed in either framework. Another possibility for recasting the LGC mediation model in the multilevel framework involves including all three variables (X, M and Y) in a single multivariate multilevel model (Hox, 2010; Raudenbush & Bryk, 2002). One way of conceptualizing such a model is as a three-level framework, with the variables distinguished at L1, repeated measures at L2, and individual subjects at L3. With three measurement occasions as shown in Figure 21.17, the data set would be organized with nine lines of data per subject, one for each of the X, M and Y variables at each of the three time points. The actual X, M, and Y values would be entered into a single generically named variable, V. Three dummy variables (D1 , D2 , and D3 ) would be created to identify which observations were associated with each of the three variables, and these dummies, in conjunction with a variable indicating the time point of measurement, would uniquely identify each of the nine unique values. The L1 model would consist of the three dummy variables, two of which will be zero for any given observation: Vvti = 𝜋1ti D1ti + 𝜋2ti D2ti + 𝜋3ti D3ti

L1:

(21.89)

𝜋1ti = 𝛽10i + 𝛽11i timeti + r1ti 𝜋2ti = 𝛽20i + 𝛽21i timeti + r2ti 𝜋3ti = 𝛽30i + 𝛽31i timeti + r3ti

(21.90)

and at L3: L3:

𝛽10i = 𝛾100 + u10i

𝛽11i = 𝛾110 + u11i

𝛽20i = 𝛾200 + u20i

𝛽21i = 𝛾210 + u21i

𝛽30i = 𝛾300 + u30i

𝛽11i = 𝛾310 + u31i

Vvti = D1ti (𝛾100 + 𝛾110 timeti + r1ti + u10i + u11i timeti ) + D2ti (𝛾200 + 𝛾210 timeti + r2ti + u20i + u21i timeti ) + D3ti (𝛾300 + 𝛾310 timeti + r3ti + u30i + u31i timeti ) (21.92) This is exactly the same as running three separate multilevel growth models on the X, M, and Y variables, but all three equations are combined into a single analysis. The advantage of the single analysis is that the L3 variance–covariance matrix of the u’s will be 6 × 6 and will contain off-diagonal elements that represent associations between growth in X and growth in M (a correlational version of the a path in Figure 21.17) and between growth in M and growth in Y (a correlational version of the b path in Figure 21.17). Because there is no error term at L1, this multivariate analysis could alternately be specified as a two-level model. This is actually the more practical specification, as it is quite difficult to suppress the L1 error term or constrain it to zero in most multilevel software packages. In separate equation form, the L1 model would be L1:

Vti = 𝛽10i D1ti + 𝛽11i D1ti timeti + 𝛽20i D2ti + 𝛽21i D2ti timeti + 𝛽30i D3ti

for variable v at time point t for subject i. Notice that there is no intercept and no error term in this equation. An intercept would imply that predictors could affect all observations equally, and we wish to keep X, M and Y distinct in this model. An error term is unnecessary because this level of the model simply sorts the variables by type, which should involve no error. The L2 models would be: L2:

These equations at each level can be algebraically combined to give the single-equation form of the model:

(21.91)

+ 𝛽31i D3ti timeti + r1ti + r2ti + r3ti Again, the specification contains no explicit intercept. The three dummy variables represent intercepts for their specific variables. Moreover, the time variable is entered in interaction with each of the dummy variables. These interactions represent the three slopes associated with the X, M, and Y variables. In addition, this equation includes three distinct error terms. This model could be estimated, for example, in SAS Proc Mixed (SAS Institute, 2014), which has the facility to estimate multiple L1 error terms corresponding to the distinct levels of a L1 variable with the group = option. A single variable distinguishing X, M, and Y with uniquely coded values would need to be created to accomplish this. The L2 specification for this model would be identical to the L3 specification of the three-level multivariate model and thus the single-equation forms of both models are identical. In the general multivariate multilevel model, associations among slopes are correlational and thus nondirectional. Mediation analyses, however, typically involve

Page 962

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

directional hypotheses (X → M and M → Y ). Examining such directional effects in the longitudinal mediation model requires the ability to include latent variables on the predictor side of the equation. The HLM program incorporates just such facility into the HLM2, HLM3, MLM, and MLM2 modules (Raudenbush, Bryk, Cheong, Congdon, & du Toit, 2011). The HLM manual provides an example in which a latent intercept is used to predict a latent slope. The method described in the manual could easily be expanded to simultaneously estimate growth parameters for more than one outcome variable and allow the slope of one to predict the slope of another. However, the latent variable modeling window in HLM currently allows each latent variable to be used as either a predictor or an outcome, but not both. So in a multivariate model including multiple measurements of X and M, this program could allow the slope of X to predict the slope of M, producing a coefficient and standard error for the a path in Figure 21.17. However, because M cannot simultaneously serve as both an outcome (on the a path from the X slope) and a predictor (on the b path to the Y slope), it would still be necessary to run a second multivariate multilevel analysis with multiple measurements of X, M, and Y to allow the slopes of X and M to predict the slope of Y to produce the b and c′ coefficients in Figure 21.17. So directional coefficients can be produced, and a mediated effect can be computed and tested from these, but incorporating the entire process into a single analysis will need to wait for further development of multilevel software. Autoregressive Latent Trajectory Mediation Models Integrating the unique strengths of the autoregressive models and the LGC models, Bollen and his colleagues (Bollen & Curran, 2004; Bollen & Zimmer, 2010) proposed the autoregressive latent trajectory (ALT) model. The ALT model parameterizes autoregressive relations between the repeated measures as in the autoregressive models, while simultaneously estimating the growth trajectories based on the pattern of changes across time. The multivariate application of the ALT model allows for parameterization of the autoregressive relations within each variable, the lagged relations between variables at specific measurement occasions, and the relations between the within-individual changes in those variables. Applying ALT models to investigation of mediation, researchers may be able to capture time-specific indirect effects as well as indirect effects based on the growth trajectories that are estimated across the entire time span the data cover.

12:38pm

963

In the ALT model, the measures of X, M, and Y at time 1 are treated as predetermined and including this predetermined time 1 measure takes care of the biases resulting from omitted variables prior to the time span the data cover. Individual trajectories in the ALT model are estimated based on the measures from time 2 to time t. The autoregressive relations within each variable are specified between the repeated measures from time 1 to time t. Thus, the measure at time t for t ≥ 2 is specified as the function of the variable itself measured at time t – 1 and the trajectory components (i.e., intercept and slope factors) estimated based on the variable measured from time 2 to time t. For example, X2 is determined by X1 and the intercept factor, while X3 is determined by X2 , the intercept factor, and the slope factor. The intercept factor in the ALT model captures the portion of the measure at time 2 that remains unexplained by the time 1 measure and the slope factor represents the growth rate estimated based on time 2 to time t measures. For an ALT model to be identified, four or more waves of data are needed (Bollen & Curran, 2006). Details about the ALT model specifications can be found elsewhere (Bollen & Curran, 2004; Bollen & Zimmer, 2010). The model in Figure 21.18 depicts an example of the ALT mediation model, where X, M, and Y are measured across four time points. The lag-1 effects between X, M, and Y at each wave are shown in the middle of the figure relating Xt−1 to Mt to Yt+1 . The mediation relations among the individual trajectories are represented by the dotted lines connecting directionally between the intercept factors and the slope factors of X, M, and Y. The time specific indirect effects are estimated based on the lagged effects between the repeated measures of X, M, and Y (e.g., X1 → M2 and M2 → Y3 ) as in the autoregressive mediation models and the indirect effect based on the within-individual changes in X, Y, and M is estimated in the relations between the slope factors (i.e., 𝜂2 → 𝜂4 → 𝜂6 ) as in the LGC mediation models. These two different types of indirect effects are independent of each other, as the autoregressive and the lagged effects are net of individual trajectory components in the ALT models. As expected, model specification can be complicated and it may take several steps to determine the appropriate final model; however, the ALT models can provide unique opportunities to simultaneously estimate and separate out the time-specific and the within-individual change based indirect effects in one comprehensive model. Given the difficulties encountered in estimating the LGC mediation models in the ECLS-K data, we did not explore ALT models for our particular example.

Page 963

Trim Size: 8.5in x 11in

964

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

M1

1

η3

η4

1 1

1

2

M2

M3

M4

Y2

Y3

Y4

X1

1 η1

X2

1 1 X3

η2

1 2

X4 Y1

1

1 1

1

η5

η6

2

Figure 21.18 An autoregressive latent trajectory mediation model across four waves.

Latent Change Score Models for Moderation and Mediation Before describing moderation and mediation in latent variable difference score models, a brief word is needed about manifest variable difference score models for longitudinal data. With two waves of data, MacKinnon (2008) suggests forming change scores for X, M, and Y, substituting the difference scores into the equations for the single mediator, and then testing for mediation using one of the cross sectional tests of mediation mentioned earlier. For three or more waves of data, difference scores may be computed and placed into an OLS autoregressive model to test for mediation. Although the steps for fitting the model and testing for mediation are the same as in the autoregressive mediation model, the interpretation of the results is quite different. Instead of a variable at one wave influencing another variable at a later wave, now it is the difference in scores on a variable across two waves influencing the difference in the scores of another variable across two later waves. With a categorical moderator variable, the simplest method to test for moderation in a LCS model would be to run a multiple group analysis and compare the parameter estimates in the different groups. Depending on the model

constraints, this would test whether the proportional or constant change between groups was different or could even test whether one group changed proportionally while another changed constantly or did not change. For observed continuous time-invariant predictors, the latent change score parameters can be regressed onto both predictors and the interaction term. Interactions between latent time-invariant predictors or between a latent time-invariant predictor and a latent change score parameter are possible as well, but require more advanced methods; Marsh, Wen, and Hau (2006) provided an overview of some of these methods. There are several ways an LCS model can be incorporated into a mediation model. Figure 21.19 shows one example of a LCS longitudinal mediation model where the mediated effect is composed of the slopes of the individual LCS models. That is, the slope of the change in X predicts the slope of the change in M, which in turn affects slope of the change in Y. This is comparable to the parallel process LGC model described by Cheong et al. (2003) because the processes are occurring at the same time. If the processes were measured after one another, the then a serial process LCS model could be specified.

Page 964

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

X1

ηX1

ΔX0

X2

ηX2

ΔX1

M1

M2

M3

M4

ηM1

ηM2

ηM3

ηM4

ΔM0

ΔM1

ΔM2

ΔM3

a

ξ1

12:38pm

965

ξ2

b c´ X3

ηX3

ΔX2

X4

ηX4

ΔX3

ξ3

ΔY0

ΔY1

ΔY2

ΔY3

ηY1

ηY2

ηY3

ηY4

Y1

Y2

Y3

Y4

Figure 21.19 Parallel process latent change score mediation model.

MacKinnon (2008) described a different LCS longitudinal mediation model, shown in Figure 21.20. In this model, rather than have the slopes of the change in individual variables predict one another, the error-free true score 𝜂 variables from one variable predict the latent change scores, Δ, in the other variable. For instance, 𝜂x1 predicts Δm1 and 𝜂m2 predicts Δy3 . This model is closer to an autoregressive model, because the variables only affect other variables at later waves. Note that the mean structure has been omitted from Figures 21.19 and 21.20 for ease of presentation. Additionally, an LCS model with a paths linking change in the predictor to change in the mediator (i.e., from Δ′x s to Δ′m s) and b paths linking change in the mediator to change in the outcome (i.e., from Δ′m s to Δ′y s) could plausibly be used to test longitudinal mediation. The process of fitting the LCS mediation models in Figures 21.19 and 21.20 is similar to fitting a LGC mediation model (MacKinnon, 2008). The first step is to fit the individual LCS models for X, M, and Y. This allows the researcher to specify the best fitting LCS model for each variable, such as determining whether a dual change or

proportional change model fits better. The second step is to add the mediation paths using the structural model. As in the LGC mediation models, other paths may be appropriate, such as a path from 𝜂x1 to the Δy2 , which is similar to the lag 1 direct effects from the autoregressive model. The Chi-square goodness of fit test, the RMSEA, and the CFI may be used to assess overall model fit. Testing for mediation can be accomplished in several ways. First, the joint significance test could be used by examining the significance of each of the individual paths. Second, specific indirect paths can be examined by using a MODEL INDIRECT command in the Mplus program and requesting bootstrap confidence intervals. As with the autoregressive models and LGC, the overall indirect effect can be examined using EQS, SAS PROC CALIS, and LISREL, but only Mplus will test specific indirect effects for significance. For in depth examples of fitting LCS models with LISREL see McArdle (2001) and Ferrer and McArdle (2003). We fit LCS mediation models to the first, third, and fifth grade ECLS-K data. Table 21.11 presents the results of the parallel process analysis similar to Figure 21.19. All six

Page 965

Trim Size: 8.5in x 11in

966

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

X1

ηX1

ΔX0

X2

ηX2

ΔX1

a1

c´1

M1

M2

M3

M4

ηM1

ηM2

ηM3

ηM4

ΔM0

ΔM1

ΔM2

ΔM3

c´2 X3

ηX3

ΔX2 a2

c´3 X4

ηX4

ΔX3

b1

b2

b3

a3 ΔY0

ΔY1

ΔY2

ΔY3

ηY1

ηY2

ηY3

ηY4

Y1

Y2

Y3

Y4

Figure 21.20 Alternate parallel process latent change score mediation model. Source: Adapted from MacKinnon (2008).

TABLE 21.11 Coefficients and Standard Errors in LCS Parallel Process Mediation Models X LEARN

M

Y

INTERP INTERN

LEARN

READ

INTERN

LEARN

MATH

INTERN

EXTERN INTERP INTERN EXTERN READ

INTERN

EXTERN MATH

INTERN

a

b

1.1302***

−0.6494***

(0.0231) 0.2639*** (0.0245) 0.2226*** (0.0130) 1.6164*** (0.1587) −0.2297*** (0.0201) −0.1864*** (0.0226)

(0.0323) 0.6345*** (0.0667) −0.0155 (0.0153) −0.0035 (0.1162) 0.2966*** (0.0302) 0.1590*** (0.0130)

Med −0.7340*** (0.0394) 0.1674*** (0.0235) −0.0035 (0.0034) −0.0057 (0.1878) −0.0681*** (0.0091) −0.0296*** (0.0043)

models included significant j and k coefficients for X, M, and Y, indicating both proportional and constant change over time for all constructs. RMSEA and CFI values for all models indicated less than ideal fit, particularly for models with interpersonal skills as the mediator. The tests of mediation in Table 21.11 indicate that the slopes of change in interpersonal skills and reading performance

significantly mediated the association between the slope of change in approaches to learning and internalizing behavior. Also, the slopes of change in both academic performance variables mediated the association between the slope of change in externalizing behavior and the slope of change in internalizing behavior. The direction of the mediated effect, however, was counterintuitive for three of the four mediated pathways. A second series of LCS mediation models was fitted to the first through fifth grade ECLS-K data in which true scores predicted latent change as in Figure 21.20. The results are reported in Table 21.12. Because three time points were included in the analysis, two distinct a and two distinct b paths were estimated in each model. a1 and b1 examine effects of first grade true scores on latent change from first to third grade, and a2 and b2 examine effects of third grade true scores on latent change from third to fifth grade. RMSEA and CFI values indicated less than ideal fit in all models, especially for those involving the interpersonal skills mediator. The tests of mediation in Table 21.12 indicate a significant a1 b1 estimate of mediation

Page 966

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

12:38pm

967

TABLE 21.12 Coefficients and Standard Errors in LCS Mediation Models with Paths from True Scores to Latent Change X

M

Y

LEARN

INTERP

INTERN

LEARN

READ

INTERN

LEARN

MATH

INTERN

EXTERN

INTERP

INTERN

EXTERN

READ

INTERN

EXTERN

MATH

INTERN

a1 1.0144*** (0.0435) 0.0004 (0.0059) −0.0170** (0.0066) −0.9607*** (0.0179) −0.0053 (0.0061) 0.0057 (0.0070)

a2

b1

0.9676*** (0.0469) 0.0770** (0.0057) 0.0143** (0.0068) −0.7689*** (0.0253) −0.0374*** (0.0064) −0.0137 (0.0074)

1.1118*** (0.1107) 0.0360 (0.0317) −0.0008 (0.0318) −0.0449** (0.0200) 0.0286 (0.0268) −0.0033 (0.0266)

for LEARN → INTERP → INTERN, but the temporally ordered a1 b2 estimate was nonsignificant. All three possible estimates of mediated effects for EXTERN → INTERP → INTERN were significant. The a2 b2 estimate in the EXTERN → READ → INTERN model was the only significant mediated effect involving the academic performance variables.

b2 0.0227 (0.0445) 0.0813 (0.0437) 0.0764** (0.0310) 0.3928** (0.1272) 0.1117** (0.0348) 0.0998*** (0.0266)

Meda1a2 1.1278*** (0.1223) 0.0001 (0.0002) 0.0001 (0.0005) 0.0431** (0.0192) −0.0002 (.0002) −0.0001 (.0002)

Meda1b2 0.0230 (0.0452) 0.0001 (0.0005) −0.0013 (0.0007) −0.3774** (0.1224) −0.0006 (.0007) 0.0006 (.0007)

Meda2b2 0.0220 (0.0431) 0.0063 (0.0034) .0011 (.0007) −0.3020** (0.0983) −0.0042** (.0015) −0.0014 (0.0008)

If the moderator is a time-invariant predictor, Zi , of one of the growth parameters, for example 𝛽0 , and interacts with another time-invariant predictor of the same growth parameter, Xi , then both the predictors and the interaction, Xi Zi , between these predictors would be added to the corresponding L2 equations in equation 21.35.

Moderation and Mediation in Exponential Decay Models

L1:

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + rti

Similar to standard multilevel models, moderation in an exponential decay multilevel growth model is accomplished by adding the moderating variable and appropriate interaction term at the correct level. If the moderator is a time-varying predictor, Zti , and interacts with another time-varying predictor, Xti , that is not time, then both the predictors and the interaction term, Xti Zti , would be added at L1 in equation 21.35.

L2:

𝛽0i = 𝛾00 + 𝛾01 Xi + 𝛾02 Zi + 𝛾03 Xi Zi + u0i

L1:

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + 𝛽3i Xti + 𝛽4i Zti + 𝛽5i Zti Xti + rti

L2:

𝛽0i = 𝛾00 + u0i 𝛽1i = 𝛾10 + u1i 𝛽2i = 𝛾20 + u2i 𝛽3i = 𝛾30 + u3i 𝛽4i = 𝛾40 + u4i 𝛽5i = 𝛾50 + u5i

(21.93)

Because of the nonlinear relation between time and the outcome variable, a linear interaction between time and a time-varying covariate is difficult to conceptualize and will not be discussed for the multilevel or latent variable exponential decay models.

𝛽1i = 𝛾10 + u1i 𝛽2i = 𝛾20 + u2i

(21.94)

L2 interactions may be of more interest for the exponential decay model as they test whether the effect of a time-invariant predictor on the initial value, rate of change, and asymptote value is dependent on the value of the moderator. For the latent variable exponential decay model with a time-invariant categorical moderator variable, the simplest method to test for moderation would be to run a multiple group analysis and compare the parameter estimates in the different groups. This would test whether the initial value, rate of change, and asymptote value were the same for the different groups. For observed continuous time-invariant predictors, the latent growth parameters can be regressed onto both the predictors and the interaction term. As with the LCS model, interactions between latent time-invariant predictors or between a latent time-invariant predictor and a latent growth parameter are possible for the latent exponential decay model, but require more advanced methods described by Marsh et al. (2006). If one of the variables is a time-varying predictor, Xti , and the other is a time-invariant predictor, Zi , then

Page 967

Trim Size: 8.5in x 11in

968

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

a cross-level interaction can be estimated by including the time-varying predictor at L1 and including the time-invariant variable as a predictor of the time-varying variable in the L2 equation in equation 21.35. L1:

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + 𝛽3i Xti + rti

L2:

𝛽0i = 𝛾00 + u0i

continuous, X and M are added at L1 of equation 21.35 and the effect of X and M on Y becomes Level1:

+ c′i Xit + rti Level2:

𝛽2i = 𝛾20 + u2i

𝛽2i = 𝛾20 + 𝛾21 Zi + u2i

L2:

Mti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + ai Xti + rti 𝛽0i = 𝛾00 + u0i 𝛽1i = 𝛾10 + u1i 𝛽2i = 𝛾20 + u2i ai = 𝛾30

bi = 𝛾30

(21.95)

This also allows for interactions between time-invariant predictors and time if Zi is added as a predictor of 𝛽1 . Note that the time-invariant predictor must also be included as an L2 predictor of 𝛽2 in order for the first-order effect Zi to be estimated in the model, which is necessary for the correct estimation of the interaction term Xti Zi . Though not technically an interaction, differential effects based on the relations between the growth parameters can be examined using the off-diagonal elements of the covariance matrix of the random effects, 𝝉. For example, if higher initial values were associated with higher asymptote values, then 𝜏02 would be positive and significantly different from zero. The same association can be examined in the latent variable exponential decay using the covariances in the 𝚽 matrix. The first step in constructing an exponential decay multilevel mediation model is to rewrite the equations from the linear single mediator model in the form of equation 21.35. When M and Y decay exponentially and X is a continuous variable, X is added at L1 of equation 21.35 and the effect of X on M becomes L1:

𝛽0i = 𝛾00 + u0i 𝛽1i = 𝛾10 + u1i

𝛽1i = 𝛾10 + u1i 𝛽3i = 𝛾30 + 𝛾31 Zi + u3i

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + bi Mti

(21.96)

where 𝛾00 , 𝛾10 , and 𝛾20 represent the mean values across subjects for the initial value, rate of change, and asymptote for M partialling out the effect of X, respectively, and 𝛾30 is the effect of X partialling out the effect of time (i.e., the natural exponential decay of M). Because M is also

c′i = 𝛾40

(21.97)

where 𝛾00 , 𝛾10 , and 𝛾20 represent the mean values across subjects for the initial value, rate of change, and asymptote for Y partialling out the effects of X and M, respectively, 𝛾30 is the effect of M partialling out the effects of time (i.e., the natural exponential decay of Y) and X, and 𝛾40 is the effect of X partialling out the effects of time and M. Krull and MacKinnon (2001) call this a 1 → 1 → 1 model because all three variables occur at L1. Here the three mediation parameters, ai , bi , and c′ i , have all been treated as fixed effects, meaning that the mediation effects do not vary between individuals. Adding random effects for these three paths would not change the interpretations of the coefficients, but it would indicate that the effect differs among individuals, an indication of the presence of moderation. For example, if an L2 predictor was added that predicted bi , such as a baseline value of M, this would create a cross-level interaction between that predictor and M, meaning the predictor of bi is a moderator of the effect of M on Y. For more information on moderation of a mediated effect see Fairchild and MacKinnon (2009); for more information on mediation at different levels see Bauer et al. (2006), Kenny et al. (2003), Krull and MacKinnon (2001), and Preacher et al. (2010). When X is a time-invariant variable, such as random assignment to a treatment, instead of entering X as an L1 covariate, X must be entered into the multilevel model at L2. The multilevel equations for the model of X on M then change from equation 21.95 to L1: L2:

Mti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i t] + rti 𝛽0i = 𝛾00 + a0 Xi + u0i 𝛽1i = 𝛾10 + a1 Xi + u1i 𝛽2i = 𝛾20 + a2 Xi + u2i

(21.98)

Page 968

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

Assuming X is a dichotomous dummy coded variable, then 𝛾00 , 𝛾10 , and 𝛾20 , are estimates of the initial value, rate of change, and asymptote of M for the group coded 0, respectively, and a0 , a1 , and a2 are the differences between the two groups in X such that 𝛾00 + a0 , 𝛾10 + a1 , and 𝛾20 + a2 are the initial value, rate of change, and asymptote of M for the group coded 1, respectively. The effects of M and X on Y change from equation 21.96 to L1:

Yti = 𝛽2i + (𝛽0i − 𝛽2i ) exp[𝛽1i tt ] + bi Mti + rti

L2:

𝛽0i = 𝛾00 + c′0 Xi + u0i 𝛽1i = 𝛾10 + c′1 Xi + u0i 𝛽2i = 𝛾20 + c′2 Xi + u0i bi = 𝛾30

(21.99)

Again assuming X is a dichotomous dummy coded variable, 𝛾00 , 𝛾10 , and 𝛾20 , are estimates of the initial value, rate of change, and asymptote of Y for the group coded 0, partialling out the effect of M, respectively, c′0 , c′1 , and c′2 are the differences between the two groups in X partialling out the effect of M, such that 𝛾00 + c′0 , 𝛾10 + c′1 , and 𝛾20 + c′2 are the asymptote value, initial value, and rate of change of Y for the group coded 1, respectively, and 𝛾30 is the effect of M partialling out the effect of time (i.e., the natural exponential decay of Y) and the effect of X on the exponential decay of Y. Krull and MacKinnon (2001) call this a 2 → 1 → 1 model because X is at L2 while M and Y are at L1. Adding M as a time-varying predictor in at L1 so that the model estimates the relation between M and Y at the same point in time violates the temporal precedence assumption. An alternate method would be to introduce a lag such that M predicts Y at later time points. With a lag of one wave, Mit in equation 21.98 would change to M(t−1)i , so that each measurement of M predicts the measurement of Y at the next time point. Then bi would be the average relation across time between M at time t and Y at time (t + 1), partialling out the effect of time. One issue with lags is that M must be measured prior to Y. Also, although the model makes no assumption that the measurements of M or Y are equally spaced, when the relation between M and Y is lagged, if the measurements are not equally spaced, bi will represent the average effect across different amounts of time. For example, if M and Y were measured at 1, 3, and 12 months, introducing a lag of one wave would mean that M1month would predict

12:38pm

969

Y3months , a lag of 2 months, while M3months would predict Y12months , a lag of 9 months, an undesirable situation. The value of M immediately following an intervention could also be entered as an L2 predictor of the rate of change and asymptote of Y. If X had a significant effect on the initial value of M, which in turn predicted the rate of change and asymptote of Y, then the effect of X on the rate of change and asymptote of Y would be mediated through its initial effect on M. This does not satisfy temporal precedence since the initial value of Y is used to estimate the rate of change and asymptote of Y, but it does make a stronger case for mediation than in equation 21.98. Using the exponential decay model shown in equations 21.95 and 21.96, when X and M are added to the model at Level 1, mediation can be tested using the joint significance test by testing a and b for significance. The mediated effect ai bi can also be tested by dividing it by the standard error Var(̂ â b+𝜎 ̂aj,bj ) = b2 Var(̂ a) + a2 Var(̂ b) + Var(̂ a)Var(̂ b) + 2abCov(̂ â b) + Cov(̂ â b)2 + Var(̂ 𝜎aj,bj )

(21.100)

that includes the covariance of the a and b parameters (Bauer et al., 2006; see also Kenny et al., 2003; Krull & MacKinnon, 1999, 2001; MacKinnon, 2008). However, this assumes that the distribution of the mediated effect is normal, which is often not the case. For this reason, creating a confidence interval around the mediated effect using the percentile bootstrap may be preferred. Testing for mediation when X is entered at L2 such as in equations 21.97 and 21.98 is slightly different than for the continuous L1 X variable model in equations 21.95 and 21.96 as there are now three individual effects of X on M, a0 , a1 , and a2 . Each of these parameters represents a different effect of X on M, and because of this, three different mediated effects can be estimated, a0 bi , a1 bi , and a2 bi , all of which may be tested for significance using one of the methods described earlier. As discussed during the presentation of autoregressive models, each of these mediated effects represents a specific indirect effect, any or all of which may be of substantive interest. Several different exponential decay mediation models are possible when using an LGC approach. The first is a direct extension of the model in equations 21.95 and 21.96. Illustrated in Figure 21.21, this model begins by fitting separate exponential decay LGC models for X, M, and Y. Then the observed M variables are regressed on the observed X variables and the observed Y variables are

Page 969

Trim Size: 8.5in x 11in

970

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

βM2

βX2 βX1

βM1 X1

βX0 X2

M1

a*

M2

a*

b*

X3

M3

a*

X4

βM0

b* M4

a* b* c´*

b*

c´* Y1

Y2

βY0

c´*

c´* Y3

βY1

Y4

βY2

Figure 21.21 An exponential decay latent growth curve model with a continuous X variable and direct paths between observed variables.

regressed on the observed X and M variables, constraining all of the paths between sets of observed variables to be equal. For the dichotomous X variable model in equations 21.97 and 21.98, separate exponential decay LGC models are fit for M and Y, and the observed Y variables are regressed on the observed M variables, again constraining these paths to be equal. The latent growth parameters from M and Y are then regressed onto X. Another form of exponential decay LGC mediation model is a parallel process model. As was described earlier, an exponential decay LGC model is equivalent to a linear LGC model, only with an exponential decay trajectory across time. Hence, the individual linear LGC models in the parallel process LGC mediation model may be replaced with exponential decay LGC models to form a parallel process LED mediation model, as shown in Figure 21.22. The parallel process mediation model in Figure 21.22 only contains paths between the rate of change growth parameters. However, depending on the theory behind the model, other paths may be appropriate. For instance, if X were dichotomous and had an effect on the initial value of M, then it may be of interest to include a path between

the initial value of M and the rate of change of Y. Or, continuing with the dichotomous X example, if M reached asymptote before Y, it could be of interest to investigate whether the asymptote of M had an effect on the asymptote of Y. Testing for mediation would be conducted similarly to the parallel process LGC by testing the significance of individual paths or by testing specific indirect paths for significance. In addition, various fit indices given by most structural equation modeling software can be used to assess overall model fit. Testing for mediation in the exponential decay LGC model is similar to testing for mediation in the multilevel exponential decay model in that the meditational parameters are tested for significance or individual indirect effects are directly tested for significance. One advantage of the latent variable framework is the addition of fit indices, which are provided by most SEM software packages and provide valuable information about the overall fit of the model and the presence of a significant mediated effect. Another advantage is the ability to model measurement error and have multiple indicators for a variable. Finally, the latent variable framework allows for the correlation of residuals and growth parameters as well as the ability to relax the equality constraints for the mediation parameters (e.g., the effect of M1 on Y1 does not have to equal the effect of M2 on Y2 ). Because the ECLS-K variables in our example do not include a treatment or intervention, we did not fit illustrative exponential decay models to this data. Moderated Mediation in SEM Sometimes a researcher may wish to test a hypothesis about whether longitudinal mediation effects differ depending on the value of a particular variable, i.e., moderated longitudinal mediation. If the longitudinal mediation is modeled using one of the SEM approaches outlined above, an appropriate test of such moderation can be conducted using techniques for incorporating moderation into this framework, the nature of which depend on the type of moderator involved. If the moderator Z represents a categorical variable with a relatively small number of levels, such as clearly defined group membership (e.g., sex, religion, grouping on age), the multiple group approach (i.e., stacked analysis) in SEM (Jöreskog, 1971) can be used for investigating the moderated mediation effects (Little, 2013; MacKinnon, 2008). Since the independent variable and the mediator may display differential effects in the different groups, stacking groups together allows for comparisons of the parameters, with some parameters

Page 970

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Models for Moderation and Mediation in Longitudinal Data

M1

M2

M3

12:38pm

971

M4

X1 βM0

βX0

βM2

βM1

a

X2 βX2

b X3 c´ βX1

X4

βY0

Y1

βY2

Y2

βY1

Y3

Y4

Figure 21.22 A parallel process exponential decay latent latent growth curve model with a dichotomous X variable and direct paths between latent growth parameters.

constrained to be equal, while other parameters specified to vary between the stacked groups. The parameters that differ between the groups indicate interactions between the variable providing the group membership and the variables in the model, such as X and M. Typically, it is assumed that the same mediation model holds for the stacked groups and the equality of parameter estimates for the relations between X and M (i.e., a paths), M and Y (i.e., b paths), or X and Y (i.e., c′ paths) are examined. Bollen (1989) suggested that researchers start with the model without restricting any parameters across the groups. A satisfactory fit would indicate that the same form of model holds across the stacked groups and provide a rationale for further testing the equivalence of the parameters. The moderated mediation can be investigated from the final model where the equivalent parameters are constrained and nonequivalent parameters are freely estimated. This approach can be applied to structural

equation models with observed variables and also models with latent variables for estimating the interaction effect between the latent variable and the categorical group variable (Rigdon, Schumacker, & Wothke, 1998). When X, M, or Y are latent variables, a series of hierarchical tests for invariance should be conducted for the measurement models before examining the equality of the structural relations, to assure that the measures used for the latent variables X, M, and/or Y are equivalent across the groups. If the growth trajectories of X, M, and Y are estimated for the LGC mediation, researchers should inspect if the trajectory shapes are equivalent or different across the stacked groups before testing the moderated mediation. The steps for testing measurement invariance can be found elsewhere (Horm & McArdle, 1992; Jöreskog, 1971; Meredith, 1993). In a stacked model for moderated mediation, several hypotheses can be tested, including a test of the equality of the relation between X and M (i.e., a paths), a test of the

Page 971

Trim Size: 8.5in x 11in

972

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

equality of the relation between M and Y (i.e., b paths), or a test of the equality of the relation between X and Y (i.e., c′ paths) across the groups. More importantly, further analysis to test the equality of the indirect effects (i.e., ab estimators) is needed. While the tests of the equality of the relations among X, M, and Y may provide the information on the source of differential indirect effects across the groups, it is possible that the estimated indirect effects obtained in the stacked groups may not statistically differ even when the magnitude of the a paths or the b paths differ across groups. The equality of the path coefficients can be tested by comparing a series of nested models where the relevant paths are constrained to be equal across groups. The equality of the estimated indirect effects can be tested by a t test by dividing the difference in the mediated effects by the pooled standard error of the mediated effect as follows: ta1 b1 −a2 b2 =

√

(a1 b1 −a2 b2 )∕

s2 pooled

(21.101)

In equation 21.100, a1 b1 and a2 b2 are the estimated √ indirect

effects from each of the stacked groups and s2 pooled is the pooled standard error of the mediated effect in each group (MacKinnon, 2008). The equality tests described for the two-group case in this section can be easily extended to cases with more than two groups. The analysis procedures described above can be carried out for each pair of groups to locate the sources of the moderation by group membership. When the moderator Z is a continuous observed variable interacting with an observed X or M variable, the moderated mediation can be examined using the product term created to reflect the moderation of Z in the relations between X and M (i.e., a path), between M and Y (i.e., b path), or between X and Y (i.e., c′ path). Again, the two variables involved in the product term (e.g., X and Z) need to be centered with respect to their means (e.g., X − X and Z − Z) before creating the product term, because a product term based on the uncentered variables will be highly correlated with the first-order effect terms and the high correlations between the predictors may introduce collinearity problems (Aiken & West, 1991; Marquardt, 1980). Centering the variables will minimize or remove the correlations between the product term and the first-order effect terms. The centered variables and their product term are then entered in the model to predict M or Y to test the moderation by Z in the specific relations in the mediation model. With significant interaction effects, further probing of the interaction patterns can be conducted by obtaining simple slopes at different levels of the moderator Z as described

earlier in this chapter. Some SEM software programs such as Mplus (Muthén & Muthén, 2014) calculate the indirect effects at different levels of Z and produce the plots of indirect effects corresponding to particular values of Z. When the moderator is a latent variable interacting with other latent variables in the model, the moderated mediation model can be investigated using the latent interaction variable. Since the use of latent variables in SEM has many advantages, including the capability of controlling for measurement errors that could obscure the interaction effects, various approaches to estimating interactions between latent variables (Arminger & Muthén, 1998; Bollen & Paxton, 1998; Jöreskog, 1998; Kenny & Judd, 1984; Klein & Moosbrugger, 2000; Marsh, Wen, & Hau, 2004; Wall & Amemiya, 2000) have been proposed in the past years. These approaches commonly specify the indicators of the latent interaction variable with the product of the indicators of the two latent variables, but they differ in the ways they specify the nonlinear constraints used for the factor loadings and variance associated with the latent interaction variable. As expected, the modeling process of these approaches may be complicated and unwieldy, and this may be the reason why such methods have not been frequently applied to the substantive research (Rigdon et al., 1998). However, the latent interaction variable approaches have the potential to be applied to estimating moderated mediation, although more research is warranted to investigate the applicability of these approaches for the models with more than one interaction term (Marsh et al., 2004). To illustrate a basic moderated longitudinal mediation analysis, we fit a multiple groups SEM for boys and girls using the autoregressive mediation model including both lagged and contemporaneous a and b paths (as in Figure 21.16 and Table 21.9) for the LEARN → INTERP → INTERN pathway, which was the best fitting longitudinal mediation model in the first to fifth grade ECLS-K data. Comparisons of the fit of nested models revealed that several parameter estimates differed by gender. Specifically, the autoregressive path from LEARN at third grade to LEARN at fifth grade and both autoregressive paths involving the INTERP mediator were larger for boys than for girls. In addition, boys had larger error variances and disturbance variances than girls for all three variables at all time points. The only mediation effect parameters that differed by gender were the lagged a paths. Boys had stronger associations between first grade LEARN and third grade INTERP (aboys = –0.1786, SE = 0.0142, p < .0001 vs. agirls = –0.1272, SE = 0.0150, p < .01) and between third grade LEARN and fifth grade INTERP than girls (aboys = –0.1686, SE = 0.0152, p < .0001 vs. agirls =

Page 972

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Modern Causal Inference and Longitudinal Mediation Models

–0.1250, SE = 0.0155, p < .01). The test of the difference between boys’ and girls’ ab estimates for the temporally ordered X 1 → M3 → Y 5 mediation effect was significant (a1 b1 – a2 b2 = –0.0019, SE = 0.0010, p < .05), indicating that the effect of approaches to learning on internalizing behavior mediated by interpersonal skill was, in fact, stronger for boys than for girls.

MODERN CAUSAL INFERENCE AND LONGITUDINAL MEDIATION MODELS There has been a dramatic increase in research on causal inference related to mediation analysis over the last decade (Imai, Keele & Tingley, 2010; MacKinnon, 2008; Pearl, 2012, Valeri & VanderWeele, 2013). In the psychology literature, causal inference has primarily focused on the need for longitudinal rather than cross sectional relations to demonstrate mediation (Cheong et al., 2003; Cole & Maxwell, 2003, Judd & Kenny, 1981; MacKinnon, 2008; Maxwell & Cole, 2007). These articles emphasize the central role of demonstrating temporal precedence such that X occurs before and causes later change in M and M occurs before and causes later change in Y, as discussed in a recent article (Maxwell, Cole, & Mitchell, 2011) and commentary on this topic (West, 2011). Temporal precedence refers to the demonstration that for a variable to be a cause of an effect, it must occur before the effect. The mediation model is a longitudinal model such that X causes M and M causes Y, so evidence for this longitudinal causal process is ideally obtained from longitudinal data. Mediation relations between variables measured at one time represent associations between levels of the variables such that a person with high scores on X also has higher scores on M and Y, as an example. In some situations, temporal order between the variables is reasonable such as when X represents random assignment to an intervention or measures of M are taken before measures of Y. Cole and Maxwell (2003) demonstrated algebraically the bias that can result when cross sectional relations are treated as if they are longitudinal relations. The notion that relations between levels of variables at a single time point differ from relations across time applies to any difference across any period of time. For example, the mediated effect for measures that differ one minute in time may differ from measures that differ by one year in time. It is not necessarily the cross sectional versus longitudinal nature of measures that is important. It is the match between when change occurs between variables in the true population mediation model and whether the variables are

12:38pm

973

measured when this mediation relation can be detected (MacKinnon, 2008). Another important causal inference topic related to temporal precedence is the influence of omitted variable bias on mediation relations (MacKinnon & Pirlott, 2015). There has been an early and sustained focus of attention in the social sciences that accurate estimation of mediation depends on the consideration of other variables that may explain or alter a mediation conclusion from a research study (James, 1980, Judd & Kenny, 1981). In the medical and epidemiology literature, causal inference for mediation has primarily addressed the issue of how confounder variables may bias mediation estimation. Omitted variable bias and confounder bias are considered synonyms in this chapter representing unmeasured variables that influence mediation relations. The potential outcome model provides a way to define causal mediation analysis in the context of confounders or omitted variables. The potential outcome model for causal inference is relevant for many fields and its application has steadily grown. The beginnings of the potential outcome approach for mediation in the social sciences can be traced to an influential paper by Holland (1988). Holland applied the Rubin causal model (Rubin, 1974) to demonstrate the challenges of identifying causal mediation effects even when X represents randomization to conditions (see also Sobel, 2008). In epidemiology, Robins and Greenland (1992) described the restrictive assumptions necessary for identification of causal mediation. Pearl (2001) described and developed a solution discussed in Robins and Greenland (1992) to obtain causal mediation effects that serves as the source for a series of important recent papers on causal mediation from the potential outcome perspective (Imai, Keele & Tingley, 2010; Imai, Keele, & Yamamoto, 2010; Lynch, Cary, Gallop, & Ten Have, 2008; Pearl, 2012; Ten Have et al., 2007; Valeri & VanderWeele, 2013; VanderWeele & Vansteelandt, 2009; Vansteelandt, 2009). These new papers provide some practical guidance and software for researchers to use. All of these previous papers have focused primarily on the cross sectional mediation case of measurement of X, M, and Y, and not the longitudinal case. The purpose of this section is to describe potential outcome model approaches for investigation of mediation effects in the context of longitudinal data. First the potential outcome model for evaluating intervention data is described, followed by applications of this model to the case of mediation. Mediation and moderation relations are described for this potential outcome model. Additional information regarding the potential outcome model for longitudinal data is outlined followed by the estimation

Page 973

Trim Size: 8.5in x 11in

974

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

of causal estimators from data. Other approaches to longitudinal mediation from a modern causal inference perspective are briefly described. Potential Outcomes Model The potential outcomes approach to estimating causal effects (Holland, 1986, 1988; Pearl, 2009; Rubin, 1974) distinguishes between an individual’s observed and counterfactual outcomes. A counterfactual outcome is an outcome for a participant in a research condition in which that participant did not serve (the condition contrary to actual condition they did serve in). For the case of two groups, let variable X be an intervention with level x (x = 1 for the intervention, x = 0 for the control) and variable Y be the outcome variable. The potential outcomes framework considers all conditions that an individual could serve in, including both intervention and control conditions, even though observed data can only be obtained for one of the two groups. In other words, these models consider all potential outcomes, so if an individual is assigned to the intervention group, the potential outcome Y(1), is the observed outcome Y. The counterfactual outcome for that individual is the value for the outcome variable if that person was assigned to the control group, Y(0). The individual causal effect is equal to Y (1) − Y(0), even though the value in the control group cannot be observed for a participant in the intervention group. Because it is not possible for the same individual to simultaneously serve in both intervention and control groups, the individual causal effect is not estimable—a problem called the fundamental problem of causal inference. However, assuming that participants are exchangeable, then it is possible to compute a causal effect averaged across individuals in each group and this average causal effect is computed as E[Y(1) – Y(0)], or the expected average difference between intervention and control groups. This average causal effect is a causal estimator under certain assumptions, primarily that individuals have been randomized to the two conditions. The potential outcome approach applied to mediation introduces a mediating variable M with level m that mediates the relation between X and Y. The potential outcome approach to mediation demonstrates several complexities in the estimation of mediation. The additional complexity is shown in the additional effects that are now relevant when including the mediator, M, in causal sequence from X to M to Y (Hogan & Lancaster, 2004; Jo, 2008; Pearl, 2001, 2012; Robins & Greenland, 1992; VanderWeele & Vansteelandt, 2009). In addition to the average causal effect now called the total effect (TE) of X on Y, there are

five additional causal effects: the controlled direct effect (CDE) which is the direct effect of X on Y at a certain value m of M, a direct effect in the control group called the pure natural direct effect (PNDE), a direct effect in the intervention group called the total natural direct effect (TNDE), the indirect effect in the control group called the pure natural indirect effect (PNIE), and the indirect effect in the intervention group called the total natural indirect effect (TNIE). Use Y(x, m) to denote the potential outcome for an individual under the intervention level x and mediator level m. X takes the value x = 0 for the control group, and x = 1 for the intervention group. If the actual value of M is m for an individual, then the counterfactual ′ value of M for that individual is denoted as m . The controlled direct effect of intervention X on outcome Y is the direct effect of the intervention on the outcome at a fixed level m of the mediator M: CDE = Y (1, m )– Y (0, m)

(21.102)

The natural direct effect of X on Y is different from the CDE in that M is set to the level M(x), the level that would have naturally occurred under one of the conditions of X, so there are now two natural direct effects corresponding to intervention and control groups. In the case of M(0), the PNDE is the effect of intervention X on outcome Y when X did not influence the mediator M (or the participants were assigned the mediator level under the control condition). PNDE = Y 1, M(0)–Y (0, M(0))

(21.103)

In the case of M(1), the TNDE is the effect of intervention X on outcome Y when X did not influence the mediator M (or the participants were assigned the mediator level under the intervention condition). TNDE = Y 1, M(1)–Y (0, M(1))

(21.104)

Following the same approach, the pure natural indirect effect is the effect of the intervention on Y when changing the level m of M in the control condition. In other words, the PNIE is the effect of the intervention on outcome when the level of M is changed when X is set to 0. PNIE = Y 0, M(1)–Y (0, M(0)).

(21.105)

The total natural indirect effect (TNIE) is the effect of the intervention Y when changing the level of M in the

Page 974

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Modern Causal Inference and Longitudinal Mediation Models

intervention condition. In other words, the TNIE is the effect of intervention on outcome when the level of M is changed when X is set to 1. TNIE = Y 1, M(1)–Y (1, M(0))

(21.106)

The total effect is equal to the sum of the pure natural direct effect plus the total natural indirect effect, and the total effect is also equal to the sum of the pure natural indirect effect plus the total direct effect. Because there are different effects for intervention and control groups, an interaction is included in this model insomuch as the direct and indirect effects may differ in the two groups. The direct and indirect effects are different in the two groups if the interaction between X and M in the prediction of Y is nonzero. If the interaction is zero, then the two direct effects are equal and the two indirect effects are equal and these results reduce to the traditional estimator of the direct and indirect effects. Importantly, Pearl (2001) demonstrates that the decomposition of the total effect into the natural direct and indirect effects holds even in models with interactions and nonlinear models such as logistic regression (i.e., the pure natural direct effect plus the Total Natural Indirect Effect equals the total effect). These estimators described already are estimators of causal effects under certain assumptions. Two assumptions are required for controlled direct effect to be identified (Pearl, 2001; VanderWeele, 2010): 1. No unmeasured confounders for the relation between X and Y. 2. No unmeasured confounders for the relation between M and Y. Two additional assumptions are required for natural direct and indirect effects to be identified: 3. No unmeasured confounders for the relation between X and M. 4. No measured or unmeasured M to Y confounders are affected by intervention. Assumptions (1) and (3) refer to the ignorability of intervention (i.e., the intervention independent of potential outcomes for the mediator and outcome) given observed pretreatment confounders. This assumption is satisfied with randomization of X. Assumption (2) refers to the ignorability of the mediator and the potential outcomes for Y given the observed intervention and pretreatment confounders. This ignorability

12:38pm

975

of the mediator assumption may be difficult to satisfy in actual research because randomization of M is usually not plausible for many studies. In other words, the value of the mediator is not randomly assigned to participants but is self-selected. Even when conditioning on observed confounders for the relation between M and Y, there can still be unobserved confounders that would lead to violation of this assumption. This assumption is strong and is often ignored in studies even though a causal interpretation of a mediated effect is not possible without its existence (VanderWeele & Vansteelandt, 2009; MacKinnon, 2008). Methods described in this article hold that assumption 3 is true and methods to investigate this assumption are only starting to be developed (Coffman & Zhong, 2012). Assumption 4 states that there are not confounders affected by the intervention that confound the M to Y relation. One goal of multiple mediator models is to assess these additional posttreatment confounders in a way that allows for causal estimation. However, approaches to deal with multiple mediators are an active area of current research in causal inference (VanderWeele & VanSteelandt, 2013). And in the application of the potential outcome model for longitudinal mediation, effects of earlier waves on later waves could be postintervention confounders that violate assumption 4. The assumptions described above have also been described as sequential ignorability assumptions in the research literature on causal mediation (Imai, Keele, & Tingley, 2010; Pearl, 2001; VanderWeele & Vansteelandt, 2009). The linear regression approach as well as potential outcome methods including the natural direct and indirect effects assumes sequential ignorability, which consists of the ignorability of the intervention assignment and the ignorability of the mediator. In other words, random assignment of X ensures that c and a represent causal effects (with assumptions 1 and 2) but b and c′ do not have a causal interpretation without further assumptions (Holland, 1988). For instance, given that the individuals are randomly assigned to conditions of an intervention, there are not confounders of the X to Y and X to M relations. However, because participants are not randomly assigned to values of M, the relation between M and Y, b, is not a causal effect because there can be confounders that account for the effect of M on Y. Comparison of Traditional and Potential Outcome Mediation Analysis For the case of continuous M and Y as in most social science studies, there is a direct correspondence between the

Page 975

Trim Size: 8.5in x 11in

976

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

estimators of the direct and indirect effects in traditional and the potential outcome methods. If all assumptions are satisfied, traditional and potential outcome methods come to the same conclusion about direct and indirect effects if the interaction of X and M on Y is zero. So the traditional estimators of ab and c′ correspond to the natural indirect and natural direct effects causal estimators. When the interaction between X and M is included in the statistical analysis, the natural indirect effects correspond to simple mediation effects in the traditional analysis where the simple mediation effects are at specific values of the mediator. However, typically the interaction of X and M is included in traditional analysis as a way to test the assumption that the causal relation of M to Y does not differ across levels of X. If the interaction is nonsignificant, it is often not included in traditional mediation analysis mainly because the relation of M to Y is often based on prior research. To clarify, if the interaction is included in the statistical model then there are simple mediation effects in traditional analysis corresponding to different b paths in the statistical analysis. These effects would reflect the mediated effect at different values of the mediator. There are corresponding effects in the potential outcome approach that refer to the effects for the control group called the pure effects and the effects for the intervention group called the total effects. For the potential outcome approach, the interaction of X and M is included in the analysis to calculate the different direct and indirect effects in each group. Moderation in the Potential Outcome Model A described already, the potential outcome model has different direct and indirect effects in the intervention and control groups. These additional estimators are only present if the interaction of X and M is nonzero. In this potential outcome framework, an interaction of indirect effects corresponds to testing the equality of the natural indirect effect in the two groups. Similarly, moderation effects could be obtained by testing the causal effects across subgroups where each subgroup could have different natural direct and indirect effects. A statistical test of the interaction across subgroups could be accomplished by dividing the difference between natural indirect effects across groups by a multivariate delta method standard error for the difference between the causal effects or by using a resampling method such as the bootstrap to create a confidence interval for the difference between the effects and testing whether zero is in the interval.

Potential Outcome Mediation Model with Longitudinal Data The causal perspective on longitudinal mediation elucidates a problem in the regression analysis of data with more than one wave. The description of this problem in the causal inference literature is in the context of time-varying interventions or situations in which earlier waves of data predict later treatment and later values of the outcome (Robins & Hernan, 2009). The problems with longitudinal data from a causal inference perspective are illustrated here using a study of the effects of loneliness on depression as described in VanderWeele, Hawkley, Thisted, and Cacioppo (2011). There is evidence that loneliness and depression are separate constructs and that loneliness leads to increased depression (see VanderWeele et al., 2011) but it would be useful to estimate the size of this effect. Say that depression at the fourth time of measurement is the focus of the analysis. Depression at the fourth time point is predicted by the following variables: depression at the first, second, and third times and loneliness at the first, second, and third times. All of these variables are likely to have some effect on Time 4 depression, but there is a problem with estimating all of these effects at once. Depression at Time 2 is on the pathway from loneliness at Time 1 to depression at Time 4, so it may mediate the effect of loneliness on later depression thereby reducing the total effect of loneliness on later depression. In other words, depression at Time 2 could be a mediator of the relation of Time 1 loneliness on depression at Time 4. Because depression at Time 2 is a potential mediator, it would not be ideal to adjust effects if a researcher was interested in estimating the total effect of earlier loneliness on later depression. Depression at Time 2 is also related to both depression and loneliness at Time 3 so it is also a confounder of these variables on Time 4 depression. Because it is a possible confounder, depression at Time 2 must be adjusted to get an accurate estimate of the relation of loneliness and depression at Time 4. So depression at Time 2 should be adjusted for to remove confounding, and it should also not be adjusted for so that it does not block the total effect of loneliness at Time 1 on depression at Time 3. The correct analysis would both adjust and not adjust for Time 2 depression, which is not possible with regression analysis. Robins, Hernan, and Brumback (2000) proposed a solution to this problem by conducting an analysis that removes confounder effects so the true effects can be estimated. In this context, a model called a marginal structural model removes effects of variables that may serve as a confounder to estimate the effects of

Page 976

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Modern Causal Inference and Longitudinal Mediation Models

loneliness on depression at later waves of measurement. The approach uses a weighting method to render an unconfounded data set. In terms of an intervention, X, and a mediation model with two waves of data for M, M1 and M2, and Y, Y1 and Y2, say we are looking at the effect of the intervention, X, on the outcome Y2. If control is made for a variable that is a consequence of the intervention, the mediator M2, then the relation of X to Y2 will be biased. By doing this, part of the effect of the intervention is blocked. For the mediation relation, X to M2 to Y2, control for X is needed to get the correct relation of M2 to Y2, that is X is a confounder of the M2 to Y2 relation. It is not possible with a single regression model to obtain an unbiased estimate of X on Y2 because it must simultaneously control for M2 and not control for M2. With a structural equation model, it is possible to estimate relevant indirect effects for effects from X to Y2 when X is uncorrelated with M1 and Y1 because the level of X is randomized to groups. The single mediated effect in this model corresponds to X to M2 to Y2. With the marginal structural model and a nonrandomized X, the time 1 measures M1 and Y1 (before the intervention is delivered) are used to weight M2 to remove confounding and then a weighted analysis is used to estimate the causal effect of the unconfounded M2 variable on Y2. A similar approach is illustrated below. Inverse Probability Weighting with Longitudinal Data An approach to obtaining effects adjusted for confounders in longitudinal mediation models is to use observed covariates to measure confounder effects and adjust analyses to remove the confounder bias (Robins et al., 2000). The goal of these analyses is to create a pseudo-population of data in which confounder bias has been removed and then conduct analyses on this new confounder-free data set. This new confounder-free data set avoids the problem of simultaneously adjusting for and not adjusting for an intermediate variable that is also a confounder. It accomplishes the goal by separately estimating the confounder effect and other effects. A weighting strategy is used to weight observations based on how much the observations are confounded to create an unconfounded data set. A statistical model is first developed for the prediction of the mediator based on covariate information from baseline measures. These baseline measures can be the measure of M and Y at baseline as well as other available variables. The weights for each participant are obtained from the residuals of two models (1) a statistical model in which only X predicts the mediator, and (2) a statistical model in which X and all covariates are used to predict the mediator.

12:38pm

977

A weight is formed by dividing the probability of the residual in the first model by the probability of the residual in the second model. A weighted regression model is used in the analysis to predict Y with M and X (Coffman, 2011; Coffman & Zhong, 2012; Robins, Hernan, & Brumbeck, 2000; Toh & Hernan, 2008). The coefficient relating M to Y is the b coefficient with confounder bias removed and sandwich estimators of standard errors are used to create confidence intervals for the effect. This method assumes that all measures of confounder bias have been measured and included in the analysis to create the weights (thereby leading to accurate estimation of the relation of M to Y, i.e., by adjustment for potential confounders). If there are repeated measures of X, M, and Y, then weights are obtained for each predicted value of M at each time with prior values of X included as covariates to adjust for confounder bias. The weights in the final analysis for the prediction of Y are the product of the weights obtained for the prediction of each wave of Y. Weights that adjust for missing data can also be included. Clearly, a substantial assumption of this method is that there are no unmeasured confounders, which may be difficult to satisfy in some situations. One important check for the model is to assess the extent to which there is enough overlap in the weights in each group with plots and tables to investigate the balance of the weighting procedure. To date the marginal structural model has not been widely used in psychology but it is more common in epidemiology and the health sciences. It is among the most straightforward of causal inference methods to apply. As described already, VanderWeele et al. (2011) applied the marginal structural model method to investigate the effect of loneliness on depression across five waves. There are several methods to estimate weights for inverse probability weighting that may be especially useful when the weights vary a great deal from very small to very large (Cole & Hernan, 2008). The weights for each participant represent how much that participant’s data are used in the final regression of X and M on Y. If the denominator in the weight equation becomes extremely small, the weights for some persons may be excessively large so that one person may overly influence the results of the analysis. One method to deal with this problem is the use of truncated weights to avoid some subjects having too much or too little influence on the analysis. A relative alternative approach avoids calculation of the weights by creating new data sets where confounder effects are removed using a method called g-estimation. G-estimation is not described here but is available in several sources (Loeys et al., 2014; Vansteelandt, 2009).

Page 977

Trim Size: 8.5in x 11in

978

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

Causal Estimators for the ECLS-K Data Set To illustrate the causal estimates, three variables were selected from the ECLS-K data set, an independent variable measure of externalizing behavior at first grade, a mediating measure of reading performance at time third grade, and an outcome measure of internalizing behavior at fifth grade. These measures were used as indicators of a mediating process such that externalizing behavior reduces reading performance, which then increases later depression. Covariates included the kindergarten fall and spring measures of externalizing, reading, and internalizing variables along with gender and age to adjust for potential confounding. The goal of the analysis was to estimate causal effects across waves adjusted for possible confounders where the confounders were measured at the first two waves of measurement. The program described in Valeri and VanderWeele (2013) was used to estimate the natural direct and indirect effects described earlier. To simplify the analysis, only complete cases were used in the analysis for a total of 7,355 individuals. First, the interaction of first grade externalizing behavior and first grade reading performance was not statistically significant in the prediction of Y (.011, SE = .032, p < .5) so separate estimates of direct and indirect effects at different levels of first grade externalizing behavior were not computed. In other words, the two different causal estimators of effects were not necessary because the nonsignificant interaction indicated that effects did not differ across levels of the mediator. As a result there is one natural direct effect (NDE = .046, SE = .007, LCL = .033, UCL = .059), one natural indirect effect (NIE = .004, SE = .0008, LCL = .002, UCL = .005) and the sum of these effects equals the total effect (TE = .049, SE = .007, LCL = .036, UCL = .062). All of the effects are statistically significant, which is perhaps not surprising given the large sample size. The metric of the effects is in terms of a one standard deviation change (.540 units) in externalizing at first grade. Dividing the indirect effect by the standard deviation of Y, suggests that a standard deviation change in externalizing led to an indirect effect via reading performance of about .04 standard deviations in internalizing symptoms. FUTURE DIRECTIONS There are many potentially fruitful directions for future research regarding longitudinal moderation and mediation. For example, Bayesian analysis of longitudinal data is one particular area where the development of new analytic methods may produce better estimates of hypothesized

effects and produce tests that more appropriately reflect substantive theory. Bayesian estimation allows researchers to incorporate prior information into the analysis, which can potentially produce more efficient estimates of moderation and mediation effects (Yuan & MacKinnon, 2009). Moreover, while many traditional estimation techniques rely on large-sample theory, Bayesian estimation is exact, making this approach particularly suitable for data with small samples (Ozechowski, 2014). This aspect of Bayesian analysis is also particularly relevant to mediation effects. Maximum likelihood estimation provides estimates of parameters and standard errors and assumes normal distributions based on large sample theory. However, the ab estimate of the mediated effect is known to be nonnormal (MacKinnon et al., 2007), and this problematic violation often necessitates the use of specialized techniques (e.g., bootstrapping or the application of the distribution of the product) for hypothesis testing. However, Bayesian analysis produces the whole (skewed) posterior distribution of the mediated effect and requires no such assumption. Muthén and Asparouhov (2012) presented a Bayesian approach to structural equation modeling, which could potentially include both mediation and moderation. Yuan and MacKinnon (2009) described a Bayesian approach to modeling multilevel mediation, and Ozechowski (2014) demonstrated the use of Empirical Bayes methods for testing mediation in a small sample. Another possible focus of future research is personcentered moderation and mediation. While more typical variable-centered approaches concentrate on examining relations among variables, person-centered approaches examine patterns of responses for individuals. The person is the unit of analysis (Bauer & Shanahan, 2007), and analytic techniques such as cluster analysis, latent profile analysis, latent class analysis, and latent transition analysis can be used to empirically sort individuals into distinct groups according to their patterns of variable values and associations. For example, individuals might be classified as showing patterns either consistent or inconsistent with mediation. Collins, Graham, and Flaherty (1998) demonstrated such an analysis and show that under particular conditions, variable-centered approaches might detect mediation when there are, in fact, no differences in the proportion of individuals showing patterns consistent with mediation in control and treatment groups. Bauer & Shanahan (2007) illustrate how a person-centered approach can be used to model complex moderated effects. The recent rapid development of new technology and methods for collecting intensive longitudinal data has resulted in a need for analytic methods suitable for

Page 978

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

Future Directions

examining moderation and mediation in such large data sets. For example, current technology allows the collection of intensive ecological momentary assessment (EMA) data, where subjects repeatedly report on key variables while they go about their daily lives. An EMA data set might include hundreds of observations obtained via smartphone from a given individual over the course of a few days or perhaps just one observation per day for several weeks, months, or an entire year. Fitness tracking devices might record several physical activity and condition variables repeatedly for a period of several hours. fMRI data contain activation information about multiple voxels over the time course of an experimental session. Analysis of such rich data sets with traditional techniques can prove cumbersome (Baraldi, Wurpts, MacKinnon, & Lockhard, 2014). As the collection of such data becomes increasingly frequent, there is a clear need for the development of specific methods to explore moderation and mediation in such big data applications. Baraldi et al. (2014) described particular multilevel models that can be used to address relevant questions in EMA data and to explore the effects of personalized, technologically delivered ecological momentary interventions. Lindquist (2012) presented an extension of the mediation model for fMRI data that allows the mediating variable to be a continuous function rather than a single scalar measure and includes causal estimators of effects. The combination of intensive longitudinal data and a person-centered analytic approach may potentially allow the examination of research questions that could not be easily addressed with smaller data sets and traditional variable-centered analytic techniques. For example, groups of subjects who confirm or do not confirm the pattern of mediation at different phases of time during a study (e.g., confirming mediation at early versus later phases) could be identified and their transition patterns over the course of time can be examined to predict their status on an outcome variable. There is also a clear need for more research on temporal design issues. Every researcher designing a longitudinal study needs to make a number of explicit decisions about the data collection schedule. This audience would benefit from clear recommendations regarding the ideal number as well as the frequency and spacing of observations for various types of research scenarios. Some research in this area has, in fact, been conducted for univariate and bivariate growth models where interest centers on testing the effects of treatment on growth (e.g., Raudenbush & Liu, 2001) or variabilily in growth across individuals and covariation among growth rates (e.g., Rast & Hofer, 2014). However, similar work focusing on the power to detect moderation and mediation effects has yet to be conducted.

12:38pm

979

In the interest of developing models that more accurately reflect developmental processes, future research should also focus on methods to parsimoniously represent nonlinear effects. Polynomial trends are, by far, the most common functions currently used to represent change over time, and linear models are the more frequently employed of these. However, most growth is probably not strictly linear (Fritz, 2014). Adding higher order polynomial terms may increase the realism of the model, but these additional factors are often difficult to interpret and incorporate into more complex analyses (such as those involving mediation and moderation). Cudeck and Harring (2007) described a number of nonlinear functions and discuss the considerations involved in choosing a particular nonlinear form to represent change. In addition to such nonlinear trends in an outcome variable over time, it is also possible that mediation or moderation effects might themselves be nonlinear. Our models typically assume that the moderated or mediated effect operates identically throughout the range of the predictor variable. However, such an assumption may be overly simplistic, and developing methods and recomendations for testing these effects would be a useful contribution to the literature. Semiparametric and related generalized additive modeling approaches are important directions in this work (Hastie & Tibshirani, 1990; Keele, 2008; Shen, Chou, Pentz, & Berhane, 2014). Though the methods described in this chapter focus primarily on interinterdividual analysis, it is potentially useful to distinguish between intra- and interindividual processes. Most psychologcal studies are focused on variation between cases (Molenaar, 2004), and explaining intraindividual processes through reference to interindividual findings is a simple generalization of the ecological fallacy (Robinson, 1950). Molenaar (2004) defines the conditions under which the structure of intra- and interindividual variation will be analogous and argues that these will rarely if ever be met in real psychological processes. If a researcher wishes to explore within-person mediation, using a 1 → 1 → 1 mulitilevel mediation model (e.g., Krull & MacKinnon, 2001), combined with judicious decisions regarding the centering of the various predictors in the model, can help to disentangle these processes. To date, however, very little work has focused on ideal methods for exploring intraindividual effects and how best to approach such questions in real world data sets. The combination of mediation and moderation into a single analytic framework is another area that would benefit from future quantitative research and application in substantive data. Currently the most frequent method for exploring whether mediated effects might, in fact, be

Page 979

Trim Size: 8.5in x 11in

980

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

moderated by a fourth variable is the rather simplistic approach of creating subgroups based on the moderator and conducting a multiple group SEM. Such an approach is well suited when the moderator is truly categorical, but artificially categorizing a continuous variable so that it can be forced into such an analysis is far from ideal. Dichotomizing a continuous variable via a median split has many undesirable consequences, including loss of power and spurious main effects and interactions (MacCallum, Zhang, Preacher, & Rucker, 2002). The development and application of more sophisticated techniques that avoid such inappropriate treatment of variables will result in much better tests of hypotheses involving mediated moderation and moderated mediation in longitudinal data. These might involve continuous variable approaches to moderation within an SEM or incorporation of person-centered methods that empirically identify subgroups of subjects based on patterns of multiple variables. Finally, causal inference approaches in general and for mediation in particular are also being rapidly developed. Several books are now appearing that should help disseminate these methods (Hernan & Robins, in press; Morgan & Winship, 2007; Pearl, 2009), though longitudinal mediation modeling has not been widely addressed in these books. There are some recent applications with longitudinal data (Ertefaie & Stephens, 2010; Lindquist, 2012; O’Malley, 2012). Some interesting recent work attempts to combine both dynamic mediation models and modern causal inference (Aalen, Roysland, & Gran 2012; Aalen, Roysland, Gran, Kouyos, & Lange, 2014). Aalen et al. (2012) described two overlapping approaches to causal inference with longitudinal mediation corresponding to interventionist and mechanistic viewpoints. The interventionist approach focuses on estimating causal effects when there is randomization or the evaluation of the status of a system when there is and when there is not an intervention. Typically this approach focuses on the potential outcomes model as described already, although some approaches do not require counterfactuals (Dawid, 2000). The mechanistic approach focuses on how effects come about rather than how an intervention would cause effects. As described by Aalen et al. (2012, p. 832) the “mechanisms in the structure of the world are present whether humans are there to intervene or not, and hence seem to be the more fundamental part.” As a compromise between interventionist and mechanistic positions, Aalen et al. (2012, p. 843) stated, “Now our modest suggestion is as follows: if a direct effect cannot reasonably be defined as a controlled or natural direct effect in the counterfactual sense because the required hypothetical manipulation of the mediator is inconceivable, then we can alternatively

view these effects as being represented by flow in a dynamic system, so that the direct effect corresponds to the flow not passing through the mediator. The indirect effect can similarly be understood as that passing through the mediator.” Aalen et al. went on to describe a stochastic differential equation model for mediation that includes dynamic models for X, M, and Y processes including a dynamic model for how X, M, and Y are related. Aalen et al. apply this idea to mediation in survival analysis data. Other recent work also addresses dynamic change based on functional data analysis and mediation (Lindquist, 2012), though this approach also targets the estimation of natural and controlled effect causal estimators described previously. More application of these methods to actual data is an important future direction of this research in psychology. Because development is longitudinal by definition, and because so many of the hypotheses developmental psychopathologists seek to test involve multiple variables interacting and affecting each other in sequence, the concepts of longitudinal moderation and mediation are particularly important in this area of research. The various analytic models presented in this chapter test how effects change over time or the mechanisms through which earlier causes produce their later effects. Application of these methods in longitudinal data will help to advance the field and achieve better understanding of the complex processes that unfold over time to produce adaptive and maladaptive behavior. REFERENCES Aalen, O. O., Roysland, K., & Gran, J. M. (2012). Causality, mediation and time: A dynamic viewpoint. Journal of the Royal Statistical Society Series A, 175, 831–861. Aalen, O. O., Roysland, K., Gran, J. M., Kouyos, R., & Lange, T. (2014). Can we believe the DAGs? A comment on the relationship between causal DAGs and mechanisms. Statistical Methods in Medical Research. doi: 10.1177/0962280213520436 Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting interactions. Thousand Oaks, CA: Sage Publications. Arminger, G., & Muthén, B. (1998). A Bayesian approach to nonlinear latent variable models using the Gibbs sampler and the Metropolis-Hastings algorithm. Psychometrika, 63(3), 271–300. Arruda, E. H., & Krull, J. L. (2013). The use of multilevel random coefficient models for the analysis of mediational change over time. Paper presented at the Biannual Multilevel Modeling Conference, Utrecht, The Netherlands. Arruda, E. H., & Krull, J. L. (2014a). The use of multilevel random coefficient models for the analysis of mediational change over time. Poster presented at the annual meeting of the American Educational Research Association, Philadelphia, PA. Arruda, E. H., & Krull, J. L. (2014b). The use of multilevel random coefficient models for the analysis of mediational change over time. Poster presented at the Annual Modern Modeling Methods Conference, Storrs, CT.

Page 980

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

References Baraldi, A. N., Wurpts, I. C., MacKinnon, D. P., & Lockhard, G. (2014). Mechanisms of behavior change from technology-based interventions. In L. Marsch, S. Lord, & J. Dallery (Eds.), Leveraging technology to transform behavioral healthcare (pp. 187–199). New York, NY: Oxford University Press.

12:38pm

981

Candel. M. J. J. M., & Winkens B. (2003). Performance of Empirical Bayes estimators of level-2 random parameters in multilevel analysis: A Monte Carlo study for longitudinal designs. Journal of Educational and Behavioral Statistics, 28, 169–194.

Barcikowski, R. S. (1981). Statistical power with group mean as the unit of analysis. Journal of Educational Statistics, 6, 267–285.

Cheong, J. (2011). Accuracy of estimates and statistical power for testing mediation in latent growth modeling. Structural Equation Modeling, 18, 195–211.

Barnard-Brak, L., & Brak, V. (2011). Pharmacotherapy and academic achievement among children with attention-deficit/hyperactivity disorder. Journal of Child and Adolescent Psychopharmacology, 21, 597–603.

Cheong, J., MacKinnon, D. P., & Khoo, S.-T. (2003). Investigation of mediational processes using parallel process latent growth curve modeling. Structural Equation Modeling, 10, 238–262. doi: 10.1207/S15328007SEM1002_5

Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173–1182.

Cicchetti, D., & Rogosch, F. A. (1996). Equifinality and multifinality in developmental psychopathology. Development and Psychopathology, 8, 598–600.

Bauer, D. J., Preacher, K. J., & Gil, K. M. (2006). Conceptualizing and testing random indirect effects and moderated mediation in multilevel models: New procedures and recommendations. Psychological Methods, 11, 142–163. doi: 10.1037/1082-989X.11.2.142 Bauer, D. J., & Shanahan, M. J. (2007). Modeling complex interactions: Person-centered and variable-centered approaches. In T. D. Little, J. A. Bovaird, & N. A. Card (Eds.), Modeling contextual effects in longitudinal studies (pp. 255–284). Mahwah, NJ: Erlbaum. Bentler, P. M. (1990). Comparative fit indexes in structural models. Psychological Bulletin, 107, 238–246. Bentler, P. M., & Wu, E. J. C. (2006). EQS for Windows v. 6.1 [computer software]. Encino, CA: Multivariate Software, Inc. Blozis, S. A. (2004). Structured latent curve models for the study of change in multivariate repeated measures. Psychological Methods, 9, 334–353. doi: 10.1037/1082-989X.9.3.334 Blozis, S. A. (2007). On fitting nonlinear latent curve models to multiple variables measured longitudinally. Structural Equation Modeling, 14, 179–201. doi: 10.1080/10705510701293445 Blozis, S. A., Conger, K. J., & Harring, J. R. (2007). Nonlinear latent curve models for multivariate longitudinal data. International Journal of Behavioral development, 31, 340–346. doi: 10.1177/0165025407077755. Bolger, N., & Laurenceau, J.-P. (2013). Intensive longitudinal methods: An introduction to diary and experience sampling research. New York, NY: Guilford. Bollen, K. A. (1989). Structural equations with latent variables. New York, NY: Wiley. Bollen, K. A., & Curran, P. J. (2004). Autoregressive latent trajectory (ALT) models: A synthesis of two traditions. Sociological Methods & Research, 32, 336–383.

Coffman, D. L. (2011). Estimating causal effects in mediation analysis using propensity scores. Structural equation modeling: A multidisciplinary journal, 18(3), 357–369. Coffman, D. L., & Zhong, W. (2012). Assessing mediation using marginal structural models in the presence of confounding and moderation. Psychological Methods, 17, 642–664. Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences (3rd ed.). Mahwah, NJ: Erlbaum. Cole, D. A., & Maxwell, S. E. (2003). Testing mediational models with longitudinal data: Questions and tips in the use of structural equation modeling. Journal of Abnormal Psychology, 112, 558–577. Cole, S. R., & Hernan, M. A. (2008). Constructing inverse probability weights for marginal structural models. American Journal of Epidemiology, 168, 656–664. Collins, L. M., Graham, J. W., & Flaherty, B. P. (1998). An alternative framework for defining mediation. Multivariate Behavioral Research, 33, 295–312. Cudeck, R. (1996). Mixed-effects models in the study of individual differences with repeated measures data. Multivariate Behavioral Research, 31, 371–403. doi: 10.1207/s15327906mbr3103_6 Cudeck, R., & du Toit, S. H.C. (2002). A version of quadratic regression with interpretable parameters. Multivariate Behavioral Research, 37, 501–519. Cudeck, R., & Harring, J. R. (2007). The analysis of nonlinear patterns of change with random coefficient models. Annual Review of Psychology, 58, 615–637. Curran, P. J. (20023) Have multilevel models been structural equation models all along? Multivariate Behavioral Research, 38, 529–569.

Bollen, K. A., & Curran, P. J. (2006). Latent curve models: A structural equation perspective. Hoboken, NJ: Wiley.

Currie, J., & Stabile, M. (2006). Child mental health in childhood and human capital accumulation: The case of ADHD. Journal of Health Economics, 25, 1094–1118.

Bollen, K. A., & Paxton, P. (1998). Two-stage least squares estimation of interaction effects. In R. E. Schumacker & G. A. Marcoulides (Eds.), Interaction and nonlinear effects in strucutral equation modeling (pp. 125–154). Mahwah, NJ: Erlbaum.

Currie, J., & Stabile, M. (2009). Mental health in childhood and human capital. In J. Gruber (Ed.), The problems of disadvantaged youth: An economic perspective (pp. 115–148). Chicago, IL: University of Chicago Press.

Bollen, K. A., & Zimmer, C. (2010). An overview of the autoregressive latent trajectory (ALT) model. In K. van Montfort, J. Oud, & A. Satora (Eds.), Longitudinal research with latent variables (pp. 153–176). Heidelberg, Germany: Springer.

Davidian, M., & Giltinan, D. M. (1995). Nonlinear models for repeated measurement data. New York, NY: Chapman and Hall/CRC Press.

Browne, M. W. (1993). Structured latent curve models. In C. M. Cuadras & C. R. Rao (Eds.), Multivariate analysis: Future directions 2 (pp. 171–198). Amsterdam, The Netherlands: North Holland.

Dearing, E., & Hamilton, L. C. (2006). Contemporary advances and classic advice for analyzing mediating and moderating variables. Monographs of the Society for Research in Child Development, 71, 88–104.

Burstein, L., & Miller, M. D. (1980). Regression-based analysis of multi-level educational data. New Directions for Methodology of Social and Behavioral Sciences, 6, 194–211.

Department of Health and Human Services, Administration for Children and Families. (1996–2001). Early Head Start research and evaluation study [computer file]. Available from http://www.icpsr.org/

Dawid, A. P. (2000). Causal inference without counterfactuals. Journal of the American Statistical Association, 95(450), 407–424.

Page 981

Trim Size: 8.5in x 11in

982

Cicchetti

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

Drabick, D. A., & Kendall, P. C. (2010). Developmental psychopathology and the diagnosis of mental health problems among youth. Clinical Psychology: Science and Practice, 17, 272–208. doi: 10.1111/j.1468-2850.2010.01219.x. Duncan, O. D. (1969). Some linear models for two-wave, two-variable panel analysis. Psychological Bulletin, 72(3), 177–182. Duncan, T. E., Duncan, S. C., & Strycker, L. A. (2006). An introduction to latent variable growth curve modeling: Concepts, issues, and applications (2nd ed.). New York, NY: Lawrence Erlbaum Associates. Edwards, J. R., & Lambert, L. S. (2007). Methods for integrating moderation and mediation: Ageneral anlytical framework using moderated path analysis. Psychological Methods, 12, 1–22. Efron, B., & Tibishirani, R. J. (1993). An introduction to the bootstrap. New York, NY: Chapman & Hall. Elder, T. E. (2010). The importance of relative standards in ADHD diagnoses: Evidence based on exact birth dates. Journal of Health Economics, 29, 641–656. doi: 10.1016/j.jhealeco.2010.06.003 Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in cross-sectional multilevel models: A new look at an old issue. Psychological Methods, 12, 121–138. Ertefaie, A., & Stephens, A. D. (2010). Comparing approaches to causal inference for longitudinal data: Inverse probability of treatment weighting versus propensity scores. International Journal of Biostatistics, 6(2), Article 14. Fairchild, A. J., & MacKinnon, D. P. (2009). A general model for testing mediation and moderation effects. Prevention Science, 10, 87–99. doi: 10.1007/s11121-008-0109-6 Ferrer, E., & McArdle, J. J. (2003). Alternative structural models for multivariate longitudinal data analysis. Structural Equation Modeling, 10, 493–524. Fletcher, J., & Wolfe, B. (2008). Child mental health and human capital accumulation: The case of ADHD revisited. Journal of Health Economics, 27, 794–800. Frazier, P. A., Tix, A. P., & Barron, K. E. (2004). Testing moderator and mediator effects in counseling psychology research. Journal of Counseling Psychology, 51, 115–134. Frazier, T. W., Youngstrom, E. A., Glutting, J. J., & Watkins, M. W. (2007). ADHD and achievement meta-analysis of the child, adolescent, and adult literatures and a concomitant study with college students. Journal of Learning Disabilities, 40, 49–65. Fritz, M. S. (2014). An exponential decay model for mediation. Prevention Science, 15, 611–622. doi: 10.1007/211121-01300390-x. Gaub, M., & Carlson, C. L. (1997). Gender differences in ADHD: A meta-analysis and critical review. Journal of the American Academy of Child & Adolescent Psychiatry, 36, 1036–1045. Gershon, J., & Gershon, J. (2002). A meta-analytic review of gender differences in ADHD. Journal of Attention Disorders, 5, 143–154. Gogineni, A., Alsup, R., & Gillespie, D. (1995). Mediation and moderation in social work research. Social Work Research, 19, 57–63. Goldstein, H. (1995). Multilevel statistical models (2nd ed.). New York, NY: Wiley. Gollob, H. F., & Reichardt, C. S. (1987). Taking account of time lags in causal models. Child Development, 58(1), 80–92. Gollob, H. F., & Reichardt, C. S. (1991). Interpreting and estimating indirect effects assuming time lags really matter. In L. M. Collins & J. L. Horn (Eds.), Best methods for the analysis of change: Recent advances, unanswered questions, future directions (pp. 243–259). Washington, DC: American Psychological Association. Gresham, F. M., & Elliott, S. N. (1990). Social skills rating scale: Elementary scale A (“How often”). Circle Pines, MN: American Guidance Service.

Hastie, T., & Tibshirani, R. (1990). Generalized additive models. London, UK: Chapman and Hall. Hayes, A. F. (2013). Introduction to mediation, moderation and conditional process analysis: A regression-based appraoch. New York, NY: Guilford Press. Hayes, A. F. (2014). Beyond Baron and Kenny: Statistical mediation analysis in the new millennium. Communication Monographs, 76, 408–420. Hedeker, D. (2004). An introduction to growth modeling. In D. Kaplan (Ed.), The SAGE handbook of quantitative methodology for the social sciences (pp. 215–234).. Thousand Oaks, CA: Sage. Hernan, M. A., & Robins, J. M. (in press). Causal inference. London, UK: Chapman & Hall. Hoffman, L., & Stawski, R. S. (2009). Personas as contexts: Evaluating between-person and within-person effects in longitudinal analysis. Research in Human Development, 6, 97–120. Hofmann, D. A., & Gavin, M. B. (1998). Centering decisions in hierarchical linear models: Implications for research in organizations. Journal of Management, 24, 123–161. Hogan, J. W., & Lancaster, T. (2004). Instrumental variables and inverse probability weighting for causal inference from longitudinal observational studies. Statistical Methods in Medical Research, 13, 17–48. Holland, P. W. (1986). Statistics and causal inference (with comments), Journal of the American Statistical Association, 81, 945–970. Holland, P. W. (1988). Causal inference, path analysis, and recursive structural equation models. Sociological Methodology, 18, 449–484. Holmbeck, G. N. (1997). Toward terminological, conceptual, and statistical clarity in the study of mediators and moderators: Examples from the child-clinical and pediatric psychology literatures. Journal of Consulting and Clinical Psychology, 5, 599–610. Horm, J. L., & McArdle, J. J. (1992). A practical and theoretical guide to measurement invariance in aging research. Experimental Aging Research, 18, 117–144. Hox, J. J. (2010). Multilevel analysis: Techniques and applications (2nd ed.). New York, NY: Routledge. Humphreys, K. L., Katz, S. J., Lee, S. S., Hammen, C., Brennan, P. A., & Najman, J. M. (2013). The association of ADHD and depression: Mediation by peer problems and parent–child difficulties in two complementary samples. Journal of Abnormal Psychology, 122, 854–867. Humphreys, L. G. (1960). Investigations of the smipex. Psychometrika, 25, 313–323. Imai, K., Keele, L., & Tingley, D. (2010). A general approach to causal mediation analysis. Psychological Methods, 15, 309–334. Imai, K., Keele, L., & Yamamoto, T. (2010). Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science, 25(1), 51–71. doi:10.1214/10-STS321 James, L. R. (1980). The unmeasured variable problem in path analysis. Journal of Applied Psychology, 65, 415–421. Jo, B. (2008). Causal inference in randomized experiments with mediational processes. Psychological Methods, 13(4), 314–336. doi: 10.1037/a0014207 Jöreskog, K. G. (1971). Simultaneous factor analysis in several populations. Psychometrika, 36, 409–426. Jöreskog, K. G. (1979). Statistical estimation of structural models in longitudinal-developmental investigations. In J. R. Nesselroade & P. B. Baltes (Eds.), Longitudinal research in the study of behavior and development (pp. 303–352). New York, NY: Academic Press. Jöreskog, K. G. (1998). Interaction and nonlinear modeling: Issues and approaches. In R. E. Schumacker & G. A. Marcoulides (Eds.), Interaction and nonlinear effects in structural equation modeling (pp. 239–250). Mahwah, NJ: Erlbaum.

Page 982

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

References Jöreskog, K. G., & Sörbom, D. (2006). LISREL for Windows v. 8.7 [computer software]. Lincolnwood, IL: Scientific Software International, Inc. Judd, C. M., & Kenny, D. A. (1981). Process analysis estimating mediation in treatment evaluations. Evaluation Review, 5(5), 602–619. Kawabata, Y., Tseng, W. L., & Gau, S. S. F. (2012). Symptoms of attention-deficit/hyperactivity disorder and social and school adjustment: The moderating roles of age and parenting. Journal of Abnormal Child Psychology, 40, 177–188. Keele, L. (2008). Semiparametric regression for the social sciences. New York, NY: Wiley. Kenny, D. A. (1979). Correlation and causality. New York, NY: Wiley. Kenny, D. A., & Judd, C. M. (1984). Estimating the nonlinear and interactive effects of latent variables. Psychological Bulletin, 96, 201–210. Kenny, D. A., Korchmaros, J. D., & Bolger, N. (2003). Lower level mediation in multilevel models. Psychological Methods, 8, 115–128. doi: 10.1037/1082-989X.8.2.115 Khoo, S. T. (2001). Assessing program effects in the presence of treatment-baseline interactions: a latent curve approach. Psychological Methods, 6(3), 234–257. Klein, A., & Moosbrugger, H. (2000). Maximum likelihood estimates of latent interaction effects with the LMS-method. Psychometrika, 65(4), 457–474. Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.). New York, NY: Guilford Press. Krause, M. R., Serlin, R. C., Ward, S. E., Rony, R. Y., Ezenwa, M. O., & Naab, F. (2010). Testing mediation in nursing research: Beyond Baron and Kenny. Nursing Research, 59, 288–294. doi: 10.1097/NNR.0b013e3181dd26b3. Kreft, I. G. G., de Leeuw, J., & Aiken, L. S. (1995). The effect of different forms of centering in hierarchical linear models. Multivariate Behavioral Research, 30, 1–21. Krull, J. L., & MacKinnon, D. P. (1999). Multilevel mediation modeling in group-based intervention studies. Evaluation Review, 23, 418–444. doi: 10.1177/0193841X9902300404 Krull, J. L., & MacKinnon, D. P. (2001). Multilevel modeling of individual and group level mediated effects. Multivariate Behavioral Research, 36, 249–277. doi: 10.1207/S15327906MBR3602_06 Lahey, B. B., Hartung, C. M., Loney, J., Pelham, W. E., Chronis, A. M., & Lee, S. S. (2007). Are there sex differences in the predictive validity of DSM–IV ADHD among younger children? Journal of Clinical Child and Adolescent Psychology, 36, 113–126. Laird, N. M., & Ware, H. (1982). Random-effects models for longitudinal data. Biometrics, 38, 963–974. Lee, S. S., Lahey, B. B., Owens, E. B., & Hinshaw, S. P. (2008). Few preschool boys and girls with ADHD are well-adjusted during adolescence. Journal of Abnormal Child Psychology, 36, 373–383. Lindquist, M. A. (2012). Functional causal mediation analysis with an application to brain connectivity. Journal of the American Statistical Association, 107, 1297–1309. Little, T. D. (2013). Longitudinal structural equation modeling. New York, NY: Guilford Press. Loeys, T., Moerkerke, B., De Smet, O., Bysse, A., Steen, J., & Vansteelandt, S. (2014). Flexible mediation analysis in the presence of nonlinear relations: Beyond the mediation formula. Multivariate Behavioral Research, 48(6), 871–894. Longford, N. T. (1993). Random coefficient models. Oxford, UK: Oxford University Press. Lynch, K. G., Cary, M., Gallop, R., & Ten Have, T. R. (2008). Causal mediation analyses for randomized trials. Health Services & Outcomes Research Methodology, 8(2), 57–76.

12:38pm

983

MacCallum, R. C., Zhang, S., Preacher, K. J., & Rucker, D. D. (2002). On the practice of dichotomization of quantitative variables. Psychological Methods, 7, 19–40. MacKinnon, D. P. (2008). Introduction to statistical mediation analysis. New York, NY: Lawrence Erlbaum Associates. MacKinnon, D. P., & Dwyer, J. H. (1993). Estimating mediated effects in prevention studies. Evaluation Review, 17, 144–158. MacKinnon, D. P., Fritz, M. S., Williams, J., & Lockwood, C. M. (2007). Distribution of the product confidence limits for the indirect effect: Program PRODCLIN. Behavior Research Methods, 39, 384–389. MacKinnon, D. P., Krull, J. L., & Lockwood, C. M. (2000). Equivalence of the mediation, confounding and suppression effect. Prevention Science, 1, 173–181. MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002). A comparison of methods to test mediation and other intervening variable effects. Psychological Methods, 7, 83–104. doi:10.1037/1082-989x.7.1.83 MacKinnon, D. P., Lockwood C. M., & Williams, J. (2004). Confidence limits for the indirect effect: Distribution of the product and resampling methods. Multivariate Behavioral Research, 39, 99–128. MacKinnon, D. P., & Pirlott, A. G. (2015). Statistical approaches for enhancing causal interpretation of the M to Y relation in mediation analysis. Personality and Social Psychology Review, 1, 30–43. doi: 10.1177/1088868314542878 MacKinnon, D. P., & Valente, M. J. (2014). Mediation from multilevel to structural equation modeling. Annals of Nutrition and Metabolism, 65, 196–202. doi: 10.1159/000362505 MacKinnon, D. P., Warsi, G., & Dwyer, J. H. (1995). A simulation study of mediated effect measures. Multivariate Behavioral Research, 30, 41–62. Marquardt, D. (1980). You should standarized the predictor variables in your regression models. Journal of the American Statistical Association, 75, 74–103. Marsh, H. W., Wen, Z., & Hau, K.-T. (2004). Structural equation models of latent interactions: Evaluation of alternative estimation strategies and indicator construction. Psychological Methods, 9(3), 275–300. Marsh, H. W., Wen, Z., & Hau, K. T. (2006). Structural equation models of latent interaction and quadratic effects. In G. R. Hancock & R. O. Mueller (Eds.), Structural equation modeling: A second course (pp. 225–265). Greenwich, CT: Information Age Publishing. Martens, M. P., & Haase, R. H. (2006). Advanced applications of structural equation modeling in counseling psychology research. Counseling Psychologist, 34, 806–838. Masten, A. S., & Cicchetti, D. (2010). Editorial: developmental cascades. Development and Psychopathology, 22, 491–495. doi:10.1017/S0954579410000222. Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitudinal mediation. Psychological Methods, 12, 23–44. Maxwell, S. E., Cole, D. A., & Mitchell, M. A. (2011). Bias in cross-sectional analyses of longitudinal mediation: Partial and complete mediation under an autoregressive model. Multivariate Behavioral Research, 46(5), 816–841. Maxwell, S. E., & Delaney, H. D. (1993). Bivariate median splits and spurious statistical significance. Psychological Bulletin, 113, 181–190. McArdle, J. J. (2001). A latent difference score approach to longitudinal dynamic structural analysis. In R. Cudeck, M. du Toit, & D. Sorbom (Eds.), Structural equation modeling: Present and future (pp. 341–380). Lincolnwood, IL: Scientific Software International. McArdle, J. J., & Hamagami, F. (2001). Latent difference score structural models for linear dynamic analyses with incomplete longitudinal data. In L. M. Collins & A. G. Sayer (Eds.), New methods for the analysis

Page 983

Trim Size: 8.5in x 11in

Cicchetti

984

c21.tex printer - Volume I - 12/10/2015

12:38pm

Moderation and Mediation in Interindividual Longitudinal Analysis

of change (pp. 138–175). Washington, DC: American Psychological Association. McArdle, J. J., & Nesselroade, J. R. (2003). Growth curve analysis in contemporary psychological research. In J. Schinka & W. Velicer (Eds.), Comprehensive handbook of psychology: Research methods in psychology (Vol. 2, p. 447–480). New York, NY: Wiley. Meeker, Jr., W. Q., Cornwell, L. W., & Aroian, L. A. (1981). The product of two normally distributed random variables. In W. J. Kennedy & R. E. Odeh (Eds.), Selected tables in mathematical statistics (Vol. 7, p. 1–256). Providence, RI: American Mathematical Society. Meredith, W. (1993). Measurement invariance, factor analysis, and factorial invariance. Psychometrika, 58, 525–543. Meredith, W., & Tisak, J. (1990). Latent curve analysis. Psychometrika, 55(1), 107–122. Molenaar, P. C.M. (2004). A manifesto on psychology as idiographic science: Bringing the person back into scientific psychology, this time forever. Measurement, 2, 201–218. Morgan, P. L., Staff, J., Hillemeier, M. M., Farkas, G., & Maczuga, S. (2013). Racial and ethnic disparities in ADHD diagnosis from kindergarten to eighth grade. Pediatrics, 132, 85–93. Morgan, S. L., & Winship, C. (2007). Counterfactuals and causal inference: Methods and principles for social research. Cambridge, UK: Cambridge University Press. Muller, D., Judd, C. M., & Yzerbyt, V. Y. (2005). When moderation is mediated and mediation is moderated. Journal of Personality and Social Psychology, 89, 852–863. Muthén, B., & Asparouhov, T. (2012). Bayesian SEM: A more flexible representation of substantive theory. Psychological Methods, 17, 313–335. Muthén, B. O., & Curran, P. J. (1997). General longitudinal modeling of individual differences in experimental designs: A latent variable framework for analysis and power estimation. Psychological Methods, 2, 371–402.

Pearl, J. (2001). Direct and indirect effects. Proceedings of the 17th Conference on Uncertainty in Artificial Intelligence (pp. 411–420). San Francisco, CA: Morgan Kaufmann. Pearl, J. (2009). Causality: Models, reasoning, and inference. Cambridge, UK: Cambridge University Press. Pearl, J. (2012). The causal mediation formula: A guide to the assessment of pathways and mechanisms. Prevention Science: The Official Journal of the Society for Prevention Research, 13(4), 426–436. Pedhazur, E. J. (1997). Multiple regression in behavioral research: Explanation and prediction (3rd ed.). Fort Worth, TX: Harcourt Brace College Publishers. Pinheiro, J. C., & Bates, D. M. (2000). Mixed-effects models in S and S-plus. New York, NY: Springer. Pituch, K. A., Stapleton, L. M., & Kang, J. Y. (2006). A comparison of single sample and bootstrap methods to assess mediation in cluster randomized trials. Multivariate Behavioral Research, 41, 367–400. Preacher, K. J., & Hayes, A. F. (2004). SPSS and SAS procedures for estimating indirect effects in simple mediation models. Behavior Research Methods, Instruments, and Computers, 36, 717–731. Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40, 879–891. Preacher, K. J., Rucker, D. D., & Hayes, R. F. (2007). Addressing moderated mediation hypotheses: Theory, methods, and prescriptions. Multivariate Behavioral Research, 42, 185–227. Preacher, K. J., Zyphur, M. J., & Zhang, Z. (2010). A general multilevel SEM framework for assessing multilevel mediation. Psychological Methods, 15, 209–233. doi: 10.1037/a0020141 R Development Core Team. (2013). R: A language and environment for statistical computing (Version 2.6.1) [Computer software]. Vienna, Austria: R Foundation for Statistical Computing.

Muthén, B., & Khoo, S. T. (1998). Longitudinal studies of achievement growth using latent variable modeling. Learning and Individual Differences, 10(2), 73–101.

Rast, P., & Hofer, S. M. (2014). Longitudinal design considerations to optimize power to detect variances and covariances among rates of change: Simulation results based on actual longitudinal studies. Psychological Methods, 19, 133–154.

Muthén, B. O., & Muthén, L. (2014). Mplus v. 7.2: The comprehensive modeling program for applied researchers [computer software]. Los Angeles, CA: Muthén & Muthén.

Raudenbush, S. W. (2001). Comparing personal trajectories and drawing causal inferences from longitudinal data. Annual Review of Psychology, 52, 501–525.

Nesselroade, J. R. (1991). Interindividual differences in intraindividual changes. In L. M. Collins & J. L. Horn (Eds.), Best methods for the analysis of change: Recent advances, unanswered questions, future directions (pp. 92–105). Washington, DC: American Psychological Association.

Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Thousand Oaks, CA: Sage.

Nosek, B. A., Graham, J., Lindner, N. M., Kesebir, S., Hawkins, C. B., Hahn, C., . . . Tenney, E. R. (2010). Cumulative and career-stage impact of social-personality psychology programs and their members. Personality and Social Psychology Bulletin, 36, 1283–1300.

Raudenbush, S. W., Bryk, A. S., Cheong, Y. F., Congdon, R. T., & du Toit, M. (2011). HLM 7: Hierarchical linear and nonlinear modeling. Chicago, IL: Scientific Software International. Raudenbush, S. W., & Liu, X. F. (2001). Effects of study duration, frequency of observation, and sample size on power in studies of group differences in polynomial change. Psychological Methods, 6, 387–401.

O’Malley, A. J. (2012). Instrumental variable specifications and assumptions for longitudinal analysis of mental health cost offsets. Health Services and Outcomes Research Methodology, 12, 254–272.

Reisby, N., Gram, L. F., Bech., P., Nagy, A., Petersen, G. O., Ortmann, J., . . . Christiansen, J. (1977). Imipramine: Clinical effects and pharmacokinetic variability. Psychopharmacology, 54, 263–272.

Owens, E. B., Hinshaw, S. P., Lee, S. S., & Lahey, B. B. (2009). Few girls with childhood attention-deficit/hyperactivity disorder show positive adjustment during adolescence. Journal of Clinical Child & Adolescent Psychology, 38, 132–143.

Rigdon, E. E., Schumacker, R. E., & Wothke, W. (1998). A comparative review of interaction and nonlinear modeling. In R. E. Schumacker & G. A. Marcoulides (Eds.), Interaction and nonlinear effects in structural equation modeling (pp. 1–16). Mahwah, NJ: Erlbaum.

Ozechowski, T. J. (2014). Empirical Bayes MCMC estimation for modeling treatment processes, mechanism of change, and clinical outcomes in small samples. Journal of Consulting and Clinical Psychology, 82, 854–867.

Robins, J. M., & Greenland, S. (1992). Identifiability and exchangeability for direct and indirect effects. Epidemiology, 3, 143–155.

Pardo, A., & Román, M. (2013). Reflections on the Baron and Kenny model of statistical mediation. Annales de Psicología, 29, 614–623.

Robins, J. M., & Hernan, M. A. (2009). Estimation of the causal effects of time-varying exposures. In G. Fitzmaurice, M. Davidian, G. Verbeke, & G. Molenberghs (Eds.), Longitudinal data analysis (pp. 553–599). New York, NY: Chapman & Hall.

Page 984

Trim Size: 8.5in x 11in

Cicchetti

c21.tex printer - Volume I - 12/10/2015

References Robins, J. M., Hernán, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. Retrieved from http://www.jstor.org/stable/3703997 Robinson, W. S. (1950). Ecological correlations and the behavior of individuals. American Sociological Review, 15, 351–357. Rogosa, D. (1988). Myths about longitudinal research. In K. W. Schaie, R. T. Campbell, W. Meredith, & S. C. Rawlings (Eds.), Methodological issues in aging research (pp. 171–209). New York, NY: Springer. Rogosa, D., Brandt, D., & Zimowski, M. (1982). A growth curve approach to the measurement of change. Psychological Bulletin, 92(3), 726–748. Rose, B. M., Holmbeck, G. N., Coakley, R. M., & Franks, E. A. (2004). Mediator and moderator effects in developmental and behavioral pediatric research. Journal of Developmental and Behavioral Pediatrics, 25, 58–67. Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66, 688–701. doi: 10.1037/h0037350 Rutter, M. (1999). Resilience concepts and findings: Implications for family therapy. Journal of Family Therapy, 21, 119–144. SAS Inc. (2014). SAS (Version 9.3) [Computer software]. Cary, NC: SAS Inc. Schneider, H., & Eisenberg, D. (2006). Who receives a diagnosis of attention deficit/hyperactivity disorder in the United States elementary school population? Pediatrics, 117(4), e601–e609. Selig, J. P., Preacher, K. J., & Little, T. D. (2012). Modeling time-dependent association in longitudinal data: A lag as moderator approach. Multivariate Behavioral Research, 47, 697–716. Shen, E., Chou, C. P, Pentz, M. A., & Berhane, K. (2014). Quantile mediation models: A comparison of methods for assessing mediation across the outcome distribution. Multivariate Behavioral Research, 49, 471–485. DOI:10.1080/00273171.2014.904221 Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling change and event occurrence. Oxford, UK: Oxford University Press. Sobel, M. E. (1982). Asymptotic confidence intervals for indirect effects in structural equation models. In S. Leinhardt (Ed.), Sociological methodology (pp. 290–312). Washington, DC: American Sociological Association. Sobel, M. E. (1986) Some new results on indirect effects and their standard errors in covariance structure models. In N. Tuma (Ed.), Sociological methodology (pp. 159–186). Washington, DC: American Sociological Association. Sobel, M. E. (2008). Identification of causal parameters in randomized studies with mediating variables. Journal of Educational and Behavioral Statistics, 33, 230–251. doi: 10.3102/1076998607307239. Steiger, J. H. (1990). Structural model evaluation and modification: An interval estimation approach. Multivariate Behavioral Research, 25, 173–180. Stone, S., Brown, T., & Hinshaw, S. (2010). ADHD-related school compositional effects: An exploration. Teachers College Record, 112(5), 3–4. Stoolmiller, M. (1994). Antisocial behavior, delinquent peer association, and unsupervised wandering for boys: Growth and change from childhood to early adolescence. Multivariate Behavioral Research, 29, 263–288. Ten Have, T. R., Joffe, M. M., Lynch, K. G., Brown, G. K., Maisto, S. A., & Beck, A. T. (2007). Causal mediation analysis with rank preserving models. Biometrics, 63, 926–934. doi: 10.1111/j.1541-0420.2007.00766.x Tofighi, D., & MacKinnon, D. P. (2011). RMediation: An R package for mediation analysis confidence intervals. Behavior Research Methods, 43(3), 692–700. doi: 10.3758/s13428-011-0076-x

12:38pm

985

Toh, S., & Hernan, M. A. (2008). Causal inference from longitudinal studies with baseline randomization. International Journal of Biostatistics, 4(1), Article 22. doi: 10.2202/1557-4679.1117. Tourangeau, K., Nord, C., Lê, T., Sorongon, A. G., & Najarian, M. (2009). Early childhood longitudinal study, kindergarten class of 1998–99 (ECLS-K), combined user’s manual for the ECLS-K eighth-grade and K–8 full sample data files and electronic codebooks (NCES 2009–004). Washington, DC: National Center for Education Statistics, Institute of Education Sciences, U.S. Department of Education. U.S. Department of Education, National Center for Education Statistics. (2009). Early childhood longitudinal study, kindergarten class of 1998–99 (ECLS-K) kindergarten through eighth grade full sample public-use data and documentation (DVD). (NCES 2009-005). Washington, DC: Author. Valeri, L., & VanderWeele, T. J. (2013). Mediation analysis allowing for exposure-mediator interactions and causal interpretation: Theoretical assumptions and implementation with SAS and SPSS macros. Psychological Methods, 18, 137–150. VanderWeele, T. J. (2010). Bias formulas for sensitivity analysis for direct and indirect effects. Epidemiology, 21, 540–551. doi: 10.1097/EDE.0b013e3181df191c VanderWeele, T. J., Hawkley, L. C., Thisted, R. A., & Cacioppo, J. T. (2011). A marginal structural model analysis for loneliness: Implications for intervention trials and clinical practice. Journal of Consulting and Clinical Psychology 79(2), 457–468. VanderWeele, T. J., & Vansteelandt, S. (2009). Conceptual issues concerning mediation, interventions and composition. Statistics and Its Interface (Special Issue on Mental Health and Social Behavioral Science), 2, 457–468. VanderWeele, T. J., & Vansteelandt, S. (2013). Mediation analysis with multiple mediators. Epidemiologic Methods, 2(1), 95–115. Vansteelandt, S. (2009). Estimating direct effects in cohort and case–control studies. Epidemiology, 20, 851–860. Voelkle, M. C., Oud, J. H. L., Davidov, E., & Schmidt, P. (2012). An SEM approach to continuous time modeling of panel data: Relating authoritarianism and anomia. Psychological Methods, 17, 176–192. Wall, M. M., & Amemiya, Y. (2000). Estimation for polytomial structural equation models. Journal of the American Statistical Association, 95, 929–940. West, S. G. (2011). Causal inference in cross sectional and longitudinal meditational models: Introduction to the special section. Multivariate Behavioral Research, 46, 812–815. Wiley, D. E., & Wiley, J. A. (1970). The estimation of measurement error in panel data. American Sociological Review, 35, 112–117. Willett, J. B. (1997). Measuring change: What individual growth modeling buys you. In E. Amsel & K. A. Renninger (Eds.), Change and development: Issues of theory, method, and application (pp. 213–243). Mahwah, NJ: Lawrence Erlbaum Associates. Willett, J. B., & Sayer, A. G. (1994). Using covariance structure analysis to detect correlates and predictors of individual change over time. Psychological Bulletin, 116, 363–381. Winship, C., & Morgan, S. L. (1999). The estimation of causal effects from observational data. Annual Review of Sociology, 25, 659–707. Yuan, Y., & MacKinnon, D. P. (2009). Bayesian mediation analysis. Psychological Methods, 14, 301–322. Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of Consumer Research, 37, 197–206.

Page 985

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

CHAPTER 22

Latent Growth Modeling and Developmental Psychopathology JUNGMEEN KIM-SPOON and KEVIN J. GRIMM

CORE THEORETICAL PRINCIPLES OF DEVELOPMENTAL PSYCHOPATHOLOGY AND LATENT GROWTH MODELING APPROACHES 986 The Pathways Framework 987 Resilience 987 Multiple Levels of Analysis 988 Normality and Psychopathology 988 Developmental Analysis 989 BASIC ASSUMPTIONS AND LONGITUDINAL DESCRIPTIVE ANALYSES 989 Primary Goals of Longitudinal Research 989 Planning and Sampling in a Longitudinal Study 989 Key Issues in Latent Growth Modeling 990 THE LATENT GROWTH MODELING 995 Historical Development 995 Latent Growth Curve as a Structural Equation Model 996 Common Latent Growth Models 997 MULTILEVEL MODELING FOR STUDYING DEVELOPMENTAL TRAJECTORIES 1000 Alternative Specification 1001 Specific Growth Models 1001 Direct Expansions of the Latent Growth Modeling Framework 1002 Multiple Group Growth Model 1002

Latent Growth Modeling Studies for Understanding Developmental Pathways in Terms of Continuity–Discontinuity and Equifinality–Multifinality 1003 Latent Growth Modeling Studies for Understanding Contextual Influences 1009 LATENT CHANGE (DIFFERENCE) SCORE MODELING 1015 Univariate Latent Change Score Models 1015 Bivariate Latent Change Score Models 1017 Latent Change Score Modeling Studies for Understanding Risk and Protective Mechanisms as Predictors, Mediators, and Moderators 1018 Latent Change Score Modeling Studies for Understanding Contextual Influences 1021 GROWTH MIXTURE MODELING 1027 Growth Mixture Model 1027 Cautions 1029 Growth Mixture Modeling Studies for Understanding Risk and Protective Mechanisms as Predictors, Mediators, and Moderators 1029 Growth Mixture Modeling Studies for Understanding Contextual Influences 1032 Growth Mixture Modeling for Multiple Levels of Analysis 1034 CONCLUSION AND FUTURE PERSPECTIVE 1036 REFERENCES 1037

CORE THEORETICAL PRINCIPLES OF DEVELOPMENTAL PSYCHOPATHOLOGY AND LATENT GROWTH MODELING APPROACHES

are reviewed. In each section, we first discuss developmental methodology and quantitative techniques for studying developmental trajectories in developmental psychopathology, focusing on latent growth modeling. We begin by reviewing core principles of developmental psychopathology, which is conceptualized as an evolving interdisciplinary science that studies the interplay among the biological, psychological, and social-contextual aspects of normal and abnormal development across the lifespan (Cicchetti, 2010, 2015). Developmental psychopathology posits the following important questions: What are the causes of a particular disorder? What are the transformations in behavioral (or phenotypic)

In this chapter, variations of latent growth modeling and their applications in developmental psychopathology Jungmeen Kim-Spoon was supported by a grant from the National Institute of Drug Abuse (DA036017). Kevin J. Grimm was supported by National Science Foundation Grants (REESE Program Grant-DRL-0815787 and REAL Program Grant-1252463). We thank Samanata Shrestha for her editorial assistance. 986

Page 986

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Core Theoretical Principles of Developmental Psychopathology and Latent Growth Modeling Approaches

manifestation? Where do changes happen in the pathway to psychopathology? Within the developmental psychopathology perspective, maladaptation is viewed as evolving through the successive adaptations in the person’s environment. That is, maladaptation is the complex result of a myriad of risk and protective factors operating over time; and it is not some sort of trait a person has or an expression of a pathogen (Sroufe, 1997). Accordingly, key research questions in developmental psychopathology focus on (1) discovery of factors that place individuals on pathways probabilistically leading to later disturbances, and (2) processes, which maintain individuals on such pathways once enjoined. Developmental psychopathology is a perspective that is especially applicable to the investigation of critical transitional turning points in development across the life span. Empirical research in developmental psychopathology involves knowing about the trajectories and pathways of disorders across the life span and asking the following questions: What happened before the onset of the disorder? How does it progress? Do the manifestations of the symptoms continue or stop at some point? As disorders may appear at any point in the life span, it is imperative to examine the course, phases, and sequelae of disorders. The goals of developmental psychopathology are broad and integrative, seeking to understand adaptation and maladaptation across the life course, and to comprehend the complex interplay among multiple levels of functioning including biological, psychological, and social processes in the development. Next, a few core theoretical principles of developmental psychopathology guide our reviews of empirical studies using latent growth modeling. The Pathways Framework The premise of developmental psychopathology is that the same laws that govern normal development also govern pathological development (Cicchetti, 2015). Therefore, pathology is viewed in terms of developmental deviation and a significant deviation in adaptation represents an increased probability of problems in dealing with (negotiating) subsequent developmental issues. A particular adaptational failure at any point in time is best viewed as placing an individual on a pathway (or trajectory) potentially leading to disorder. For example, anxious attachment (maladaptive pattern of attachment) is not viewed as psychopathology per se, but in terms of developmental risk for disturbance. Developmental psychopathology also believes that change is possible at any points on the pathway. Even

12:38pm

987

when a maladaptive pathway is enjoined, return to positive functioning often remains possible. Thus, this perspective emphasizing the concept of pathways is closely related to the concept of resilience. Developmental psychopathology also emphasizes that the same rules of development do not necessarily apply to all children and families. In this regard, two theoretical concepts are relevant. First, equifinality suggests that there are multiple pathways to similar manifest outcomes. For example, a pattern of maladaptation that has many features in common, such as depressive mood, lack of social engagement, and low self-esteem, may be the result of distinctly different developmental pathways, one rooted in alienation, and one rooted in anxiety and helplessness. Second, multifinality suggests that a single pathway can lead to multiple outcomes. For example, not all maltreated children develop similar disorders or any disorders. Some maltreated children show depression, others show conduct disorder, and still others show substance use problems. Yet there still are some maltreated children who do not show any critical psychopathology. Resilience Developmental psychopathologists are interested in studying pathways to competent adaptation despite exposure to conditions of adversity because the results will help obtain a better understanding of both normal development and psychopathology. Resilience refers to the fact that some children facing adversity nonetheless do well or return to positive functioning following a period of maladaptation (Sroufe, 1997). Resilience is shown in the positive end of the distribution of developmental outcomes among individuals at high risk. These resilient individuals maintain adaptive functioning in spite of serious risk hazards (Rutter, 1990). Developmental psychopathologists see resilience as a dynamic process encompassing positive adaptation within the context of significant adversity. According to Luthar, Cicchetti, and Becker (2000), two critical conditions are implicit in this notion: (1) exposure to significant threat or severe adversity; and (2) the achievement of positive adaptation despite major assaults on the developmental process. Luthar and colleagues (2000) further differentiated diverse developmental processes determined by protective and vulnerability factors. Protective factors have direct ameliorative effects at both high- and low-risk conditions. Thus, protective effects are exhibited when the main overall effects on the at-risk individuals are promoting their positive functioning. Such main effects of protective factors can be described as

Page 987

Trim Size: 8.5in x 11in

988

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

protective-stabilizing processes where individuals with the attribute show stability in competence despite increasing risk, whereas individuals without the attribute decrease in competence. Protective factors may operate to have interactive or moderating effects. First, protective-enhancing processes may be seen when the attribute (moderator) allows individuals to interact with stress in such a way that their competence is augmented with increasing risk. Second, protective-reactive processes may be described when the attribute generally confers advantages but less so when risk is high. In contrast to protective factors, vulnerability factors are identified when the main overall effects on the at-risk individuals are negative, thus demoting their positive functioning. Those individuals with the attribute are expected to manifest greater maladjustment at both high- and low-risk conditions. Such main effects of vulnerability factors are described as vulnerable-stable processes, where the general disadvantages of individuals with the attribute remain stable despite changing levels of stress. Interactive or moderating effects of vulnerability factors can be shown in vulnerable and reactive processes where the overall disadvantage linked with the attribute is heightened with increasing levels of stress. Overall, these vulnerability and protection processes illustrate that the impact of stress on competence levels depends on protective or vulnerability factors. Indeed, prior research in developmental psychopathology suggests that many risk factors do not have direct, additive effects but rather indirect, interactive effects. Rutter (1990) demonstrated that risk factors did not lead to an increased rate of psychiatric disorder if they occurred truly in isolation. Rather, the rate was increased if there were two or more concurrent risk variables. For example, family discord has an adverse effect only when it is accompanied by other adversities. Research on the impact of parental divorce suggests that divorce accompanies many stressors that may take a toll on children, including economic hardship (decreased income for the single-mother family) and maternal depression (Hetherington, Bridges, & Insabella, 1998). Researchers in developmental psychopathology should be reminded that one should consider the impact of a particular risk factor in the presence of other independent or related risk factors. This perspective is related to the multiple level approach principle in developmental psychopathology discussed in the following section. Multiple Levels of Analysis Most developmental psychopathologists will agree that singular, linear cause of psychopathology will rarely be

obtained. More often, adequate prediction of either disturbance or resilience requires considering multiple risks and protective factors and their interplay (Cicchetti & Sroufe, 2000). In addition, within a developmental psychopathology perspective, the development of psychopathology is viewed as unfolding along different developmental pathways among different individuals. Thus, it is expected that relevant causal processes vary among individuals who show the same pattern of disorders (i.e., equifinality), and that there is heterogeneity in the expression of disorders (i.e., multifinality) (Cicchetti, 2010; Cicchetti & Rogosch, 1996). Therefore, in studying developmental pathways and processes of psychopathology, it is important to investigate simultaneously different psychopathological outcomes as well as individuals’ functioning in multiple domains of development and across multiple settings. Developmental psychopathology studies involving multiple levels of analyses have been increasing. For example, recent empirical studies of resilience have incorporated neurobiological and molecular genetic analyses in addition to the traditional measures of psychosocial contributors (Cicchetti & Curtis, 2007). The integrated knowledge gained by employing multiple biological and psychological levels of analyses is particularly beneficial for contributing to the ability to design individualized preventive interventions to facilitate the development of resilient functioning (Cicchetti & Toth, 2006). Contextual factors play an important role in shaping the processes of maladaptation, psychopathology, and resilience. For example, prior research demonstrates that social-contextual experiences can influence neurobiological structure and functioning (Cicchetti & Tucker, 1994; Grossman et al., 2003). Within a developmental psychopathology perspective, researchers are goaded to move away from the search for single pathogens and to move toward the search for a complex of influences that initiate a developmental pathway that is probabilistically associated with disturbances. Thus, systematic examination of multiple levels of contexts, including community-, institutional-, and societal-level influences, is encouraged in developmental psychopathology research (Cicchetti & Rogosch, 1999). Normality and Psychopathology Why do developmental psychopathologists study atypical, high-risk populations? It is because they believe not only that knowledge from the investigation of normal development can inform the study of high-risk conditions and psychopathology but also that the investigation of

Page 988

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Basic Assumptions and Longitudinal Descriptive Analyses

risk and pathology can enhance our comprehension of normal development (Cicchetti, 2010, 2015). Developmental psychopathology is studying how individuals function through examinations of extremes in the distribution. For example, as will be illustrated later, studies of maltreatment indicate that children who have experienced the extremes of dysfunction in parenting tend to show a broad range of psychological maladaptation. Such a finding indicates the criticality of adaptive parenting practices in the development of normal populations. Developmental Analysis In developmental psychopathology research, it is necessary to examine the relationships between predictors and problem behaviors over time and across ages (Cicchetti, 2015; Cicchetti & Rogosch, 1999). Accordingly, longitudinal research beginning even prior to the onset of disorder is important since such research is necessary for untangling causal mechanisms and processes. Some variables may have stronger links with onset of problems, while others may be more tied to persistence or desistance of the symptoms. Only longitudinal research can identify risk and protective factors associated with different stages in the progression of psychopathology. More importantly, examining longitudinal trajectories (as opposed to a snapshot) is essential for identifying developmental pathways to maladaptation, psychopathology, and resilience, as well as understanding diversity in process and outcomes of maladaptation, psychopathology, and resilience. As such, the core theoretical principles of developmental psychopathology demand empirical evaluation based on longitudinal data.

BASIC ASSUMPTIONS AND LONGITUDINAL DESCRIPTIVE ANALYSES Primary Goals of Longitudinal Research In their seminal paper, Baltes and Nesselroade (1979) described the history and rationale for conducting longitudinal research. In their work, they outlined five rationales for collecting longitudinal data, which could be studied only with longitudinal, repeated-measures data: identifying (1) intraindividual change; (2) interindividual variability in intraindividual change; (3) interrelationships among intraindividual changes; (4) determinants of intraindividual change; and (5) interindividual variability in the determinants of intraindividual change. Additionally, Baltes and Nesselroade indicated that the primary reason

12:38pm

989

for collecting longitudinal data was the identification of intraindividual change. Since their seminal paper, the primary approach used to evaluate and model intraindividual change is the latent growth model (McArdle, 1986; Meredith & Tisak, 1984, 1990). The latent growth model not only allows for the study of intraindividual or within-person change but also allows for the study of interindividual variability in intraindividual change (or between-person differences in within-person change) and determinants of intraindividual change (or causes/mechanisms of within-person change). The latent growth model has grown out of several modeling traditions, but is now dominated by the structural equation modeling (SEM) and multilevel modeling (MLM; also referred to as the mixed-effects and random coefficient model) frameworks. The SEM and MLM frameworks each have benefits and limitations (Ghisletta & Lindenberger, 2004); however, over time, the two frameworks have grown to be more similar, and the relative advantages of each approach have been minimized. Planning and Sampling in a Longitudinal Study The study of intraindividual change should be approached with careful planning and design. The design of a longitudinal study requires thoughtful consideration regarding the sampling of individuals, the sampling of measurement occasions, and the sampling of constructs (Cattell, 1952). The sampling of individuals is a first obvious consideration when designing any research study; however, additional considerations need to be made when planning a longitudinal study because the sampling of individuals can impact the time metric (sampling of measurement occasions) as well as the sampling of constructs. In a longitudinal study, the time metric, the time scale over which change has occurred, can be referenced by the beginning of the study. Example time metrics may be measurement occasion number (e.g., 1, 2, 3) and chronological time (e.g., 1 month, 12 months, 24 months). Such time metrics are common in psychopathology when studying intraindividual change subsequent to the beginning of an intervention (e.g., baseline, 1 month follow-up, 12-month follow-up). Alternatively, the time metric can be referenced at the individual level. Example time metrics may be chronological time since birth (age), time since an individual event (e.g., birth of child, marriage, death of family member), and time to an individual event (e.g., time to crawling, time until death). When planning of structuring change over an individual time metric, the sampling of individuals requires careful thought.

Page 989

Trim Size: 8.5in x 11in

990

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

For example, a study of intraindividual changes in depression during adolescence could be structured as a cohort-based study where 11-year-olds are sampled at the first occasion and then followed for the next five years at regular or irregular intervals. Alternatively, a sample of 11-year-olds, 12-year-olds, and 13-year-olds can be sampled at the first occasion and the followed for the next 3 years at regular or irregular intervals. This type of accelerated longitudinal study (Bell, 1954) requires a different sampling of individuals. The sampling of individuals also influences the sampling of constructs and the scales intended to measure the construct. Often, researchers will choose scales that are appropriate for measuring interindividual differences within the sample at the first measurement occasion without considering whether the scales are sensitive to expected intraindividual changes or whether the scales will be appropriate for the sample at the conclusion of the study. Measures must be sensitive to intraindividual change over the time period studied; however, most psychological measures are not evaluated for this ability—the majority of scales are psychometrically evaluated with a cross sectional sample. It is important to note that reliable and valid measures for studying interindividual differences are not necessarily reliable and valid for measuring intraindividual change. A second consideration for the sampling of scales is to look for scales that remain appropriate for the sample at the conclusion of the study. Often, the same measurement scales are utilized over the entire course of the study because researchers want to use the observed changes in the scores as a dependent variable. If this is done, it is important for the measures to be valid and reliable for all individuals at all time points and sensitive to change over the entire range of the time that the scale is utilized. For example, utilizing a child behavior problem scale for 3- to 5-year-olds may not be optimal when studying development from ages 4 to 7 because the sample ages out of the age range for which the scale was developed. Alternatively, researchers can plan for changes in the measurement of constructs over the course of the study. For some studies, it is unreasonable to have measures that are reliable, valid, and sensitive to intraindividual change over the entire course of the longitudinal study due to the length of the study and the sample characteristics and how the sample characteristics are expected to change over the course of the study. For example, studies of lifespan development may need to utilize different assessments over the course of the study as the sample ages (e.g., McArdle et al., 2009). In these cases, it is essential to build in scale and/or item overlap across time to create

scores that are quantitatively comparable. If this is not done, another approach is to use a separate sample to do equating/linking. There are upsides (e.g., do not over burden longitudinal sample) and downsides (e.g., inability to examine invariance within longitudinal sample) of this approach. Thoughtful consideration for the sampling of measurement occasions is also key to gaining the most information from the longitudinal study. The sampling of measurement occasions includes the number of measurement occasions as well as the timing or spacing of measurement occasions. However, a third consideration is whether or not the number of measurement occasions and the timing of assessments is the same or varies over individuals (McArdle & Woodcock, 1997). For this we simply note that much can be gained by intentionally building flexibility and variability into the measurement schedule of individuals and note that all aspects of the sampling of measurement occasions can be studied through simulation prior to any data collection (see Ferrer & Grimm, 2012). The sampling of measurement scales was discussed above, so here we simply reiterate the importance of utilizing measurement instruments that are sensitive to change over the planned data collection period and that many validated instruments may be unable to track the growth and changes taking place within the individual. Key Issues in Latent Growth Modeling Primary issues to the successful application of latent growth models include issues related to the measurement of the outcome of interest and issues related to the time scale over which the developmental process occurs and over which change is statistically modeled (Grimm & Ram, 2011). Measurement Latent growth models are typically applied to repeatedmeasures data obtained from multiple individuals. Among the first questions when approaching such data are how the construct was measured. Specific questions that arise are as follows: 1. What are the measurement properties of the instrument? 2. Are the measurement properties equivalent over time? 3. Was the instrument sensitive to and able to capture within-person change? 4. Are there distributional issues that need to be considered, such as ceiling effects, floor effects, skew, and

Page 990

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Basic Assumptions and Longitudinal Descriptive Analyses

whether the variable provides limited information (e.g., binary, ordinal variables)? 5. Most importantly, is the instrument measuring the same construct in the same metric across time? Ideally, the measurement instrument should have strong measurement properties (reliability and validity) at each measurement occasion. Additionally, the measurement instrument should be sensitive to change over the time scale of the study and at the level of aptitude (ability, trait-level) of the population. For example, utilizing the Woodcock-Johnson Applied Problems subtest may be useful for measuring changes in mathematics-related skills over a 2-year period for a sample of third to fifth grade students. However, the same measure is inappropriate for measuring mathematics-related skills over a 2-month period in a sample of preschool children. The reasons for the difficulty level of the test and the implied level of aptitude of the samples as well as the time span of the study must be thoughtfully considered. Moreover, certain instruments that are ideal for measuring relatively large changes in aptitude are poor at measuring relatively small changes in aptitude. A check on the distribution of scores from the instrument over time can highlight potential issues when approaching the data from a modeling standpoint. Ideally, scores will be more or less normally distributed at each measurement occasion or point along the time metric (e.g., each grade, each year). Deviations from normal distributions can indicate potential issues as well as potential opportunities. Truncated distributions, such as those with floor or ceiling effects, may require statistical adjustment when modeling. Depending on the nature of truncation, models can be fit to account for the truncation—for example, ceiling or floor effects can be statistically modeled as well as zero-inflated distributions (see Olsen & Schafer, 2001). If these distributional anomalies are not accounted for within the statistical model, incorrect solutions and conclusions can be drawn (Wang, Zhang, McArdle, & Salthouse, 2008). Nonnormal distributions, such as those that arise from count data (e.g., How many cigarettes smoked within the last 2 weeks?) or binary data (e.g., Have you smoked within the past 2 weeks?) should be modeled with the appropriate statistical distribution (e.g., Poisson, logit). Otherwise, model fit and parameter estimates can be biased. Alternatively, a skewed distribution may indicate the need for a mixture distribution (McLachlan, & Peel, 2000), which can arise because the sample is actually comprised from multiple subpopulations. In these situations, a growth mixture model (Muthén & Muthén, 2000; Muthén

12:38pm

991

& Shedden, 1999) or latent class trajectory model (Nagin, 1999) is needed. One of the most important considerations is that the instrument measures the same construct in the same metric at each point along the time metric. This enables the scores obtained from the instrument to be compared quantitatively. Thus, intraindividual change can be tracked. The question of whether the same construct was measured in the same metric is a question of measurement invariance (Meredith, 1964a, 1964b, 1965, 1993). Statistically, this question can be examined through fitting a series of item factor analyses or item response models (see Meredith & Horn, 2001; Widaman & Reise, 1997). However, when a single score is obtained from an instrument, measurement invariance is often assumed without thorough investigation. In these cases, researchers run the risk of drawing incorrect conclusions. It is not appropriate to assume measurement invariance holds simply because the same questions were asked at each measurement occasion; however, this is often done. Additionally, it is incorrect to assume measurement invariance does not hold when different questions were asked at different measurement occasions (see McArdle et al., 2009). Time Metrics A second key issue for understanding intraindividual change is the time metric, over which intraindividual change is thought to occur and the time scale for statistically tracking longitudinal change. Within the growth curve model, there is a single dominant time scale used to track intraindividual change. This dominant time scale serves as a proxy for tracking the time over which the underlying change process occurs (Grimm & Ram, 2011). Examples of time scales are chronological age, gynecological or maturational age, time since the beginning of the study, measurement occasions, time until death or other event (e.g., marriage, graduation), and grade in school. To further illustrate this point, we have plotted physical height for a sample of females from the National Longitudinal Survey of Youth–Children and Young Adults (NLSY-CYA; Center for Human Resource Research, 2004) in Figure 22.1 with different time scales. We chose the NLSY-CYA because this study began in 1986 with a sample of children from mothers who participated in the National Longitudinal Survey of Youth in 1979 (NLSY-79). Thus, children were different ages in 1986, and children were also added to the study as the study progressed. Furthermore, we chose physical height because its developmental progression is strongly influenced by chronological and maturational age.

Page 991

Trim Size: 8.5in x 11in

Cicchetti

992

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

(a)

(b)

(c)

(d)

(e)

Figure 22.1 Longitudinal plots of individual changes in height based on different time metrics for a sample of female participants of the National Longitudinal Survey of Youth–Children and Young Adults. Source: Bureau of Labor Statistics, U.S. Department of Labor, and National Institute for Child Health and Human Development. Children of the NLSY79, 1979–2010. Produced and distributed by the Center for Human Resource Research, The Ohio State University. Columbus, OH: 2012.

Page 992

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Basic Assumptions and Longitudinal Descriptive Analyses

Figure 22.1a is a plot of physical height against year in study from 1986 through 2002. This is obviously an inappropriate time scale to model changes in physical height given what we know about the developmental process and given the lack of systematicity in the individual trajectories of the longitudinal plot. At the same time, this plot is helpful because it highlights how different participants began the study in different years. Figure 22.1b is a plot of the same data against measurement occasion. Measurement occasion is a commonly utilized time scale for many longitudinal studies; however, for modeling changes in physical height in the NLSY-CYA, it is also inappropriate because participants entered the study at different ages. Additionally, measurement occasion, as a researcher-defined time scale, is unlikely to be the time scale over which intraindividual change is dependent upon (Grimm & Ram, 2011). On the positive side, this plot highlights how participants have a different number of measurement occasions. Figure 22.1c is a plot of physical height against year in school or grade. Grade is closely aligned with chronological age and maturational age, but is also inappropriate for studying individual changes in physical height. One issue that is evident in this plot is that children may repeat a grade (or two) and this can be seen as vertical lines within an individual trajectory. That is, the value of the time scale remains the same at two different measurement occasions. Obviously, grade could be measured more finely, such as grade plus the percent of school completed since the beginning of the school year or the amount of time since the child began kindergarten for the first time. Ideally, the chosen time metric will accurately represent the amount of time that has elapsed between assessments. For these data, year in school or grade does not adequately do this. Figure 22.1d is a plot of physical height against chronological age. Here, we see that children began the study at different ages and begin to see the developmental process that we know—stable increases in height during childhood, rapid increases during the pubertal years, and little to no changes in height during late adolescence/early adulthood. This is the time metric often utilized when modeling changes in physical height (see Grimm et al., 2011). We note that chronological age is a common time-metric when studying intraindividual changes in many types of biological and psychological constructs; however, other scalings of age may be as, or even more, appropriate (Nesselroade & Featherman, 1997). As a final illustration relevant to this point, we have plotted physical height against gynecological age in Figure 22.1e. Gynecological age was defined as time to or time since the individual’s

12:38pm

993

age of menarche. In this figure, we see that changes in physical height were somewhat more rapid near age of menarche (gynecological age = 0) compared with changes that occurred after or before menarche. Certain research questions and theories posit alternative scalings of age as the ideal time scale to model intraindividual change and these should be utilized to the extent possible (e.g., terminal decline; Gerstorf et al., 2008). Ideally, the chosen time metric will adequately represent the amount of time that has passed between assessments as well as represent a dominant reason why the intraindividual changes have occurred as they have. One approach is to consider the time metric, which is able to align participants, such that the shape of developmental changes is most similar. In practice, we recommend examining multiple feasible time metrics before proceeding with analyses. This can be done with a series of plots, as in Figure 22.1, and done conceptually and theoretically. In the majority of situations, we acknowledge that multiple ongoing time metrics are likely to impact intraindividual change (Grimm & Ram, 2011). For example, changes in depression over adolescence may be related to schooling demands and peers, which change with year in school, as well as demands at home, which are likely associated with chronological age, as well as biological changes taking place within the individual associated with maturational age. These time metrics are strongly associated with one another (highly collinear), forcing researchers to choose between them or model them jointly and deal with the presence of collinearity. We refer readers to McArdle and Bell (1954), Ram et al. (2010), and Wohlwill (1973) for discussions regarding the choice of time metric in latent growth curve models. Modeling Intraindividual Change Once the time metric is decided upon, the next step is to determine, at the individual level, how the observed trajectories are associated with the chosen time metric. The majority of applications in growth curve modeling focus on simple functional forms of change, such as linear and quadratic change functions. However, the true underlying change process is likely to be more complicated. Here, a balance must be struck between the complexity of the model and the amount of information present in the data. There are several approaches to choosing an appropriate individual change model. The first, and most common, approach is to begin with simple change functions and progress through a series of viable alternatives chosen based upon an individual plot, such as those presented in Figure 22.1. The simplest change function is a no-change, no-growth, or intercept-only model, where each individual

Page 993

Trim Size: 8.5in x 11in

994

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

has an intercept score that is maintained across the observation period. This is often a model researchers would like to reject, but is a common starting point because of its simplicity and because the model yields fit statistics that can be statistically compared with more complex models because the intercept only model is nested under most growth models. Next, a linear model is fit, which is followed by a quadratic model and higher order polynomials if necessary. Finally, a few inherently nonlinear models (e.g., exponential, Gompertz, logistic) are fit that may be able to mimic the observed overall change pattern (see Grimm & Ram, 2009). A second approach is to consider the important characteristics of the developmental process and fit a model with those properties, and if this model fits the data well then use the model. For example, important aspects (or parameters where individuals are likely to differ) in the developmental process related to changes in physical height may be rates of change in infancy, childhood, and adolescence along with the timing of puberty and a final adult height. Preece and Baines (1978) derived multiple functional forms with many of these important aspects and applied them to longitudinal height data and evaluated model fit (see also Grimm et al., 2011). We do, however, acknowledge that this approach is challenging when considering the range of psychological constructs and the general lack of strong theories describing important changes in such constructs. A third approach is to model intraindividual change patterns with a series of splines (see Cudeck & Klebe, 2002; Ram & Grimm, 2007)—a series of lines or simple curves (e.g., quadratic) that connect at specific points in time (knot points) that can be chosen by the researcher or estimated from the data. These models are sometimes referred to as multiphase models because the various splines can represent phases of development that have beginning and end points. These models are especially useful when the underlying developmental process is thought to be guided by shifts or when the developmental process is interrupted by an external event, such as a death in the family. In this situation, changes prior to the event and after the event can be thought of as having distinct developmental processes. A fourth approach is to let the data speak. There is a variety of what we would call exploratory approaches for modeling change (see Grimm, Steele, Ram, & Nesselroade, 2013). For example, Meredith and Tisak (1990) fit a latent basis or unstructured growth model, where an ideal structure of change was derived from the data. Functional forms of change that mimicked this derived change pattern were fit to the data and evaluated. Grimm et al. (in press) advocated a completely exploratory approach

to growth modeling where a series of models are fit with an increasing number of latent variables. The associations between the time-metric and the observed trajectories are unstructured (similar to the latent basis model). Thus, in this approach the number of latent variables and the individual patterns of change are unknown, which is similar to how exploratory factor analysis models are fit to cross sectional multivariate data. This approach can yield an appropriate number of latent variables as well as potential forms of change that may be suitable for the data. Additional exploratory approaches are based upon principal component analysis (Davison, 2008; Tucker, 1958, 1966), multidimensional scaling (Ding, Davison, & Petersen, 2005), and smoothing spline models (Guo, 2002). Interindividual Variability in Intraindividual Change In latent growth curve models, the intraindividual change pattern or the functional form of change is invariant over participants. For example, if a linear model is chosen as the intraindividual change function, then all individuals are expected to follow a linear change pattern. Interindividual variability in intraindividual change, within the latent growth modeling framework, enters in the parameters, which describe the individual change pattern. That is, even though individuals follow the same functional form (e.g., linear), the parameters of the functional form (e.g., intercept and slope) are allowed to vary over individuals. Allowing the parameters of the functional form to vary over individuals provides for a wide variety of potential developmental patterns, especially with a functional form with multiple parameters (e.g., cubic and spline models). Theoretically, any parameter of the intraindividual change function can vary over individuals (e.g., intercept and slope of linear model are allowed to vary over individuals); however, there are situations where theory or a lack of information from the data leads certain parameters to be invariant over individuals such that interindividual variability is not manifested. One of these situations was evident when studying intraindividual changes in reading ability (Grimm, Ram, & Estabrook, 2010). The chosen intraindividual change model was a Gompertz model. The Gompertz model has four parameters: a lower asymptote, an upper asymptote, the location of the inflection point, and a rate of change from the lower to the upper asymptote. For reading ability, the lower asymptote is the predicted reading ability when time approached negative infinity. Theoretically, it is logical that interindividual differences in reading ability at this point in time were minimal or nonexistent. Thus, this parameter was not allowed to vary over individuals. Ideally, theoretical notions of the

Page 994

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

The Latent Growth Modeling

developmental change process should drive and inform the nature and location of interindividual variability in intraindividual change; however, again we recognize the limitations of the data that are brought to the analysis. Determinants of Intraindividual Change As mentioned earlier, one of the rationales for collecting longitudinal data proposed by Baltes and Nesselroade (1979) is the identification of determinants of intraindividual change. The latent growth curve model accommodates this rationale by allowing for interindividual difference variables (e.g., gender and intervention type) to predict the parameters of the growth curve, which describe interindividual variability in intraindividual change. Ideally, these associations will indicate the degree to which the interindividual difference variables are associated with the rate of intraindividual change; however, this is not necessarily the case with complex models that have multiple parameters impacting intraindividual change (Grimm, Castro-Schilo, & Davoudzadeh, 2013). Finally, we note that the identification of causes of interindividual differences in intraindividual change can only be identified with appropriate controls, such as random assignment. THE LATENT GROWTH MODELING Historical Development Early work in the modeling and understanding of intraindividual change was very descriptive before the rise of fitting mathematical models to individual-level data. In this approach, linear and a variety of complex nonlinear models were considered and the researchers had multiple goals. A first goal was to minimize the size of the residuals. Models with smaller residuals better represented to data and were seen as superior. A second goal was to have a mathematical function with parameters that had meaning in relation to the intraindividual change process under consideration. These, sometimes contrasting, goals are evident in the study of physical height where certain mathematical models were specified to capture key aspects of development, such as the timing of puberty (e.g., Preece & Baines, 1978), whereas other models were developed to simply map onto the observed data with little attention paid to the interpretability of parameters (e.g., reference). This phase along the developmental continuum, is often referred to as individual growth modeling (Willett & Sayer, 1994), continued for a number of years with few modifications (e.g., Berkey, 1982), and remains a useful approach to studying intraindividual change. Furthermore, interindividual

12:38pm

995

variability in intraindividual change could be examined by relating parameters of the mathematical function to between-person variables (e.g., Wishart, 1938). During the individual growth modeling phase, alternative approaches to understanding intraindividual change surfaced. For example, Tucker (1958, 1966) fit principal component analysis models to the matrix of sums of squares and cross-products to decompose the matrix into generalized learning curves and individual weights. The generalized learning curves represent dominant trends within the data and the individual weights represented the extent to which an individual’s trajectory was impacted by each generalized learning curve. Tucker used the term generalized learning curves due to his work with learning data; however, this approach can be applied to a variety of types of data and is commonly referred to as Tuckerized curves (see Grimm, Steele, et al., 2013; Meredith & Tisak, 1984). Rao (1958) described several approaches for understanding intraindividual change and interindividual variability in intraindividual change. First, Rao (1958) discussed calculating an average rate of change—in essence decomposing the repeated measures into an initial state and a rate of change. In this approach, he also discussed transformations of the time scale to make changes more or less constant (linear change). Rao (1958) also discussed the possibility of using factor analytic methods, specified the growth curve model, and utilized principal components analysis of the uncorrected sum of squares and products matrix to estimate a model. This last routine is equivalent to Tucker’s (1958) generalized learning curves and was developed independently at the same time. In 1984, Meredith and Tisak presented their paper, Tuckerizing curves, at the International Meeting of the Psychometric Society. In the presentation, Meredith and Tisak described the latent curve model and how the model can be specified within the SEM framework. They discussed multiple types of latent growth models including the free curve or latent basis model as well as a negative exponential model and this work was published in 1990 (Meredith & Tisak, 1990). During this time, McArdle expanded on Meredith and Tisak’s work by exploiting the flexibility of the structural equation modeling framework (McArdle & McDonald, 1984) to specify combinations of latent growth curves and behavior genetic models (McArdle, 1986), multivariate latent growth models (McArdle, 1988), and multiple group latent growth models (McArdle & Hamagami, 1996). Additionally, Browne and du Toit (1991) and Browne (1993) highlighted how the SEM framework could be used to approximate growth models that followed inherently nonlinear mathematical functions.

Page 995

Trim Size: 8.5in x 11in

996

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

At a similar time, Laird and Ware (1982) published a paper in biometrics on random-effects models for longitudinal data describing models for growth and for repeated measures. An early article that reached psychologists was written by Bryk and Raudenbush (1987), who published a Psychological Bulletin article on the use of hierarchical linear model (i.e., multilevel model, mixed-effects model, and random coefficient model) to assess change. This spawned the use of this framework for studying growth within the psychological and educational disciplines. The work that followed was focused on how the mixed-effects modeling framework could be used to study inherently nonlinear processes. This work appeared in books by Davidian and Giltinan (1995), Littell et al. (1996), and Pinheiro and Bates (2000). Additional work by Singer (1998) and Willett and Sayer (1994) was prominent in the development of latent growth models and spread the use of the multilevel approach to studying change. By the late 1990s, it was clear that multilevel models for assessing change and structural equation models for assessing change were equivalent in many cases (Curran, 2003; McArdle & Hamagami, 1996), especially for the most commonly fit models (e.g., polynomial models). Papers followed describing the various ways to fit latent growth models (e.g., Ferrer, Hamagami, & McArdle, 2004) and the relative advantages and disadvantages of each approach (Ghisletta & Lindenberger, 2004). At first, the multilevel model had advantages related to individually spaced measurements and the flexibility of fitting inherently nonlinear models, whereas the structural equation modeling approach had advantages related to indices of model fit, flexibility in the residual structure, and the inclusion of measurement models. Several of these relative advantages have been reduced because each framework has increased in flexibility. Furthermore, general modeling programs (e.g., WinBUGS) have further blurred the differences between the multilevel and structural equation modeling approaches for understanding change. In the late 1990s and early 2000s, there were two additional important expansions of the general latent curve modeling framework. The first development was the introduction of the growth mixture model by Muthén and Shedden (1999). The growth mixture model allowed for an exploration of growth trajectories and for an additional way to incorporate between-person differences in growth. The growth mixture model combined the finite mixture model and the latent growth model and allowed researchers to cluster individuals based upon their individual growth trajectories. Thus, between-person differences in growth were allowed to manifest themselves

by different classes and between-person differences within each class. Additionally, we note a similar development appeared in the latent class trajectory model, which was introduced by Nagin (1999). The difference between the growth mixture model and the latent class trajectory model was that between-person differences were allowed for within each latent class in the growth mixture model but not in the latent class trajectory model. The second development was the introduction of the latent difference or latent change score modeling framework by McArdle and Hamagami (2001; see also McArdle, 2001). The latent change score model allowed for dynamic associations between multiple outcomes while simultaneously modeling growth trajectories. This model placed an emphasis on time-sequential within-person change and led to subsequent developments in differential models (see Boker, Neale, & Rausch, 2004) and latent acceleration models (Hamagami & McArdle, 2007). More recently there have been developments that combine several of the models discussed above. For example, Grimm and Ram (2009b) proposed the second-order growth mixture model—a combination of the second-order growth model (McArdle, 1988) and the growth mixture model (Muthén & Shedden, 1999). Grimm, Zhang, Hamagami, and Mazzocco (2013) combined the latent difference score model (McArdle & Hamagami, 2001) with the estimation of inherently nonlinear growth models within the SEM framework (Browne & du Toit, 1991), which allowed for an emphasis on within-person change when modeling nonlinear growth trajectories. Grimm (2006) combined the latent difference score model with the finite mixture model, and Grimm, Ram, and Estabrook (2010) presented a combination of the finite mixture model and the estimation of inherently nonlinear growth models within the SEM framework to examine clustering of nonlinear growth trajectories. Finally, Zhang, McArdle, and Nesselroade (2011) presented growth rate models—rotations of growth models where a latent variable representing the rate of change at a given time point is included in the model to highlight the individual rate of change in nonlinear growth models. Latent Growth Curve as a Structural Equation Model Within the SEM framework, the latent growth curve is fit as a restricted common factor model. The growth factors (e.g., intercept and linear slope in a linear growth model) of the growth model are latent variables within the SEM framework. In the specification that follows and all model specifications included in this chapter, we use Greek letters for parameters that are fixed or estimated (e.g., factor

Page 996

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

The Latent Growth Modeling

loading matrix) and Latin letters for entities that are not estimated nor fixed (e.g., factor scores). The general latent growth model can be expressed as Yi = Λhi + ui

(22.1)

where Yi is a T × 1 vector of the repeatedly measured observed scores for individual i;, where T represents the number of repeated assessments based on the chosen time metric; Λ is a T × R factor loading matrix, which serves to define the growth factors of the growth curve (e.g., intercept and slope); R is the number of growth factors; hi is an R × 1 vector of latent factor scores (e.g., intercept and slope scores) for individual i; and ui is a T × 1 vector of unique scores for individual i. The latent factor scores can be written as deviations from the sample-level mean written as hi = 𝛂 + di

997

structure for the latent growth model is simply the factor loading matrix multiplied by the mean vector of the latent variables, and the expected covariance structure follows the common factor analysis model of the latent variable covariance matrix pre and post multiplied by the factor loading matrix plus the residual or unique covariance matrix. A path diagram of a latent growth curve model with two latent variables is given in Figure 22.2. In such diagrams, squares represent observed variables, circles represent latent variables, and the triangle denotes the constant. One-headed arrows represent directional relationships, such as regression paths and two-headed arrows represent nondirectional or symmetric relations, such as covariances. The labels of the paths in the path diagram are derived from the previous specification. With the path diagram in Figure 22.2 and the path tracing rules derived by Wright (1921), model expectations can be calculated, which replicate the matrix algebra given already.

(22.2)

where 𝛂 is an R × 1 vector of latent variable means, and di is an R × 1 vector of mean deviations for individual i. The covariance matrix of mean deviations is 𝛙, an R × R matrix indicating the magnitude of between-person differences in the latent growth factors as well as the degree of linear association among the latent growth factors. The unique scores have expected means equal to 0 (i.e., E(ui ) = 0) and covariance matrix 𝚯 (i.e., COV(ui ) = 𝚯), where is a T × T matrix. Often, only the diagonal elements of are estimated, which forces the unique scores to be time-specific. However, additional types of structures of can be examined, such as first-order autoregressive and block-diagonal (see Kwok, West, & Green, 2007). Additionally, many researchers force the diagonal elements to be equal across time (i.e., 𝚯 = I𝜃, where is a T × T identity matrix and is a scalar). The latent growth model places a set of expectations on the repeated measures and it is important to understand these expectations to evaluate model fit and misfit. Additionally, understanding model expectations can guide the use of model modifications. The model expectations of observed variables based upon the latent growth model is 𝛍 = Λ𝛂 ∑ = 𝚲𝛙Λ′ + 𝚯

12:38pm

Common Latent Growth Models To further show how the SEM framework can be used to specify latent growth models, we present a series of specific latent growth models that are commonly considered in psychology and psychopathology. In each description, we present specifications of the required matrices. In all specifications, we assume there are four equally spaced measurement occasions (T = 4).

1 1,1

1

h1

2,2

2 h2

2,1

𝛬

Y1

Y2

u1

Y3

u2

Y4

u3

u4

(22.3) 

where is an T × 1 vector of expected means and is a T × T expected covariance matrix. The expected mean

Figure 22.2





A latent growth curve model.



Page 997

Trim Size: 8.5in x 11in

998

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

Linear Change The linear change model is the most commonly fit growth curve in applied research. The linear change model has two latent variables (R = 2): an intercept, often centered at the first measurement, and a linear slope representing the individual constant rate of change over the observation period. The intercept and linear slope are allowed to vary over persons and allowed to covary with each other. Conceptually, the linear growth model is similar to fitting a linear regression model to each individual’s data, retaining individual parameters for the intercept and slope, and summarizing those estimated parameters to calculate their means to get a sense of the average growth trajectories as well as their variances and covariance to understand between-person differences in the growth parameters and their degree of linear association. The linear growth model is specified as ⎛ ⎜1 ⎜ ⎛Y1 ⎞ ⎜ ⎜ ⎟ ⎜1 ⎜Y2 ⎟ = ⎜ ⎜Y3 ⎟ ⎜ ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎜ ⎜ ⎜1 ⎝

1 − k1 ⎞ k2 ⎟⎟ 2 − k1 ⎟ ⎛u ⎞ ( ) ⎜ 1⎟ ⎟ k2 ⎟ h1 u + ⎜ 2⎟ 3 − k1 ⎟ h2 ⎜u3 ⎟ ⎟ ⎜u ⎟ k2 ⎟ ⎝ 4⎠ 4 − k1 ⎟ ⎟ k2 ⎠

(22.4)

where the repeated measures for individual i is a product of the factor loading matrix multiplied by the factor scores for individual i, plus the unique scores for individual i. In the factor loading matrix, the first column contains fixed values set equal to 1 to define the intercept factor, h1 , and the second column will contain fixed values set equal to change linearly with time to define the linear slope, h2 . The constants are chosen to center the intercept and scale the slope. Often, is set to 1 to center the intercept at the first occasion and is set to 1 to scale the slope in terms of the unit of time used in the selected time metric. Quadratic Change The quadratic change model is commonly considered when curvature is evident in the growth pattern and when the linear model does not adequately capture the intraindividual change pattern. The quadratic growth model is specified as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

0 1 2 3

⎛u1 ⎞ 0⎞ h ⎞ ⎛ ⎜ ⎟ ⎟ 1 1∕2 ⎟ ⎜h ⎟ + ⎜u2 ⎟ 2 2 ⎟ ⎜ ⎟ ⎜u3 ⎟ ⎝h3 ⎠ ⎜ ⎟ 9∕2⎟ ⎝u4 ⎠ ⎠

(22.5)

The quadratic model has three latent variables, an intercept, h1 , a linear slope, h2 , and a quadratic slope, h3 . The factor loadings for the intercept are all fixed at 1, the factor loadings for the linear slope change linearly with time, and the factor loadings for the quadratic slope are simply the square of the factor loadings for the linear slope divided by 2 (this is done to scale the quadratic slope in terms of acceleration). We note the same potential to center the intercept and scale the linear and quadratic slope. In this example, the centering and scaling constants, and k2 , are both set equal to 1. With this centering, the intercept represents the predicted value of the outcome at the first occasion; the linear slope represents the individual rate of change at the first occasion; and the quadratic slope represents acceleration, curvature, or how quickly the rate of change is changing, which is constant across time. The quadratic model is often considered; however, the parameters of the quadratic model can be difficult to map onto theoretical notions of change (Cudeck & du Toit, 2002) and difficult to interpret in isolation because both the linear and quadratic slopes impact the intraindividual rate of change (see Grimm, Castro-Schilo, & Davoudzadeh, 2013). To combat these issues, Cudeck and du Toit (2002) reparameterized the quadratic model to emphasize the time when the maximum (or minimum) value is reached and the maximum (or minimum) value. These parameters are easily interpretable, highlighting interesting aspects of quadratic curve. Unfortunately, Cudeck and du Toit’s (2002) approach is rarely considered when fitting the quadratic growth model in applied settings. Transformations of Time Polynomial transformations of time, such as in the quadratic model described above, are common in the application of latent growth models but are not the only viable transformations of time. For example, square root or log transformations of time may yield useful results concerning the growth of the construct under study. In essence, the growth process may be linear with respect to a different time scale. If this alternative time scale is utilized, then the process would conform to a linear growth model (see also Rao, 1958). As an example, we present the square root growth model, which can be specified as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜1 ⎜Y2 ⎟ = ⎜ ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜ ⎝ 4 ⎠ ⎝1

0 ⎞ ⎛u ⎞ ( ) ⎜ 1⎟ ⎟ 1 √ ⎟ h1 + ⎜u2 ⎟ ⎜u3 ⎟ 2⎟ h √ ⎟ 2 ⎜u ⎟ 3⎠ ⎝ 4⎠

(22.6)

Page 998

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

The Latent Growth Modeling

where the first column of the factor loading matrix defines the intercept and the second column defined the linear slope in square root time. Additionally, if the second √ column of the factor loading matrix was divided by 3, then the slope can be discussed as the amount of change the has occurred from the first to the last measurement occasion; however, the shape of intraindividual change would still follow square root time. Latent Basis or Unstructured Change The latent basis or unstructured change model is a very flexible model that borders on exploratory change because the pattern of change is derived from the data. That is, instead of imposing a structure on change, such as linear or quadratic, within the factor loading matrix, the factor loading matrix has elements that are freely estimated. The latent basis model can be written as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

⎛u1 ⎞ 0 ⎞ ⎟( ) ⎜ ⎟ u 𝜆2,2 ⎟ h1 + ⎜ 2⎟ ⎜u3 ⎟ 𝜆3,2 ⎟ h2 ⎜u ⎟ 1 ⎟⎠ ⎝ 4⎠

(22.7)

The first column of the factor loading matrix is set equal to 1 to define the intercept latent variable. The second column, on the other hand, has two fixed elements and two estimated elements. The first element is set equal to 0, the second and third are estimated from the data, and the last element is set equal to 1. This pattern of loadings define, what most call, a shape factor because the estimated factor loadings define the structure or individual change pattern and this pattern of change can vary greatly in shape. The two fixed factor loadings are needed for identification; however, their values and locations within the second column are arbitrary. This configuration centers the intercept at the first measurement occasion and scales the shape factor in terms of change from the first to the last occasion. Additionally, the estimated factor loadings represent the proportion of total change that has occurred since the first measurement occasion. Such interpretations are straightforward when change is monotonic; however, the latent basis model can structure change that is not monotonic (e.g., rise and fall). Even though the interpretation of the estimated factor loadings remain, changing the identification constraints to emphasize change to the maximum (or minimum) point can make interpretation more straightforward (see Ram & Grimm, 2007). Finally, we note that the latent basis model has been considered a model that optimizes the rescaling of time due to the estimated factor loadings.

12:38pm

999

Inherently Nonlinear Change Patterns An additional collection of latent growth models follows patterns of change that are inherently nonlinear. In such models, a parameter is contained within the function of time—usually in an exponent or raised to an exponent. That is, within the factor loading matrix a nonlinear function is specified where an estimated parameter is contained within that function. Such models have several benefits because of the wide variety of shapes of development that can be modeled and the flexibility for modeling between-person differences in the shape of development (Burchinal & Appelbaum, 1991; Grimm, Ram, & Hamagami, 2011). Two such inherently nonlinear models are the power model and the exponential growth model. The power model can be expressed as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

𝛽0 𝛽1 𝛽2 𝛽3

⎛u1 ⎞ −1⎞ ⎟( ) ⎜ ⎟ −1⎟ h1 u + ⎜ 2⎟ ⎜u3 ⎟ −1⎟ h2 ⎜u ⎟ −1⎟⎠ ⎝ 4⎠

(22.8)

where the factor loading matrix has a series of ones in the first column to define the intercept and the second column defines a power function where an estimated parameter 𝛽 is raised to the power of time (0, 1, 2, 3) and then one is subtracted from this value. The variety of developmental patterns is determined by the estimated value of 𝛽. The exponential model can be written as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

⎛u1 ⎞ 1 − exp (−𝛽 ⋅ 0)⎞ ⎟( ) ⎜ ⎟ 1 − exp(−𝛽 ⋅ 1) ⎟ h1 u + ⎜ 2⎟ ⎜u3 ⎟ 1 − exp(−𝛽 ⋅ 2) ⎟ h2 ⎜u ⎟ 1 − exp(−𝛽 ⋅ 3) ⎟⎠ ⎝ 4⎠

(22.9)

where the second column of the factor loading matrix now changes as a function of 1 − exp(−𝛽 ⋅ t), where 𝛽 is an estimated parameter and t is the value to time based upon the chosen time scale. In the power model and the exponential model, we note that the factor loadings are not fixed values as in the linear, quadratic, or square root models but are also not freely estimated as in the latent basis model. Instead, the values are structured to change according to a specific function. The utility and flexibility of inherently nonlinear models is shown when allowing the parameters within the factor loading matrix to vary over individuals. This type of model is not directly estimable within the SEM framework; however, this type of model can be approximated and in many

Page 999

Trim Size: 8.5in x 11in

1000

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

cases the approximation is very good (see Browne, 1993; Browne & du Toit, 1991; Grimm, Ram, & Hamagami, 2011, for details on this approximation). Multiphase Change Multiphase change models (see Cudeck & Klebe, 2002) are a type of latent growth model where the repeated measures data are compartmentalized and change is more or less modeled within discrete phases of time. These models also go by the name spline or piecewise growth models. Change within each phase can be modeled with any of the above mentioned growth models; however, relatively simple models, such as a linear or quadratic change model, are often fit within each phase. For example, a bilinear model can be specified as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

0 1 1 1

⎛u1 ⎞ 0⎞ ⎟ ⎛h1 ⎞ ⎜ ⎟ 0⎟ ⎜ ⎟ ⎜u2 ⎟ h + 1⎟ ⎜ 2 ⎟ ⎜u3 ⎟ ⎝h ⎠ ⎜ ⎟ 2⎟⎠ 3 ⎝u4 ⎠

(22.10)

where the first column of the factor loading matrix defines the intercept, which is centered at the first measurement occasion; the second column defines the first linear phase, which represents intraindividual change during the first two measurement occasions; and the third column defines the second linear phase, which represents intraindividual change from the second to fourth measurement occasions. These multiphase models can accommodate a variety of nonlinear change patterns and tend to have parameters that are easily interpretable.

MULTILEVEL MODELING FOR STUDYING DEVELOPMENTAL TRAJECTORIES The multilevel model for studying developmental trajectories was presented by Laird and Ware (1982) and Bryk and Raudenbush (1987). Growth models can be fit in the multilevel modeling framework as a two-level model because the repeated measurements are nested within individuals. There are several ways to present the multilevel model—we present two common approaches, highlight how the specifications are equivalent, and how the specifications can map onto the structural equation modeling approach for latent growth analyses. Following Laird and Ware (1982), the multilevel model for change can be specified as Yi = Xi 𝛂 + Zi di + ei

(22.11)

where Yi is a ni × 1 vector of observed scores for individual i, Xi (chi) is a known ni × R design matrix linking 𝛂, an R × 1 of unknown fixed effects estimates, to Yi , Zi (zeta) is a known ni × K design matrix linking di , a K × 1 vector of unknown individual effects, to Yi , and ei is an ni × 1 vector of residuals. The individual effects, di , are assumed to be multivariate normal with means equal to 0 and covariance matrix 𝛙 (i.e., di ∼ MVN(0, 𝛙)). The residuals are assumed to be normally distributed with mean 0 and covariance matrix 𝚯i (i.e., ei ∼ N(0, 𝚯i )). We note that the dimensions of 𝚯i are dependent on i, but the unknown parameters do not depend on i and 𝚯i = 𝜃I, where I is an identity matrix, and 𝜃 is a common residual variance. This model yields a mean vector of Yi equal to Xi 𝛂 and covariance matrix Zi 𝛙Z′i + 𝚯i . The matrix specification of the multilevel growth model (equation 22.11) follows closely the specification within the SEM framework (equation 22.1). Of note, in the growth model without covariates and traditionally specified, the two design matrices, Xi and Zi , are equivalent and are individual functions of time. With these matrices being equivalent, the model more closely mimics the structural equation modeling specification with a separation of the mean vector for the latent variables (i.e., 𝛂) from the individual deviations (i.e., di ). Noting these similarities, we now highlight the main differences. The first difference is the dimension of several matrices. In the multilevel modeling framework, the dimensions of several matrices vary over individuals. The vector of observed scores, the design matrices, and the residual vector have dimensions that vary at the individual level, whereas the comparable matrices within the structural equation modeling framework have dimensions that do not vary over the individual level and have dimensions based on the number of occasions from the chosen time metric. The next major difference is that the design matrices within the multilevel modeling framework (Xi and Zi ) vary at the individual level (subscripted by i), whereas the factor loading matrix from the structural equation modeling framework (Λ) does not vary at the individual level. Thus, the multilevel model is able to accommodate individually varying measurement schedules, whereas the structural equation modeling framework is for a fixed measurement schedule that all individuals follow. However, we note two things. First, although the measurement schedule is fixed, all participants do not need to be measured at each of the fixed measurement occasions. Second, recent advances within the SEM framework have allowed for the factor loading matrix to be individually specified

Page 1000

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Multilevel Modeling for Studying Developmental Trajectories

(essentially turning the structural equation model into a multilevel model). The third major difference between the frameworks is the ways in which the latent variables or random coefficients combine within the model. The structural equation modeling framework is a linear modeling framework, which means that the latent variables can only combine in a linear or additive fashion (i.e., latent variables multiply the factor loading matrix and are then added together to yield predicted values). In the multilevel modeling framework, the random coefficients can also combine in a nonlinear or multiplicative fashion. Therefore, there are models that can be fit within the multilevel, or more specifically the nonlinear multilevel, modeling framework that cannot be fit within the structural equation modeling framework. Alternative Specification An alternative way to present the multilevel model is Yit = f (timeit , hi ) + eit

(22.12)

12:38pm

1001

the interpretation of hi2 , and eit is a time-specific residual score for individual i at timeit . The constants, k1 and k2 , can be set to any value; however, the values for these constants impact interpretation and should be carefully considered. For example, to center the intercept at the zero point of the original time metric, k1 = 0. Thus, the intercept, hi1 , would reflect an individual’s true level at the attribute when timeit = 0. This may be reasonable for many time scales, but is inappropriate for other time scales and constructs. For example, when examining individual changes in depression as a function of age during adolescence, it might be reasonable to center the chosen time metric, age, around the age when measurements were first taken (e.g., k1 = 13 years). In a similar vein, the choice of k2 should be carefully considered to scale the time metric in an appropriate way. We note that k2 is often set equal to 1 to scale the slope, h2i , in terms of the original time metric, timeit ; however, for certain constructs that change more rapidly or more slowly, it is reasonable to choose a different value of k2 . Quadratic Growth The quadratic model can be specified as

where Yit is the observed response for the i-th individual at time t, f (timeit , bi ) is a linear or nonlinear function of time with random coefficients, h′i = (hi1 , hi2 , . . . , hiR ), that combine linearly or nonlinearly, and eit is the residual for individual i at time t, which is assumed to follow a normal distribution with constant variance (i.e., eit ∼ N(0, 𝜃)). The random coefficients are often assumed to follow a multivariate normal distribution, such that hi ∼ MVN(𝛂, 𝛙), where 𝛂 is an R × 1 vector of fixed-effects parameters and 𝛙 is a R × R covariance matrix of random-effects parameters. Specific Growth Models

Yit = hi1 + hi2 ⋅

timeit − k1 k2

(

) + hi3 ⋅

timeit − k1 2k2

)2 + eit

(22.14) where hi1 is the intercept for individual i, hi2 is the linear slope for individual i, hi3 is the quadratic slope for individual i, and eit is the time-specific residual score for individual i at timeit . Similar to the linear growth curve, the intercept remains the predicted score when timeit = k1 . The linear slope represents the rate of change when timeit = k1 and the quadratic slope represents acceleration or curvature. Inherently Nonlinear Models

Linear Growth The linear growth model is written as ( Yit = hi1 + hi2 ⋅

(

timeit − k1 k2

) + eit

(22.13)

where Yit are the repeated measures for individual i at time t, hi1 is the intercept or predicted score when timeit = k1 for individual i, k1 is a constant used to center the intercept, hi2 is the linear slope or the predicted amount of change time in Yit for a one unit change in k it for individual i , k2 is 2 a scaling coefficient used to adjust the metric of timeit and

One of the benefits of the multilevel modeling framework is the straightforward extension to the nonlinear multilevel modeling framework. That is, the multilevel model can be used to fit any type of nonlinear change model. This flexibility is important because development is often highly nonlinear and linearity often serves as an approximation for a truly nonlinear process. In this section, we present two nonlinear models, both of which follow an exponential trajectory, that vary in complexity. The first exponential model can be written as Yit = hi1 + hi2 ⋅ (1 − exp(−𝛽 ⋅ timeit )) + eit

(22.15)

Page 1001

Trim Size: 8.5in x 11in

Cicchetti

1002

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

where hi1 is the predicted score when timeit = 0 for individual i, hi2 is the total amount of change from hi1 to the asymptotic level for individual i, 𝛽 is the rate of approach to the asymptotic level, and eit is the time-dependent residual. In this model, 𝛽 is an estimated (fixed effect) parameter and does not vary across participants, whereas hi1 and hi2 are random coefficients that vary over participants. This form of the exponential is additive because the random coefficients enter the model in a linear/additive fashion because 𝛽 is an estimated parameter. We note that this model can be fit within the SEM framework (see equation 22.9) as long as nonlinear constraints are available, and the majority of structural equation modeling programs do allow for such constraints. The second exponential model is written as Yit = hi1 + hi2 ⋅ (1 − exp(−hi3 ⋅ timeit )) + eit

(22.16)

where hi1 is the predicted score when timeit = 0 for individual i, hi2 is the total amount of change from hi1 to the asymptotic level for individual i, hi3 is the rate of approach to the asymptotic level for individual i, and eit is the time-dependent residual. Now, the rate of approach varies over individuals. This simple change makes the model multiplicative because hi2 is a multiplier of hi3 and therefore, these random coefficients enter the model in a nonlinear fashion. Thus, the structural equation modeling framework is unable to estimate this model (see Browne, 1993); however, the multilevel model can estimate the model directly—a big advantage of the multilevel modeling framework. Direct Expansions of the Latent Growth Modeling Framework The specifications described are focused on the study of intraindividual change and interindividual variability in intraindividual change. A common next step involves the evaluation of determinants of interindividual variability in intraindividual change. As noted already, this is often accomplished with the inclusion of time-invariant covariates as predictors of the latent growth variables. Moving back to the linear growth model for simplicity, this model can be specified as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

⎛u ⎞ 0⎞ ( )] ⎜ 1 ⎟ ⎟ [( ) ( ) u 𝛾 d 1⎟ 𝛼1 + 1 ⋅ (X ) + 1 + ⎜ 2⎟ ⎜u3 ⎟ 𝛾2 d2 2⎟ 𝛼2 ⎜u ⎟ 3⎟⎠ ⎝ 4⎠ (22.17)

where 𝛼1 and 𝛼2 is the predicted value of h1 and h2 when the covariate X is 0, 𝛾1 and 𝛾2 are regression coefficients of the impact of X on h1 and h2 , and d1 and d2 are latent variable disturbances representing unmodeled variability in h1 and h2 . Additionally, time-varying covariates are modeled as having an instantaneous impact on the observed score. Such a model can be expressed as ⎛Y1 ⎞ ⎛1 ⎜ ⎟ ⎜ ⎜Y2 ⎟ = ⎜1 ⎜Y3 ⎟ ⎜1 ⎜Y ⎟ ⎜1 ⎝ 4⎠ ⎝

⎛𝛾1 0⎞ ⎟( ) ⎜ 0 1⎟ h1 +⎜ ⎜0 2⎟ h2 ⎜0 3⎟⎠ ⎝

0 𝛾2 0 0

0 0 𝛾3 0

0 ⎞ ⎛X1 ⎞ ⎛u1 ⎞ ⎟⎜ ⎟ ⎜ ⎟ 0 ⎟ ⎜X2 ⎟ ⎜u2 ⎟ + 0 ⎟ ⎜X3 ⎟ ⎜u3 ⎟ 𝛾4 ⎟⎠ ⎜⎝X4 ⎟⎠ ⎜⎝u4 ⎟⎠

(22.18) where X1 through X4 is a time-varying covariate—a single variable that varies across time—and 𝛾1 through 𝛾4 are the regression coefficients for the time-varying covariate describing the impact the time-varying covariate has on the outcome variable. Often these effects are constrained to be equal across time (e.g., 𝛾1 = 𝛾2 = 𝛾3 = 𝛾4 ) to mimic the way the model is typically fit within the multilevel modeling framework; however this constraint is not required. Multiple Group Growth Model The latent growth model of equation 22.1 can be extended to accommodate multiple observed groups of individuals. In this approach, the data are split based upon the grouping variable and separate growth models are specified for each group. Typically, a series of models are evaluated, which test the invariance of specific collections of parameters, such as the latent variable means, the latent variable covariances, and the structure of change. Extending equation 22.1, we can write the multiple group growth model as (g)

(g)

(g)

Yi = Λ(g) hi + ui

(22.19)

(g)

where Yi is a T × 1 vector of the repeatedly measured observed scores for individual i in group g, Λ(g) is a T × R (g) factor loading matrix for group g, hi is an R × 1 vector of (g) latent factor scores for individual i in group g, and ui is a T × 1 vector of unique scores for individual i in group g. The individual factor scores are written as deviations from the group-level means, expressed as (g)

(g)

hi = 𝛂(g) + di

(22.20)

where 𝛂(g) is an R × 1 vector of latent variable means for (g) group g, and di is an R x 1 vector of mean deviations

Page 1002

Trim Size: 8.5in x 11in

Cicchetti

c22.tex printer - Volume I - 12/10/2015

Multilevel Modeling for Studying Developmental Trajectories

for individual i in group g. In the multiple group growth model, the covariance matrix of mean deviations is 𝛙(g) , a group-specific R x R covariance matrix indicates the magnitude of between-person differences in the latent growth factors within each group as well as the degree of linear association among the latent growth factors within each group. Finally, the covariance matrix for unique scores, 𝚯(g) , is also group specific. This specification of the multiple group model is equivalent to separately specifying a growth model for each of the g = 1 to G groups. In practice, a series of models are fit to evaluate where group differences manifest. There are a variety of approaches to systematically test for group differences, and here we follow the approach advanced by Ram and Grimm (2009). In this approach, the first model is the invariant model where all parameter estimates are constrained to be equal over group (i.e., Λ(g) = Λ, 𝛂(g) = 𝛂, 𝛙(g) = 𝛙, and 𝚯(g) = 𝚯). This model is equivalent to the latent growth model without accounting for any grouping variable. The second model is the Means Model where the latent variable means are the only parameter estimates allowed to vary over group (i.e., Λ(g) = Λ, 𝛂(g) , 𝛙(g) = 𝛙, and 𝚯(g) = 𝚯). Next, is the Means+Variances Model where the latent variable means, the latent variable variances, and the residual variances are allowed to vary over group (i.e., Λ(g) = Λ, 𝛂(g) , 𝛙(g) and 𝚯(g) ). The fourth and final model is the configural invariance model (i.e., curves model), where the factor loading matrix is also allowed to vary over group (i.e., Λ(g) , 𝛂(g) , 𝛙(g) and 𝚯(g) ). In this model, all parameter estimates are separately estimated for each group. Nested chi-square difference tests can be used to compare each model with the more restricted models to determine whether and where the groups differ in estimated parameters. Latent Growth Modeling Studies for Understanding Developmental Pathways in Terms of Continuity–Discontinuity and Equifinality–Multifinality Understanding Developmental Pathways Using Latent Growth Modeling In this section, we illustrate several studies that demonstrate how growth curve models can be used to inform substantive questions in developmental psychopathology research focusing on the core principles of developmental psychopathology. A latent growth modeling study by Kim and Cicchetti (2006) demonstrates how growth curve models can be used to describe differential pathways to psychopathology. This study used latent growth modeling to

12:38pm

1003

investigate longitudinal relationships between self-system processes and depressive symptoms among maltreated and nonmaltreated children. One of the main goals was to describe developmental trajectories in self-system processes and depressive symptoms during the elementary school years among maltreated and nonmaltreated children. From a methodological viewpoint, an important advance of this study was to implement a latent growth curve model using SEM to examine interindividual differences in intraindividual variability (McArdle & Bell, 2000) in both initial levels and rates of change in children’s self-system processes and depressive symptoms. The participants included 251 children (142 maltreated and 109 nonmaltreated) who attended a summer day camp research program. Children ranged in age from 6 to 11 years (M = 8.46, SD = 1.11) at Wave 1 and 64% of the children were boys. The four-wave longitudinal data involved 7- to 12-year-olds. To investigate developmental trajectories of self-system processes and depressive symptoms, we conducted growth curve analyses with a maximum likelihood estimation method, which allows for inclusion of respondents with missing data by using full information maximum likelihood (FIML) estimation (Arbuckle, 1996). In evaluating the overall goodness-of-fit of each model, the root mean square error of approximation (RMSEA; Browne & Cudeck, 1993) index assesses the degree of lack of fit for a model and values less than .05 and .08 are taken to reflect a close fit and a reasonable fit, respectively. The comparative fit index (CFI; Bentler, 1990) varies along a 0–1 continuum in which values greater than .90 and .95 are considered to reflect acceptable and excellent fits to the data, respectively. We first fit a two-factor growth model over the four time points. The first latent factor was the intercept (or level) representing the initial starting point of the growth function, and the four factor loadings for the latent intercept factor were fixed to one. The second latent factor was the slope of the growth function, representing the rate of change in the growth trajectory over time. The latent intercept and slope factors were freely correlated. To determine the shape of trajectories of study variables for the entire sample, three alternative models were fitted separately for depressive symptoms (measured by the Children’s Depression Inventory, CDI; Kovacs, 1985) and self-esteem (measured by the Self Esteem Inventory, SEI; Coopersmith, 1981). The first model was a no-growth model that assumed no slope component (i.e., thus assumed no growth or change). The second model was a linear growth model that assumed a linear pattern of change over time and fixed values of slope parameters as

Page 1003

Trim Size: 8.5in x 11in

1004

Cicchetti

c22.tex printer - Volume I - 12/10/2015

12:38pm

Latent Growth Modeling and Developmental Psychopathology

[B(T1) = 0, B(T2) = 1, B(T3) = 2, B(T4) = 3]. The third model was a latent basis growth model that estimated slope parameters from the data to yield an overall group shape. The slope parameter for the first occasion was fixed to zero [B(T1) = 0] to allow a separation of the intercept and slope components, and the slope parameter for the last occasion was fixed to one [B(T4) = 1] to provide a scale of measurement for the slopes. The slope parameters for the second and the third occasions were freely estimated from the data. Model comparison was facilitated by positing a nested ordering of models in which the parameter estimates for a more restrictive model were a proper subset of those in a more general model. As can be seen in Table 22.1, the chi-square difference test indicated that a linear growth model provided the best fit to the data for the SEI. For the SEI, significant variance of both intercept (𝜎 2 = .01, SE = .002, p < .05) and slope (𝜎 2 = .001, SE = .000, p < .05) indicated the presence of significant individual differences in initial levels and change in children’s self-esteem. Both intercept mean (M = .68, SE = .01, p < .05) and slope mean (M = .03, SE = .004, p < .05) were significantly different from zero, showing that children’s self-esteem increased over time. In contrast, a latent basis growth model was the best-fitting model for the CDI. The coefficients of the slope factor loadings represented the percentage of growth relative to the total change occurring over all time points. The pattern of CDI slope factor loadings (i.e., .00, .59, .89, 1.00) reflected a nonlinear change in the CDI scores. Significant variance existed in both intercept (𝜎 2 = 38.45, SE = 9.42, p < .05) and slope (𝜎 2 = 25.55, SE = 10.93, p < .05), reflecting meaningful individual variability in average level and change in depressive symptoms over time. The intercept mean was significantly different from zero (M = 9.47, SE = .48, p < .05), and the slope mean was negative and significantly different from

zero (M = −3.55, SE = .51, p < .05), suggesting that children’s depressive symptoms decreased over time. The intercept and the slope factors were negatively correlated (r = −.74, p < .05), suggesting that children who reported higher initial levels of depressive symptoms tended to show a more rapid decrease in depressive symptoms over time compared to children who reported lower initial levels of depressive symptoms. Therefore, the data indicated a significant decrease in depressive symptoms and linear increases in self-esteem over the four time points. There are no particular theories to delineate particular developmental trajectory patterns of depressive symptoms and self-systems among maltreated and nonmaltreated school-age children. However, this finding was consistent with some previous research that reported decreases in self-reported internalizing problems and increases in perceptions of internal control in similar samples (e.g., Bolger & Patterson, 2001). In estimating developmental trajectories using latent growth modeling, the main goal is to estimate unobserved or latent trajectory as a function of the set of repeated measures that were observed to describe how the developmental process unfolds within each person—that is, intraindividual change. As shown in this example, the growth curve model can further inform the interindividual changes (the means and variances of the individual trajectory parameters across individuals) based on intraindividual trajectories. Within latent growth modeling, each individual is characterized by his or her own unique trajectory that best represents their observed data over multiple time points. The advantage of latent growth modeling is clear especially when we compare latent growth modeling to more traditional models (such as autoregressive regression models) in which a single parameter estimate represents the relation between two repeated measures pooled over all subjects over time (Curran & Willoughby, 2003).

TABLE 22.1 Comparisons of Fitted Growth Curve Models for Self-Esteem and Depressive Symptoms of Maltreated and Nonmaltreated Children1 Model Label2

𝜒2

df

p(exact)

CFI

RMSEA

p(close)

Δ𝜒 2

Δdf

p(d)

Self-Esteem (SEI) 1-a. No-growth model 1-b. Linear growth model 1-c. Latent growth model

70.70 3.52 .76

8 5 3

.00 .62 .86

.78 1.00 1.00

.18 .00 .00

.00 .85 .94

67.18 2.76

3 2

0) are carrying copies of the actor’s own genes.2 For this reason, inclusive fitness theory has sometimes been presented as a theory of the “selfish gene” (Dawkins, 1976)—a potentially misleading label, given its original focus on the evolution of altruism. Inclusive fitness theory is also equivalent to multilevel selection theory, an approach to social evolution that focuses on group rather than individual dynamics (e.g., Sober & Wilson, 1998). While multilevel selection has often been viewed as alternative to inclusive fitness, it has since become clear that the two theories are mathematically interchangeable (see Marshall, 2011; West et al., 2007), and differ only in how they partition the costs and benefits of social traits. Whereas inclusive fitness partitions fitness 1 More precisely, relatedness is a regression coefficient that predicts the recipient’s genotype from the actor’s genotype. Relatedness can become negative if two individuals can be expected to be genetically less similar than two randomly selected members of the population (see Grafen, 2009; West et al., 2007). 2 The existence of epigenetic inheritance does not fundamentally change this picture. If epigenetic markings are reliably transmitted across generations, they are equivalent to genetic alleles from the standpoint of natural selection. If epigenetic markings are reversible and environmentally induced, they mediate short-term developmental plasticity and are irrelevant to inclusive fitness computations (see Shea, Pen, & Uller, 2011).

Toward an Evolutionary-Developmental Framework for Psychopathology

effects between actors and recipients, multilevel selection partitions them between individuals and their broader social groups. Thus, in a multilevel framework, altruism toward group members (to the point of self-sacrifice or sterility) can be selected for if it is counterbalanced by an appropriate benefit to the group as a whole. The equivalent explanation in terms of inclusive fitness is that group formation mechanisms typically increase relatedness within groups relative to that between groups. Thus, helping group members leads to an indirect fitness benefit that can be so strong as to override large individual costs. Special Metatheoretical Assumptions Psychological Mechanisms. Psychological adaptations, which govern mental and behavioral processes, are referred to by evolutionary psychologists as psychological mechanisms. Most research in evolutionary psychology focuses on identifying evolved psychological mechanisms because it is at this level where invariances occur. Indeed, evolutionary psychologists assert that there is a core set of universal psychological mechanisms that comprise our shared human nature (see Buss, 2005). The move to the level of psychological mechanisms is important to avoid a common fallacy—that of assuming that human behavior (1) has the conscious goal of maximizing inclusive fitness, and (2) actually maximizes inclusive fitness in current environments. At a very general level, natural selection does tend to produce organisms that behave as if they were trying to maximize their expected fitness (see above). However, actual behavior is ultimately mediated by a host of psychological mechanisms with local and sometimes conflicting goals (e.g., learning a language, finding and attracting mates, choosing food, avoiding diseases). There is no general “fitness maximization mechanism” anywhere in the brain. Each mechanism works and evolves within constraints (e.g., information availability, time constraints, coordination and conflict with other mechanisms, previous evolutionary history); as a result, the overall structure of the mind–brain is more akin to a gerrymandered contraption than an optimal, omniscient decision maker. Even more importantly, the fact that a given adaptation was produced through differential reproduction does not imply that either (1) selection is currently favoring that adaptation or (2) variation in the expression of that adaptation will be associated with current reproductive success. For example, the dopamine-mediated reward mechanisms found in the mesolimbic system in the brain appear to have evolved to provide a pleasurable reward in the presence of adaptively relevant stimuli such as food or sex. In contemporary environments, however, these same mechanisms are

7

subverted by the use of psychoactive drugs such as cocaine and amphetamines, which deliver huge dollops of pleasurable reward in the absence of the adaptively relevant stimuli, often to the user’s detriment (Durrant & Ellis, 2003). The concept of a psychological mechanism was updated by Bjorklund and colleagues (2007) to make it more consistent with EDP’s metatheoretical assumption of probabilistic epigenesis. These authors proposed a definition of evolved probabilistic cognitive mechanisms (p. 22): [Psychological] mechanisms that are functionally organized to solve recurrent problems faced by ancestral populations, are highly probable when species-typical environments are encountered (i.e., when the developmentally relevant features of the environment are in the range typically encountered during a species’ evolution), and are products of emerging developmental systems that have evolved over the course of the ontogenies of our ancestors.

This definition stresses the probabilistic nature of the ontogenetic processes responsible for building psychological mechanisms; it also makes it clear that, while evolved mechanisms prepare an organism for life in a species-typical environment, they are not preformed or specified in advance by a rigid genetic program. Domain Specificity. As is apparent from the preceding paragraphs, evolutionary psychology views psychological mechanisms as having some degree of functional specialization. More specifically, psychological mechanisms are composed of structures that (1) exist in the form they do because they recurrently solved specific problems of survival and reproduction over evolutionary history; (2) are designed to take only certain kinds of information from the world as input; (3) process that information according to a specific set of rules and procedures; (4) generate output in terms of information to other psychological mechanisms and physiological activity or manifest behavior that is directed at solving specific adaptive problems (see Buss, 2012). In short, psychological mechanisms are designed by selection to address specific domains of the physical and social world. Although evolutionary psychologists assert that the mind is not comprised primarily of content-free (domain-general) psychological mechanisms, it is likely that different mechanisms differ in their levels of specificity, and that there are some higher level mechanisms that function to integrate information across more specific lower level mechanisms. In addition, some general-purpose abilities (e.g., associative learning) may be co-opted in the context of different specialized functions. It is important to stress that functional specialization of a psychological

8

Evolutionary Foundations of Developmental Psychopathology

mechanism does not imply clear-cut anatomical localization in the brain, nor complete functional independence from other mechanisms. Indeed, psychological mechanisms are expected to show a considerable degree of integration and reciprocal interaction. The rationale behind the domain specificity argument is fairly straightforward: What counts as adaptive behavior differs markedly from domain to domain. The sort of adaptive problems posed by food choice, mate choice, and social exchange often require different kinds of solutions. A clear analogy can be drawn with the functional division of labor in human physiology. Different organs have evolved to serve different functions and possess properties that allow them to fulfill those functions efficiently, reliably, and economically: the heart pumps blood, the kidneys excrete urine, and so on. A super, all-purpose, domain-general internal organ faces the impossible task of serving multiple, incompatible functions. Analogously, a super, all-purpose, domain-general mind/brain mechanism faces the impossible task of efficiently and reliably solving the plethora of behavioral problems encountered by humans in ancestral environments. Thus, neither an all-purpose physiological organ nor an all-purpose psychological mechanism is likely to evolve. Environment of Evolutionary Adaptedness. Biological adaptation is necessarily a historical concept, and all claims about adaptation are claims about the past. The environment in which a given trait evolved is termed its environment of evolutionary adaptedness (EEA). When we claim that the thick insulating coat of the polar bear is as an adaptation, we are claiming that possession of that trait advanced reproductive success in ancestral environments. However, traits that served adaptive functions and thus were selected for in past environments may not still be adaptive in present or future environments. In a globally warmed world, for example, the polar bear’s pelt may become a handicap that reduces the fitness of its owner. While natural selection is expected—all else being equal—to weed out traits that have become detrimental to fitness, the process may often take a long time. This generates the potential for mismatch between an organism’s adaptations and its present environment. The possibility of mismatch raises a subtle but crucial point regarding the meaning of adaptive. Broadly speaking, psychological and physiological processes can be described as adaptive if they result from the unimpaired functioning of adaptations. Thus, adaptive in the broad sense is a shorthand to describe the functioning of naturally selected processes and mechanisms, regardless of whether they are

currently promoting reproductive success (i.e., adaptive in the narrow sense). For example, pursuit of mating relationships with fertile partners is guided by adaptive psychological processes, regardless of whether contraceptive technology prevents reproduction in present-day societies. Within the same organism, different adaptations will often have different EEAs (for extended discussion see Durrant & Ellis, 2003). Consider the human adaptations of language and infant attachment. While the origin of language is firmly anchored in approximately the last 2 million years, infant attachment reflects a much lengthier evolutionary history and a shared heritage with other mammalian and primate species. While evolutionary timing helps define the EEA of a trait, the EEA itself is not a specific time or place; rather, EEAs capture the statistical regularities of the environment in which the trait evolved (Tooby & Cosmides, 1990). Environmental variation itself can be part of an EEA; for example, metabolic processes can evolve so as to maximize survival in an unpredictable environment, whereby food abundance is suddenly followed by starvation. In this case, metabolic adaptations evolve in an EEA characterized by a consistent pattern of unpredictable variation. Over the last few millennia—the span of a few hundred generations—humans have experienced rapid and constantly accelerating rate of change in health, nutritional, social, and technological conditions. While genetic evolution has been accelerating as well (e.g., Hawks, Wang, Cochran, Harpending, & Moyzis, 2007), many of our evolved adaptations can be expected to be at least partly mismatched to modern lifestyles. At the same time, many adaptation-relevant aspects of our environment have probably remained the same: humans everywhere, for example, still find and attract mates, have sex, raise families, make friends, compete for status, and gossip (Crawford, 1998). Most important is that current and ancestral environments do not have to be identical in every respect to sustain the normal development and expression of evolved psychological mechanisms. Probabilistic Epigenesis. The concept of probabilistic epigenesis has a long history in embryology and is one of the central assumptions of DST (see Gottlieb, 2007). Probabilistic epigenesis holds that development involves continuous bidirectional influences between genetic activity, neural activity, behavior, and the physical and social environment (similar interactions take place in the development of non-neural mechanisms). In this view, neural structures begin to function when they are still developing, and their activity—both spontaneous and evoked by the

Toward an Evolutionary-Developmental Framework for Psychopathology

environment—plays an important role in the ontogenetic process. This is contrasted with predetermined models in which genetic programs build neural structures, that begin to function and interact with the environment only when they are mature. The reciprocal, bidirectional interaction between multiple levels introduces a probabilistic element in the outcomes of developmental processes. A key implication of probabilistic epigenesis is that genetic activity is influenced and regulated by neural, behavioral, and external events. Gene–environment (GxE) interactions in development, whereby the effects of an allelic variant are contingent on contextual variables, are prime examples of probabilistic epigenesis (Gottlieb, 2007). Probabilistic epigenesis provides reasons for expecting widespread plasticity in the outcomes of developmental processes; however, it is not sufficient to explain adaptive plasticity and phenotype-environment matching (Bjorklund et al., 2007). Understanding adaptive plasticity requires a synthesis between the proximate and ultimate level of analysis—where development meets adaptation. Developmental Systems Theory: An Alternative Metatheory? Researchers in developmental psychopathology often refer to developmental systems theory (DST) as a metatheoretical framework for the discipline. DST is a general approach to development and evolution rooted in the organismic concepts of embryology and developmental psychobiology. The major themes of DST are probabilistic epigenesis and developmental plasticity, with a strong emphasis on bidirectional interplay between genes and environment; an extended view of inheritance that goes beyond DNA to include epigenetic processes, cellular structures, scaffolded developmental environments (e.g., nests), and culturally transmitted information; and a view of the developing organism as actively involved in shaping its environment (see Griffiths & Gray, 2004; Oyama et al., 2001). Consistent with the metatheoretical framework we have presented, DST emphasizes the multiplicity of factors that jointly determine phenotypic outcomes and stresses the contextual, contingent nature of development. Soft Versus Hard DST Much of the difficulty in discussing the role of DST stems from the fact that DST is not a single, unified theory; in fact, it is possible to recognize at least two versions of DST—a “soft” version and a “hard” version—with vastly different implications for developmental science (Frankenhuis et al., 2013; Robert et al., 2001). Soft DST is

9

essentially a theory of development; in this view, a developmental system comprises all the resources (e.g., genes, cellular structures, sensory experiences, physical parameters of the environment) that contribute to the ontogeny of the individual organism. However, the organism remains the main biological entity, and evolutionary processes acts on populations of organisms. In other words, soft DST reconceptualizes the causal structure of development—for example by placing genetic inheritance in a broader perspective and emphasizing bidirectional effects—but is otherwise consistent with inclusive fitness theory and the logic of individual adaptation (Pradeu, 2010). Indeed, many developmentally oriented extensions of evolutionary biology already incorporate the main tenets of soft DST (e.g., West-Eberhard, 2003). In contrast, hard DST is not so much a theory of development as a radical alternative to mainstream evolutionary theory. In hard DST, a developmental system comprises all the resources that produce the developmental outcomes that are stably replicated in that lineage. As a consequence, it is impossible to meaningfully distinguish between organism and environment, and what evolves are not populations of organisms but populations of replicating organism–environment systems. Such holistic reconceptualization of natural selection breaks down the individual-as-maximizing agent analogy and makes adaptationist analysis all but impossible (Pradeu, 2010). This is because hard DST is inconsistent with inclusive fitness theory: selection is no longer assumed to act on individuals that can be more or less genetically related with one another, but on whole developmental systems (comprising every recurring influence on development, including social and biogeographical factors) for which there is no meaningful definition of reproductive success or relatedness. In addition, hard DST not only objects to the concept of genetic programs, but—in a further break from mainstream biology—also rejects the very idea that genes store information as unacceptable preformationism (Oyama et al., 2001). In summary, DST comprises two related but partially distinct approaches. Soft DST is a developmentally oriented extension of mainstream evolutionary theory, and is fully consistent with the metatheoretical framework of EDP. In contrast, hard DST advances a radically novel theory of evolution, and constitutes an alternative metatheoretical framework with little overlap with that of EDP. Embracing soft DST does not commit one to also adopt the assumptions of hard DST. Unfortunately, the distinction between the soft and hard version of DST is often obscured in the literature, leaving many researchers

10

Evolutionary Foundations of Developmental Psychopathology

confused as to the exact implications of the theory (see Frankenhuis et al., 2013; Pradeu, 2010). Implications for Developmental Science The distinction between soft and hard DST provides insight in the current theoretical status of developmental psychology and psychopathology. We surmise that, when developmental scientists embrace a DST perspective, they usually reason in terms of soft DST. However—and possibly without realizing it—they end up adopting the whole metatheoretical package of hard DST, with the added baggage of antiadaptationism and a priori rejection of mainstream evolutionary thinking. As a result, developmental science is deprived of some of the most powerful tools in biology, such as inclusive fitness theory and the concept of adaptation. By contrast, we contend that the metatheoretical framework of EDP—a synthesis of adaptationism and soft DST—provides a suitable evolutionary foundation for developmental psychopathology. In the remainder of the chapter we demonstrate the heuristic and integrative power of this approach.

BEYOND PATHOLOGY: ADAPTATION, MALADAPTATION, AND DISORDERS In an evolutionary framework, the terms adaptive and maladaptive denote the effect of a trait or behavior on biological fitness. From the standpoint of the individual organism, adaptive traits are those that enhance inclusive fitness compared with potential alternatives. However, all adaptations have fitness costs as well as benefits; to be adaptive a trait does not have to be cost free but it only needs to yield a positive overall contribution to the organism’s fitness. This notion of adaptation and maladaptation contrasts sharply with how the same terms are usually employed in developmental psychology and psychopathology. In these disciplines, adaptive traits and behaviors are those that promote health, safety, subjective well-being, and mutually rewarding social relations. Socially undesirable, aversive, or health-damaging traits are viewed as maladaptive. These definitions of adaptation and maladaptation are conceptually orthogonal and ought to be carefully differentiated. In this chapter we always refer to adaptation and maladaptation in the biological sense, and employ the terms desirable and undesirable to denote the implications of a trait for health, safety, well-being, and social values. Unsettling as it may be, the logic of natural selection promotes reproductive success, not happiness or health

(see Cosmides & Tooby, 1999; Gluckman et al., 2011; Nesse, 2004a). Thus, biologically adaptive traits may or may not be socially desirable or conducive to health and well-being; conversely, a trait is not maladaptive just because it has negative effects on an individual’s welfare. Traits that consistently reduce well-being and adversely impact an individual’s health can be selected for as long as they result in enhanced reproduction—a highly counterintuitive notion in mainstream psychology. At the same time, adaptiveness and desirability—though conceptually distinct—are functionally connected to some degree. This is because positive emotions such as joy, excitement, and pride and are generally aroused by the fulfillment of fitness-enhancing goals, while threats to fitness are generally met with negative feelings such as sadness, anger, and shame (Nesse, 2004a). The functional connection between threats to fitness and negative emotions lends intuitive plausibility to the implicit assumption—firmly entrenched in psychopathology and psychiatry—that aversive traits are by default pathological or “maladaptive” (see Nesse & Jackson, 2006). The evolutionary approach challenges this assumption, and unpacks the intuitive concept of disorder by separating adaptation from health and desirability. The result is a general framework for thinking about pathology that can be applied to both medical and psychopathological conditions. What Is a Disorder? Mental disorder is a central concept of psychopathology, yet a satisfactory definition of disorder is notoriously difficult to achieve. In an influential paper, Jerome Wakefield (1992) built on previous biologically informed approaches to advance a definition of disorder as a harmful dysfunction. According to this definition, conditions are recognized as disorders when they (1) are caused by the failure of a biological mechanism to perform its evolved function, and (2) inflict some harm or damage on the affected person, as judged by sociocultural standards. This is a hybrid definition that combines the objective dysfunction criterion with the subjective, culturally bound harm criterion. In Wakefield’s account, people evaluate a condition as a disorder when the subjective perception of harm or undesirability is coupled with the idea that something in the body or mind is not working properly. In line with an evolutionary approach, what constitutes proper functioning of a biological mechanism can be correctly evaluated only by considering the evolved functions of that mechanism. To understand pathology, one needs to understand the function of the relevant biological mechanisms, as well as

Beyond Pathology: Adaptation, Maladaptation, and Disorders

the structure of the environment in which they evolved (Nesse, 2001; Troisi & McGuire, 2002). We believe that, correctly understood, the harmful dysfunction analysis is a useful heuristic for reasoning about pathology and disorders. Although Wakefield’s proposal has been hotly debated, its core propositions have withstood criticism (see Wakefield, 1999, 2011). To avoid common misgivings, the following points should be kept in mind. First, evolved mechanisms are defined broadly; major organs like the heart are mechanisms, but so are specialized brain areas, microscopic cellular structures, and biochemical pathways. Accordingly, dysfunctions can occur in many ways and for a wide variety of causes (e.g., deleterious genetic mutations, pathogen infections, injuries and wounds, side effects of evolved defenses). Second, dysfunctional is not synonymous with maladaptive. Since dysfunctions interfere with evolved design, they can often be expected to reduce an individual’s fitness; however, a reduction in fitness is not required to identify a dysfunction. It is quite possible for a dysfunction to be selectively neutral—for example, because it occurs too late in life to impact an individual’s reproductive success, or because changes in the environment reduce its damaging effects. This is why myopia—a failure of the crystalline lens to project a focused image on the retina—remains a dysfunction even if glasses and contact lenses eliminate its negative effects on survival. Third, the concept of dysfunction is a fuzzy one rather than all-or-none, and evolved mechanisms can show varying degrees of functionality (Wakefield, 1999). Thus, obvious instances of dysfunction are going to be surrounded by borderline cases for which there is no clear-cut demarcating criterion—as for example in the case of hypertension, extreme variation in height, and personality disorders. A Taxonomy of Undesirable Conditions Despite their theoretical significance, harmful dysfunctions are only a fraction of what people regard as diagnosable problems or seek treatment for (Cosmides & Tooby, 1999; Lilienfeld & Marino, 1999). Fever is an evolved defense against pathogens; with rare exceptions, it reflects a well-functioning system rather than a dysfunction—yet it is often treated with drugs. More generally, conditions that are not harmful dysfunctions in Wakefield’s sense may nevertheless be labeled and treated as disorders, especially if their etiology and functional implications are incompletely understood. For example, it has been hypothesized that some forms of psychopathy are adaptive behavioral phenotypes that exist at a low frequency and thrive by

11

exploiting others (e.g., Mealey, 1995). If this hypothesis were correct, a number of apparent dysfunctions (e.g., reduced empathy, lack of guilt, impulsivity) would be better understood as design features of the psychopathic strategy. Still, psychopathy is a source of trouble for society at large, and would be legitimately regarded as a condition in need of treatment even if it were established as a biologically adaptive variant rather than a disorder in the strict sense. This example also illustrates how conflicts of interest between social actors modulate the perception and definition of a problematic condition; obsessional jealousy may appear desirable and useful to the affected individual, but harmful and undesirable to his/her partners (for a detailed analysis of this issue see Cosmides & Tooby, 1999). We now take a wider perspective and consider the many ways evolutionary and developmental processes may result in undesirable conditions, including—but not limited to—harmful dysfunctions in the narrow sense (see Cosmides & Tooby, 1999; Gluckman et al., 2011). The taxonomy we present combines Wakefield’s dysfunction criterion with the effects of a given condition on biological fitness (Figure 1.1). When considering the adaptiveness of a condition, we further distinguish between the fitness contribution of a trait or mechanism—averaged across all the individuals who express it—and the fitness of a particular individual. The distinction is useful because a mechanism may be fitness-enhancing on average, while imposing fitness costs on some individuals (e.g., Cosmides & Tooby, 1999). Distinguishing between individual and average fitness permits a fine-grained analysis of the interplay between adaptation and maladaptation in psychopathology (Frankenhuis & Del Giudice, 2012). Although we discuss them separately, the following categories are not mutually exclusive; a given condition or class of conditions may well reflect the interplay of multiple factors and require overlapping evolutionary explanations. Harmful Dysfunctions All biological and artificial mechanisms—no matter how well designed—are vulnerable to malfunctions, failures, and breakdowns. Developmental pathways typically evolve canalization properties (e.g., biochemical buffering mechanisms) that confer them robustness against accidents and perturbations. However, the accumulation of such events over time can affect development, resulting in random deviations from the target phenotype (developmental instability). More dramatically, an evolved mechanism may cease to perform its functions because of accidents or environmental insults beyond its regulatory

12

Evolutionary Foundations of Developmental Psychopathology

Figure 1.1 An evolutionary taxonomy of undesirable conditions.

capacity (e.g., brain injury, exposure to toxins), deleterious genetic/epigenetic changes (e.g., mutations and deletions), and attacks or manipulations by pathogens (see Crespi, 2000, 2010). New deleterious mutations arise at every generation and they may be passed down to descendants, persisting for some time in a population until they are weeded out by natural selection. Harmful variants may be especially difficult to eliminate if they have recessive effects (i.e., they are expressed only when an individual inherits two copies of the same allele). The continuous process of creation and elimination of deleterious mutations is called mutation–selection balance; its dynamics determine the frequency and persistence of harmful variants in a population. Sometimes, a single mutation in a critical pathway is sufficient to cause a disorder; more often, disorders may result from the cumulative effect of many slightly deleterious mutations (mutation load), each with a small impact on phenotypic function. Mutation–selection balance has been proposed as a likely explanation for the persistence of common, heritable, and harmful mental disorders like autism, schizophrenia, bipolar disorder, and mental retardation (Keller & Miller, 2006). Since a large proportion of human genes are expressed in brain development, the likelihood that mutation load will have negative consequences on mental functioning is especially high. The role of mutation load in autism and schizophrenia is consistent with the

high rate of new mutations in people with these disorders (e.g., Sanders et al., 2012). Exposure to pathogens (harmful viruses, bacteria, and other parasites) is another common cause of biological dysfunction. Infectious diseases—especially when they occur in early development—have been associated with increased risk for a number of mental disorders including autism, schizophrenia, and depression (see Patterson, 2011; Benros, Mortensen, & Eaton, 2012). The role of pathogens in the etiology of mental disorders does not contradict that of genetic mutations. Infections, like mutations, can perturb developmental processes at critical stages; accordingly, mutation load and pathogen load may ultimately converge on the same neurobiological pathways and exert a cumulative effect on the risk for psychopathology. In addition to their direct effects on individual organisms, pathogens may indirectly contribute to the risk of harmful dysfunctions through their effect on the evolution of defenses. Pathogens and hosts are constantly involved in coevolutionary arms races, so that for every improvement in defensive mechanisms, new means of offense are going to be selected for on the other side (and vice versa). Coevolutionary arms races tend to produce increasingly complex offense/defense mechanisms (consider the intricacy of the immune system); in turn, increased complexity may render those mechanisms more vulnerable to failures and dysfunctions (Nesse, 2001).

Beyond Pathology: Adaptation, Maladaptation, and Disorders

Evolutionary Mismatches Because of natural selection, evolving organisms tend to become progressively more successful at surviving and reproducing in their environment, broadly conceived to include not only physical factors but also social relations with conspecifics as well as interactions with other species (predators, prey, pathogens, and so forth). The environment, however, is not a static background: environments change all the time because of external events (e.g., geological change), social evolution within a species (e.g., increased population density), and coevolutionary processes between species (e.g., new pathogens). When environmental changes are rapid and extensive, previously adaptive mechanisms may suddenly become maladaptive and generate all sorts of unintended and/or undesirable consequences. Thanks to cultural and technological evolution, humans have gained an unprecedented power to alter their social and physical environment, and in so doing have created enormous opportunity for evolutionary mismatch. Evolutionary mismatch occurs when an organism encounters a novel environmental context (outside of the range that was recurrently encountered over its evolutionary history) that disrupts normal development or impairs adult functioning. Evolutionary mismatches are likely to be implicated—to various degrees—in the etiology of mental disorders. In modern societies, for example, the media expose girls and women to a relentless stream of images of unrealistically attractive “competitors”—an artificial, evolutionarily novel kind of social stimulus. It has been hypothesized that such exposure hyperactivates the evolved mechanisms that regulate female competition for attractiveness and status, contributing to the rising incidence of eating disorders (e.g., Abed, 1998). Other instances of potential mismatch are less obvious. For example, sanitation in developed countries determines a lack of exposure to common microorganisms (“old friends”) during development. These novel hygienic conditions appear to interfere with the early ontogenetic processes that train the immune system and set its overall functioning parameters. The resulting states of chronic inflammation may increase the risk for a range of physical and mental disorders, especially depression (Raison, Lowry & Rook, 2010). Although we have emphasized its negative consequences, evolutionary mismatch is an unavoidable and often vital aspect of evolution. By definition, all evolving organisms exhibit some degree of mismatch to their present environment—otherwise they would stop evolving altogether. The very process of adaptation generates subtle forms of mismatch that may contribute to the etiology of

13

undesirable conditions. When a trait has been subjected to strong recent selection, the resulting adaptive changes may co-occur with maladaptive side effects on other traits that are genetically and/or developmentally linked to the selected trait. Similarly, recently evolved adaptations are likely to show increased scope for dysregulation because they have yet to be fine-tuned by natural selection (see Crespi, 2010). Maladaptive Outcomes of Adaptive Mechanisms So far, we have reviewed case in which undesirable conditions are caused by failures of evolved design. Harmful dysfunctions occur when a biological mechanism fails to perform its evolved functions; conversely, evolutionary mismatches occur when an intact mechanism becomes maladaptive because of novel environmental conditions. However, maladaptive outcomes at the individual level may systematically occur even when adaptive mechanisms perform their evolved functions in an environment that matches the EEA on the relevant dimensions. This is one of the central insights of evolutionary psychopathology: observing maladaptive outcomes at the individual level is not sufficient to infer maladaptation at the level of evolved mechanisms. We now review some important reasons why adaptive mechanisms may systematically yield maladaptive outcomes (for in-depth discussion see Cosmides & Tooby, 1999; Crespi, 2010; Frankenhuis & Del Giudice, 2012). Maladaptive Outcomes of High-Risk Strategies. An important source of maladaptation at the individual level is the evolution of risky adaptive strategies3 . Risky behavior is part and parcel of daily life: many activities that contribute to survival and reproduction also increase the probability of harm, injury, loss, or death. Searching for food and competing for mates are both fraught with danger, but potential dangers are compensated by the potential fitness advantages of these activities. From an evolutionary perspective, we would expect natural selection to favor mechanisms that produce risk taking when the fitness benefits outweigh the costs. Further insight in the dynamics of risky strategies can be gained by defining risk in its technical sense of unpredictable variation in 3

In an evolutionary framework, the term strategy denotes an organism’s realized phenotype among a set of possible phenotypes. Adoption of a given strategy can depend on both environmental and genetic factors. It should be stressed that the term does not imply conscious planning, deliberation, or even awareness; an organism’s choice between alternative strategies can be implemented by low-level physiological means, such as a hormonal switch or a change in genetic expression.

14

Evolutionary Foundations of Developmental Psychopathology

outcomes (see Frankenhuis & Del Giudice, 2012; Ellis et al., 2012). Whereas some behavioral decisions offer a narrow range of possible outcomes (low-risk), others entail widely variable outcomes (high-risk), with the potential for large gains as well as large losses. Consider a predator that can choose between two types of prey: larger and hard-to-catch animals versus smaller and easily caught ones. Imagine also that the expected energetic returns associated with hunting each prey type are identical: one results in a high reward with a low probability, the other in a low reward with a high probability. In this scenario, hunting larger prey qualifies as more risky, because it entails more variable outcomes. Broadly speaking, natural selection favors risk aversion when the relationship between behavioral outcomes and fitness is characterized by diminishing returns. For instance, a well-fed animal should look for low-risk food items (or not forage at all) when additional calories only slightly improve its condition. Conversely, when better outcomes yield increasing fitness returns, organisms may become risk-prone. For example, an animal on the brink of starvation may choose to forage in a nutrient-rich habitat, even if it is infested with predators, because it has so much to gain from additional calories (discussed in Frankenhuis & Del Giudice, 2012). An analogous logic applies to competition for reproduction. In some mating systems, reproductive benefits are highly skewed towards top-ranking individuals (i.e., winner-takes-all systems). In such conditions, males are intensely selected to compete for top rank, even if this implies a greater risk of injury; for instance, male elephant seals engage in ferocious fights that often cause harm and sometimes result in death. Still, males benefit—on average—from participating in fights because not participating implies being shut out from reproduction. The logic of risky strategies can shed light on the interplay between adaptation and maladaptation in development and psychopathology. For example, externalizing behavior can be interpreted as a high-risk tactic of social competition (Del Giudice et al., 2011; Ellis et al., 2012; Martel, 2013). In some cases, aggressive children and adolescents become dominant, respected, and popular leaders in their peer groups; in other cases, they do not succeed and become unpopular or rejected, incurring physical and psychological harm. These outcomes can be individually maladaptive even if they result from an adaptive strategy designed to achieve dominance and social status. A similar logic may apply to schizotypal personality traits (e.g., the tendency to experience unusual perceptions, bizarre ideation, and reference thoughts). While schizotypal traits

increase the risk of schizophrenia (a severe, harmful disorder), when schizotypal individuals do not develop a disorder, their enhanced creativity may facilitate high mating success (e.g., Nettle, 2001; reviewed in Del Giudice, Angeleri, Brizio, & Elena, 2010); this would amount to a high-risk strategy with widely variable outcomes. Maladaptive Outcomes of Evolutionary Conflict. Conflicts of interest between individuals abound in nature—for example between mating rivals, or between dominants and subordinates in a hierarchy. While cooperation and even altruism can be favored by natural selection, it is often the case that a given individual can maximize its own fitness only at the expense of another individual’s reproduction. An especially intriguing kind of evolutionary conflict is that between parents and their offspring (Trivers, 1974; see Schlomer, Del Giudice, & Ellis, 2011). While a parent is equally related to all its offspring (r = 0.5), each offspring is more related to itself (r = 1) than to any of its present or future siblings (r = 0.5 in the case of full siblings). Offspring are thus selected to demand more than their “fair share” of their parent’s investment in time, food, protection; conversely, parents are selected to resist such attempts, setting the stage for parent–offspring conflict over the distribution of parental investment. Of course, parents and offspring also have a lot of evolutionary interests in common, so conflict is tempered with a substantial amount of cooperation and altruism. Although parent–offspring conflict is not maladaptive in itself—parents and offspring are both acting so as to maximize their own fitness—the dynamics of conflict often result in nontrivial costs for both parties. Furthermore, conflict may occasionally escalate to dangerous levels, yielding maladaptive outcomes for the parent, the offspring, or both. A dramatic example is provided by prenatal conflicts about fetal nutrition (Haig, 1993; reviewed in Schlomer et al., 2011). During pregnancy, the placenta—a fetal organ that only expresses the fetal DNA—releases massive amounts of hormones in the maternal bloodstream. These hormones affect maternal metabolism so as raise the nutrient content of maternal blood and increase the supply of blood to fetal circulation. The interplay between fetal hormones and maternal countermeasures may produce a range of undesirable side effects, including gestational hypertension and gestational diabetes. In rare cases, the physiological tug-of-war between mother and fetus may become dysregulated and result in life-threatening conditions such as preeclampsia (severe maternal hypertension). In a recent paper, one of us (Del Giudice, 2012) speculated that a similar

Beyond Pathology: Adaptation, Maladaptation, and Disorders

conflict may arise about fetal exposure to maternal stress hormones, with mothers favoring higher levels of exposure than fetuses. Indeed, several puzzling features of stress regulation in pregnancy could be explained by the interplay between fetal attempts at manipulation and maternal countermeasures (for a detailed exposition see Del Giudice, 2012). Elevated prenatal stress has been associated with increased risk for a broad range of psychopathological outcomes in children—including anxiety, hyperactivity, autism, and schizophrenia (reviewed in Glover, 2011). To some extent, these undesirable outcomes may arise as maladaptive side effects of parent–offspring conflict in pregnancy. Evolutionary conflict usually takes place between different individuals, but this is not always the case. Indeed, conflicts of interest can also arise between different genes within the same individual (intragenomic conflict; for a thorough review see Burt & Trivers, 2006). Intragenomic conflicts may involve sexual chromosomes, mitochondrial genes, or “selfish” strands of DNA that—for various reasons—follow inheritance rules that differ from those of the rest of the genome. Most relevant to the present discussion, the maternally and paternally inherited halves of an individual’s genome may have divergent fitness interests when parental investment is involved. In species that are not perfectly monogamous—that is, most sexually reproducing species including humans—the occurrence of multiple paternity increases the chance that siblings in the same family have the same mother but different fathers. As a result, the genes inherited from the father are—on average—less strongly related to those of one’s siblings than the genes inherited from the mother. This generates complex dynamics in which maternal and paternal genes may favor opposite traits in the offspring (e.g., maternal genes may benefit from less demanding offspring, while paternal genes may benefit from more demanding offspring; see Schlomer et al., 2011). This latent conflict between the paternal and maternal genome is played out by imprinted genes, that is, genes that are differentially expressed depending on whether they were inherited from the mother or from the father (see Burt & Trivers, 2006; Schlomer et al., 2011; Wilkins & Haig, 2003). Not surprisingly, imprinted genes have been found to be involved in prenatal conflicts about fetal nutrition (reviewed in Schlomer et al., 2011). In addition, many imprinted genes are expressed in the brain, and parent-of-origin effects have been detected in the key signaling pathways that mediate social behavior—including the dopaminergic, serotonergic, and oxytocinergic pathways (see e.g., Davies, Lynn, Relkovic, & Wilkinson, 2008).

15

It is quite possible that intragenomic conflict between imprinted genes may contribute to the development of psychopathology. For example, Crespi and Badcock (2008) hypothesized that autistic spectrum conditions are characterized by overexpression of paternal genes, whereas psychosis spectrum conditions are characterized by overexpression of maternal genes. This diametrical model of autism and psychosis was revised and extended by Del Giudice and colleagues (2010) to account for nonclinical variation in autistic-like and schizotypal personality traits. Misfiring Defenses. Adaptive defenses are mechanisms designed to protect individuals from physical and/or social harm. Most negative emotions—including fear, anxiety, disgust, and shame—can be conceptualized as defensive mechanisms, as they play crucial protective roles against physical danger, contamination by pathogens, social exclusion, and so forth (see Nesse, 2004a; Nesse & Jackson, 2006). The calibration of defenses involves a trade-off between the rate of false negatives (failing to activate a defense mechanism when a threat is present) and that of false positives (mistakenly activating the mechanism when no threat is present). Defensive mechanisms are usually designed by natural selection to accept a high rate of false positives so as to avoid catastrophic false negatives; this is known as the smoke detector principle (Nesse, 2005). The smoke detector principle suggests that defensive mechanisms will often misfire or activate with excessive intensity, even when no actual threat is present. Adaptive defenses—like fever, cough, and anxiety—are usually aversive and often disabling; occasionally, inappropriate activation of a defensive mechanism may cause serious harm to the individual. For this reason, misfiring defenses are a likely source of undesirable conditions, ranging from benign “false alarms” to dangerous overreactions. The crucial point is that inappropriate activation of a defensive mechanism does not necessarily imply that the mechanism is dysfunctional or dysregulated—even optimally functional defenses may be designed to misfire from time to time. The logic of the smoke detector principle can be employed to shed light on the etiology of emotional symptoms such as panic attacks, anxiety, and phobic symptoms (Nesse, 2005; Nesse & Jackson, 2006). Developmental Mismatches. Conditional adaptation is the process by which developing organisms make use of contextual cues to direct their developmental trajectory, so as to increase the likelihood that their future phenotype will match the state of the environment. Conditional adaptation is a manifestation of adaptive plasticity, and—when successful—it can dramatically increase the reproductive

16

Evolutionary Foundations of Developmental Psychopathology

success of an organism across a broad range of environments. However, the predictive accuracy (i.e., validity) of contextual cues is usually far from perfect; even when accuracy can be improved by sampling the environment more thoroughly, the potential benefit must be balanced against the required investment of time and effort. For these reasons, conditional adaptation is a fallible process, and a proportion of individuals end up developing a mismatched phenotype. Natural selection can favor conditional adaptation even if the fitness costs of mismatch are high, as long as the average benefits of plasticity are larger than the average costs across individuals. We will deal more extensively with the costs and benefits of developmental plasticity in a later section. Constraints and Trade-offs. The design of an organism is always shaped by countless physical constraints that limit the range of phenotypic change and burden evolved adaptations with undesirable side effects. For example, the erect posture of humans necessarily increases the impact of falling; a larger body size makes organisms more vulnerable to starvation, and so forth. Physical constraints are compounded by the legacy of previous evolutionary history: natural selection builds incrementally on previous designs, and its inability to start from scratch introduces further constraints on adaptive design (path dependence). For example, the fact that human babies are delivered through the pelvic canal poses severe constraints on head size at birth; conversely, selection for larger head size at birth is the biggest ultimate source of maladaptive obstetrical complications. Equally important is the ubiquity of design trade-offs: increasing the functionality of one system may interfere with the functionality of another; increasing the efficiency of a system early in life may lead to decreased efficiency when the organism gets older; enhanced defenses against a given disease may increase vulnerability to another; and so forth. Specifically, a “risk factor” for disorder A may often protect the individual from disorder B. For example, it has been suggested that the long allele of the serotonin transporter gene promoter (5-HTTLPR) may offer protection against depression but increase the risk for psychopathy (Glenn, 2011). Undesirable Adaptations The last category in this taxonomy is also the most intriguing from an evolutionary standpoint. As we just discussed, undesirable conditions often reflect the individually maladaptive outcomes of otherwise adaptive mechanisms. However, it may also be the case that adaptive outcomes are perceived as undesirable conditions, or even classified

as bona fide disorders (Nesse, 2004a; Nesse & Jackson, 2006). Distinguishing undesirable adaptations from maladaptive outcomes can be theoretically and empirically challenging (see Nesse, 2011), but is an essential step to correctly understand the meaning and etiology of the relevant conditions. Antisocial, Exploitative, or Socially Devalued Strategies. In complex social species like ours there are many potential routes to reproductive success, and not all of them involve cooperation and prosociality. Individuals who develop antisocial, exploitative behavioral strategies may often reap considerable rewards—especially in harsh and unpredictable social contexts. Of course, the enhanced reproductive success of (some) antisocial individuals may come at a cost to their own emotional well-being as well as the welfare of their victims. We already mentioned the hypothesis that some types of psychopathy represent an adaptive strategy of this kind; the hypothesis is supported by the robust association between psychopathic traits and a pattern of precocious sexuality, promiscuity, and sexual coercion (see Barr & Quinsey, 2004; Del Giudice, 2014a; Glenn, Kurzban, & Raine, 2011; Mealey, 1995). A similar case has been made for borderline personality disorder, a pervasive pattern of impulsivity and emotional, affective, and relational instability that is more common in females (Brüne, Ghiassi, & Ribbert, 2010). The heterogeneous category of personality disorders is likely to include other biologically adaptive behavioral variants that are treated as problematic, for example because they cause harm or distress to an individual’s social partners. Aversive Defenses. When defenses activate inappropriately or respond with excessive intensity, the outcome may be correctly recognized as maladaptive. However, many protective mechanisms have strongly aversive effects (e.g., vomiting, panic); for this reason, they may give rise to undesirable conditions not only when they misfire but also when they respond appropriately in presence of actual threats. Sometimes, defensive processes can be altogether mistaken for disorders, especially if their logic is incompletely understood and if the correspondence between threat and response is imperfect (because of the smoke detector principle). Indeed, the “fallacy of mistaking defenses for diseases” is a pervasive feature of current psychopathological approaches (Nesse & Jackson, 2006). Many diagnosable instances of emotional disorders—involving low mood, anxiety, and so forth—may be better understood as unpleasant but adaptive responses to contextual factors. As already noted, distinguishing adaptive defensive reactions from maladaptive outcomes or dysfunctional

Beyond Mental Health: Conditional Adaptation and Life History Theory

responses is not an easy task (Nesse, 2011). This is exemplified by the debate on evolutionary models of depression. Some authors have argued that major depression can be adaptive as a mechanism of motivational disengagement from unproductive goals, signaling of social submission, and solicitation of help from family and friends (e.g., Sloman & Price, 1987; Watson & Andrews, 2002). However, while low mood has a number of crucial adaptive functions, the available evidence is more consistent with the idea that major depression usually reflects a maladaptive dysfunction of the systems involved in mood regulation (e.g., Nesse, 2006; Nettle, 2004). Health–Reproduction Trade-Offs. Antisocial strategies and aversive defenses do not exhaust the potential range of undesirable adaptations. The more general point is that, since natural selection maximizes fitness rather than health, traits that increase reproductive success may often have substantial health costs. For example, many health problems associated with aging are the price we pay for more efficient functioning earlier in life (see Nesse, 2001). In developmental psychology, risk taking and impulsivity in adolescence are often viewed as dysfunctional; however, they are better explained as behavioral adaptations to the stronger mating competition faced by human males (e.g., Ellis et al., 2012; Nesse, 2001). Implications for the Core Points of Developmental Psychopathology The mutual interplay between normal and pathological development is one of the core points of developmental psychopathology. An evolutionary perspective offers a deeper understanding of how “normality” and “pathology” can be defined in the first place and provides researchers with a conceptual toolkit for analyzing the full spectrum of undesirable conditions—from harmful and/or maladaptive dysfunctions to adaptive but undesirable mechanisms that may be erroneously mistaken for disorders. In between lies a range of explanatory categories in which adaptation and maladaptation coexist to various degrees. All too often, models in developmental psychopathology converge on dysregulation as the default explanation of undesirable conditions (see the next section). As we have shown here, dysregulation is only one of many potential explanations of psychopathological outcomes; a biologically informed taxonomy like the one we presented (see also Cosmides & Tooby, 1999; Nesse, 2001, 2011) can be a useful guide to formulate alternative hypotheses and build more sophisticated explanatory models.

17

An especially intriguing case is that of high-risk strategies characterized by unpredictable outcomes. Strategic risk provides a powerful explanation of multifinality, since—by definition—individual variables associated with risky strategies can be expected to predict positive outcomes in some individuals and negative outcomes in others. Furthermore, the outcomes of high-risk strategies are often determined in large part by chance factors, highlighting the connection between multifinality and probabilistic causality in developmental trajectories. A similar picture emerges if one considers the calibration of adaptive defenses and the probabilistic trade-offs involved in the balance between misfiring and appropriate responding. An evolutionary approach also provides a nuanced view of the interplay between risk and protective factors— another defining point of developmental psychopathology. In particular, the logic of constraints and trade-offs suggests that some putative “risk factors” for a given condition may actually protect individuals from other (and perhaps more severe) conditions. Similarly, the logic of adaptive defenses should alert researchers to the possibility that some putative “protective factors” involving defense downregulation may actually interfere with an individual’s ability to protect itself from rare but potentially severe threats. In sum, the approach we advocate goes beyond intuitive notions of risk and resilience and contributes to draw a more realistic picture of the complex, layered relations between health and pathology. BEYOND MENTAL HEALTH: CONDITIONAL ADAPTATION AND LIFE HISTORY THEORY A widespread set of assumptions in developmental psychology is that children raised in supportive and wellresourced environments (e.g., who live in communities with social networks and resources for young people; who have strong ties to schools and teachers; who benefit from nurturing and supportive parenting; who are exposed to prosocial peers) tend to develop normally and express optimal trajectories and outcomes. By contrast, developmental processes among children raised in high-stress environments (e.g., who experience poverty, discrimination, low neighborhood attachment, and community disorganization; who feel disconnected from teachers and schools; who experience high levels of family conflict and negative relationships with parents; who are exposed to delinquent peers) put them at risk for dysregulation, leading to impaired functioning and problem behaviors that are destructive to themselves and others. This set of assumptions is powerful and pervasive, if usually implicit,

18

Evolutionary Foundations of Developmental Psychopathology

and underlies what we call the mental health model of developmental psychopathology. In contrast with the mental health model, theory and research in evolutionary biology have come to acknowledge that, in most species, single best strategies for survival and reproduction are unlikely to evolve. Instead, the locally optimal strategy normally varies as a function of three overarching parameters. First, the costs and benefits of different strategies depend on the physical, economic, and social parameters of an organism’s environment (e.g., food availability, mortality rates, quality of parental investment, social competition). This context-dependency means that a strategy that promotes success in some environmental contexts may lead to failure in others. Second, the success and failure of different strategies depends on an organism’s condition or relative competitive abilities in the population (e.g., age, body size, health, history of wins and losses in agonistic encounters); that is, the cost–benefit trade-offs of different strategies varies depending on an organism’s internal condition and competitive status. Third, an organism’s sex often has important implications for the range of available strategies and their relative costs and benefits. In this section we discuss how developmental processes increase adaptation by matching an organism’s phenotype to local environmental conditions and individual characteristics. We begin by reviewing the general concepts of plasticity and conditional adaptation. We then introduce evolutionary life history theory and show how it provides a general framework for adaptive plasticity, as well as an integrative understanding of the development of individual differences in physiology, growth, and behavior. Developmental Plasticity and Conditional Adaptation Because the viability of different survival and reproductive strategies is so context- and condition–dependent, natural selection tends to maintain adaptive developmental plasticity: biological systems that reliably guide the development of alternative phenotypes (including anatomy, physiology, and behavior) to match an organism’s internal condition and external environments (see West-Eberhard, 2003). Developmental plasticity involves “durable biological change in the structure or function of a tissue, organ, or biological system” (Kuzawa & Quinn, 2009, p. 132). Importantly, adaptive developmental plasticity is a nonrandom process; it is the outcome of structured interplay between the organism and its environment, shaped by natural selection to increase the capacity and tendency of individuals to track both their internal condition and external environments and adjust the development of

their phenotypes accordingly. Developmental plasticity is ubiquitous throughout the animal world (see extensive reviews in DeWitt & Scheiner, 2004; West-Eberhard, 2003). Developmental plasticity is critically important for enabling organisms to adapt to stress, which has always been part of the human experience. Indeed, almost half of children in hunter–gatherer societies—the best model for human demographics before the agricultural revolution—die before reaching adulthood (e.g., Volk & Atkinson, 2013). Thus, from an evolutionarydevelopmental perspective, stressful rearing conditions, even if those conditions engender sustained stress responses that must be maintained over time, should not so much impair neurobiological systems as direct or regulate them toward patterns of functioning that are adaptive under stressful conditions (see Ellis et al., 2012; Frankenhuis & de Weerth, 2013). Because developmental plasticity involves durable change, it is inherently forward-looking; that is, it involves predicting—and preparing—for future environments. Boyce and Ellis (2005) make this explicit in their definition of conditional adaptation: “Evolved mechanisms that detect and respond to specific features of childhood environments, features that have proven reliable over evolutionary time in predicting the nature of the social and physical world into which children will mature, and entrain developmental pathways that reliably matched those features during a species’ natural selective history” (p. 290). During fetal development and infancy, important features of the environment are communicated to the child via the placenta and lactation in nutrients, metabolites, hormones, growth factors, and immune factors that reflect the mother’s current and past experiences (Kuzawa & Quinn, 2009). Beyond these molecular signals from the mother, relevant features of the environment are detected and encoded through the child’s ongoing experiences. Developmental plasticity necessitates developmental trade-offs. For example, tadpoles (rana sylvatica) alter their size and shape based on the presence of dragonfly larvae in their rearing environment (Van Buskirk & Relyea, 1998). These alterations involve development of smaller and shorter bodies and deep tail fins. Although tadpoles that do not undergo these morphological changes are highly vulnerable to predation by dragonflies, those that do but end up inhabiting environments that are not shared with dragonflies have relatively poor developmental and survival outcomes. In short, the predator-induced phenotype is only conditionally adaptive. This process highlights that in many cases, natural selection favors a primary phenotype that yields high payoffs under favorable

Beyond Mental Health: Conditional Adaptation and Life History Theory

circumstances and a secondary phenotype that “makes the best of a bad situation” (West-Eberhard, 2003). The Role of Genotypic Variation As should be clear from the tadpole example, in addition to the apparent benefits of developmental plasticity, there can be substantial costs. On the one hand, there is the cost of producing and maintaining the appropriate regulatory and assessment mechanisms to support alternative patterns of development. On the other hand, environmental cues may have limited validity, and thus developmental plasticity in response to current conditions may fail to correctly predict future environmental conditions. Consequently, while adaptive developmental plasticity is widespread, it is not always be the best or only option. As an alternative to adaptive developmental plasticity, or in conjunction with it, natural selection may also maintain genetic variation as a solution to the critical adaptive problem of matching phenotypes to heterogeneous environments. There are a variety of circumstances in which genetic contributions to alternative phenotypes are likely to be favored by natural selection. When individuals inhabit multi-niche environments, and they are able to choose the niche that best fits their phenotype, it may partly or fully obviate the need for developmental plasticity. Instead, a diversity of genetically regulated phenotypes that are specialized to the different social or physical niches can thrive in this context (see Wilson & Yoshimura, 1994). In addition, genetic variation can be maintained through balancing selection, whereby selection for alternative phenotypes systematically changes across time, space, population states, and so forth. A common type of balancing selection is frequency-dependent selection, which occurs when the fitness of different phenotypes changes as a function of their frequency in a population. The most viable form of frequency-dependent selection is negative, selecting against a given phenotype as it becomes more common. For example, aggressive individuals may be very successful when they are surrounded by tame individuals; however, as they multiply and begin to “invade” the population, their reproductive success may drop as they now compete mainly with other aggressive individuals. Balancing selection can also result from heterozygote advantage (when individuals who are heterozygous at a certain locus have higher fitness than either of the homozygous types) or from changes in selection pressures over time and space (fluctuating selection). Fluctuating selection pressures, by definition, weaken directional selection and therefore enable higher rates of genetically-regulated phenotypic variation (including neutral and deleterious forms of variation).

19

A crucial question is, to what degree should phenotypic variation be more developmentally contingent and plastic versus more strongly regulated by genotypic variation? The answer is not simple; indeed, what is typically found in organisms is a mixture of the two. Theoretical models suggest that one should often expect a balance between genetic and environmental determination of phenotypic individual differences. Depending on the structure of environmental variation, the costs and benefits of plasticity, and the life history of an organism, a given selection regime—for example one of temporally fluctuating selection—may maintain different proportions of developmental plasticity and genotypic variation. The reproductive strategies of the male swordtail fish provide an example of this complexity, demonstrating the importance of adaptive genetic variation, adaptive developmental plasticity, and their interplay (reviewed in Ellis, Jackson, & Boyce, 2006). In the swordtail, three alleles at the P locus on the Y chromosome correspond to three modes in size distribution of mature males (small, intermediate, and large). Although all three genotypes perform the range of species-typical mating strategies, they do so at different size-related frequencies. Specifically, small, intermediate, and large males generally sneak, sneak and court, and court females, respectively. Size is the primary mediating mechanism in this species through which allelic variations influence mating strategies. In determining alternative mating strategies, the key developmental event in male swordtail fish is gonadarche (maturation of the gonads). Specifically, the three alleles at the P locus differentially influence timing of gonadarche, which occurs earlier in genotypically small than in genotypically large males. In addition to these genetic influences, timing of gonadarche is also sensitive to a number of environmental factors, such as temperature and agonistic interactions with other males. These environmental influences can result in genotypically small males that are larger than genotypically intermediate males, and alternative mating strategies correlate more strongly with size than with genotype. In addition, mating strategies of male swordtail fish are competition-dependent in relation to interaction with other males. For example, males of intermediate size will sneak and chase females rather than court when in the presence of larger males. In sum, both genomic and environmental factors influence timing of gonadarche, which in turn coordinates patterns of gene expression involved in the developmental cascade that induces sexual maturation and halts or dramatically reduces growth. Timing of gonadarche strongly influences size, and size is a major developmental

20

Evolutionary Foundations of Developmental Psychopathology

factor in entrainment of alternative mating strategies. At the same time, mating strategies are conditionally adjusted in response to current physical and social dimensions of the environment. Thus, although there are strong genotypic influences on size and developmentally-linked mating strategies, the development of the alternative phenotypes in fact emerges through a complex series of gene-environment interplay. Importantly, these developmental interactions occur through integrated effects of gene products and environmental conditions on the developing phenotype.

variable, and denotes the individual’s overall potential for plasticity. When the reaction norms of different individuals are not parallel (different slopes; Figures 1.2b and 1.2c), the result is a statistical genotype × environment (GxE) or phenotype–environment (PxE) interaction, whereby the effect of the environment is moderated by an individual’s genotype/phenotype (and vice versa).

Reaction Norms

From Life History Trade-offs to Life History Strategies

A useful tool for thinking about developmental plasticity is the concept of a reaction norm. A reaction norm is a function describing how a single individual may express different phenotypes in response to a range of environmental conditions. While reaction norms are often treated as a property of genotypes (see Schlichting & Pigliucci, 1998), genotypic effects on development—including individual differences in plasticity—are always mediated by the preexisting phenotype. Moreover, genetically different individuals may develop the same phenotype following different developmental trajectories. Thus, reaction norms may be legitimately employed to map phenotypic change on preexisting phenotypic (rather than genotypic) differences. Figure 1.2 illustrates individual differences in developmental plasticity in the simple case of linear reaction norms. As can be seen in the figure, individuals may differ in the elevation and/or slope of their reaction norms. A steeper slope indicates higher susceptibility to environmental factors, as the same amount of variation in environmental conditions results in a larger change in the expressed phenotype. The reaction range of an individual is the difference between the minimum and maximum phenotypic score over a fixed range of the environmental

A major framework in evolutionary biology for explaining patterns of developmental plasticity and individual differences is life history theory (see Kaplan & Gangestad, 2005; Stearns, 1992). All organisms live in a world of limited resources; for example, the energy that can be extracted from the environment in a given amount of time is intrinsically limited. Time itself is a limited good; the time spent by an organism looking for mates cannot be used to search for food or care for extant offspring. Due to these structural and resource limitations, organisms cannot maximize all components of fitness simultaneously and instead are selected to make trade-offs that prioritize resource expenditures so that greater investment of time and/or resources in one domain occurs at the expense of investment in competing domains. For example, resources spent on mounting a robust inflammatory response to fight infection cannot be spent on reproductive effort. Thus, the benefits of inflammatory response are traded off against the costs of lower ovarian function in women and reduced musculoskeletal function in men (Clancy et al., 2013; Muehlenbein & Bribiescas, 2005). Trade-offs between reproductive effort and health go in the opposite direction as well, as early reproductive maturation is linked to more physical health problems in

Adaptive Plasticity in the Development of Life History Strategies

Figure 1.2 (1) Individual differences in phenotypic elevation but not in slope and reaction range; all genotypes have the same plasticity. (2) Individual differences in slope and reaction range (differential plasticity); all genotypes have the same elevation at the environmental mean. (3) Individual differences in elevation, slope, and reaction range.

Beyond Mental Health: Conditional Adaptation and Life History Theory

adulthood (e.g., Allsworth, Weitzen, & Boardman, 2005). Each trade-off constitutes a decision node in allocation of resources, and each decision node influences the next decision node (opening up some options, foreclosing others) in an unending chain over the life course (Ellis, Figueredo, Brumbach, & Schlomer, 2009). This chain of resource-allocation decisions—expressed in the development of a coherent, integrated suite of physiological and behavioral traits—constitutes the individual’s life history strategy. Life history strategies are adaptive solutions to fitness trade-offs within the constraints imposed by social conditions, physical laws, phylogenetic history, and developmental mechanisms. An organism’s life history strategy coordinates morphology, physiology, and behavior in a way that maximizes expected fitness in a given environment (Braendle, Heyland, & Flatt, 2011; Réale et al., 2010). At the most basic level, the resources of an organism must be distributed between somatic effort and reproductive effort. Somatic effort can be further subdivided into growth, survival and body maintenance, and developmental activity (Geary, 2002). Developmental activity includes play, learning, exercise, and other activities that contribute to building and accumulating embodied capital—strength, coordination, skills, knowledge, and so forth (Kaplan & Gangestad, 2005; Kaplan, Hill, Lancaster, & Hurtado, 2000). Reproductive effort can be subdivided into mating effort (finding and attracting mates, conceiving offspring), parenting effort (investing resources in already conceived offspring), and nepotistic effort (investing in other relatives, for example siblings and grandoffspring). The critical decisions involved in a life history strategy can be summarized by the fundamental trade-offs between current and future reproduction, between quality and quantity of offspring, and—in sexually reproducing species—between mating and parenting effort (see Ellis et al., 2009). By delaying reproduction, an organism

Figure 1.3 The fast–slow continuum of life history variation.

21

can accumulate resources and/or embodied capital, thus increasing the quality and fitness of future offspring; however, the risk of dying before reproducing increases concomitantly. When reproduction occurs, the choice is between many offspring of lower quality and fewer offspring of higher quality. Although intensive parental investment is a powerful way to increase the embodied capital (and long-term prospects) of one’s descendants, the fitness gains accrued through parenting must be weighed against the corresponding reduction in mating opportunities. Different life history strategies solve these problems in different ways by determining how organisms allocate effort among fitness-relevant traits. The same basic framework can be used to describe differences between species, as well as differences between individuals of the same species. At the broadest level of analysis, life history traits covary along a dimension of slow versus fast life history strategies. Variation along the slow-fast continuum is observed both between related species and between individuals of the same species (see Ellis et al., 2009; Réale et al., 2010). Slow growth and late reproduction correlate with long life span, high parental investment, fewer offspring of higher quality, and low juvenile mortality. Conversely, fast growth and early reproduction correlate with high juvenile mortality, short life span, larger numbers of offspring, and reduced parental investment in each (Figure 1.3). Fast life history strategies are comparatively high risk, focusing on mating opportunities (which typically involves more risky and aggressive behavior), reproducing at younger ages, and producing a greater number of offspring with more variable outcomes. Sex Differences in Life History Trade-offs The asymmetries introduced by sexual reproduction have important implications for the life histories of males and females. For example, in most species males tend to engage

22

Evolutionary Foundations of Developmental Psychopathology

in higher mating effort and lower parental effort than females (Geary, 2002; Trivers, 1972). In addition, males usually undergo stronger sexual selection, that is, their reproductive success is more variable than that of females; they also tend to mature more slowly, in order to gain the competitive abilities and qualities needed for successful competition for mates. Sexual asymmetries in life history strategies can be attenuated in species with monogamous mating systems and when both parents contribute to offspring care. Compared with other mammals, humans show an unusually high degree of paternal investment; we are clearly adapted for the possibility of monogamous, long-term relationships. However, human paternal care is also highly variable and facultative (e.g., Geary, 2005; Quinlan, 2008), and strict monogamy is rarely if ever found. Overall, human mating is best characterized as strategically flexible (Gangestad & Simpson, 2000), with a widely documented tendency for men to engage in higher mating effort than women. As a result, the trade-off between current and future reproduction is more pressing for women than for men: women’s reproductive rate is limited by the long duration of gestation and the considerable energetic investment of pregnancy and lactation, and their window for successful reproduction necessarily ends with menopause. In contrast, men can potentially sire many offspring in a very short time, as well as for a more extended period of their lives. Men’s crucial trade-off is the one between mating and parenting: the payoffs of high mating effort are potentially much larger for males, who can benefit directly from having access to a large number of partners; women can usually have only one child at a time, and thus benefit comparatively less from mating with multiple partners. Environmental Determinants of Life History Strategy Developmental calibration of slow versus fast life history strategies is a prototypical case of developmental plasticity. Key dimensions of the environment that regulate the development of life history strategies include energy availability, extrinsic morbidity–mortality, and predictability of environmental change (Ellis et al., 2009; Kuzawa & Bragg, 2012). Energetic resources—caloric intake, energy expenditures, and related health conditions—set the baseline for many developmental processes. Energy scarcity slows growth and delays sexual maturation and reproduction, resulting in a “slow” life history strategy. However, when bioenergetic resources are adequate to support growth and development, then proximal cues to extrinsic morbidity–mortality and unpredictability generally promote faster life history strategies.

Extrinsic morbidity–mortality refers to external sources of disability and death that are relatively insensitive to the adaptive decisions of the organism. Environmental cues indicating high levels of extrinsic morbidity–mortality cause individuals to develop faster life history strategies. Faster strategies in this context—a context that devalues future reproduction—function to reduce the risk of disability or death prior to reproduction. Moreover, high extrinsic morbidity–mortality means that investing in parental care has quickly diminishing returns, which favors reduced parental investment and offspring quantity over quality. Accordingly, exposure to environmental cues indicating extrinsic morbidity–mortality (i.e., observable cues that reliably covaried with morbidity–mortality risks during evolutionary history) can be expected to shift life history strategies toward current reproduction by anticipating maturation and onset of sexual activity. In humans, these cues may include exposure to violence, harsh childrearing practices, premature disability and death of other individuals in one’s local ecology, and so forth. In addition to extrinsic morbidity–mortality, environmental unpredictability—stochastic changes in ecological and familial conditions also regulates development of life history strategies (Ellis et al., 2009). In environments that fluctuate unpredictably (e.g., changing randomly between Conditions A and B, so exposure by parents or their young offspring to Condition A does not reliably forecast whether offspring will mature into Condition A or B), long-term investment in a development of a slow life history strategy does not optimize fitness; all of the energy invested in the future is wasted if the individual matures into an environment where life expectancy is short. Instead, individuals should detect and respond to signals of environmental unpredictability by adopting faster life history strategies. In humans, cues of unpredictability may include erratic neighborhood conditions, frequent residential changes, fluctuating economic conditions, changes in family composition, and so forth. Belsky and colleagues (1991) were the first to hypothesize that harsh parenting, conflictual family relations, and insecure attachment would predict early sexual maturation, impulsivity, reduced cooperation, and exploitative interpersonal styles—the expected correlates of a fast life history strategy in humans. Empirical studies have confirmed these associations and detailed how harsh family relations, insecure attachment, and high levels of mortality in the immediate environment predict early puberty (in females), precocious sexuality, unstable couple relationships, and promiscuous mating styles (see special section of Developmental Psychology; Ellis & Bjorklund, 2012).

Beyond Mental Health: Conditional Adaptation and Life History Theory

Key psychological mediators of fast life history strategies include insecure attachment styles, impulsivity, present orientation (the inability to delay gratification and/or wait for larger rewards in the future), and a short subjective life expectancy. These variables are reliably associated with earlier onset of sexual activity, unrestricted sociosexuality (a desire for short-term, promiscuous sexual relationships), larger number of sexual partners, earlier age at first birth in women, increased risk taking, reduced cooperation, and antisocial behavior (reviewed in Belsky, 2012; Chisholm, 1999; Del Giudice, 2009; Del Giudice, 2014a; Figueredo & Jacobs, 2010; Figueredo et al., 2006). At the level of personality traits, slow life history strategies are robustly associated with agreeableness and conscientiousness (Del Giudice, 2014a). Taken together, these results strongly support the existence of a fast-slow dimension underlying a broad spectrum of individual differences in humans. Because extrinsic morbidity–mortality and unpredictability are distinct, developmental exposures to each of these environmental factors should uniquely contribute to variation in life history strategy (Ellis et al., 2009). Longitudinal analyses of the National Longitudinal Study of Adolescent Health, the National Institute of Child Health and Human Development (NICHD) Study of Early Child Care and Youth Development, and the Minnesota Longitudinal Study of Risk and Adaptation (MLSRA) support this prediction (Belsky, Schlomer, & Ellis, 2012; Brumbach, Figueredo, & Ellis, 2009; Simpson, Griskevicius, Kuo, Sung, & Collins, 2012). For example, in the NICHD and MLSRA studies, exposures to environmental unpredictability in the first 5 years of life (e.g., parental changes, residential changes) uniquely predicted faster life history strategies in adolescence and emerging adulthood, independent of the effects of unpredictability in later childhood and indicators of extrinsic morbidity–mortality. The Centrality of the Phenotype All developmental processes are ultimately the product of structured organism–environment interplay. Development is always modulated by the organized phenotype, which is initially provided by the parents in the form of a zygote and then changes during ontogeny in response to both genetic and environmental influences. Consider a central life history trait: timing of sexual maturation. As discussed above, sexual maturation is regulated by energetic conditions, so that—on average—individuals in well fed populations experience early puberty and poorly fed populations experience late puberty. The effects of energetic conditions, however, are modulated by the

23

organized phenotype. For example, food-getting ability (a behavioral phenotype), metabolic efficiency (a physiological phenotype), and energy stores in the form of body fat (a morphological phenotype) all contribute to regulation of puberty; that is, these phenotypic traits modulate the effects of the critical environmental factor (energy availability) on maturation and functioning of the reproductive axis. The same logic applies to genetic effects: genes provide templates for the production of particular molecules that become incorporated into the phenotype, depending on the responsivity of the phenotype to those molecules and the presence of the necessary environmental building blocks (substances from outside the organism) to support gene expression (West-Eberhard, 2003). The effects of genes, environments, and phenotypes are hierarchically organized: The preexisting phenotype is the transducer of both genetic and environmental sources of information. Specifically, genetic and environmental effects depend on the phenotype being organized to accept them, and the modified phenotype retains these effects as development proceeds. In this sense, the phenotype embodies one’s own particular history of genetic and environmental effects. The organizational role of the phenotype is critical to understanding the development of life history strategies. As we will discuss in detail in the next section, Del Giudice and colleagues (2011) proposed that one of the key functions of the stress response system is to regulate an organism’s life history strategy. According to the adaptive calibration model (ACM; Del Giudice et al., 2011; Ellis, Del Giudice, & Shirtcliff, 2013), the stress response system coordinates the development of alternative life history strategies by affecting a broad suite of physiological and psychological traits, including growth and maturation, sexual and reproductive functioning, social learning, aggression, competition and risk taking, pair bonding, and related factors. This occurs in part through extensive physiological linkages between the stress response system and the reproductive axis (Ellis, 2004). The key idea is that activation of stress, metabolic, and immune system responses during childhood provides crucial information about threats and opportunities in the environment, their type, and their severity. Over time, this information becomes biologically embedded in the parameters—recurring set points and reactivity patterns—of these systems. These parameters provide the developing person with statistical summaries of key dimensions of the environment. An alternative pathway for the effects of stress may revolve around somatic damage: if early stress causes permanent damage to the organism and thus reliably reduces life expectancy, it may be adaptive for individuals exposed to stress early in life to engage in faster

24

Evolutionary Foundations of Developmental Psychopathology

life history strategies even if the environment improves later on (Nettle, Frankenhuis, & Rickard, 2013). In total, the stress response system operates as a mechanism of conditional adaptation: it collects and biologically embeds information from the environment, and makes use of that information to match the developing phenotype to local environmental conditions. In this manner, the environment becomes instantiated in the phenotype. At the same time, the phenotype modulates environmental effects at all points in development. As a result of differences in extant phenotypes, individuals differ in their reaction norms (Schlichting & Pigliucci, 1998). Because reaction norms differ in slope across individuals, some people are more likely than others to experience sustained developmental change in response to environmental exposures, including change in the physiological parameters that mediate development of alternative life history strategies. Moreover, as a result of differences in the organized phenotype, life history-relevant physiological parameters already differ across individuals at birth (cf. temperament). Stated differently, people differ in the elevation of their reaction norms. This means that developmental change in the physiological bases of life history strategies and their behavioral outcomes are likely to occur around different points on the life history spectrum (i.e., around the faster range of life history in some individuals and the slower range in others). In total, the organized phenotype is present from conception and can be described in terms of the steepness and location of its reaction norms along various dimensions. These reaction norms, which have already undergone significant development by the time a child is born, are both regulated by and constrain the effects of environmental and genetic factors. The organized phenotype incorporates and biologically embeds environmental and genetic inputs throughout the life course. This ongoing process translates into individual differences in such critical traits as body size, energy reserves, metabolic efficiency, susceptibility to environmental influence, immune function, fecundity, mate value, and fighting ability. Differences between individuals in these phenotypic traits influence the cost–benefit trade-offs of different life history strategies and thus play a central role in regulating the development of these strategies. Consider the trade-off between mating effort and parenting effort in men. Sexual selection models, such as Gangestad and Simpson’s (2000) strategic pluralism theory, emphasize social and sexual competition as important factors shaping adaptive variation in reproductive strategies. According to this perspective, individuals who

are competitively advantaged relative to peers (i.e., who possess social and physical attributes that make them successful in same-sex competition and targets of choice by the other sex) have more mating opportunities. These enhanced opportunities tend to bias resource allocations toward more mating effort at the expense of parental effort. Because male reproductive success is ultimately constrained by the ability to access, attract, and retain mates, alternative male mating strategies should be especially attuned to the demands and desires of women and the ability of men to successfully engage in intrasexual competition. To a large extent, this variable success arises from phenotypic traits that facilitate gaining status and attracting mates (e.g., size, aggressiveness, physical attractiveness, social relations with others). This leads to the hypothesis that intrasexual competitive abilities in men will regulate life history strategies, especially the mating-parenting trade-off. This hypothesis has been supported by a large empirical literature showing that men who achieve high social status or who possess honest indicators of genetic quality (e.g., physical attractiveness, bilateral symmetry of body parts) engage in more mating effort. For example, anthropological evidence indicates that social status is directly related to male reproductive success in horticultural, agricultural, and pastoral societies (see Pérusse, 1993, for an extensive review). Men with higher status in industrial societies, as measured by education, occupation, and income, also report a greater number of sex partners than men of lower social status (Pérusse, 1993). In addition, men who report that they are more attractive to the opposite sex also report having sex at an earlier age, a greater number of sex partners, and an unrestricted sociosexual orientation (reviewed in James & Ellis, 2013). Finally, men who are more symmetrical in bilateral traits have been found to have more lifetime sexual partners as well as more extrapair sexual encounters during ongoing relationships (controlling for physical attractiveness). In contrast, no consistent relations have been found between women’s symmetry and number of lifetime sexual partners or extrapair sexual relationships (reviewed in Thornhill & Gangestad, 2004). Another critical life history trade-off that is regulated by phenotypic condition is current versus future reproduction. Effort put into reproducing now will use energy or resources that cannot be used or saved for future reproduction. The costs of current reproduction may be paid in terms of reduced number, quality, or survival of future offspring, as well as reduced growth and survival of the parent. This decision whether to pay these costs critically

Beyond Allostatic Load: The Stress Response System as a Mechanism of Conditional Adaptation

25

depends of the physical condition of the individual. If either external conditions (e.g., infectious diseases, warfare) or internal state (e.g., poor health, oxidative stress) indicate a heightened probability of premature disability or death, then individuals should shift resource allocations toward current reproduction (presuming adequate bioenergetic resources to support a fast strategy). Research examining relations between exposures to stress, biological aging or health, and reproductive strategies has provided preliminary support for this hypothesis. For example, Bleil and colleagues (2013) found that heightened psychosocial stress was associated not only with ovarian reserve depletion in older women but also with earlier puberty and higher antral follicle count in younger women, indicating a faster life history strategy. Likewise, women who were exposed in utero to the Dutch famine of 1944–1945 not only had increased risk of chronic degenerative disease but also started reproducing at a younger age, had more offspring, had more twins, and were less likely to remain childless (Painter et al., 2008), again indicating a faster life history strategy. An important qualification to these findings is that individuals with life-expectancy-reducing chronic disease diagnosed during childhood also shift toward current reproduction (e.g., early age at first reproduction), even though the incidence of serious chronic disease was uncorrelated with family and ecological stressors (Waynforth, 2012). These data indicate that internal factors, such as compromised phenotypic condition (i.e., damage), in and of itself, can contribute to accelerate life history strategies. Once again, this underscores the centrality of the phenotype in organizing adaptive plasticity through the reciprocal interplay of environmental and genetic factors, either of which can have more decisive effects in different developmental contexts (see extended discussion in West-Eberhard, 2003).

The framework of life history theory adds a layer of specificity to this general picture. Life history concepts can be employed to make remarkably accurate predictions about the structure of individual differences in physiology, growth, and behavior, and the environmental factors that shift development along alternative trajectories. In particular, life history theory delineates basic dimensions of environmental stress and support that underlie the multitude of risk and protective factors described in developmental psychopathology—resource availability, morbidity/mortality risk, and unpredictability. This is especially useful given the confusing abundance of environmental/contextual variables that might be measured and correlated with developmental outcomes. It is also important to stress how adaptive plasticity and life history theory offer a thoroughly contextualist view of development, though one that is compatible with a major role of genetic factors and genotype × environment interactions. Finally, the centrality of the phenotype in enabling and structuring adaptive plasticity has a number of implications for the prospect of integrating developmental psychopathology and EDP. First of all, it does away with the notion that EDP is wedded to any sort of genetic determinism and shows how it is possible to integrate a sophisticated view of developmental mechanisms within an explicitly evolutionary framework. Second, it suggests a deep theoretical rationale for the prevalence of probabilistic causality in development. Third, it affords a principled way to investigate the connection between behavioral strategies and brain and neurobiological factors, thanks to the concept of biological embedding. In the next section we will further illustrate this point by reviewing the role of the stress response system in collecting environmental information and regulating physical and behavioral development.

Implications for the Core Points of Developmental Psychopathology

BEYOND ALLOSTATIC LOAD: THE STRESS RESPONSE SYSTEM AS A MECHANISM OF CONDITIONAL ADAPTATION

While the mental health model of developmental processes resonates with the intuitions of many researchers, its narrow view of adaptation and maladaptation is an obstacle toward the goal of synthesizing normal and pathological development in a single framework. The concepts of adaptive plasticity and conditional adaptation offer a better appreciation of the logic of individual differences and that of gene-environment interplay in development. The crucial idea is that adverse environmental conditions often direct developmental processes along alternative adaptive pathways, rather than simply impair or dysregulate them.

How does repeated or chronic childhood adversity shape biobehavioral development and, through it, mental and physical health? Consistent with the mental health model, there is a widely accepted answer to this question in the field of developmental psychopathology. Instantiated in models of “toxic stress” (Shonkoff et al., 2012) and “allostatic load” (Lupien et al., 2006; McEwen & Stellar, 1993), that answer posits a striking duality: biological responses to stress are usually adaptive in the short term, but protracted activation of stress response systems is

26

Evolutionary Foundations of Developmental Psychopathology

maladaptive and toxic in the long term. Toxic stress causes disruptions of brain structure and function, resulting in dysregulation of physiological mediators—autonomic, neuroendocrine, metabolic, and immune—“that are the precursors of later impairments in learning and behavior as well as the roots of chronic, stress-related physical and mental illness” (Shonkoff et al., 2012, p. e236). As eloquently stated by Juster and colleagues (2011), the wear and tear of toxic stress and altered stress hormone functioning “inexorably strains interconnected biomarkers that eventually collapse like domino pieces trailing toward stress-related endpoints” (p. 725). These models of toxic stress and allostatic load, however, only tell half of the story. The other half is the central role of the stress response system (SRS) in orchestrating physical and psychosocial development of both humans and nonhuman species (Ellis et al., 2006; Korte, Koolhaas, Wingfield, & McEwen, 2005), both in terms of species-typical development and individual differences. One of the most remarkable features of the SRS is the wide range of individual variation in its physiological parameters. Some individuals respond quickly and strongly even to minor events, whereas others show flat response profiles across most situations. Furthermore, the balance of activation among primary SRS subsystems—the sympathetic nervous system (SNS), the parasympathetic nervous system (PNS), and hypothalamus-pituitary-adrenal (HPA) axis—can vary considerably across individuals. In developmental psychopathology, the standard framework for understanding the development and meaning of individual differences in stress responsivity is that of the allostatic load model (ALM; see Juster et al., 2011; Lupien et al., 2006; McEwen & Stellar, 1993). A guiding assumption of the ALM is that there is an optimal level of stress responsivity; accordingly, both “hyperarousal” and “hypoarousal”—recurring over or under activity of physiological mediators—are routinely described as dysfunctional deviations from the norm, usually caused by a combination of excessive stress exposure and genetic or epigenetic vulnerability. In this framework, environmental stress is treated as a risk factor for all kinds of symptoms and disorders (Compas & Andreotti, 2013). While some authors have argued that optimal adaptation is fostered by environments that contain moderate amounts of stressors (e.g., Rutter, 1993; Seery, 2011), the underlying assumption remains that a single best environment exists, and that deviations from that optimum cause dysregulation and pathology. In this section we argue that acceptance of these assumptions, without placing them in a larger evolutionary-

developmental framework, has impeded our understanding of the role of stress response systems in adaptively regulating development (for a detailed exposition see Ellis & Del Giudice, 2014). Specifically, models of allostatic load focus on the long-term costs of childhood stress and adversity—the “wear and tear” on multiple organ systems induced by chronic stress—but do not address the benefits of calibrating autonomic, neuroendocrine, metabolic, and immune systems to match current and future environments. We argue that this overemphasis on costs misses something fundamental and thus weakens the conceptual power of the ALM perspective. The result has been an imbalanced approach to research that has yielded dramatically more empirical knowledge about dysfunction than adaptive function, making it difficult to gain a coherent big picture of the subject matter. A promising alternative to the ALM is provided by the adaptive calibration model (ACM; Del Giudice et al., 2011), a theory of individual differences in stress responsivity that builds on the concepts of life history theory and developmental plasticity. The ACM supplements the ALM and revises some of its key assumptions, thus laying the foundation for a broad theory of individual differences in stress responsivity. In this section we summarize the key tenets of the ACM, explicitly compare the ACM with the ALM, and discuss the implications of the two models for understanding adaptive and maladaptive developmental responses to stress (for more extended discussion, see Ellis & Del Giudice, 2014). Besides offering a broader perspective on the role of stress in development, the ACM exemplifies how the principles of EDP can be leveraged to achieve theoretical integration across multiple levels of analysis, from social behavior to neurobiology. The Adaptive Calibration Model The ACM is a theory of developmental programing focusing on calibration of the SRS and associated life history strategies to local environmental conditions. The ACM has its main theoretical foundations in life history theory and the theory of adaptive developmental plasticity (West-Eberhard 2003); it integrates and extends previous evolutionary models of stress (e.g., Boyce & Ellis 2005; Flinn, 2006; Korte et al., 2005; Porges, 2007) into a coherent theoretical framework. For a detailed presentation of the model, see Del Giudice and colleagues (2011). The central tenet of the ACM is that the SRS operates as a mechanism of conditional adaptation, with a key role in

Beyond Allostatic Load: The Stress Response System as a Mechanism of Conditional Adaptation

27

Figure 1.4 Core theoretical structure of the adaptive calibration model. SRS: stress response system; LH: life history; OT: oxytocin; 5-HT: serotonin, DA: dopamine. Source: Reprinted from M. Del Giudice, B. J. Ellis, & E. A. Shirtcliff, The adaptive calibration model of stress responsivity, Neuroscience & Biobehavioral Reviews, 35, 1564, 2011.

regulating the development of individual life history strategies (Figure 1.4). In the ACM, the activation of autonomic, neuroendocrine, metabolic, and immune system responses during childhood provides crucial information about threats and opportunities in the environment, their type, and their severity. Over time, this information becomes embedded in the parameters—recurring set points and reactivity patterns—of these systems. These parameters provide the developing person with statistical summaries of key dimensions of the environment. For example, sustained activation of the HPA axis is generated by exposures to danger, unpredictable or uncontrollable contexts, and social evaluation, as well as energetic stress (see Dickerson & Kemeny, 2004; Gunnar et al., 2009); thus, the HPA axis tracks the key environmental variables involved in regulation of alternative life history strategies. Analogous arguments have been made regarding mesolimbic dopamine (Gatzke-Kopp, 2011). In turn, individual differences in SRS functioning regulate the coordinated development of a broad cluster of life history-relevant traits (Figure 1.4). Although the ACM focuses on developmental plasticity, all developmental processes are the product of systematic organism–environment interplay. Because some individuals have steeper reaction norms than others, some individuals are more likely than others to experience sustained developmental change in response to environmental exposures. Further, as a result of differences

in the organized phenotype, SRS parameter values already differ across individuals at birth. Consequently, individuals differ in the location of their reaction norms along SRS dimensions (Boyce & Ellis, 2005), with change more likely to occur for different individuals around higher versus lower ends of responsivity. Within these reaction norm constraints, the ACM articulates a theory of environmental regulation. In total, the SRS (1) collects and biologically embeds information from the environment and (2) makes use of that information to match the developing phenotype to local environmental conditions. A crucial aspect of this matching process is the (iterative) calibration of the SRS itself in the service of life history goals (curved arrow in Figure 1.4). SRS activity also feeds back on the system itself, resulting in responsivity patterns that are adaptively calibrated to current environmental conditions and the individual’s overall strategy. This underscores the fact that responsivity patterns develop over time, and may change—within limits—if the local environment undergoes prolonged changes in safety or predictability (i.e., recalibration). Changes in responsivity are also expected to occur in tandem with key hormonal switches such as adrenarche, gonadarche, childbirth, and menopause. More details on pathways and transitions in development of responsivity patterns can be found in Del Giudice and colleagues (2011).

28

Evolutionary Foundations of Developmental Psychopathology

The Role of the SRS in Allostasis and in Transduction of Environmental Information Environmental events signaling threats to survival or well-being produce a set of complex, highly orchestrated responses within the neural circuitry of the brain and peripheral neuroendocrine pathways regulating metabolic, immunologic, and other physiological functions. The SRS comprises primarily three anatomically distinct yet integrated and cross-regulated circuits: the PNS, SNS, and HPA axis. The general function of the PNS is to promote vegetative functions in the absence of stress (i.e., rest and restorative behavior) and reduce or downregulate cardiac activity. When a stressor is encountered, the PNS responds quickly by withdrawing this inhibitory influence (i.e., vagal withdrawal), allowing the SNS to operate unopposed and thus causing rapid increases in physiological arousal. The PNS promotes flexible responding to stress, sustained attention, and coping with mild to moderate stressors (such as solving a difficult puzzle). More extreme defense reactions associated with freeze/hide behaviors also involve PNS activation, albeit via different efferent fibers (Porges, 2007). If parasympathetic deactivation is not sufficient to cope with the present challenge, activation of the SNS occurs within seconds, providing a second layer of response in this hierarchy. Sympathetic activation mediates fight–flight responses following a fast, direct pathway via the noradrenergic innervation of visceral organs and a slower, hormonal pathway through innervation of the adrenal medulla (e.g., Gunnar & Vazquez, 2006). Following SNS activation, the adrenal medulla secretes epinephrine (E) and norepinephrine (NE) to increase heart rate, respiration, blood supply to skeletal muscles, and glucose release in the bloodstream. The third component of the SRS is the HPA axis, which mounts more delayed, long-term responses to environmental challenge. The end point of the HPA response is cortisol release by the adrenal cortex, typically within 5 minutes after the triggering event, with a cortisol peak between 10 and 30 minutes (Gunnar & Vazquez, 2006). The main effects of cortisol are to (1) mobilize physiological and psychological resources (e.g., energy release, alertness and vigilance, memory sensitization), and (2) counter-regulate physiological effects of SNS activation, facilitating stress recovery. The process by which the regulatory parameters of the SRS (as well as other neurobiological systems) are modified in the face of challenge is termed allostasis (i.e., “stability through change”). Allostasis refers to the moment to moment process of increasing or decreasing

vital functions (i.e., adaptively adjusting physiological parameters within the organism’s operating range) to new steady states in response to the demands of the environment and the organism’s resources (McEwen & Stellar, 1993; see also Lupien et al., 2006). Allostasis functions to help the organism cope with challenging events or stressors, enabling short-term adaptation to environmental perturbations. However, the term allostasis is not always used consistently; for example, some authors restrict the meaning of allostasis to long-term, potentially permanent changes in the system’s parameters in contexts of protracted stress (what McEwen and Wingfield [2003] labeled allostatic states and is now more commonly referred to as biological embedding). The SRS orchestrates whole-organism reactions to challenge through a suite of coordinated responses (i.e., allostatic adjustments). Depending on the intensity and duration of a stressor, SRS activation can reorient attentional focus, increase the organism’s readiness for action (e.g., by increased heart/respiratory rate and changes in blood flow to various organs), shift the balance between different memory- and learning-related processes, release glucose into the bloodstream, suppress (or enhance) reproductive functioning, regulate immune function, and so on. The concept of allostasis represents a significant point of convergence between the ACM and the ALM. The ACM explicitly embraces the concept of allostasis and describes the coordination of allostatic responses as one of the main biological functions of the SRS. The SRS responds not only to threats and challenges in the environment but also to novelties and positive social opportunities (e.g., unexpected or exciting rewards, opportunities for status enhancement, potential sexual partners). More generally, the SRS appears to mediate susceptibility to both cost-inflicting and benefit-conferring features of the environment, operating as an amplifier (when highly responsive) or filter (when unresponsive) of various types of contextual information (see extended discussion in Ellis, Del Giudice, & Shirtcliff, 2013). As we will discuss in the next section, this dual function of the SRS is captured by the concept of biological sensitivity to context (Boyce & Ellis, 2005), which posits that a highly responsive SRS increases the organism’s openness to environmental influence. The Role of the SRS in Regulating Development of Life History Strategies The ACM proposes that, across development, the environmental information collected by the SRS (in interaction with the child’s genotype) canalizes physiological and

Beyond Allostatic Load: The Stress Response System as a Mechanism of Conditional Adaptation

behavioral phenotypes to match local ecological contexts (Figure 1.3). The SRS coordinates the development of alternative life history strategies by affecting a broad suite of life history-related physiological and psychological traits, including growth and maturation, sexual and reproductive functioning, social learning, aggression, competition and risk taking, pair bonding, and related factors. The assumption is that these traits and trade-offs are regulated in ways that once—even if possibly no longer—reliably enhanced fitness across different environmental contexts. First of all, the SRS is crucially involved in the regulation of growth and metabolism, and chronic stress has been linked to individual differences in physical growth patterns. The SRS also modulates learning in a number of different ways: HPA and autonomic profiles have been associated with individual differences in cognitive functioning, memory, and self-regulation. Second, the SRS is functionally implicated in all the components of mating and parenting, beginning with sexual maturation. The autonomic systems, HPA, and gonadal axes are connected by extensive functional cross-talk, and HPA activity is linked to variation in pubertal maturation and fecundity. Variation in SRS functioning is also associated with romantic attachment styles. In turn, attachment styles predict relationship stability, commitment, and investment—all key determinants of parenting effort in humans. More directly, SRS functioning affects parenting behavior, including controlling and intrusive parenting practices, inconsistent discipline, and parental sensitivity to children’s needs and demands. In men, cortisol and testosterone work together to direct somatic and behavioral effort toward mating or parenting. Finally, sexual competition is a crucial aspect of mating effort. Dominance-seeking, aggression, and risk taking are all functionally connected to mating competition, and all are associated with SRS functioning in synergy and interaction with testosterone, serotonin, and dopamine. Furthermore, stress exposure regulates mating behavior by altering mate preferences and affecting the perceived attractiveness of potential sexual partners (reviewed in Del Giudice et al., 2011; Ellis & Del Giudice, 2014). In summary, the SRS not only collects and encodes crucial life history-relevant information but is also involved in the regulation of all the major aspects of human life history strategies. Other systems that contribute to life history regulation include the hypothalamic-pituitary-gonadal axis, the serotonergic, dopaminergic, and oxytocinergic systems, and the immune system. Not coincidentally, all of these systems engage in extensive bidirectional cross-talk with the SRS (see e.g., Gatzke-Kopp, 2011; Miller,

29

Chen, & Parker, 2011; van Goozen, Fairchild, Snoek, & Harold, 2007). Patterns of Responsivity Del Giudice and colleagues (2011) provide an extended theoretical and empirical treatment of the logic underlying the development of alternative responsivity patterns in the ACM (see Figure 1.5), including predicted relations between SRS physiology and serotonergic, dopaminergic, and oxytocinergic functioning. Here we briefly summarize this logic. In safe, low-stress environments, a highly responsive SRS enhances social learning and engagement with the external world, allowing the child to benefit more fully from social resources and opportunities (Boyce & Ellis, 2005), thus favoring development of a sensitive phenotype (pattern I). The association between high parental sensitivity, positive family relations, and the development of a highly responsive SRS is supported by a number of studies (e.g., Ellis et al., 2005; Evans et al. 2013; Hackman et al., 2013). According to the ACM, sensitive patterns should be characterized by moderate HPA/SNS responsivity and high PNS responsivity. A sensitive phenotype makes children better at detecting positive opportunities and learning to capitalize on them. For example, high PNS responsivity has been linked to socio-emotional competence, engagement, and self-regulation (e.g., Stifter & Corey, 2001). Social learning and sensitivity to context are especially adaptive in the context of slow life history strategies, as a form of protracted somatic investment (Kaplan & Gangestad, 2005). In very safe and protected settings, sensitive individuals will rarely experience strong, sustained activation of the SNS and HPA systems; thus, the individual enjoys the benefits of responsivity without paying significant health costs (e.g., immune, energetic). At moderate levels of environmental stress, the cost– benefit balance begins to shift as the potential advantages of high sensitivity decrease and the physiological and health costs of maintaining a hyper-responsive SRS increase. The optimal level of SRS responsivity is predicted to fall downward, favoring development of buffered phenotypes (pattern II) characterized by moderately low reactivity and a slow life history strategy. Buffered responsivity is expected to be the modal pattern in most populations (with most SRS parameters set around the mean), particularly in the low-risk, middle-class populations that provide a majority of research participants in psychology and neuroscience. The emergence of buffered responsivity patterns under conditions of moderate environmental stress is empirically consistent with the stress inoculation hypothesis, the idea that early exposure to

30

Evolutionary Foundations of Developmental Psychopathology

Figure 1.5 Predicted curvilinear relation between developmental context and optimal levels of stress responsivity. The figure does not imply that all components of the SRS will show identical responsivity profiles, nor that they will activate at the same time or over the same time course (see Del Giudice et al., 2011). Male and female symbols indicate sex-typical patterns of responsivity, but the model also predicts substantial within-sex variation. Source: Adapted from M. Del Giudice, B. J. Ellis, & E. A. Shirtcliff, The adaptive calibration model of stress responsivity, Neuroscience & Biobehavioral Reviews, 35, 1577, 2011.

repeated mild stressors downregulates the SRS and leads to increased resistance to stress later on (e.g., Rutter, 1993). However, the ACM places stress inoculation in a broader theoretical perspective, in which moderate responsivity is only one out of many locally adaptive patterns of SRS functioning. The benefits of increased responsivity rise again when the environment is perceived as dangerous and/or unpredictable. A responsive SRS enhances the individual’s ability to react appropriately to dangers and threats while maintaining a high level of engagement with the social and physical environment. Moreover, engaging in fast life history strategies should lead the individual to allocate resources in a manner that discounts the long-term physiological costs of the stress response in favor of more immediate advantages. In this context, the benefits of successful defensive strategies outweigh the costs of frequent, sustained HPA and SNS activation, leading to vigilant phenotypes (pattern III). The predicted profile of vigilant individuals includes high HPA/SNS responsivity and low PNS responsivity (see Del Giudice et al. 2011). In turn, this physiological profile should be associated with fast life history-related traits such

as fast maturation and high mating effort. Increased SRS responsivity in dangerous environments can be expected to go together with increased responsivity in other neurobiological systems; for example, hyper-dopaminergic function may contribute to the vigilant phenotype by boosting attention to threat-related cues and fast associative learning (Gatzke-Kopp, 2011). Compared with their sensitive counterparts, vigilant individuals may show slower HPA recovery (i.e., they may take longer to return to baseline) and slower habituation (see Gunnar & Vazquez, 2006). For this reason, they are also likely to show stronger hypocortisolism following prolonged periods of stress. Although we argue that vigilant phenotypes represent biologically adaptive responses to early stress, their “hair trigger” regulation may render them especially vulnerable to breakdown or persistent dysregulation following extreme stressors. On average, males and females are expected to differ in the predominant behavioral correlates of vigilant patterns— aggressive and agonistic behavior versus anxiety and withdrawal—because of the different costs and benefits of aggression, impulsivity, and risk taking in the two sexes

Beyond Allostatic Load: The Stress Response System as a Mechanism of Conditional Adaptation

(Archer, 2009; see also Martel, 2013). Indeed, several studies suggest that acute HPA activation—typical of vigilant patterns—tends to promote risk taking in males and risk aversion in females (see Mather & Lighthall, 2012; Starcke & Brand, 2012). Accordingly, high SRS responsivity can be associated with both internalizing and externalizing symptoms, especially in younger children (see Alink et al., 2008; Gunnar & Vazquez, 2006; van Goozen et al., 2007). In very dangerous environments characterized by severe or traumatic stress, the balance shifts again toward low responsivity, especially for males who adopt a fast, mating-oriented life history strategy characterized by antagonistic competition and extreme risk taking. Such a strategy requires outright insensitivity to threats, dangers, social feedback and the social context. For an extreme risk-taker, informational insulation from environmental signals of threat is an asset, not a weakness (see also Korte et al., 2005). In particular, adopting an exploitative/antisocial interpersonal style requires one to be shielded from social rejection, disapproval, and feelings of shame (all amplified by heightened HPA responsivity; reviewed in Del Giudice et al., 2011). An unemotional pattern of generalized low responsivity (pattern IV) can be evolutionarily adaptive at the high-risk end of the environmental spectrum – especially in males – despite its possible negative consequences for the social group and for the individual’s subjective well-being. This type of chronic low responsivity should be carefully distinguished from temporary exhaustion periods, usually arising after prolonged SRS activation in highly responsive individuals exposed to enduring stressors. The association of risk taking with low levels of SRS responsivity and basal activity is well documented, especially in males (e.g., Bubier & Drabick 2008; Halpern et al. 2002). Unemotional profiles should be associated with high mating effort and early sexual maturation; this is consistent with the robust association between low SRS responsivity, externalizing behaviors (especially in male adolescents and adults), and callous-unemotional traits. As we discuss in a later section, this constellation of traits is associated with early maturation, precocious sexuality, and sexual promiscuity. Finally, the ACM hypothesizes two developmental pathways leading to unemotional responsivity patterns. In the first pathway, an initially responsive phenotype shifts toward unresponsivity following chronic severe stress. In particular, some children are expected to shift from pattern III to pattern IV during middle childhood or adolescence (Del Giudice et al. 2011). This prediction is consistent with the finding that associations between HPA activity and aggressive/externalizing behavior tend to be positive

31

in preschoolers but negative in middle childhood and adulthood (Alink et al., 2008). In the second pathway, unresponsivity may develop even in low-stress environments because of strong genetic predispositions, and may be apparent already in early childhood. The logic of sex differences in responsivity patterns is based on sexual selection theory informed by life history considerations. At the slow end of the life history continuum, both sexes engage in high parental investment, and male and female interests largely converge on long-term, committed pair bonds; sex differences in behavior are thus expected to be relatively small. As environmental danger and unpredictability increase, males benefit by shifting to low-investment, high-mating strategies; females, however, do not have the same flexibility since they benefit much less from mating with multiple partners and incur higher fixed costs through childbearing. Thus, male and female strategies should increasingly diverge at moderate to high levels of environmental danger/unpredictability. In addition, sexual competition takes different forms in males and females, with males engaging in more physical aggression and substantially higher levels of risk taking (e.g., Archer, 2009; Wilson, Daly, & Pound, 2002). As mating effort increases, sexual competition becomes stronger and sex differences in competitive strategies become more apparent. For these reasons, sex differences in responsivity patterns and in the associated behavioral phenotypes should be relatively small at low to moderate levels of environmental stress (patterns I and II) and increase in stressful environments (pattern III). Finally, we expect males to be overrepresented as high-risk, low-investment strategists (pattern IV). The shifting equilibrium between costs and benefits of responsivity is predicted to result in a complex curvilinear relation between environmental stress and SRS responsivity (Figure 1.5). The ACM taxonomy of responsivity patterns is supported by preliminary empirical evidence on autonomic responsivity in middle childhood (Del Giudice, Hinnant, Ellis, & El-Sheikh, 2012). While the SRS has a crucial role in directing the development of alternative life history strategies, it does so in interaction with other neurobiological systems. Accordingly, the four responsivity patterns described in the ACM are characterized not just by different profiles of SRS functioning, but also by functionally organized individual differences in behavior, in the physiology of the hypothalamic-pituitary-gonadal axis, in serotonergic and dopaminergic pathways, and other systems (Del Giudice et al., 2011). Each pattern reflects a unique combination of costs, benefits, and specific vulnerabilities to pathology. Finally, it should be noted that

32

Evolutionary Foundations of Developmental Psychopathology

Figure 1.5 depicts an idealized, population-level reaction norm. In practice, different individuals may respond differently to the same level of environmental stress and show different degrees of plasticity—owing to different environmental sampling histories and the effects of genetic, epigenetic, and endophenotypic variation—as discussed in the section on differential susceptibility. ALM and ACM: A Comparison Allostatic load is a label for the long-term costs of allostasis; it is often described as the “wear and tear” that results from repeated allostatic adjustments (i.e., adaptation to stressors), exposing the organism to adverse health consequences. The ALM emphasizes that biological responses to threat, while essential for survival, have negative long-term effects that promote illness. The benefits of mounting biological responses to threat are traded off against costs to mental and physical health, and these costs (allostatic load) increase as the organism ages. Basic tenets of the ACM an ALM are compared in detail in Ellis and Del Giudice (2014) and juxtaposed in Table 1.1. Both the ACM and ALM emphasize the

adaptive nature of short-term physiological responses to stress (Table 1.1, row labeled Activation of autonomic, neuroendocrine, metabolic, and immune systems). Further, the ACM concurs with the ALM regarding the effects of childhood stress and adversity on regulation of allostatic mechanisms. Indeed, a substantial body of research has now documented biological embedding of early life stress in SRS parameter values (Table 1.1, row labeled Changes in allostatic mechanisms). In the ALM, however, this biological embedding is construed negatively, as the result of cumulative stress exposures that predispose the individual to morbidities and premature mortality. As shown in Table 1.1 (row labeled Cognitive, behavioral, and emotional impairments in children), some of these outcomes include lower performance on standard tests of intelligence and executive functions and increased mental health problems (reviewed in Ellis & Del Giudice, 2014). The ACM also acknowledges that chronic SRS activation carries substantial costs, in terms of biological fitness as well as health and well-being. While the ACM stresses conditional adaptation, it leaves open the possibility that, for a number of reasons, some developmental outcomes are biologically maladaptive. In terms of proximal responses

TABLE 1.1 Comparison of Adaptive Calibration Model and Allostatic Load Model1 Responses to psychosocial stress/ unpredictability

Examples of response

ACM

ALM

Activation of autonomic, neuroendocrine, metabolic, and immune systems

• Acute SNS and HPA responses mobilize energy reserves; protect against septic shock and nutrient deprivation; permit fight or flight responses that are normally protective against danger. • Inflammation accelerates the healing of wounds.

Central to theory

Central to theory

Changes in allostatic mechanisms

• Increased inflammatory tone • Elevated cortisol and catecholamines • Muted cardiovascular responses to stress

Central to theory

Central to theory

Cognitive, behavioral, and emotional impairments in children

• Reduced scores on standard tests of intelligence, language, memory, and other abilities • Early onset and increased prevalence of psychopathology

Not inconsistent with theory

Central to theory

Cognitive, behavioral, and emotional adaptations to stress in children

• Tailoring of emotion systems, arousal responses, and perceptual abilities to the detection and monitoring of danger • Development of insecure attachments, mistrustful internal working models, opportunistic interpersonal orientations, oppositional-aggressive behavior

Central to theory

Not inconsistent with theory

Long-term deleterious outcomes

• Cognitive and physical impairments • Depression • Increased risk of cardiovascular disease and all-cause mortality

Not inconsistent with theory

Central to theory

Long-term adaptive changes in biobehavioral systems

• Adaptive calibration of autonomic, neuroendocrine, metabolic, and immunological systems • Regulation of alternative life history strategies to match ecological conditions

Central to theory

Beyond the scope of the theory

1 Light shading indicates a difference in emphasis between the ACM and ALM. Dark shading indicates a qualitative divergence between the two theories. Reprinted from B. J. Ellis & M. Del Giudice, Beyond allostatic load: Rethinking the role of stress in regulating human development, Development and Psychopathology, 20, 9, 2014.

Beyond Allostatic Load: The Stress Response System as a Mechanism of Conditional Adaptation

to childhood adversity, The ACM and ALM mainly differ in their emphasis on the benefits (ACM) versus the costs (ALM) of allostatic adjustments (shown in italics in Table 1.1). Cost–Benefit Trade-Offs in the Development of Alternative Phenotypes. The ACM and ALM diverge considerably in how they deal with cost–benefit trade-offs, individual differences, and long-term developmental changes. From an evolutionary standpoint, the ALM makes no distinction between the two meanings of adaptive and maladaptive, as conceptualized from a mental health versus evolutionary perspective. Indeed, maladaptation is typically inferred whenever there are substantial costs to the organism. For example, if elevated cortisol levels in adolescents are associated with an undesirable outcome, such as reduced working memory, then elevated cortisol is classified as a marker of allostatic load (see Juster et al., 2011). This reasoning ignores the crucial fact that biological processes are maintained by natural selection when their fitness benefits outweigh the costs, not when they are cost-free; indeed, even large costs can be offset by large enough expected benefits. Although there are practical reasons for identifying allostatic load biomarkers, this approach alone is incomplete because it only specifies dysfunction and not the adaptive functions of developmentally calibrated biological parameters. Because of the failure to distinguish between (mal)adaptive and (un)desirable outcomes, most applications of the ALM do not adequately address the trade-offs involved in the development of physiological and behavioral phenotypes; as a consequence, the ALM literature often lacks a theory of adaptive individual variation in stress responsivity (see Korte et al., 2005, for a notable exception). Instead the ALM focuses on optimal SRS parameter values, as defined by covariation with desirable health outcomes; deviations from these optimal settings form the basis of dysregulation. The applied goal of the ALM is to identify nonoptimal autonomic, neuroendocrine, metabolic, and immune profiles that predict psychiatric and biomedical disorder (Table 1.1). In contrast, the ACM emphasizes adaptation in context and posits that optimal SRS parameter values vary as a function of environmental conditions. From this perspective, the notion of globally optimal baseline or responsivity levels for SRS parameters is highly problematic; indeed, the entire literature on biological sensitivity to context demonstrates that the value of hypo-responsivity versus hyper-responsivity is context dependent (see extended

33

discussion that follows). The ACM gives full consideration to the costs and benefits of SRS responsivity. For example, consider heightened stress responsivity in a dangerous, unpredictable environment. In the ACM, it is hypothesized that the costs of repeated SRS activation are offset by improved management of danger. Although the system is on a hair trigger, with a resulting increase in anxiety and/or aggression, few instances of actual danger will be missed. In addition, engaging in a fast, present-oriented life history strategy makes it optimal to discount the long-term health costs of chronic SRS activation if the immediate benefits are large enough (for in-depth discussion, see Del Giudice et al., 2011). In the ALM framework, the same pattern of responsivity would be treated as dysfunctional because the stress response is deployed even in absence of true dangers (e.g., “excessive” response, “unnecessary” triggering; e.g., Lupien et al., 2006) and because of the associated undesirable states and health risks (e.g., interpersonal distress). However, this approach fails to consider that biological defenses are usually designed by natural selection to accept a high rate of false positives. In most instances, unnecessary responding is an adaptive feature of the system—though a costly one—rather than a sign of dysregulation or malfunction. Long-Term Adaptations to Stress: The Developmental Regulation of Alternative Life History Strategies According to the ACM, childhood adaptations to stress may eventuate in long-term adaptive changes in biobehavioral systems. Herein lies the key difference between the ACM and ALM (shown in bold in Table 1.1). In the ALM, energy devoted to mounting autonomic, neuroendocrine, metabolic, and immune responses to threat is traded off against wear and tear on multiple organ systems. The ACM extends this logic by conceptualizing these trade-offs as decision nodes in allocation of resources. It is through this chain of resource-allocation decisions—instantiated in the regulatory parameters of the SRS and related biological systems—that the developing organism adapts to local conditions. Thus, the ACM shifts the emphasis from dysregulation to conditional adaptation (Table 1.1). From an evolutionary perspective, increased wear and tear is a cost of pursuing a fast life history strategy. The fast strategy is instantiated in a chain of resource allocation decisions over the life course that “make the best of a bad situation” by trading off survival for reproduction. Thus, many biologically embedded changes that the ALM conceptualizes as costs (e.g., heightened HPA reactivity)

34

Evolutionary Foundations of Developmental Psychopathology

the ACM views as decision nodes in development of a faster strategy. Conversely, slower life history strategies involve greater allocation of resources toward enhancing growth, vitality, and long-term survival (e.g., DNA repair). Development of a fast life history strategy in dangerous and unpredictable contexts is not impairment or dysfunction; it is a coherent, organized response to stress that has been shaped by a natural selective history of recurring exposures to harsh and unpredictable environments. Implications for the Core Points of Developmental Psychopathology The integrative perspective on stress responsivity presented in this section touches on many core points of developmental psychopathology. First of all, the ACM illustrates that evolutionary principles are not just relevant to the behavioral level of analysis. On the contrary, they can be fruitfully applied to understand the role of brain and neurobiological factors in development, and employed to craft detailed models of neurobiological functioning. Like the ALM, the ACM helps synthesize normal and pathological development in a single framework; however, the ACM goes beyond the ALM with a broader view of costs and benefits and a detailed theory of adaptive matching between environment and phenotype. The ACM fully incorporates the contextualism of life history theory; in this perspective there is no single optimal phenotype or developmental trajectory—only locally adaptive, contextually sensitive strategies with their baggage of costs and trade-offs. An important feature of the logic of responsivity patterns is that it invites—indeed, virtually requires—a person-centered approach to individual differences in stress physiology and neurobiology (e.g., Del Giudice et al., 2012). Even the schematic model depicted in Figure 1.5 shows that causal relations between environmental variables and developmental outcomes are predicted to be highly nonlinear; if the model is correct, attempts to model stress responsivity with standard linear models are doomed to yield inconsistent and misleading results. Responsivity patterns also provide a striking illustration of equifinality and multifinality in the interplay between environment, neurobiology, and behavior. In the ACM, dramatically different environments may entrain the development of similar responsivity profiles; for example, sensitive and vigilant patterns are both hypothesized to imply high HPA responsivity. In turn, similar responsivity profiles may predict remarkably different behavioral correlates—for example low impulsivity in sensitive patterns versus high impulsivity in vigilant patterns. While the resulting

developmental trajectories may look empirically baffling, they become tractable when they are framed in the appropriate functional perspective. Finally, the ACM potentially explains why, in the empirical literature, heightened and dampened stress responsivity seem to work as risk factors in some studies and as protective factors in others. Both sensitive and buffered phenotypes are associated with lower levels of psychosocial stress and concomitant development of slower life history strategies, whereas both vigilant phenotypes and unemotional phenotypes are associated with higher levels of psychosocial stress and concomitant development of faster life history strategies (Figure 1.5). Depending on the specific outcomes under investigation, heightened reactivity may look like a protective factor in sensitive phenotypes and a risk factor in vigilant phenotypes, while dampened reactivity may look like a protective factor in buffered phenotypes and a risk factor in unemotional phenotypes. As we discuss in a later section, the picture is further complicated by the possibility that some disorders are actually more likely to occur at the slow end of the life history continuum. Taming this complex interplay of cause and effect will only be possible with the help of a detailed theory of biological and developmental trade-offs.

BEYOND DIATHESIS-STRESS: DIFFERENTIAL SUSCEPTIBILITY TO ENVIRONMENTAL INFLUENCES As discussed in the preceding sections, early experiences can have profound and lasting effects on psychological development. Nevertheless, people vary dramatically in the extent to which they respond to their social and physical environment. Similar developmental experiences may have profound effects on some individuals and slight or even negligible effects on others. The idea that some individuals are more susceptible than others to environmental adversity has a long history in psychology, and is captured by the complementary concepts of vulnerability and resilience (for a recent review see Compas & Andreotti, 2013). Individual differences in vulnerability may be determined by a combination of genetic factors and previous experiences, and explain why, for example, only some children develop such undesirable outcomes as aggression, insecure attachment, and cognitive difficulties in response to stressful or impoverished rearing experiences. The concept of vulnerability lies at the heart of the diathesis-stress model of psychopathology, which is arguably the dominant paradigm in the field.

Beyond Diathesis-Stress: Differential Susceptibility to Environmental Influences

Against this background, recent empirical and theoretical advances have brought about a conceptual revolution in the study of organism × environment interactions. As has become increasingly apparent, many of the same factors that determine increased vulnerability to stress and adversity also confer enhanced responsivity to the positive, supportive aspects of the environment (Belsky, Bakermans-Kranenburg, & Van IJzendoorn, 2007; Boyce & Ellis, 2005). No longer confined to vulnerability, susceptibility to environmental influence can be understood as a generalized trait that increases the range of potential developmental outcomes in a bivalent fashion. While some individuals are disproportionately likely to suffer damage if exposed to harshness and adversity, they are also disproportionately likely to benefit from nurturance and support. Here we introduce the concept of differential susceptibility, discuss its importance for developmental processes, and review the models that have been proposed to explain the evolution of systematic individual differences in susceptibility to the environment. For further discussion see Ellis, Boyce, Belsky, Bakermans-Kranenburg, & Van IJzendoorn (2011a). Differential Susceptibility: Orchids and Dandelions In the differential susceptibility literature, highly susceptible children are referred to by the shorthand designation of orchid children, signifying their special sensitivity to both highly stressful and highly nurturing environments. For example, highly irritable infants tend to become less sociable than moderately irritable infants when they are insecurely attached, but more sociable if they experience secure attachment relationships (Stupica, Sherman, & Cassidy, 2011). Similarly, children with difficult temperament develop more behavior problems when they experience low-quality care, but fewer behavior problems when reared in a high-quality context (Pluess & Belsky, 2009). Children who are low in susceptibility to environmental influence, on the other hand, are designated as dandelion children, reflecting their relative ability to function adequately in species-typical circumstances of all varieties (Boyce & Ellis, 2005). Differential susceptibility can be distinguished from both vulnerability (specific sensitivity to negative environmental factors) and the more recent concept of vantage sensitivity (specific sensitivity to positive environmental factors; see Pluess & Belsky, 2013). Converging evidence from different areas of research indicates that highly susceptible children share a cluster of interrelated traits including high physiological reactivity across multiple systems (including the SRS),

35

negative emotionality, and difficult temperament (reviewed in Belsky et al., 2007; Boyce & Ellis, 2005; Ellis et al., 2011a). While the genetics of susceptibility is still incompletely understood, genes involved in serotonergic and dopaminergic pathways appear to play a central role (e.g., Bakermans-Kranenburg & Van IJzendoorn, 2011; Belsky & Beaver, 2011; Heiming & Sachser, 2011); as discussed above, serotonergic and dopaminergic pathways interact bidirectionally with the SRS. Together, these interconnected systems contribute to a general phenotypic dimension of neurobiological susceptibility to the environment (see Boyce & Ellis, 2005; Ellis et al., 2011a). Figure 1.6 illustrates the concept of differential susceptibility by showing the reaction norms of high susceptibility and low susceptibility individuals in a population. Differential susceptibility requires that reaction norms cross somewhere in the middle range of environmental variation. Figure 1.6 may represent either a genotype × environment (GxE) or phenotype × environment (PxE) interaction, depending on whether individuals are distinguished by their genotype (e.g., variants of the serotonin transporter gene) or phenotype (e.g., high stress responsivity, negative emotionality). In summary, differential susceptibility is a special case of developmental plasticity; its defining characteristics are (1) a crossover interaction in which more susceptible individuals have a broader reaction range and steeper reaction norms (compare with Figure 1.2b), and (2) a clear positive–negative polarity of the relevant environmental dimension. These two factors together result in systematic variation in the strength of relations between environmental exposures and developmental outcomes that maps on to the orchid–dandelion distinction. Whether environmental conditions can be labeled as “positive” or “negative” in a biological sense depends on their likely effects on an individual’s fitness. As we discussed in the section on life history theory, many of the negative environmental factors investigated by developmental psychologists are ultimately correlated with danger, unpredictability, and resource scarcity. Differential susceptibility has far-reaching implications for developmental psychology and psychopathology; it moderates the effects of environmental exposures on developmental and life outcomes. Ultimately, this means that the development of some individuals, more than others, will be influenced by their experiences and environments—even if these were exactly the same. Differential susceptibility to the environment, therefore, constitutes a central mechanism in the regulation of alternative patterns of human development. Over the last few years, the concept

36

Evolutionary Foundations of Developmental Psychopathology

Figure 1.6 Reaction norms showing differential susceptibility. More susceptible individuals are disproportionally affected by both negative and positive environments.

of differential susceptibility has generated considerable enthusiasm in developmental psychology and has rapidly become the foundation of an expanding empirical literature. Evidence of differential susceptibility exists for a number of traits and outcomes including prosociality, sociability, externalizing and internalizing symptoms, depression, and timing of puberty (see Ellis & Boyce, 2011, and other papers in that special issue).

Evolutionary Models of Differential Susceptibility Individual differences in neurobiological susceptibility to the environment raise many important questions. Why do they exist in the first place? Are they adaptive—and if so, what evolutionary processes are responsible for their maintenance in human populations? When do they emerge in development, and by what interplay of genetic and environmental factors? Evolutionary models of differential susceptibility (Belsky, 1997, 2005; Boyce & Ellis, 2005; Ellis et al., 2006) have explored different explanations for the maintenance of individual differences—balancing selection, selection for conditional adaptation, and selection for diversified bet-hedging. These explanations are not mutually exclusive, and indeed they can potentially be integrated within a single theoretical framework.We now briefly review each in turn.

Maintenance of Differential Susceptibility Through Balancing Selection There is considerable evidence that individual differences in susceptibility can be attributed—at least in part—to individual differences in genotype. The best examples include allelic variation in the serotonin-transporter gene promoter (5-HTTPLR) and the dopamine D4 receptor gene (DRD4). Recent meta-analyses have demonstrated that these polymorphisms moderate environmental effects in a “for better and for worse” manner; that is, different variants of these genes are generally associated with differential susceptibility to both stressful and nurturing environmental conditions (Bakermans-Kranenburg & van IJzendoorn, 2011; van IJzendoorn, Belsky, & Bakermans-Kranenburg, 2012). Other genetic variants involved in differential susceptibility have been identified, not only in dopaminergic and serotonergic pathways but also in the oxytocinergic system and the HPA axis (see Cicchetti & Rogosch, 2012; Ellis et al., 2011a). As discussed in a previous section, adaptive genotypic differences in a population may be maintained by various forms of balancing selection, including negative frequency-dependent selection. Building on the work of Wilson and Yoshimura (1994), Ellis and colleagues (2006) proposed a negative frequency-dependent model of the maintenance of genetically regulated variation

Beyond Diathesis-Stress: Differential Susceptibility to Environmental Influences

in differential susceptibility to the environment. In this model, developmental specialists are defined by relatively flat reaction norms, such that phenotypic development is minimally sensitive to normal environmental variation. Different developmental specialists are instead characterized by different genetically regulated behavioral types (e.g., shy vs. bold, slow vs. fast life history strategy). Although developmental specialists lack plasticity, their specialized personalities enable them to thrive in the niche that they are specialized to exploit. Developmental generalists, by contrast, are defined by relatively steep reaction norms, such that phenotypic development is highly context-dependent. Different developmental generalists thus experience different developmental outcomes based on their rearing experiences. All else being equal, the presence of multiple niches in a single environment will favor developmental specialists over developmental generalists when individuals can evaluate and select niches that increase their fitness. This is because specialists outperform generalists in their preferred niche. However, multi-niche environments are often characterized by negative density-dependence, meaning that as a given niche becomes more crowded (i.e., overexploited relative to its size), the fitness benefits of specializing in that niche decrease. This is the cost of specialization. Models of genotypic variation in susceptibility are still in the formative stages. The general hypothesis that frequency-dependent selection can maintain individual variation in susceptibility or responsiveness to the environment is well supported by mathematical models (Wilson & Yoshimura, 1994; see also Wolf, van Doorn, & Weissing, 2008) but has not yet generated empirical work in humans. Another plausible but less explored possibility is that other types of balancing selection—for example spatially and temporally fluctuating selection—may coexist with frequency-dependent selection and contribute to maintain genotypic variation in human populations (see Ellis et al., 2011a). Development of Differential Susceptibility Through Conditional Adaptation Depending on the structure of the environment and of an organism’s life cycle, conditional adaptation can be maintained alongside genetic variation by spatially and temporally fluctuating selection pressures (see Del Giudice, 2012, under review). The theory of Biological Sensitivity to Context (Boyce & Ellis, 2005) posits a nonrandom distribution of neurobiological susceptibility to the environment within populations that emerges through conditional adaptation to variable environmental conditions. From

37

this perspective, differential susceptibility results, in part, from individuals tracking different environmental conditions and altering their development to match those conditions. The assumption is that this matching process promoted fitness across heterogeneous environmental contexts over human evolution. Biological sensitivity to context theory identified physiological mechanisms of differential susceptibility— autonomic, adrenocortical, and immune reactivity to psychosocial stressors—and proposed that this psychobiologic reactivity moderated the effects of early environmental exposures on physical and mental health outcomes in the bivalent manner discussed above (reviewed in Pluess & Belsky, 2009; Ellis et al., 2011a). This dual function signified the need to conceptualize stress reactivity more broadly as biological sensitivity to context, which Boyce and Ellis (2005) defined as neurobiological susceptibility to both cost-inflicting and benefit-conferring features of the environment and operationalized as heightened reactivity in one or more of the stress response systems. In total, biological sensitivity to context theory proposed that individual differences in the magnitude of biological stress responses function to regulate openness or susceptibility to environmental influences, ranging from harmful to protective. As discussed in the previous section, the ACM explicitly builds on the logic of biological sensitivity to context. In a nutshell, the theory of biological sensitivity to context proposes that heightened susceptibility is adaptive at both ends of the environmental continuum—in both highly stressful and highly protected environments. This results in a U-shaped curvilinear relation between early stress/adversity and susceptibility to later environmental effects (Boyce & Ellis, 2005). Consistent with the focus of biological sensitivity to context on adaptive plasticity, the ACM views early experience as a crucial determinant of individual differences in neurobiological susceptibility to the environment. Potential cues of danger and adversity versus safety and support include prenatal stress (e.g., fetal exposure to stress hormones), parenting quality, family stability and conflict, attachment security, and neighborhood quality. These contextual factors are hypothesized to affect development mainly by upregulating or downregulating the activity of infants’ and children’s stress response systems, as shown in Figure 1.5. Whereas the original conceptualization of biological sensitivity context focused primarily on the functions of high stress responsivity, the ACM expanded this conceptualization to encompass the functions of low stress responsivity—and thus added

38

Evolutionary Foundations of Developmental Psychopathology

the unemotional pattern shown in Figure 1.5. This addition turned the original U-shaped curve into a more complex function (see discussion of the ACM responsivity patterns in the previous section). Like the theory of biological sensitivity to context, the ACM conceptualizes variation in neurobiological susceptibility to the environment as part of a conditional adaptation process that matches stress response profiles to local conditions. This conceptualization is supported by a large empirical literature demonstrating changes in SNS, PNS, and HPA parameter values over development in response to different developmental experiences (reviewed in Boyce & Ellis, 2005; Del Giudice et al., 2011), including prenatal experiences such as infections, fetal undernutrition, and exposure to maternal stress hormones (Pluess & Belsky, 2011). Differential Susceptibility as Adaptive Stochastic Variation According to the theory of differential susceptibility advanced by Belsky (1997, 2005), the main adaptive function of differential susceptibility is spreading the risk of mismatch by making some individuals resistant to environmental influences, including—but not limited to—the influences of parental behavior on children’s development. When early cues correctly predict future states of the environment, susceptible individuals benefit from enhanced phenotype-environment matching; but when prediction fails (for example because the environment undergoes sudden and unpredictable changes), they end up developing a mismatched phenotype and potentially suffering large fitness costs as a result. Individuals who follow a fixed developmental trajectory may avoid the fate of their more susceptible counterparts when ecological cues fail to predict later environmental states. Thus, for a parent, producing offspring with varying degrees of susceptibility works as an insurance against future unpredictability. In evolutionary terms, this is an example of reproductive bet-hedging—a strategy that reduces average individual fitness in the short term, but enhances the long-term reproductive success of the genetic lineage by decreasing fitness variance across generations (Bull, 1987; Seger & Brockmann, 1987; Starrfelt & Kokko, 2012). More specifically, differential susceptibility is conceptualized as an instance of diversified bet-hedging, that is, a strategy that reduces fitness variance between generations by increasing phenotypic variability among offspring (Philippi & Seger, 1989). Such a strategy could be maintained by an evolutionary history of exposure to environments that fluctuated unpredictably over time. Bet-hedging strategies

increase the probability of achieving some reproductive success in every generation while limiting success in good conditions and shielding against total failure in bad (for a detailed treatment see Starrfelt & Kokko, 2012). Theory and data from evolutionary biology indicate that fluctuating selection pressures, if sufficiently strong, can support variable or random generation of offspring phenotypes (“adaptive coin-flipping”) arising from a monomorphic genetic structure. This strategy can be implemented through a stochastic developmental switch (presumably instantiated by epigenetic mechanisms), which generates one of several alternative phenotypes according to a probabilistic rule. For example, in a range of animal species, when mothers cannot forecast the likely environment of their offspring, or environmental cues in the maternal generation suggest that the offspring environment is likely to vary unpredictably, mothers hedge their bets by increasing variation in offspring phenotypes (Crean & Marshall, 2009). Although temporally fluctuating selection pressures can select for stochastic phenotypic variation, diversified bet-hedging cannot be instantiated through genetic polymorphisms because the resulting phenotypes do not all have the same average fitness (Bull, 1987; Philippi & Seger, 1989). Thus, the bet-hedging hypothesis can explain adaptive stochastic variation in susceptibility but not genotypic variation, in contrast with early formulations of the theory (Belsky, 1997). More recent formulations of the theory (e.g., Belsky, 2005; Pluess & Belsky, 2011) also acknowledge the role of early environmental influences—especially prenatal exposure to stress—in shaping individual levels of susceptibility. In total, the bet-hedging hypothesis of differential susceptibility is valuable in that it provides a plausible adaptive explanation for stochastic variation in susceptibility to environmental influence. More generally, the larger framework advanced by Belsky (2005) emphasizes the potential adaptive significance of unsystematic within-family variability in susceptibility to the environment. However, adaptive stochastic variation coexists with many sources of neutral or maladaptive unsystematic variation, which result from the random effects of sexual recombination, nonadaptive phenotypic plasticity, random perturbations in developmental processes, harmful genetic and epigenetic mutations, and so forth. The empirical challenge for the theory advanced by Belsky (1997, 2005) is that it may be extremely difficult to distinguish between adaptive within-family variation that has been shaped by natural selection and within-family variation that is truly random and non-adaptive (see Ellis et al., 2011a).

Beyond Diathesis-Stress: Differential Susceptibility to Environmental Influences

Differential Susceptibility as a Model of Organism–Environment Interplay: The Case of Pubertal Development The differential susceptibility framework returns us to the centrality of the phenotype and the complexity of organism–environment interplay. Susceptibility to the environment is instantiated in the biology of the nervous system; it is neurobiological susceptibility. Genetic susceptibility factors operate through neurobiological processes, and behavioral indicators of susceptibility are grounded in neurobiology. Neurobiological susceptibility itself is not a static trait; it develops and changes in response to genetic and environmental influences, which become incorporated into the phenotype over time. Genetic factors, environmental factors, and the extant phenotype all interact in intricate ways that are not yet entirely understood, including both systematic effects and stochastic processes. A good demonstration of the real-life complexity of differential susceptibility is the interplay of early life stress, stress responsivity, and polymorphic variation in the estrogen receptor 𝛼 gene (ESR1) in the regulation of pubertal development. The timing of pubertal development is a central life history trait; as discussed in a previous section, Belsky and colleagues (1991) first predicted that early stress would entrain fast life history strategies, leading to accelerated puberty timing (the psychosocial acceleration theory; see Ellis, 2004). Specifically, the theory posits that ecological stressors in and around the family create conditions that undermine parental functioning and lower the quality of parental investment—such as by escalating marital conflict, increasing negativity and coercion in parent–child relationships, and reducing positivity and support in parent–child relationships. According to the theory, children respond to these familial and ecological conditions (particularly those experienced in the first 5–7 years of life) by developing in a manner that speeds up pubertal maturation, anticipates the onset of sexual activity, and promotes the development of a cluster of behavioral traits associated with fast life histories, including impulsivity and unstable pair bonds. A study by Ellis and Essex (2007) investigated the psychosocial acceleration theory in the Wisconsin Study of Families and Work. Consistent with predictions, higher quality parent–child relationships in preschool—more parental warmth and family positivity, less parent–child stress and conflict—forecast slower pubertal maturation in daughters. Although this association proved robust, the unique effect of parent–child relationships on puberty was

39

relatively small. However, theories of differential susceptibility suggest that the weak main effects of environmental variables on many developmental outcomes may often reflect the fact the children differ in whether, how, and how much they are affected by rearing experiences. As articulated by Belsky (2000), the weak main effects of parenting variables on pubertal timing may overestimate the impact of family environments in some children (dandelions, more fixed reproductive development) and underestimate it in others (orchids, more plastic reproductive development). This hypothesis was supported in a later study of the same data set (Ellis, Shirtcliff, Boyce, Deadorff, & Essex, 2011b). Specifically, higher SRS responsivity in first grade—a key facet of neurobiological susceptibility— moderated the effects of parental behavior on maturation: lower quality parent–child relationships forecast faster initial tempo of puberty and earlier pubertal timing, but only among children showing heightened SNS/HPA reactivity. In other words, the data revealed a clear PxE interaction between parent–child relationships and SRS responsivity, consistent with the biological sensitivity to context account of differential susceptibility (Boyce & Ellis, 2005). This, however, is only part of the story. Neurobiological susceptibility is not a fixed trait but develops over time through the interplay of various causal factors—including genetic variation and early environmental effects. In the same dataset, growing up in either highly supportive or highly stressful home environments predicted development of high SNS reactivity (consistent with the U-shaped curve predicted by biological sensitivity to context theory; see previous discussion), whereas growing up in a highly stressful home environment predicted heightened HPA activation (Ellis, Essex, & Boyce, 2005). Thus, under conditions of early environmental stress and uncertainty, indexed by coercive and unsupportive family relationships, individuals developed heightened neurobiological susceptibility to the environment (as indexed by SNS and HPA activation) and subsequently accelerated pubertal maturation in early adolescence, with family stress and neurobiological susceptibility acting synergistically in this process. Heightened neurobiological susceptibility thus enabled a stronger pubertal response to adversity. According to psychosocial acceleration theory, this response may represent a strategic—that is, functional—way of developing under stress. The other side of the coin is that under conditions of early environmental protection and stability, indexed by positive and supportive family relationships, individuals also developed heightened neurobiological susceptibility,

40

Evolutionary Foundations of Developmental Psychopathology

which in combination with high-quality parent–child relationships forecast slower initial pubertal tempo and later pubertal timing (Ellis et al., 2011b). In this case, heightened neurobiological susceptibility enhanced responsiveness to environmental resources and support. As suggested by life history models, the resulting pattern of late sexual maturation may also constitute adaptive variation. Specifically, Ellis (2004) hypothesized that children have been selected to capitalize on the benefits of high quality parental investment and reduce the costs of low quality parental investment by contingently altering the length of childhood. Given high family resources and support and biological sensitivity to these development-enhancing contexts, extending childhood by delaying onset of puberty or slowing pubertal tempo may function to improve socio-competitive competencies (i.e., embodied capital) that ultimately increase reproductive potential. Timing of puberty depends not only on the environment; indeed, all measures of pubertal development also show heritability values indicative of robust genotypic effects (see Ellis, 2004). As can be expected from the vantage point of differential susceptibility theory, recent evidence indicates that GxE interactions are also involved in the regulation of pubertal development (Manuck, Craig, Flory, Halder, & Ferrell, 2011). Consistent with past research (reviewed in Ellis, 2004), Manuck and colleagues (2011) found that women who reported being raised in families characterized by distant interpersonal relationships and high levels of conflict tended to reach menarche earlier than women raised in close families with little discord. However, this effect was moderated by variation in the gene coding for estrogen receptor 𝛼 (ESR1). Among women who were homozygous for minor alleles of the two ESR1 polymorphisms examined in the study, a childhood history of low-quality family relationships was associated with earlier age of menarche compared with a childhood history of high-quality family relationships; no such effect was found among women with other ESR1 genotypes. Intriguingly, estrogen receptor 𝛼 is highly expressed in the hypothalamus, where it regulates the activity of the HPA axis (Bao, Meynen, & Swaab, 2008). Thus, among its other possible effects, variation in ESR1 may affect pubertal development by contributing to the development of higher or lower levels of neurobiological susceptibility. In summary, regulation of pubertal development depends on genetic factors and GxE interactions, such as between ESR1 variation and family stress; on environmental factors, such as energetics and psychosocial stress; on interactions between these environmental factors and extant phenotypic characteristics that modulate

neurobiological susceptibility to the environment, such as SRS responsivity; on the developmental calibration of these stress response systems, which filter and embed information about environmental stress and support, mediating the organism’s openness to environmental inputs; and on genetic influences and GxE interactions in the regulation of neurobiological susceptibility to the environment (see extended discussion in Ellis, 2013). At the nexus of all of these processes is the organized phenotype, which exists from conception; modulates, integrates, and retains genetic and environmental effects; and is the basis of differential susceptibility. Yet even if all of these factors are taken into account, there is still be much unexplained variation due to stochastic developmental processes. This is the tangled web of development. Theoretical models propose that differential susceptibility develops through a mixture of genetic regulation—maintained by balancing selection—and environmental regulation that enables conditional adaptation. Whereas balancing selection and conditional adaptation result in systematic variation in neurobiological susceptibility to the environment, unsystematic variation may also be maintained by natural selection as an insurance policy against unpredictably changing environments (diversified bet-hedging). Taken together, these evolutionary processes may result in large individual differences in whether, how, and how much people are affected by their experiences. Differential susceptibility, therefore, can be expected to play a key role in moderating the effects of environmental conditions on developmental outcomes, including the development of life history strategies. Implications for the Core Points of Developmental Psychopathology The theory of differential susceptibility brings about a fundamental change in the way one thinks about risk and resilience in development. First, the same genetic or phenotypic factor can behave as both a risk and a protective factor depending on the ecological context. Second, the very children whose heightened responsivity appears to make them vulnerable to developing psychopathology (orchid children) may also be most able to benefit from positive, supportive environments and interventions. The divergent outcomes associated with highly susceptible phenotypes also set the stage for pervasive, systematic manifestations of multifinality in development. As we have seen, the development of susceptibility is itself a dynamic process, taking place through an intricate interplay of genetic, environmental, and stochastic

Beyond the DSM: A Life History Framework for Mental Disorders

effects. This perspective combines the contextualism of conditional adaptation with new and potentially fruitful explanations of probabilistic causality. Whereas standard approaches tend to see probabilistic causality as opposed to biological adaptation, the theory of bet-hedging shows how natural selection can make adaptive use of developmental randomness as a response to unpredictable environmental change. Still other implications for developmental psychopathology stem from the centrality of the phenotype in regulating differential susceptibility. First, a view of the phenotype as the result of sequential organism–environment interplay blurs the distinction between variable-centered and person-centered approaches. Second, the concept of neurobiological susceptibility is a powerful reminder that genetic, epigenetic, and environmental effects must ultimately be understood in terms of their influence on brain and neurobiological pathways; it also illustrates how multiple pathways and molecules—for example serotonin, dopamine, cortisol, and sex hormones—may exert synergistic effects by converging on general, adaptive dimensions of phenotypic variation. BEYOND THE DSM: A LIFE HISTORY FRAMEWORK FOR MENTAL DISORDERS In the preceding sections we showed that EDP provides an integrative biological perspective on human development, and sheds light on the origin and function of individual differences in life history strategy, developmental plasticity, and physiological responsivity to stress. We now take this approach one step further and show how the principles of life history theory can be employed to outline a unifying framework for the analysis and classification of mental disorders (for a detailed exposition see Del Giudice, 2014a, 2014b). For the sake of simplicity and consistency with standard usage, in this section we employ the term disorder as a synonym for diagnosable condition—regardless of whether the condition represents a harmful dysfunction in the narrow sense (Wakefield, 1992). Limitations of Current Taxonomic Approaches The dominant approach to the classification of mental disorders is that of the Diagnostic and Statistical Manual (DSM) of the American Psychiatric Association (2013). In the DSM, disorders are defined by lists of symptoms, and grouped together mainly on the basis of symptom similarity—in keeping with the atheoretical stance embodied by the Manual. Thus, DSM-5

41

categories include, for example, anxiety disorders; disruptive, impulse-control, and related disorders; depressive disorders; obsessive-compulsive and related disorders; and feeding and eating disorders (American Psychiatric Association, 2013). While the DSM system has many undisputable qualities—diagnostic reliability above all—it also has a number of significant problems (see, e.g., Beauchaine, Klein, Erickson, & Norris, 2013; Nesse & Jackson, 2006). The main limitation of the DSM is also its defining feature: the deliberate absence of a theoretical model of mental disorders. Since disorders and disorder categories are primarily defined by symptom similarity, many diagnostic classes are likely to include a heterogeneous mix of conditions with different etiological, developmental, and functional characteristics. More troubling, the DSM lacks a theory of normative mental functioning, and is therefore ill equipped to discriminate between adaptive defensive responses and disordered functioning (Nesse & Jackson, 2006). The main alternative to the DSM system comes from a family of empirical approaches based on patterns of genetic and phenotypic correlations between disorders (e.g., Kendler, Prescott, Myers, & Neale, 2003; Lahey, Van Hulle, Singh, Waldman, & Rathouz, 2011; Lahey et al., 2008; Watson, 2005). Empirical taxonomic studies suggest the existence of broad, hierarchically organized clusters of disorders that overlap only in part with DSM categories. The fundamental distinction in empirical taxonomies is that between internalizing and externalizing disorders. Externalizing disorders are characterized by impulsivity, disinhibition, and high levels of aggressive, antisocial, and/or disruptive behavior. Internalizing disorders are marked by high levels of anxiety and negative emotionality. The internalizing spectrum comprises a cluster of distress disorders (depression, generalized anxiety disorder [GAD], post-traumatic stress disorder [PTSD]) and a cluster of fear disorders (panic disorder, agoraphobia, social phobia, and specific phobias; Clark & Watson, 2006). Obsessive-compulsive spectrum disorders are typically treated as a separate internalizing cluster; other disorders that are usually included in the internalizing spectrum are bipolar disorders and borderline personality disorder (BPD), although their exact placement is more problematic (see Beauchaine & Hinshaw, 2013; Lahey et al., 2008; Watson, 2005). A recent factor-analytic study by Caspi and colleagues (2014) supplemented the internalizing and externalizing categories with a thought disorder factor comprising schizophrenia, mania (bipolar spectrum), and obsessive-compulsive disorder (OCD). Moreover,

42

Evolutionary Foundations of Developmental Psychopathology

the authors identified a general, higher-order factor of psychopathological risk they labeled the p factor (see Caspi et al., 2014). The internalizing–externalizing distinction has received considerable empirical support, and has become a standard tool in developmental psychopathology (Beauchaine & Hinshaw, 2013). In the latest edition of the DSM, the authors explicitly recognized the usefulness of the broad distinction between internalizing and externalizing disorders (American Psychiatric Association, 2013, p. 13). However, this approach also has a number of important limitations. First of all, disorders are grouped based on their emotional and affective characteristics; however, emotions can serve multiple motivational goals, and associations between emotions and motivational processes are often remarkably nonspecific (see Nesse, 2004a). For example, anger can be triggered by aggressive competition, by threats to one’s dominance or status, by suffering or witnessing acts of injustice, by separation from an attachment figure, and so forth. Anxiety, shame, and sadness are prominently associated with psychopathology, but their motivational specificity is also extremely low. Thus, classifications of disorders based on emotion and affect are unlikely to reliably capture the underlying motivational structure. Moving to the empirical level, there is mounting evidence that the interalizing–externalizing dichotomy is riddled with ambiguities and inconsistencies. To begin with, depression and GAD—often regarded as prototypical internalizing disorders—are in fact bridge diagnoses that overlap with both internalizing and externalizing disorders at the phenotypic, genetic, and developmental level (e.g., Lahey et al., 2008, 2011). Also, some disorders that are usually considered part of the internalizing spectrum—such as obsessive-compulsive disorder (OCD) and BPD—show atypically large correlations with externalizing disorders (Crowell, Kaufman, & Lenzenweger, 2013; Lahey et al., 2008). Finally, the interalizing–externalizing taxonomy excludes many important pathological conditions—notably schizophrenia, autism, and most personality disorders—because they are not primarily characterized by mood/emotional alterations and do not fit the conceptual distinction between internalization and externalization (the recent analysis by Caspi and colleagues [2014] is a partial exception). A Life History Framework for Psychopathology In the preceding sections we discussed how life history strategies play a central role in the organization

of physiology and behavior. They define an organism’s priorities and determine the allocation of effort and resources toward competing biological goals. Differences in life history strategy are the joint product of genetic and environmental influences on development, and are reflected in organized patterns of individual differences in motivation, affect, self-regulation, and personality. By organizing individual differences on such a broad scale, life history strategies set the stage for the development of psychopathology. More precisely, individual differences in life history strategy can be expected to determine individual differences in risk profiles for a broad range of mental disorders. As one moves along the fast–slow continuum of life history variation, some disorders and symptoms should become more frequent, while others should become less likely to occur. The predictable association between life history strategy and risk for psychopathology offers a high-level functional criterion for the classification of mental conditions. This leads to the novel distinction between fast spectrum and slow spectrum disorders—that is, disorders that cluster at the fast or slow end of the life history continuum (Del Giudice, 2014a, 2014b). Until recently, life history approaches to psychopathology have focused almost exclusively on the fast end of the fast-slow continuum. As widely recognized in the literature, fast life history strategies can predispose individuals to a variety of disorders, either as maladaptive outcomes of life history-related traits or potentially adaptive but undesirable behavioral strategies (e.g., Belsky et al., 1991; Brüne et al., 2010; Figueredo & Jacobs, 2010; Jonason, Li, Webster, & Schmitt, 2009; Salmon, Figueredo, & Woodburn, 2009). The framework advanced by Del Giudice (2014a) extends this approach by addressing the role of slow strategies in setting the stage for the development of mental disorders. It is important to stress that the functional connection between life history strategy and psychopathology is usually an indirect one. Causal pathways to psychopathology involve a multiplicity of traits and mechanisms—including temperament and personality, self-regulatory processes, and so forth. The general idea is that an individual’s configuration of life history-related traits may increase the likelihood of developing a certain disorder or cluster of disorders—often in interaction with other causal factors including developmental insults, deleterious genetic or epigenetic mutations, infections, nutritional deficits, and psychosocial stressors. The power of life history theory lies in the ability to integrate these diverse etiological processes within a common frame of reference. The result is a large-scale map of the psychopathological landscape

Beyond the DSM: A Life History Framework for Mental Disorders

organized along the fast-slow axis of life history variation. Such a map is an invaluable guide in understanding comorbidity patterns, since functionally related disorders—for example different disorders in the slow spectrum—can be expected to co-occur more frequently in the same individual. At the same time, the fast-slow distinction can be used to tease apart functionally distinct conditions that coexist within the same descriptive category because of their phenotypic similarity. In total, a functional analysis based on life history principles helps to “carve nature at its joints” by revealing commonalities between separate categories and suggesting important distinctions between phenotypically similar disorders (Keller & Nesse, 2006). Of course, mental disorders are complex biosocial phenomena, and as such they can be analyzed at many different levels. A life history analysis is only the first step toward a comprehensive functional account of psychopathology: the broad perspective afforded by the fast–slow distinction should be complemented by narrower accounts focusing on specific motivational/behavioral systems, cognitive mechanisms, genetic pathways, and so forth. Four Pathways From Life History Strategy to Psychopathology The general statement that life history strategies set the stage for the development of psychopathology can be supplemented by a finer-grained analysis of the causal pathways that lead to the onset of mental disorders. First of all, some adaptive life history-related traits may be regarded as symptoms. This is most likely to happen with fast life history strategies characterized by impulsive, exploitative, or aggressive tendencies. The resulting phenotype may be classified as a disorder, even if it does not reflect maladaptive or dysfunctional processes. Even if they are biologically adaptive, or used to be adaptive in ancestral environments, such strategies may often involve substantial costs in terms of health and emotional well-being. Another important category of adaptive traits that may be diagnosed as symptoms of a disorder is that of aversive defenses. When defenses activate inappropriately or respond with excessive intensity, the outcome may be correctly recognized as maladaptive. However, many protective mechanisms have strongly aversive effects, and can be occasionally harmful to the individual. For this reason, they may give rise to undesirable conditions not only when they misfire but also when they respond appropriately in presence of actual threats. The correlates of life history strategies often include upregulation or downregulation of psychological and

43

physiological defensive mechanisms. Upregulated defenses have a lower threshold for activation and/or respond with higher intensity when they activate. Defense upregulation can be associated with both fast and slow strategies, although the specific type of mechanism involved is likely to differ between the two. In the context of fast life histories, upregulated defenses help protect the individual from immediate danger in risky, unpredictable environments. In the context of slow strategies, upregulated defenses may help the individual prevent dangerous events and avoid potentially risky situations, even if the current environment is reasonably safe. Moreover, protecting oneself from even minor damages and losses contributes to the long-term maintenance of the soma (i.e., somatic effort)—a key priority for slow life history individuals. In contrast, downregulation of defenses is most likely in the context of fast life history strategies, especially those involving a high degree of risk taking. As discussed in a previous section, the underlying logic is that, to fulfill their purpose, such strategies require insensitivity to threats, dangers, and so forth. The second pathway from life history strategy to psychopathology derives from the fact that life history-related traits may be expressed at maladaptive levels. Even phenotypic traits that are biologically adaptive within a certain range may become maladaptive if they exceed that range. Sometimes, the expected fitness associated with a trait may slowly increase up to an optimal level, then decrease abruptly following a “cliff-edged” function. In such cases, selection for optimal trait levels may result in a high frequency of maladaptive phenotypes that overshoot the fitness optimum (Nesse, 2004b). A trait can reach maladaptive expression levels owing to a combination of genetic, epigenetic, and environmental factors that contribute to push the phenotype in the same direction. In the simplest case, extreme levels of a trait may appear in the offspring of two individuals who are both high on that trait, yet still within the adaptive range. Thus, assortative mating—the tendency for sexual partners to be more similar than average on a certain trait—can increase the risk for psychopathology due to extreme trait values. Parent–offspring conflict and intragenomic conflict are other likely causes of maladaptive trait expression. When conflict is present, phenotypic development can be conceptualized as the result of opposing forces, much like a game of tug-of-war. If for any reason this dynamic equilibrium is broken (for example, a mutation in the offspring may make it unable to counteract parental manipulation), the resulting imbalance may easily determine dysregulated or pathological outcomes.

44

Evolutionary Foundations of Developmental Psychopathology

In principle, the pathway leading from maladaptive trait expression levels to psychopathology may involve traits associated with both fast and slow life histories. However, there is some evidence that assortative mating on life history-related traits in humans tends to become stronger toward the slow end of the continuum (Figueredo & Wolf, 2009). If so, disorders that involve maladaptive expression levels of adaptive traits should occur more frequently in association with slow strategies, as similarity between parents increases the likelihood that offspring will inherit extreme genotypic combinations. Third, as noted in a previous section, even adaptive strategies may yield individually maladaptive outcomes. Risky strategies are a prime candidate as a systematic source of individually maladaptive outcomes. Life history-related traits can steer individuals on high-risk pathways, thus increasing the likelihood of maladaptive and/or undesirable outcomes in case of strategy failure— even when the strategy is adaptive on average. This is more likely to happen in the context of fast life history strategies, which tend to promote risk taking and favor the pursuit of large, immediate returns regardless of the potential costs. While some individuals engaging in high-risk strategies may end up developing mental disorders, other individuals expressing the same traits may enjoy desirable or biologically adaptive outcomes, often depending on chance and unpredictable contextual factors. Another important category of adaptive traits that systematically produce maladaptive outcomes is that of defensive mechanisms. Following the logic of the smoke detector principle, defensive mechanisms are usually designed to misfire occasionally, even in absence of threats. Individual differences in life history strategy are reflected in the calibration of behavioral and physiological defenses (see earlier discussion of the ACM), and indirectly affect the risk of inappropriate defense activation. Fourth and last, adaptive life history-related traits may increase vulnerability to dysfunction. While life history traits are designed to promote adaptation, they can nevertheless increase vulnerability to some types of dysfunction as a side effect. For example, some configurations of personality traits within the adaptive range (for example schizotypy or autistic-like personality) may become especially conducive to psychopathology when they are coupled with high mutation load or brain-damaging infections (see Del Giudice, 2010). Also, fast life history-related traits such as risk proneness and future discounting may indirectly increase an individual’s exposure to environmental factors such as pathogens. Finally, upregulated defensive systems are not only more prone to misfiring but also more vulnerable to actual instances of malfunction and

dysregulation (Nesse, 2001). These four pathways from life history strategy to psychopathology are logically distinct but not mutually exclusive and may coexist in the etiology of any given disorder. Correlates of Fast and Slow Spectrum Psychopathology The conceptual distinction between fast and slow spectrum pathology provides a powerful heuristic criterion for the functional classification of mental disorders. Whatever the specific causal pathway (or combination of pathways) that determine the onset of a given disorder, fast spectrum conditions will be associated with traits such as low agreeableness and conscientiousness, impulsivity, disinhibition, and early sexual maturation (especially in females). Conversely, slow spectrum conditions will exhibit a signature of slow life history-related traits in the areas of motivation, self-regulation, personality, and sexual maturation. Correlations between life history–related traits and specific disorders may or may not imply a causal role of those traits in the etiology of the disorders. However, regardless of their role in the etiology of a given disorder, life history correlates can be employed as convergent markers of the underlying life history strategy. In principle, this approach can be extended to include genetic, epigenetic, and neurobiological markers (e.g., Del Giudice et al., 2011; Figueredo et al., 2004, 2006). A nonexhaustive list of markers of fast and slow spectrum psychopathology is presented in Table 1.2. A life history perspective yields novel predictions about the environmental correlates of mental disorders. Ecological harshness and unpredictability tend to entrain development of fast life history strategies, while slow strategies are favored in safe and predictable contexts. As a result, many classic risk factors for psychopathology—such as stressful life events, low socioeconomic status, negative family relationships, trauma, and abuse—are predicted to increase the occurrence of fast spectrum disorders, but not that of slow spectrum disorders. On the contrary, slow spectrum disorders should be associated—at least on average—with safe, predictable environments, higher socioeconomic status, and reduced exposure to ecological and family stressors. Sex Differences If life history strategies set the stage for psychopathology, sexual asymmetries in life history trade-offs should produce consistent patterns of sex differences in the epidemiology of mental disorders. The first key asymmetry concerns the mating versus parenting trade-off. On average, human males invest more in mating effort and less in parenting effort than females. The intensity

Beyond the DSM: A Life History Framework for Mental Disorders TABLE 1.2 Correlates of fast and slow spectrum psychopathology

Motivation

Fast spectrum psychopathology

Slow spectrum psychopathology

Social antagonism

Social compliance, conformity Stable attachments Delayed sexuality Sexual restraint, low sex drive Preference for routines

Unstable attachments Precocious sexuality Sexual promiscuity, high sex drive Sensation/novelty seeking Risk taking

Self-regulation

Personality traits

Risk aversion, harm prevention

Disinhibition, impulsivity Discounting of future rewards

Inhibition, restraint

Low conscientiousness Low agreeableness

High conscientiousness High agreeableness

Discounting of immediate rewards

Sexual maturation Early, fast maturation

Late, slow maturation

Environment

Safe, predictable Low exposure to stressors

Harsh, unpredictable High exposure to stressors

Source: Reprinted from M. Del Giudice, An evolutionary life history framework for psychopatholgy, Psychological inquiry, 25, 274, 2014a.

of mating effort increases sexual selection for competitive traits such as risk taking, dominance-seeking, and physical aggression (see Archer, 2009; Wilson et al., 2002). In total, higher mating effort in males should predispose them to fast spectrum disorders characterized by high levels of risk taking, such as those in the externalizing cluster (Martel, 2013). In contrast, females have generally less to gain and more to lose from high-risk strategies than males and can be expected to invest more effort in somatic maintenance and protection. As a consequence, they should be more prone to develop disorders that involve the upregulation of protective defenses, and to exhibit more psychological and physiological symptoms reflecting defense upregulation. This prediction applies to disorders across the fast-slow continuum, since upregulated defenses can be functionally associated with both fast and slow life history strategies. The higher incidence of anxiety disorders in females (see Martel, 2013) is consistent with this prediction. Another important asymmetry in life history strategy concerns the trade-off between current and future reproduction. As already discussed in the section on life history theory, this trade-off plays a more critical role in the organization of female life history strategies, since decisions concerning reproductive timing are more critical for females than for males. As a consequence, the timing of sexual maturation in females should be more sensitive

45

to cues of danger and unpredictability (James et al., 2012). Indeed, the available data suggest that ecological stress in the first years of life anticipates gonadal puberty in girls, but not in boys (reviewed in Belsky, 2012). In addition, indices of sexual maturation in females can be expected to form a tighter cluster with other life history-related traits including motivation, personality, self-regulation, and so forth. It follows that maturation timing and rate should be stronger predictors of psychopathology in females than in males. This prediction is well supported by empirical research; the bulk of evidence indicates that individual differences in sexual maturation are more robustly associated with psychopathology in girls than in boys (see Ge & Natsuaki, 2010; Mendle, Turkheimer, & Emery, 2007). Predictions about sex differences based on life history theory (Del Giudice, 2014a) can be integrated with those from a recent evolutionary model advanced by Martel (2013). Martel employed sexual selection theory to explain the male-biased prevalence of childhood-onset externalizing disorders and the symmetrical, female-based prevalence of adolescent-onset internalizing disorders. Because of differential sexual selection for social dominance versus interpersonal competence in males and females, she also predicted that males should be more sensitive to early environmental stressors related to the broader ecological conditions (including those occurring prenatally), whereas females should be more sensitive to interpersonal stressors occurring around puberty. These predictions are supported by considerable empirical evidence, and can be extended to the neurobiological level to yield insight in the role of prenatal testosterone, dopamine, and serotonin in the etiology of common mental disorders in the two sexes (reviewed in Martel, 2013). Toward a Life History Taxonomy of Mental Disorders The general framework outlined in this section can be applied to individual disorders and categories of disorders, yielding an initial life history taxonomy based on the fast-slow continuum. In what follows, we briefly discuss how six common categories of mental disorders relate to the fast–slow distinction. For an extended analysis and review of the relevant empirical literature, see Del Giudice (2014a, 2014b). Externalizing disorders The externalizing spectrum comprises various disorders marked by aggressive, antisocial, and/or disruptive behavior, including oppositional defiant disorder (ODD), conduct disorder (CD), and antisocial personality disorder (APD). Externalizing disorders are also associated with

46

Evolutionary Foundations of Developmental Psychopathology

high risk for substance abuse. Disorders in the externalizing spectrum are strongly male-biased and show high phenotypic and genetic correlations with one another, indicating the existence of a coherent, heritable dimension of externalizing behavior. At the same time, the development of externalizing behaviors is strongly conditioned by environmental factors, such as harsh-rejecting parenting, aggregation of high-risk youth in after-school programs, and exposure to neighborhood violence and criminality (reviewed in Beauchaine, Hinshaw, & Pang, 2010). In a life history perspective, externalizing spectrum disorders are prototypical instances of fast spectrum psychopathology. Externalizing symptoms are associated with impulsivity and undercontrol, early puberty timing and fast sexual maturation in both sexes, earlier onset of sexual activity, and larger numbers of partners in adolescence and young adulthood. Low socioeconomic status, harsh or unpredictable parental discipline, parental conflict, family disruption, and child abuse—all cues of danger and unpredictability—are consistent predictors of externalizing behavior, consistent with predictions derived from life history theory. Evolutionary models of externalizing spectrum disorders tend to stress the potential biological adaptiveness of aggressive, exploitative, and risky behavior—especially when coupled with promiscuous short-term sexuality (e.g., Barr & Quinsey, 2004; Figueredo & Jacobs, 2010; Jonason et al., 2009; Mealey, 1995). Accordingly, many evolutionary scholars see externalizing disorders as adaptive but undesirable constellations of traits. In some instances, externalizing disorders may represent maladaptive extremes of potentially adaptive traits (see MacDonald, 2012). It should be stressed that externalizing disorders can be adaptive even if their social outcomes are negative on average. This can happen if successful outcomes yield disproportionate fitness returns, even in a minority of cases. Finally, high-risk behavioral strategies are likely to involve downregulation of defensive mechanisms; indeed, externalizing disorders in adolescents and adults are often associated with reduced anxiety, fearlessness, and dampened responsivity of the SRS. Schizophrenia Spectrum Disorders Schizophrenia is a family of mental disorders characterized by delusions, hallucinations, and cognitive disorganization. Given the severe reduction in reproductive success associated with a schizophrenia diagnosis, most evolutionary scholars regard this disorder as a maladaptive outcome of dysregulated sociocognitive processes (e.g., Burns, 2004; Keller & Miller, 2006). While schizophrenia spectrum

disorders (SSDs) are highly heritable, schizophrenia risk is also increased by adverse environmental factors such as nutritional deficiencies, infections, and birth complications. This suggests that accumulated deleterious mutations and environmental insults may converge on common neurobiological pathways, increasing the risk of cognitive breakdown. Even if SSDs are biologically maladaptive conditions, there may be evolutionary advantages associated with schizotypal traits—a constellation of personality traits associated with increased risk of psychosis. Most individuals who have psychotic experiences at some point in their life recover completely, and never transition to a diagnosable SSD. Various authors have proposed that schizotypal traits may be maintained by sexual selection processes based on mate choice. According to the sexual selection model of schizotypy (Nettle, 2001, 2006; Shaner, Miller, & Mintz, 2004), schizotypy-increasing alleles affect brain processes so as to increase traits such as verbal and artistic creativity, thus conferring mating advantages on those individuals who do not develop a psychiatric condition. However, the outcomes of schizotypy may be either beneficial (mating success) or harmful (schizophrenia), depending in part on the individual’s genetic quality (i.e., lack of deleterious mutations) and developmental condition (e.g., good nutrition and low exposure to pathogens). Consistent with the sexual selection model, positive schizotypal traits—unusual cognitive and perceptual experiences, tendency to magical ideation, reference and paranoid thoughts—are associated with verbal and artistic creativity, larger numbers of sexual partners, unrestricted sociosexuality, and reduced investment in long-term couple relationships. Large-scale studies of patients and their relatives show a robust familial association between schizophrenia and creativity. Schizotypal traits peak in adolescence/young adulthood and show a marked decline with age, mirroring typical changes in mating effort. In addition, positive schizotypy is associated with lower agreeableness and higher levels of aggression, suggesting a degree of overlap between the schizophrenia spectrum and the externalizing spectrum. In light of this convergent evidence, SSDs can be provisionally classified as belonging to the fast spectrum of psychopathology, although there are reasons to expect a degree of functional heterogeneity (see Del Giudice, 2014a). According to sexual selection models, schizotypy can be understood as a high-risk strategy oriented toward short-term mating, whose negative outcomes become manifest as schizophrenia and other SSDs. Alternatively,

Beyond the DSM: A Life History Framework for Mental Disorders

the milder disorders of the schizophrenia spectrum (e.g., schizotypal personality disorder, brief psychotic disorder) may result from maladaptive levels of expression of potentially adaptive traits associated with fast life history strategies. Autism spectrum disorders The autism spectrum comprises disorders of variable severity characterized by impairments in social interaction, communication problems, and restricted and repetitive behaviors/interests. Severe autism is almost certainly maladaptive, and some theorists have focused specifically on the negative aspects of autism spectrum disorders (ASDs). For example, Shaner and colleagues (Shaner, Miller, & Mintz, 2008) hypothesized that autism—like schizophrenia—may represent the negative extreme of a fitness indicator, a hypothesis consistent with the large number of deleterious mutations found in ASD patients. This negative emphasis should be balanced by accumulating evidence that autistic-like traits in the normative range—also known as the broader autistic phenotype— have a number of desirable and potentially adaptive correlates. Specifically, autistic-like traits predict higher systemizing abilities and attention to detail, better visuospatial skills, and enhanced low-level sensory processing in the visual and auditory domains. The autistic facets of repetitive behaviors, restricted interests, and detail-oriented cognitive style are associated with the development of outstanding talents in children. More generally, autistic-like traits are higher in people with technical-scientific interests and careers. Accordingly, several theorists have argued that ASDs can be seen as extreme and usually maladaptive manifestations of otherwise adaptive traits (e.g., Baron-Cohen, 2003; Crespi & Badcock, 2008). In this perspective, Del Giudice and colleagues (2010) hypothesized that sexual selection may contribute to maintain autistic-like traits in the population despite the fitness costs of severe ASDs. Specifically, they argued that autistic-like traits in their non-pathological form contribute to a male-typical strategy geared toward high parental investment, low mating effort, and long-term allocation of resources—in other words, a male-typical manifestation of slow life history strategy. This hypothesis offers a parsimonious explanation of the male-biased distribution of both autistic-like traits and ASDs. In support of this hypothesis, autistic-like traits predict lower interest in short-term mating, increased investment of time and resources in one’s partner, and stronger commitment to long-term romantic relations. People high in autistic-like traits report shorter duration of friendships but longer

47

duration of romantic relationships, and their partners are on average just as satisfied as those of people low in autistic-like traits. In a life history perspective, ASDs are thus likely candidates for inclusion in the slow spectrum of psychopathology. Further evidence comes from the finding that sexual maturation is delayed in women high in autistic-like traits as well as in women with ASD. Autistic-like traits may function adaptively as part of a slow life history strategy—especially in males—and become maladaptive only when they cross a certain threshold. Given the remarkable heterogeneity of ASDs, this functional explanation is likely to apply only to a subset of people diagnosed with autistic disorders. Different ASD subtypes may well require different explanations. The existence of functionally distinct subtypes of ASDs may explain the inconsistent correlation of autism risk with socioeconomic status in epidemiological studies. Obsessive-Compulsive Spectrum Disorders Disorders in the obsessive-compulsive spectrum are primarily characterized by patterns of compulsive, repetitive thoughts and/or behaviors, usually associated with worry and anxiety. In addition to obsessive-compulsive disorder (OCD), the OC spectrum includes body dysmorphic disorder, hoarding disorder, grooming disorders (skin picking and hair pulling), and obsessive-compulsive personality disorder (OCPD)—a pervasive profile of orderliness, rigid perfectionism, and need to control one’s self and environment. In the evolutionary literature, OCD is usually treated as a maladaptive exaggeration of an adaptive trait or the result of a dysfunction in precautionary cognitive systems. However, the milder forms of the disorder are not necessarily maladaptive in the biological sense. Current models converge on the idea that the main functional substrate of OCD is an adaptive mechanism—the hazard-precaution system or security motivation system—specialized for dealing with potential low frequency threats such as food poisoning (e.g., Boyer & Lienard, 2006; Szechtman & Woody, 2004; Woody & Szechtman, 2011). The peculiar logic of potential threats explains many features of compulsions (see Woody & Szechtman, 2011); obsessions can be explained as the involuntary generation of potential risk scenarios, a mechanism designed to increase future harm avoidance. Consistent with a threat prevention account and with the prediction that females should be more likely to develop symptoms reflecting upregulated defenses, adult OCD patients are overwhelmingly women.

48

Evolutionary Foundations of Developmental Psychopathology

A life history analysis indicates that the OC spectrum is best understood as a functionally heterogeneous category comprising two clusters of disorders—a slow spectrum one and a fast spectrum one. Slow-spectrum OCD is marked by reactive obsessions (Lee & Kwon, 2003); reactive obsessions concern realistic fears of contamination, mistakes, accidents, or disarray. They are triggered by cues of potential threats and are typically followed by preventive behaviors such as ordering or cleaning; anxiety is directed at the possible consequences of one’s actions rather than at the obsession itself. Reactive obsessions are associated with high conscientiousness, perfectionism, heightened responsibility and personal standards, normal levels of motor and cognitive inhibition, and a prevalence of contamination/cleaning symptoms. Reactive OCD fits straightforwardly in the slow spectrum of psychopathology, as a combination of exaggerated trait expression, upregulation of adaptive defenses, and dysfunctional protective responses. Obsessive-compulsive personality disorder (OCPD) also fits this classification, given its many overcontrol features and strong association with conscientiousness. Fast-spectrum OCD is characterized by autogenous obsessions—obsessions with sexual, aggressive, or blasphemous content. Autogenous obsessions tend to be bizarre, ego-dystonic, and threatening. They often have no apparent trigger, or are triggered by remote/bizarre thought associations (Lee & Kwon, 2003). Autogenous obsessions are associated with positive schizotypy, indices of psychotic thought disorganization, low conscientiousness, and reduced inhibitory control. The heterogeneous nature of the OC spectrum explains why OCD shows high comorbidity with both ASDs and SSDs. The two OC clusters can be expected to show markedly different epidemiological profiles; for example, traumatic events and low SES should be more strongly associated with fast-spectrum OCD, whereas slow-spectrum OCD should often arise in safe and predictable environments. Consistent with placement in the slow-spectrum cluster, OCPD is uniformly associated with high education levels, and OCDP patients have the highest socioeconomic status of all personality disorders (for extended discussion see Del Giudice, 2014b). Eating Disorders Eating disorders (EDs) are defined by heightened concern with body shape and weight and associated behaviors such as dieting, binge eating, purging, and exercising. Eating disorders occur almost exclusively in females, and their age of onset peaks in adolescence. Most evolutionary models

of eating disorders focus on the connection between dieting behavior and female reproduction. Two main alternative hypotheses have been proposed so far. First, dieting may work as a means to suppress fertility and delay or forego reproduction when the social environment is not optimal—for example, when social support by relatives and partners is low, or when social competition is too harsh (e.g., Mealey, 2000; Surbey, 1987). Second, dieting may work primarily as a female strategy in mating and status competition (e.g., Abed, 1998; Ferguson, Winegard, & Winegard, 2011; Salmon et al., 2009). Thinness is a reliable signal of youth, and dieting can increase one’s attractiveness because of men’s strong preference for younger partners; in addition, dieting can enhance status in female groups (thus indirectly influencing mating success), especially when cultural emphasis on thinness is strong. This hypothesis is supported by the robust pattern of associations among perceived sexual competition, dieting behavior, and eating symptoms. Under both hypotheses, the psychological processes that underlie dieting behavior are fundamentally adaptive, and lead to maladaptive outcomes (such as severe EDs) only when they become dysregulated or get trapped in vicious cycles. The mating competition hypothesis of eating disorders can be easily reframed in a life history perspective. Both fast and slow strategists can face intense competition for mates; the main difference is that fast strategists compete primarily to become desirable sexual partners, whereas slow strategists compete primarily to be chosen as long-term partners in committed relationships. Thus, eating disorders can arise at both ends of the fast-slow continuum. A life history analysis confirms that EDs are indeed a functionally heterogeneous category. Slow spectrum EDs are associated with high-functioning/perfectionist personality profiles, low comorbidity rates (mostly with OCD and OCPD), and the most favorable clinical outcomes (see e.g., Thompson-Brenner et al., 2008; Westen & Harnden-Fischer, 2001). This cluster of eating disorders is associated with high self-esteem, relatively intact family and couple relationships, and a history of fewer stressful life events. Another profile that can be included in the slow spectrum is that of overcontrolled patients. This profile is associated with high rates of depression, low self-esteem and passivity, restricted emotionality, and comorbidity with OCPD. Overcontrolled ED patients might be engaging in reproductive suppression—an intrinsically future-oriented strategy— following loss of status or social support, as suggested by their depressed mood, low self-esteem, and acute sense of social exclusion.

Beyond the DSM: A Life History Framework for Mental Disorders

In contrast, fast spectrum EDs are associated with dysregulated personality profiles (see Thompson-Brenner et al., 2008). Dysregulated ED patients show high levels of impulsivity and antisocial/externalizing behavior, high comorbidity (especially with borderline personality disorder), and many stressful life events including high rates of sexual abuse. While patients in the high-functioning/ perfectionist and overcontrolled groups can be diagnosed with either anorexia nervosa (AN) or bulimia nervosa (BN), the dysregulated subtype is strongly associated with BN (Westen & Harnden-Fischer, 2001). As a result, patients with BN—considered as a whole—show higher average levels of impulsivity, earlier sexual maturation, and earlier sexual debut than AN patients. Depression Depression is characterized by protracted episodes of distress and low, dejected mood. The clinical presentation of depression is quite heterogeneous; attempts to subtype depressive disorders based on empirical patterns of symptom co-occurrence consistently identify (1) a subtype characterized exclusively by depressed mood and feelings of worthlessness; (2) one or more subtypes characterized by somatic symptoms in absence of depressed mood; and (3) one or more subtypes in which depressed mood and somatic symptoms coexist. Somatic symptoms of depression include sleep disturbances (insomnia or hypersomnia), appetite disturbances (increased or decreased appetite), psychomotor disturbances (agitation or retardation), fatigue, and pain. All these symptoms are functionally related to the SRS, and in particular the HPA axis. Most evolutionary theories of depression focus on low mood and its motivational and behavioral correlates. In the prevailing view, depressed mood is an adaptive defensive mechanism, whereas clinical depression is usually maladaptive and reflects a dysfunction of the same mechanism (e.g., Allen & Badcock, 2003; Nesse, 2006; Nettle, 2004, 2012). Some theorists have argued that clinical depression may be an adaptation itself (e.g., Price, Sloman, Gardner, Gilbert, & Rohde, 1994; Watson & Andrews, 2002). Although this hypothesis appears reasonable in the specific case of postpartum depression (Hagen, 1999), there are many reasons to doubt its general applicability. The function of low mood as a protective mechanism is twofold. First, low mood helps people disengage from the pursuit of central life goals that have become unproductive. Second and more specifically, it promotes a risk-averse approach in unfavorable social circumstances—especially following losses in social support (typically in females), close relationships, and social status or dominance (typically

49

in males). While affective reactivity determines one’s susceptibility to episodes of low mood, stress reactivity is the crucial factor in the development of somatic symptoms. Thus, a complete evolutionary account of depression cannot be separated from evolutionary models of SRS functioning. By synthesizing the ACM with evolutionary models of depressed mood, it is possible to predict a complex relation between depression and life history strategy. Both fast and slow strategists can fail to obtain or maintain crucial social resources—status, dominance, and support—resulting in episodes of depressed mood and risk for clinical depression. At the slow end of the continuum, males and females are both expected to develop relatively high levels of stress responsivity (Del Giudice et al., 2011), even if the actual intensity of stress responses is buffered by the availability of social support and lack of chronic stressors. As a result, symptom profiles at the slow end of the spectrum should not differ greatly between the sexes. Intriguingly, some subtypes of depression—in particular those characterized by pure depressed mood or pure somatic symptoms—are associated with very low rates of trauma, neglect, and abuse. Moving toward the fast end of the continuum, both sexes face increasing threats to their ability to gain and maintain social resources. The availability of social support and stable, intimate relationships declines rapidly as environments become dangerous and unpredictable, exposing females to increased risk for depressed mood. At the same time, sex differences in stress responsivity can be expected to become proportionally larger, as more males develop unemotional responsivity patterns. Vigilant SRS profiles can be adaptive in dangerous and unpredictable contexts, especially in females; however, they also increase the risk of SRS dysregulation and dysfunction. In total, fast life history strategies should lead to increased risk for depression in both sexes, with females showing the highest rates of depressed mood and somatic symptoms. Consistent with these predictions, early and/or fast sexual maturation is a risk factor for depression in both sexes, with stronger effects in females. In addition, depression subtypes involving a combination of low mood and somatic symptoms are overwhelmingly more common in females, and are also associated with the highest rates of early trauma, neglect, and abuse. In conclusion, depression may occur at both ends of the fast–slow continuum, suggesting the existence of functionally distinct clusters of depressive disorders. Unfortunately, the current literature defines depression subtypes exclusively in terms of symptom co-occurrence; further research in a life history framework should attempt

50

Evolutionary Foundations of Developmental Psychopathology

to identify functional subtypes of depression based on motivation, personality, self-regulation, and comorbidity with other fast and slow spectrum disorders. Summary and Integration A life history analysis of mental disorders reveals a coherent picture of associations between individual differences in life history strategy and specific patterns of risk for psychopathology. The constellation of fast spectrum conditions includes externalizing disorders, schizophrenia spectrum disorders, OCD with autogenous obsessions, the dysregulated subtype of eating disorders (typically expressed as BN), and depressive disorders characterized by a combination of mood and somatic symptoms. These disorders tend to co-occur, both within families and within individuals; many of them share elements of impulsivity, disinhibition, and/or bizarre ideation. Slow spectrum psychopathology includes OCPD, OCD with reactive obsessions, autism spectrum disorders, the perfectionist and overcontrolled subtypes of eating disorders, and a cluster of depressive disorders of lesser severity. These comorbid disorders tend to share elements of inhibition, overcontrol, and cognitive rigidity. They are also characterized by lack of association with standard risk factors for psychopathology such as stressful life events, low SES, and early abuse; in some cases, they are actually associated with more favorable ecological and socio-economic conditions. The same approach can be easily extended to other disorders. For example, borderline personality disorder (BPD) bears the hallmarks of fast life history strategies—impulsivity, unstable attachments, risk taking, promiscuous sexuality, antisocial and paranoid personality features, and high comorbidity with externalizing disorders (Brüne et al., 2010; see also Crowell et al.,

2013). Similarly, disorders in the bipolar spectrum show substantial genotypic and phenotypic overlap with schizotypy and schizophrenia, including a familial association with enhanced creativity (see Del Giudice, 2014a, 2014b). A provisional classification of slow and fast spectrum disorders is shown in Figure 1.7. This classification is still tentative and incomplete, and many gaps and questions remain—for example about the possible functional heterogeneity of autism and schizophrenia, the role of reproductive suppression in disordered eating, or the identification of fast and slow spectrum subtypes of depression (see Del Giudice, 2014b). However, even this initial analysis illustrate how a life history framework can bring an integrative perspective to psychopathology, highlight connections between previously separate models, and suggest novel empirical questions. Even more importantly, this approach has the potential to overcome the limitations of current taxonomic systems and offer a more solid foundation for the classification of mental disorders. In particular, the fast-slow distinction is both more inclusive and more accurate than the interalizing–externalizing distinction. It is more inclusive because it integrates mood and anxiety disorders with personality disorders, schizophrenia spectrum disorders, and autism spectrum disorders—all within the same conceptual framework. It is more accurate because it resolves many inconsistencies inherent in the basic interalizing–externalizing distinction. For example, the ambiguous placement of OCD in the internalizing spectrum is explained by the heterogeneity of OCD; specifically, the autogenous subtype of OCD is a fast spectrum disorder with strong functional connections with externalizing symptoms. More generally, the interalizing–externalizing distinction may be problematic because it is in large part

Figure 1.7 Provisional life history taxonomy of common mental disorders. BPD = borderline personality disorder, OCD = obsessive-compulsive disorder, OCPD = obsessive-compulsive personality disorder. Source: Reprinted from M. Del Giudice, An evolutionary life history framework for psychopatholgy, Psychological inquiry, 25, 285, 2014a.

Conclusion

illusory. The obvious genotypic and phenotypic coherence of the externalizing spectrum may have led researchers to assume that internalizing disorders must form a symmetrical category with similar properties of coherence. A life history perspective suggests that this assumption is probably mistaken, and that the internalizing spectrum may turn out to be a largely artificial collection of disorders with divergent functional properties. Implications for the Core Points of Developmental Psychopathology The life history framework discussed in this section has important implications for developmental psychopathology. Most crucially, it shows how equifinality and multifinality in the development of mental disorders can be explained in a functional perspective. On one hand, the same kind of symptom—for example eating symptoms and obsessions—can arise in relation to different life history strategies. As a result, phenotypically similar disorders can be associated with opposite profiles of personality, sexual maturation, ecological factors, and so forth. On the other hand, the same basic dimensions of life history strategy can play a role in the etiology of functionally related but superficially different disorders. While these manifestations of equifinality and multifinality may be problematic in the context of the standard externalizing–internalizing distinction, they can be easily understood in terms of the fast–slow distinction advanced by Del Giudice (2014a). A closely related point is that, in this perspective, not all disorders are expected to arise in association with typical risk factors such as early stress, negative family relationships, and low SES. This helps make sense of the puzzling fact that some disorders seem to develop more frequently in safe, predictable ecologies and families with high socioeconomic status. The model presented here may also explain why insecure attachment—a robust psychological correlate of fast life history strategy—is consistently associated with externalizing symptoms but only weakly predictive of internalizing symptoms (Groh, Roisman, van IJzendoorn, Bakermans-Kranenburg, & Fearon, 2012). Finally, the coexistence of fast and slow spectrum subtypes within the same diagnostic category may go a long way toward explaining inconsistent or contradictory patterns of epidemiological findings. CONCLUSION We started this chapter with a promise—to show how an evolutionary approach can help developmental

51

psychopathology realize its full potential, and to demonstrate how EDP provides an integrative, powerful metatheory for the field. We hope we fulfilled our promise and succeeded in arousing the reader’s interest in the EDP approach. Throughout the chapter, we sought to illustrate how an evolutionary-developmental perspective supports and extends the core points of developmental psychopathology. Consider for example the concept of multifinality. In the standard view, multifinality is a ubiquitous, general property of complex developmental systems. However, the mechanisms that generate of multifinality are usually left unspecified; as a result, the concept is often used to redescribe empirical findings rather than explain them. The EDP approach demystifies multifinality by grounding the concept in adaptive function, and provides the tools for predicting when multifinality should apply (or not) to a given developmental process, stage, or outcome. Thus, multifinality can be understood as a necessary consequence of differential susceptibility; the logic of the ACM predicts when similar physiological profiles may predict widely divergent behavioral outcomes; and life history theory (together with sexual selection theory) explains why similar developmental experiences may set the stage for phenotypically different but functionally related disorders. Even more importantly, these apparently disparate aspects of development can be understood in relation to one another and unified within an integrative theoretical framework. Even if we covered a lot of ground, we barely scratched the surface of our topic. The evolutionary approach to development and psychopathology is a growing multidisciplinary enterprise, and new theories, models, and findings are published at an ever increasing pace. We therefore conclude by suggesting a reading list for further explorations of the field. Ellis and Bjorklund (2005) and Burgess and MacDonald (2005) offer a more complete overview of EDP, including discussion of cognitive processes such as memory and language. Special sections of Development and Psychopathology on differential susceptibility (Ellis & Boyce, 2011) and Developmental Psychology on conditional adaptation (Ellis & Bjorklund, 2012) provide useful collections of relevant work. An introduction to theories and models in evolutionary psychopathology can be found in Brüne (2008) and McGuire and Troisi (1998). Finally, Ellison and Gray (2009) show how evolutionary thinking can be applied to neurobiological and endocrinological processes. Explaining the development of psychopathology is a formidable task, calling for convergence and integration across myriad disciplines and levels of analysis. We believe that evolutionary theory offers invaluable tools for

52

Evolutionary Foundations of Developmental Psychopathology

this task, and hope that developmental psychopathology will join forces with EDP toward a common understanding of human development in all its living complexity.

REFERENCES Abed, R. T. (1998). The sexual competition hypothesis for eating disorders. British Journal of Medical Psychology, 71, 525–547. doi: 10.1111/j.2044–8341.1998.tb01007.x Alink, L. R. A., van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., Mesman, J., Juffer, F., & Koot, H. M. (2008). Cortisol and externalizing behavior in children and adolescents: Mixed meta-analytic evidence for the inverse relation of basal cortisol and cortisol reactivity with externalizing behavior. Developmental Psychobiology, 50, 427–450. doi: 10.1002/dev.20300 Allen, N. B., & Badcock, P. B. T. (2003). Darwinian models of depression: A review of evolutionary accounts of mood disorders. Progress in Neuro-Psychopharmacology & Biological Psychiatry, 30, 815–26. doi: 10.1016/j.pnpbp. 2006.01.007 Allsworth, J. E., Weitzen, S., & Boardman, L. A. (2005). Early age at menarche and allostatic load: Data from the third National Health and Nutrition Examination Survey. Annals of Epidemiology, 15, 438–444. doi: 10.1016/j.annepidem.2004.12.010 Archer, J. (2009). Does sexual selection explain human sex differences in aggression? Behavioral and Brain Sciences, 32, 249–266. doi: 10.1017/S0140525X09990951 Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to rearing environment depending on dopaminerelated genes: New evidence and a meta-analysis. Development and Psychopathology, 23, 39–52. doi: 10.1017/S0954579410000635 Bao, A.-M., Meynen, G., & Swaab, D. F. (2008). The stress system in depression and neurodegeneration: Focus on the human hypothalamus. Brain Research Reviews, 57, 531–553. doi: 10.1016/j.brainresrev .2007.04.005 Baron-Cohen, S. (2003). The essential difference. London, UK: Penguin. Barr, K. N., & Quinsey, V. L. (2004). Is psychopathy pathology or a life strategy? Implications for social policy. In C. Crawford & C. Salmon (eds.), Evolutionary psychology, public policy and personal decisions (pp. 293–317). Mahwah, NJ: Lawrence Erlbaum. Beauchaine, T. P., & Hinshaw, S. P. (Eds.). (2013). Child and adolescent psychopathology (2nd ed.). Hoboken, NJ: Wiley. Beauchaine, T. P., Klein, D. N., Erickson, N. L., & Norris, A. L. (2013). Developmental psychopathology and the Diagnostic and Statistical Manual of Mental Disorders. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (2nd ed., pp. 29–110). Hoboken, NJ: Wiley. Belsky, J. (1997). Variation in susceptibility to rearing influences: an evolutionary argument. Psychological Inquiry, 8, 182–186. doi: 10.1207/s15327965pli0803_3 Belsky, J. (2000). Conditional and alternative reproductive strategies: individual differences in susceptibility to rearing experiences. In J. Rodgers, D. Rowe, & W. Miller (Eds.), Genetic influences on human fertility and sexuality (pp. 127–146). Boston, MA: Kluwer. Belsky, J. (2005). Differential susceptibility to rearing influences: An evolutionary hypothesis and some evidence. In B. Ellis & D. Bjorklund (Eds.), Origins of the social mind: Evolutionary psychology and child development (pp. 139–163). New York, NY: Guilford. Belsky, J. (2012). The development of human reproductive strategies: Promises and prospects. Current Directions in Psychological Science, 21, 310–316. doi: 10.1177/0963721412453588

Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2007). For better and for worse: differential Susceptibility to environmental influences. Current Directions in Psychological Science, 16, 300–304. doi: 10.1111/j.1467–8721.2007.00525.x Belsky, J., & Beaver, K. M. (2011). Cumulative-genetic plasticity, parenting, and adolescent self-regulation. Journal of Child Psychology and Psychiatry, 52, 619–626. doi: 10.1111/j.1469–7610.2010.02327.x Belsky, J., Schlomer, G. L., & Ellis, B. J. (2012). Beyond cumulative risk: Distinguishing harshness and unpredictability as determinants of parenting and early life history strategy. Developmental Psychology, 48, 662–673. doi: 10.1037/a0024454 Belsky, J., Steinberg, L., & Draper, P. (1991). Childhood experience, interpersonal development, and reproductive strategy: An evolutionary theory of socialization. Child Development, 62, 647–670. doi: 10.1111/j.1467–8624.1991.tb01558.x Benros, M. E., Mortensen, P. B., & Eaton, W. W. (2012). Autoimmune diseases and infections as risk factors for schizophrenia. Annals of the New York Academy of Sciences, 1262, 56–66. doi: 10.1111/j.1749–6632 .2012.06638.x Bjorklund, D. F., & Ellis, B. J. (2005). Evolutionary psychology and child development: an emerging synthesis. In B. J. Ellis & D. F. Bjorklund (Eds.), Origins of the social mind: Evolutionary Psychology and Child Development (pp. 3–18). New York, NY: Guilford. Bjorklund, D. F., Ellis, B. J., & Rosenberg, J. S. (2007). Evolved probabilistic cognitive mechanisms: an evolutionary approach to gene × environment × development interactions. Advances in Child Development and Behavior, 35, 1–39. doi: 10.1016/B978–0–12–009735–7 .50006–2 Bjorklund, D. F., & Pellegrini, A. D. (2002). The origins of human nature: evolutionary developmental psychology. Washington, DC: American Psychological Association. Bleil, M. E., Adler, N. E., Pasch, L. A., Sternfeld, B., Gregorich, S. E., Rosen, M. P., & Cedards, M. I. (2013). Psychological stress and reproductive aging among pre-menopausal women. Human Reproduction, 27, 2720–2728. doi: 10.1093/humrep/des214 Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary-developmental theory of the origins and functions of stress reactivity. Development and Psychopathology, 17, 271–301. doi: 10.1017/S0954579405050145 Boyer, P., & Lienard, P. (2006). Why ritualized behavior? Precaution systems and action parsing in developmental, pathological and cultural rituals. Behavioral and Brain Sciences, 29, 595–613. doi: 10.1017/S0140525X06009332 Braendle, C., Heyland, F., & Flatt, T. (2011). Integrating mechanistic and evolutionary analysis of life history variation. In T. Flatt & F. Heyland (Eds.), Mechanisms of life history evolution. The genetics and physiology of life history traits and trade-offs (pp. 3–10). New York, NY: Oxford University Press. Brumbach, B. H., Figueredo, A. J., & Ellis, B. J. (2009). Effects of harsh and unpredictable environments in adolescence on development of life history strategies: a longitudinal test of an evolutionary model. Human Nature, 20, 25–51. doi: 10.1007/s12110–009–9059–3 Brüne, M. (2008). Textbook of evolutionary psychiatry: The origins of psychopathology. New York, NY: Oxford University Press. Brüne, M., Ghiassi, V., & Ribbert, H. (2010). Does borderline personality reflect the pathological extreme of an adaptive reproductive strategy? Insights and hypotheses from evolutionary life-history theory. Clinical Neuropsychiatry, 7, 3–9. Bubier, J. L., & Drabick, D. A. G. (2008). Affective decision-making and externalizing behaviors: the role of autonomic activity. Journal of Abnormal Child Psychology, 36, 941–953. doi: 10.1007/s10802– 008–9225–9

References

53

Bull, J. J. (1987). Evolution of phenotypic variance. Evolution, 41, 303–315.

Crespi, B. J. (2000). The evolution of maladaptation. Heredity, 84, 623–629. doi: 10.1046/j.1365–2540.2000.00746.x

Burgess, R. L., & MacDonald, K. (Eds.). (2005). Evolutionary perspectives on human development (2nd ed.). Thousand Oaks, CA: Sage.

Crespi, B. J. (2010). The origins and evolution of genetic disease risk in modern humans. Annals of the New York Academy of Sciences, 1206, 80–109. doi: 10.1111/j.1749–6632.2010.05707.x

Burns, J. K. (2004). An evolutionary theory of schizophrenia: Cortical connectivity, metarepresentation, and the social brain. Behavioral and Brain Sciences, 27, 831–885. doi: 10.1017/S0140525X04000196 Burt, A., & Trivers, L. R. (2006). Genes in conflict: the biology of selfish genetic elements. Cambridge, MA: Harvard University Press. Buss, D. M. (2005). The handbook of evolutionary psychology. New York, NY: Wiley. Buss, D. M. (2012). Evolutionary psychology: the new science of the mind (4th ed.). Boston, MA: Allyn & Bacon. Buss, D. M., & Hawley, P. H. (Eds.). (2011). The evolution of personality and individual differences. New York, NY: Oxford University Press. Carere, C., & Maestripieri, D. (Eds.). (2013). Animal personalities: Behavior, physiology, and evolution. Chicago, IL: Chicago University Press. Caspi, A., Houts, R. M., Belsky, D. W., Goldman-Mellor, S. J., Harrington, H., Israel, S., . . . , Moffitt, T. E. (2014). The p factor: One general psychopathology factor in the structure of psychiatric disorders? Clinical Psychological Science, 2, 119–127. doi: 10.1177/2167702613497473 Chisholm, J. S. (1999). Death, hope and sex: Steps to an evolutionary ecology of mind and morality. Cambridge, UK: Cambridge University Press. Cicchetti, D. (1990). A historical perspective on the discipline of developmental psychopathology. In J. Rolf, A. Masten, D. Chicchetti, K., Nuechterlein, & S. Weintraub (Eds.), Risk and protective factors in the development of psychopathology (pp. 2–28). New York, NY: Cambridge University Press.

Crespi, B., & Badcock, C. (2008). Psychosis and autism as diametrical disorders of the social brain. Behavioral and Brain Sciences, 31, 241–261. doi:10.1017/S0140525X08004214 Crowell, S. E., Kaufman, E. A., & Lenzenweger, M. F. (2013). The development of borderline personality and self-inflicted injury. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (pp. 577–609). Hoboken, NJ: Wiley. Davies, W., Lynn, P. M. Y., Relkovic, D., & Wilkinson, L. S. (2008). Imprinted genes and neuroendocrine function. Frontiers in Neuroendocrinology, 29, 413–427. doi: 10.1016/j.yfrne.2007.12.001 Dawkins, R. (1976). The selfish gene. Oxford, UK: Oxford University Press. Del Giudice, M. (2009). Sex, attachment, and the development of reproductive strategies. Behavioral and Brain Sciences, 32, 1–21. doi: 10.1017/S0140525X09000016 Del Giudice, M. (2012). Fetal programming by maternal stress: Insights from a conflict perspective. Psychoneuroendocrinology, 37, 1614–1629. doi: 10.1016/j.psyneuen.2012.05.014 Del Giudice, M. (2014a). An evolutionary life history framework for psychopathology. Psychological Inquiry, 25, 261–300. doi: 10.1080/1047840X.2014.884918 Del Giudice, M. (2014b). A tower unto Heaven: Toward an expanded framework for psychopathology. Psychological Inquiry, 25, 394–413. doi: 10.1080/1047840X.2014. 925339

Cicchetti, D. (2006). Development and psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology (2nd ed., Vol. 1, pp. 1–23). Hoboken, NJ: Wiley.

Del Giudice, M., Angeleri, R., Brizio, A., & Elena, M. R. (2010). The evolution of autistic-like and schizotypal traits: a sexual selection hypothesis. Frontiers in Psychology, 1, 41. doi: 10.3389/fpsyg.2010.00041

Cicchetti, D., & Rogosch, F. A. (2012). Gene by environment interaction and resilience: Effects of child maltreatment and serotonin, corticotropin releasing hormone, dopamine, and oxytocin genes. Development and Psychopathology, 24, 411–427. doi: 10.1017/S0954579412000077

Del Giudice, M., Ellis, B. J., & Shirtcliff, E. A. (2011). The adaptive calibration model of stress responsivity. Neuroscience & Biobehavioral Reviews, 35, 1562–1592. doi: 10.1016/j.neubiorev.2010.11.007

Clancy, K. H. B., Klein, L. D., Ziomkiewicz, A., Nenko, I., Jasienska, G., & Bribiescas, R. G. (2013). Relationships between biomarkers of inflammation, ovarian steroids, and age at menarche in a rural polish sample. American Journal of Human Biology, 25, 389–398. doi: 10.1002/ajhb.22386 Clark, L. A., & Watson, D. (2006). Distress and fear disorders: An alternative empirically based taxonomy of the “mood” and “anxiety” disorders. British Journal of Psychiatry, 189, 481–483. doi: 10.1192/bjp.bp. 106.03825 Compas, B. E., & Andreotti, C. (2013). Risk and resilience in child and adolescence psychopathology. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (pp. 143–169). Hoboken, NJ: Wiley. Cosmides, L., & Tooby, J. (1999). Toward an evolutionary taxonomy of treatable conditions. Journal of Abnormal Psychology, 108, 453–464. doi: 10.1037/0021–843X.108.3.453 Crawford, D. (1998). Environments and adaptations: Then and now. In C. C. Crawford & D. L. Krebs (Eds.), Handbook of evolutionary psychology: Ideas, issues, and applications (pp. 275–302). Mahwah, NJ: Erlbaum. Crean, A. J., & Marshall, D. J. (2009). Coping with environmental uncertainty: Dynamic bet hedging as a maternal effect. Philosophical Transactions of the Royal Society B, 364, 1087–1096. doi: 10.1098/rstb.2008.0237

Del Giudice, M., Hinnant, J. B., Ellis, B. J., & El-Sheikh, M. (2012). Adaptive patterns of stress responsivity: A preliminary investigation. Developmental Psychology, 48, 775–790. doi: 10.1037/a0026519 DeWitt, T. J., & Scheiner, S. M. (2004). Phenotypic plasticity: Functional and conceptual approaches. New York, NY: Oxford University Press. Dickerson, S. S., & Kemeny, M. E. (2004). Acute stressors and cortisol responses: A theoretical integration and synthesis of laboratory research. Psychological Bulletin, 130, 355–391. doi: 10.1037 /0033–2909.130.3.355 Durrant, R., & Ellis, B. J. (2003). Evolutionary Psychology. In M. Gallagher & R. J. Nelson (Eds.), Comprehensive handbook of psychology (Vol. 3, pp. 1–33). New York, NY: Wiley. Ellis, B. J. (2004). Timing of pubertal maturation in girls: An integrated life history approach. Psychological Bulletin, 130, 920–958. doi: 10.1037/0033–2909.130.6.920 Ellis, B.J. (2013). The hypothalamic-pituitary-gonadal axis: A switchcontrolled, condition-sensitive system in the regulation of life history strategies. Hormones and Behavior, 64, 215–225. doi: 10.1016 /j.yhbeh.2013.02.012 Ellis, B. J., & Boyce, W. T. (2011). Differential susceptibility to the environment: Toward an understanding of sensitivity to developmental experiences and context. Development and Psychopathology, 23, 1–5. doi:10.1017/S095457941000060X Ellis, B. J., Boyce, W. T., Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011a). Differential susceptibility to the

54

Evolutionary Foundations of Developmental Psychopathology environment: An evolutionary-neurodevelopmental theory. Development and Psychopathology, 23, 7–28. doi: 0.1017/S0954579410000611

Ellis, B. J., & Bjorklund, D. (2005) (Eds.), Origins of the social mind: Evolutionary psychology and child development. New York, NY: Guilford. Ellis, B. J., & Bjorklund, D. (2012). Beyond mental health: An evolutionary analysis of development under risky and supportive environmental conditions: an introduction to the special section. Developmental Psychology, 48, 591–597. doi: 10.1037/a0027651 Ellis, B. J., & Del Giudice, M. (2014). Beyond allostatic load: Rethinking the role of stress in regulating human development. Development and Psychopathology, 26, 1–20. doi:10.1017/S0954579413000849. Ellis, B. J., Del Giudice, M., Dishion, T. J., Figueredo, A. J., Gray, P., Griskevicius, V., . . . , & Wilson, D. S. (2012). The evolutionary basis of risky adolescent behavior: implications for science, policy, and practice. Developmental Psychology, 48, 598–623. doi: 10.1037/a0026220

Flinn, M. V. (2006). Evolution and ontogeny of the stress response to social challenges in the human child. Developmental Review, 26, 138–174. doi: 10.1016/j.dr.2006.02.003 Frankenhuis, W. E., & Del Giudice, M. (2012). When do adaptive developmental mechanisms yield maladaptive outcomes? Developmental Psychology, 48, 628–642. doi: 10.1037/a0025629 Frankenhuis, W. E., & de Weerth, C. (2013). Does early-life exposure to stress shape or impair cognition? Current Directions in Psychological Science, 22, 407–412. doi: 10.1177/0963721413484324 Frankenhuis, W. E., Panchanathan, K., & Barrett, H. C. (2013). Bridging developmental systems theory and evolutionary psychology using dynamic optimization. Developmental Science, 16, 584–598. doi: 10.1111/desc.12053 Gangestad, S. W., & Simpson, J. A. (2000). The evolution of human mating: Trade-offs and strategic pluralism. Behavioral and Brain Sciences, 23, 573–587. doi: 10.1017/S0140525X0000337X

Ellis, B. J., Del Giudice, M., & Shirtcliff, E. A. (2013). Beyond allostatic load: The stress response system as a mechanism of conditional adaptation. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (2nd ed., pp. 251–284). Hoboken, NJ: Wiley & Sons.

Gatzke-Kopp, L. M. (2011). The canary in the coalmine: The sensitivity of mesolimbic dopamine to environmental adversity during development. Neuroscience and Biobehavioral Reviews, 35, 794–803.

Ellis, B. J., & Essex, M. J. (2007). Family environments, adrenarche, and sexual maturation: A longitudinal test of a life history model. Child Development, 78, 1799–1817. DOI: 10.1111/j.1467–8624.2007 .01092.x

Ge, X., & Natsuaki, M. N. (2010). In search of explanations for early pubertal timing effects on developmental psychopathology. Current Directions in Psychological Science, 18, 327–331. doi: 10.1111/j.1467–8721.2009.01661.x

Ellis, B. J., Essex, M. J., & Boyce, W. T. (2005). Biological sensitivity to context: II. Empirical explorations of an evolutionary-developmental theory. Development and Psychopathology, 17, 303–328. doi: 10.1017/S0954579405050157

Geary, D. C. (2002). Sexual selection and human life history. Advances in Child Development and Behavior, 30, 41–101. doi: 10.1016/S0065– 2407(02)80039–8

Ellis, B. J., Figueredo, A. J., Brumbach, B. H., & Schlomer, G. L. (2009). Fundamental dimensions of environmental risk: The impact of harsh versus unpredictable environments on the evolution and development of life history strategies. Human Nature, 20, 204–268. doi: 10.1007/s12110–009–9063–7 Ellis, B. J., Jackson, J. J., & Boyce, W. T. (2006). The stress response system: Universality and adaptive individual differences. Developmental Review, 26, 175–212. doi: 10.1016/j.dr.2006.02.004 Ellis, B. J., Shirtcliff, E. A., Boyce, W. T., Deardorff, J., & Essex, M. J. (2011b). Quality of early family relationships and the timing and tempo of puberty: Effects depend on biological sensitivity to context. Development and Psychopathology, 23, 85–99. doi: 10.1017/S0954579410000660 Ellison, P. T., & Gray, P. B. (2009). Endocrinology of social relationships. Cambridge, MA: Harvard University Press. Ferguson, C. J., Winegard, B., & Winegard, B. M. (2011). Who is the fairest one of all? How evolution guides peer and media influences on female body dissatisfaction. Review of General Psychology, 15, 11–28.

Geary, D. C. (2005). Evolution of paternal investment. In D. Buss (Ed.), The handbook of evolutionary psychology (pp. 483–505). Hoboken, NJ: John Wiley & Sons. Glenn, A. L. (2011). The other allele: Exploring the long allele of the serotonin transporter gene as a potential risk factor for psychopathy: A review of the parallels in findings. Neuroscience and Biobehavioral Reviews, 35, 612–620. doi: 10.1016/j.neubiorev.2010.07.005 Glenn, A. L., Kurzban, R., & Raine, A. (2011). Evolutionary theory and psychopathy. Aggression and Violent Behavior, 16, 371–380. doi: 10.1016/j.avb.2011.03.009 Glover, V. (2011). Prenatal stress and the origins of psychopathology: an evolutionary perspective. Journal of Child Psychology and Psychiatry, 52, 356–367. doi: 10.1111/j.1469–7610.2011.02371.x Gluckman, P. D., Low, F. M., Buklijas, T., Hanson, M. A., & Beedle, A. S. (2011). How evolutionary principles improve the understanding of human health and disease. Evolutionary Applications, 4, 249–263. doi: 10.1111/j.1752–4571.2010.00164.x Gottlieb, G. (2007). Probabilistic epigenesis. Developmental Science, 10, 1–11. doi: 10.1111/j.1467–7687.2007.00556.x

Figueredo, A. J., & Jacobs, W. J. (2010). Aggression, risk-taking, and alternative life history strategies: The behavioral ecology of social deviance. In M. Frias-Armenta & V. Corral-Verdugo (Eds.), Biopsychosocial perspectives on aggression (pp. 3–27). Hauppauge, NY: Nova Science Publishers.

Grafen, A. (1999). Formal Darwinism, the individual-as-maximizingagent analogy and bet-hedging. Proceedings of the Royal Society of London B, 266, 799–803. doi: 10.1098/rspb.1999.0708

Figueredo, A. J., Vásquez, G., Brumbach, B. H., & Schneider, S. M. R. (2004). The heritability of life history strategy: The K-factor, covitality, and personality. Social Biology, 51, 121–143. doi: 10.1080/19485565.2004.9989090

Grafen, A. (2009). Formalizing Darwinism and inclusive fitness theory. Philosophical Transactions of the Royal Society B, 364, 3135–3141. doi: 10.1098/rstb.2009.0056

Grafen, A. (2006). Optimisation of inclusive fitness. Journal of Theoretical Biology, 238, 541–563. doi:10.1016/j.jtbi.2005.06.009

Figueredo, A. J., Vásquez, G., Brumbach, B., Schneider, S. M. R., Sefcek, J. A., Tal, I. R., . . . Jacobs, W. J. (2006). Consilience and life history theory: From genes to brain to reproductive strategy. Developmental Review, 26, 243–275. doi: 10.1016/j.dr.2006.02.002

Griffiths, P., & Gray, R. (2004). The developmental systems perspective: organism–environment systems as units of development and evolution. In M. Pigliucci & K. Preston (Eds.), Phenotypic integration: Studying the ecology and evolution of complex phenotypes (pp. 335–339). Oxford, UK: Oxford University Press.

Figueredo, A. J., & Wolf, P. S. A. (2009). Assortative pairing and life history strategy: A cross-cultural study. Human Nature, 20, 317–330. doi: 10.1007/s12110–009–9068–2

Groh, A. M., Roisman, G. I., van IJzendoorn, M. H., BakermansKranenburg, M. J., & Fearon, R. P. (2012). The significance of insecure and disorganized attachment for children’s internalizing

References

55

symptoms: A meta-analytic study. Child Development, 83, 591–610. doi: 10.1111/j.1467–8624.2011.01711.x

depressive symptom patterns. Journal of Personality and Social Psychology, 91(2), 316–330.

Gunnar, M. R., Talge, N. M., & Herrera, A. (2009). Stressor paradigms in developmental studies: What does and does not work to produce mean increases in salivary cortisol. Psychoneuroendocrinology, 34, 953–967. doi: 10.1016/j.psyneuen.2009.02.010

Kendler, K. S., Prescott, C. A., Myers, J., & Neale, M. C. (2003). The structure of genetic and environmental risk factors for common psychiatric and substance use disorders in men and women. Archives of General Psychiatry, 60, 929–937. doi:10.1001/archpsyc.60 .9.929

Gunnar, M. R., & Vazquez, D. (2006). Stress neurobiology and developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology: Developmental neuroscience (2nd ed., Vol. 2, pp. 533–568). Hoboken, NJ: Wiley & Sons. Hagen, E. (1999). The functions of post-partum depression. Evolution and Human Behavior, 20, 325–359. doi:10.1016/S1090–5138(99)00016–1 Haig, D. (1993). Genetic conflicts in human pregnancy. Quarterly Review of Biology, 68, 495–532. Halpern, C. T., Campbell, B., Agnew, C. R., Thompson, V., & Udry, J. R. (2002). Associations between stress reactivity and sexual and nonsexual risk taking in young adult human males. Hormones and Behavior, 42, 387–398. doi: 10.1006/hbeh.2002.1831 Hamilton, W. D. (1964). The genetical evolution of social behavior. Journal of Theoretical Biology, 7, 1–52. doi:10.1016/0022– 5193(64)90038–4 Hawks, J., Wang, E. T., Cochran, G. M., Harpending, H. C., & Moyzis, R. K. (2007). Recent acceleration of human adaptive evolution. Proceedings of the National Academy of Sciences, 104, 20753–20758. doi: 10.1073/pnas.0707650104 Heiming, R. S., & Sachser, N. (2011). Consequences of serotonin transporter genotype and early adversity on behavioral profile— Pathology or adaptation? Frontiers in Neuroscience, 4, 187. doi: 10.3389/fnins.2010.00187 Hinshaw, S. P. (2013). Developmental psychopathology as a scientific discipline. In T. P. Beauchaine & S. P. Hinshaw (Eds.), Child and adolescent psychopathology (2nd ed., pp. 3–28). Hoboken, NJ: Wiley. James, J., & Ellis, B. J. (2013). The development of human reproductive strategies: Toward an integration of life history and sexual selection models. In J. A. Simpson & L. Campbell (Eds.), The Oxford handbook of close relationships (pp. 771–794). New York, NY: Oxford University Press. James, J., Ellis, B. J., Schlomer, G. L., & Garber, J. (2012). Sex-specific pathways to early puberty, sexual debut, and sexual risk taking: Tests of an integrated evolutionary—developmental model. Developmental psychology, 48, 687–702.

Ketelaar, T., & Ellis, B. J. (2000). Are evolutionary explanations unfalsifiable? Evolutionary psychology and the Lakatosian philosophy of science. Psychological Inquiry, 11, 1–21. doi: 10.1207 /S15327965PLI1101_01 Korte, S. M., Koolhaas, J. M., Wingfield, J. C., & McEwen, B. S. (2005). The Darwinian concept of stress: Benefits of allostasis and costs of allostatic load and the trade-offs in health and disease. Neuroscience and Biobehavioral Reviews, 29, 3–38. doi: 10.1016 /j.neubiorev.2004.08.009 Kuzawa, C. W., & Bragg, J. M. (2012). Plasticity in human life history strategy: Implications for contemporary human variation and the evolution of genus Homo. Current Anthropology, 53, S369–S382. doi: 10.1086/667410 Kuzawa, C. W., & Quinn, E. A. (2009). Developmental origins of adult function and health: Evolutionary hypotheses. Annual Review of Anthropology, 38, 131–147. doi: 10.1146/annurev-anthro-091908– 164350 Lahey, B. B., Rathouz, P. J., Applegate, B., Hulle, C. V., Garriock, H. A., Urbano, R. C., . . . Waldman, I. D. (2008). Testing structural models of DSM-IV symptoms of common forms of child and adolescent psychopathology. Journal of Abnormal Child Psychology, 36, 187–206. Doi: 10.1007/s10802–007–9169–5 Lahey, B. B., Van Hulle, C. A., Singh, A. L., Waldman, I. D., & Rathouz, P. J. (2011). Higher-order genetic and environmental structure of prevalent forms of child and adolescent psychopathology. Archives of General Psychiatry, 68, 181–189. doi: 10.1001/archgenpsychiatry.2010.192 Lakatos, I. (1970). Falsificationism and the methodology of scientific research programmes. In I. Lakatos & A. Musgrave (Eds.), Criticism and the growth of knowledge (pp. 91–196). Cambridge, UK: Cambridge University Press. Lee, H.-J., & Kwon, S.-M. (2003). Two different types of obsession: Autogenous obsessions and reactive obsessions. Behaviour Research and Therapy, 41, 11–29. doi:10.1016/S0005–7967(01)00101–2

Jonason, P. K., Li, N. P., Webster, G. D., & Schmitt, D. P. (2009). The dark triad: Facilitating a short-term mating strategy in men. European Journal of Personality, 23, 5–18. doi: 10.1002/per.698

Lickliter, R., & Honeycutt, H. (2003). Developmental dynamics: Toward a biologically plausible evolutionary psychology. Psychological Bulletin, 129, 819–835. doi: 10.1037/0033–2909.129.6.819

Juster, R. P., Bizik, G., Picard, M., Arsenault-Lapierre, G., Sindi, S., Trepanier, L., . . . Lupien, S. J. (2011). A transdisciplinary perspective of chronic stress in relation to psychopathology throughout life span development. Development and Psychopathology, 23, 725–776. doi: 10.1017/S0954579411000289

Lupien, S. J., Ouellet-Morin, I., Hupbach, A., Tu, M. T., Buss, C., Walker, D., Pruessner, J., & McEwen, B. S. (2006). Beyond the stress concept: Allostatic load—A developmental biological and cognitive perspective. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology: Developmental neuroscience (2nd ed., Vol. 2, pp. 578–628). Hoboken, NJ: Wiley & Sons.

Kaplan, H. S., & Gangestad, S. W. (2005) Life history theory and evolutionary psychology. In D. M. Buss (Ed.), The handbook of evolutionary psychology (pp. 68–95). New York, NY: John Wiley & Sons. Kaplan, H., Hill, K., Lancaster, J., & Hurtado, A. M. (2000). A theory of human life history evolution: Diet, intelligence, and longevity. Evolutionary Anthropology, 9, 156–185. Keller, M. C., & Miller, G. (2006). Resolving the paradox of common, harmful, heritable mental disorders: which evolutionary genetic models work best? Behavioral and Brain Sciences, 29, 385–452. doi: 10.1017/S0140525X06009095 Keller, M. C., Nesse, R. M. (2006). The evolutionary significance of depressive symptoms: different adverse situations lead to different

MacDonald, K. B. (2012). Cutting nature at its joints: Toward an evolutionarily informed theory of natural types of conduct disorder. Journal of Social, Evolutionary, and Cultural Psychology, 6, 260–291. Manuck, S. B., Craig, A. E., Flory, J. D., Halder, I., & Ferrell, R. E. (2011). Reported early family environment covaries with menarcheal age as a function of polymorphic variation in estrogen receptor-𝛼. Development and Psychopathology, 23, 69–83. doi: 10.1017/S0954579410000659 Marshall, J. A. R. (2011). Group selection and kin selection: formally equivalent approaches. Trends in Ecology and Evolution, 26, 325–332. doi: 10.1016/j.tree.2011.04.008

56

Evolutionary Foundations of Developmental Psychopathology

Martel, M. M. (2013). Sexual selection and sex differences in the prevalence of childhood externalizing and adolescent internalizing disorders. Psychological Bulletin, 139, 1221–1259. doi: 10.1037/a0032247

Nettle, D., Frankenhuis, W. E., & Rickard, I. J. (2013). The evolution of Predictive Adaptive Responses in human life history. Proceedings of the Royal Society B, 280, 20131343. doi: 10.1098/rspb.2013.1343

Mather, M., & Lighthall, N. R. (2012). Risk and reward are processed differently in decisions made under stress. Current Directions in Psychological Science, 21, 36–41. doi: 10.1177/0963721411429452

Olson, M. E. (2012). The developmental renaissance in adaptationism. Trends in Ecology and Evolution, 27, 278–287. doi: 10.1016/j.tree.2011.12.005

Mayr, E. (1963). Animal species and evolution. Cambridge, MA: Harvard University Press.

Overton, W. F. (2006). Developmental psychology: Philosophy, concepts, methodology. In W. Damon & R. M. Lerner (Eds.), Handbook of child psychology (Vol. 1, pp. 18–88). Hoboken, NJ: Wiley.

McEwen, B. S., & Stellar, E. (1993). Stress and the individual: Mechanisms leading to disease. Archives of Internal Medicine, 153, 2093–2101. doi: 10.1001/archinte.1993.00410180039004 McEwen, B. S., & Wingfield, J. C. (2003). The concept of allostasis in biology and biomedicine. Hormones and Behavior, 42, 2–15. doi: 10.1016/S0018–506X(02)00024–7 McGuire, M., & Troisi, A. (1998). Darwinian psychiatry. New York, NY: Oxford University Press. Mealey, L. (1995). The sociobiology of sociopathy: An integrated evolutionary model. Behavioral and Brain Sciences, 18, 523–541. doi: 10.1017/S0140525X00039595 Mealey, L. (2000). Anorexia: A “losing” strategy? Human Nature, 11, 105–116. doi: 10.1007/s12110–000–1005–3 Mendle, J., Turkheimer, E., & Emery, R. E. (2007). Detrimental psychological outcomes associated with early pubertal timing in adolescent girls. Developmental Review, 27, 151–71. doi: 10.1016/j.dr.2006.11.001 Miller, G. E., Chen, E., & Parker, K. J. (2011). Psychological stress in childhood and susceptibility to the chronic diseases of aging: Moving toward a model of behavioral and biological mechanisms. Psychological Bulletin, 137, 959–997. doi: 10.1037/a0024768 Muehlenbein, M. P., & Bribiescas, R. G. (2005). Testosterone-mediated immune functions and male life histories. American Journal of Human Biology, 17, 527–558. doi: 10.1002/ajhb.20419 Nesse, R. M. (2001). On the difficulty of defining disease: a Darwinian perspective. Medicine, Health Care and Philosophy, 4, 37–46. doi: 10.1023/A:1009938513897 Nesse, R. M. (2004a). Natural selection and the elusiveness of happiness. Philosophical Transactions of the Royal Society of London B, 359, 1333–1347. doi: 10.1098/rstb.2004.1511 Nesse, R. M. (2004b). Cliff-edged fitness functions and the persistence of schizophrenia. Behavioral and Brain Sciences, 27, 862–863. doi: 10.1017/S0140525X04300191 Nesse, R. M. (2005). Natural selection and the regulation of defenses: A signal detection analysis of the smoke detector principle. Evolution and Human Behavior, 26, 88–105. doi: 10.1016/j.evolhumbehav .2004.08.002 Nesse, R. M. (2006). Evolutionary explanations of mood and mood disorders. In D. J. Stein, D. J. Kupfur, & A. F. Schatzberg (eds.), American Psychiatric Publishing textbook of mood disorders (pp. 159–175). Arlington, VA: American Psychiatric Publishing. Nesse, R. M. (2011). Ten questions for evolutionary studies of disease vulnerability. Evolutionary Applications, 4, 264–277. doi: 10.1111/j.1752–4571.2010.00181.x Nesse, R. M., & Jackson, E. D. (2006). Evolution: psychiatric nosology’s missing biological foundation. Clinical Neuropsychiatry, 3, 121–131. Nettle, D. (2001). Strong imagination: Madness, creativity and human nature. Oxford, UK: Oxford University Press. Nettle, D. (2004). Evolutionary origins of depression: A review and reformulation. Journal of Affective Disorders, 81, 91–102. doi: 10.1016/j.jad.2003.08.009 Nettle, D. (2006). Reconciling the mutation–selection balance model with the schizotypy-creativity connection. Behavioral and Brain Sciences, 29, 418. doi: 10.1017/S0140525X06359092

Oyama, S., Griffiths, P., & Gray, R. (2001) (Eds.), Cycles of contingency. Cambridge, MA: MIT Press. Painter, R. C., Westendorp, R. G. J., de Rooij, S. R., Osmond, C., Barker, D. J. P., & Rosenboom, T. J. (2008). Increased reproductive success of women after prenatal undernutrition. Human Reproduction, 23, 2591–2595. doi: 10.1093/humrep/den274 Patterson, P. H. (2011). Infectious behavior: Brain-immune connections in autism, schizophrenia, and depression. Cambridge, MA: MIT Press. Pérusse, D. (1993). Cultural and reproductive success in industrial societies: Testing the relationship at the proximate and ultimate levels. Behavioral and Brain Sciences, 16, 267–322. doi:10.1017 /S0140525X00029939 Philippi, T., & Seger, J. (1989). Hedging one’s evolutionary bets, revisited. Trends in Ecology and Evolution, 4, 41–44. doi: 10.1016 /0169–5347(89)90138–9 Pluess, M., & Belsky, J. (2009). Differential susceptibility to rearing experience: the case of childcare. Journal of Child Psychology and Psychiatry, 50, 396–404. doi: 10.1111/j.1469–7610.2008.01992.x Pluess, M., & Belsky, J. (2011). Prenatal programming of postnatal plasticity? Development and Psychopathology, 23, 29–38. doi: 10.1017/S0954579410000623 Pluess, M., & Belsky, J. (2013). Vantage sensitivity: individual differences in response to positive experiences. Psychological Bulletin, 139, 901–916. doi:10.137/a0030196 Porges, S.W. (2007). The polyvagal perspective. Biological Psychology, 74, 116–143. doi: 10.1016/j.biopsycho.2006.06.009 Pradeu, T. (2010). The organism in developmental systems theory. Biological Theory, 5, 216–222. Price, J. S., Sloman, L., Gardner, R., Gilbert, P., & Rohde, P. (1994). The social competition hypothesis of depression. British Journal of Psychiatry, 164, 309–315. doi: 10.1192/bjp. 164.3.309 Quinlan, R. J. (2008). Human pair-bonds: Evolutionary functions, ecological variation, and adaptive development. Evolutionary Anthropology, 17, 227–238. doi: 10.1002/evan.20191 Raison, C. L., Lowry, C. A., & Rook, G. A. W. (2010). Inflammation, sanitation, and consternation: Loss of contact with coevolved, tolerogenic microorganisms and the pathophysiology and treatment of major depression. Archives of General Psychiatry, 67, 1211–1224. doi: 10.1001/archgenpsychiatry.2010.161 Réale, D., Garant, D., Humphries, M. M., Bergeron, P., Careau, V., & Montiglio, P.-O. (2010). Personality and the emergence of the pace-of-life syndrome concept at the population level. Philosophical Transactions of the Royal Society B, 365, 4051–4063. doi: 10.1098/rstb.2010.0208 Reese, H. W., & Overton, W. F. (1970). Models of development and theories of development. In L. R. Goulet & P. B. Baltes (Eds.), Life span developmental psychology: Research and theory (pp. 115–145). New York, NY: Academic Press. Robert, J. S., Hall, B. K., & Olson, W. M. (2001). Bridging the gap between developmental systems theory and evolutionary developmental biology. BioEssays, 23, 954–962. doi: 10.1002/bies.1136 Rutter, M. (1993). Resilience: Some conceptual considerations. Journal of Adolescent Health, 14, 690–696.

References Salmon, C., Figueredo, A. J., & Woodburn, L. (2009). Life history strategy and disordered eating behavior. Evolutionary Psychology, 7, 585–600. Sanders, S. J., Murtha, M. T., Gupta, A. R., Murdoch, J. D., Raubeson, M. J., Willsey, A. J., . . . , & State, M. W. (2012). De novo mutations revealed by whole-exome sequencing are strongly associated with autism. Nature, 485, 237–241. doi:10.1038/nature10945

57

Szechtman, H., & Woody, E. (2004). Obsessive-compulsive disorder as a disturbance of security motivation. Psychological Review, 111, 111–127. doi: 10.1037/0033–295X.111.1.111

Schlichting, C. D., & Pigliucci, M. (1998). Phenotypic evolution: A reaction norm perspective. Sunderland, MA: Sinauer Associates.

Thompson-Brenner, H., Eddy, K. T., Franko, D. L., Dorer, D. J., Vashchenko, M., Kass, A. E., & Herzog, D. B. (2008). A personality classification system for eating disorders: A longitudinal study. Comprehensive Psychiatry, 49, 551–560. doi: 10.1016/j.comppsych .2008.04.002

Schlomer, G. L., Del Giudice, M., & Ellis, B. J. (2011). Parent–offspring conflict theory: An evolutionary framework for understanding conflict within human families. Psychological Review, 118, 496–521. doi: 10.1037/a0024043

Thornhill, R., & Gangestad, S. W. (2004). The evolutionary psychology of human physical attraction and attractiveness. In A. Moya & E. Font (Eds.), Evolution: From molecules to ecosystems (pp. 247–259). New York, NY: Oxford University Press.

Scott-Phillips, T. C., Dickins, T. E., & West, S. A. (2011). Evolutionary theory and the ultimate-proximate distinction in the human behavioral sciences. Perspectives on Psychological Science, 6, 38–47. doi: 10.1353/pbm.2011.0020

Tinbergen, N. (1963). On the aims and methods of ethology. Zeitschrift für Tierpsychologie, 20, 410–433. doi: 10.1111/j.1439– 0310.1963.tb01161.x

Seery, M. D. (2011). Resilience: A silver lining to experiencing adverse life events? Current Directions in Psychological Science, 20, 390–395. doi: 10.1177/0963721411424740

Tooby, J., & Cosmides, L. (1990). The past explains the present: Emotional adaptations and the structure of ancestral environments. Ethology and Sociobiology, 11, 375–424. doi: 10.1016/0162– 3095(90)90017-Z

Segerstrale, U. (2000). Defenders of the truth: The battle for science in the sociobiology debate and beyond. Oxford, UK: Oxford University Press.

Tooby, J., Cosmides, L., & Barrett, H. C. (2003). The second law of thermodynamics is the first law of psychology: Evolutionary developmental psychology and the theory of tandem, coordinated inheritances: Comment on Lickliter and Honeycutt (2003). Psychological Bulletin, 129, 858–865. doi: 10.1037/0033–2909.129.6.858

Shaner, A., Miller, G. F., & Mintz, J. (2004). Schizophrenia as one extreme of a sexually selected fitness indicator. Schizophrenia Research, 70, 101–109. doi:10.1016/j.schres.2003.10.005

Trivers, R. L. (1972). Parental investment and sexual selection. In B. Campbell (Ed.), Sexual selection and the descent of man 1871–1971 (pp. 136–179). Chicago, IL: Aldine.

Shea, N., Pen, I., & Uller, T. (2011). Three epigenetic information channels and their different roles in evolution. Journal of Evolutionary Biology, 24, 1178–1187. doi: 10.1111/j.1420–9101.2011.02235.x

Trivers, R. L. (1974). Parent–offspring conflict. American Zoologist, 14, 249–264. doi: 10.1093/icb/14.1.249

Seger, J., & Brockmann, H. J. (1987). What is bet-hedging? Oxford Surveys in Evolutionary Biology, 4, 182–211.

Shonkoff, J. P., Garner, A. S., Siegel, B. S., Dobbins, M. I., Earls, M. F., Garner, A. S., . . . Wood, D. L. (2012). The lifelong effects of early childhood adversity and toxic stress. Pediatrics, 129, e232–e246. doi: 10.1542/peds.2011–2663 Simpson, J. A., Griskevicius, V., Kuo, S. I.-C., Sung, S., & Collins, W. A. (2012). Evolution, stress, and sensitive periods: The influence of unpredictability in early versus late childhood on sex and risky behavior. Developmental Psychology, 48, 674–686. doi: 10.1037/a0027293 Sloman, L., & Price, J. S. (1987). Losing behavior (yielding subroutine) and human depression: proximate and selective mechanisms. Ethology and Sociobiology, 8, 99–109. doi: 10.1016/0162–3095(87)90022–7

Troisi, A., & McGuire, M. T. (2002). Darwinian psychiatry and the concept of mental disorder. Neuroendocrinology Letters, 23, 31–38. Van Buskirk, J., & Relyea, R. A. (1998) Selection for phenotypic plasticity in Rana sylvatica tadpoles. Biological Journal of the Linnean Society, 65, 301–328. doi: 10.1111/j.1095–8312.1998.tb01144.x van Goozen, S. H. M., Fairchild, G., Snoek, H., & Harold, G. T. (2007). The evidence for a neurobiological model of childhood antisocial behavior. Psychological Bulletin, 133, 149–182. doi: 10.1037/0033–2909.133.1.149

Sober, E., & Wilson, D. S. (1998). Unto others: the evolution and psychology of unselfish behavior. Cambridge, MA: Harvard University Press.

Van IJzendoorn, M. H., Belsky, J., & Balermans-Kranenburg, M. J. (2012). Serotonin transporter genotype 5HTTLPR as a marker of differential susceptibility? A meta-analysis of child and adolescent gene-by-environment studies. Translational Psychiatry, 2, e147. doi:10.1038/tp. 2012.73

Starcke, K., & Brand, M. (2012). Decision making under stress: a selective review. Neuroscience and Biobehavioral Reviews, 36, 1228–1248. doi: 10.1016/j.neubiorev.2012.02.003

Volk, A. A., & Atkinson, J. A. (2013). Infant and child death in the human environment of evolutionary adaptation. Evolution and Human Behavior, 34, 182–192. doi: 10.1016/j.evolhumbehav.2012.11.007

Starrfelt, J., & Kokko, H. (2012). Bet-hedging—a triple trade-off between means, variances and correlations. Biological Reviews, 85, 742–755. doi: 10.1111/j.1469–185X.2012.00225.x

Wakefield, J. C. (1992). The concept of mental disorder: On the boundary between biological facts and social values. American Psychologist, 47, 373–388. doi: 10.1037/0003–066X.47.3.373

Stearns, S. (1992). The evolution of life histories. Oxford, UK: Oxford University Press.

Wakefield, J. C. (1999). Evolutionary versus prototype analyses of the concept of disorder. Journal of Abnormal Psychology, 108, 374–399. doi: 10.1037/0021–843X.108.3.374

Stifter, C. A., & Corey, J. M. (2001). Vagal regulation and observed social behavior in infancy. Social Development, 10, 189–201. doi: 10.1111 /1467–9507.00158 Stupica, B., Sherman, L. J., & Cassidy, J. (2011). Newborn irritability moderates the association between infant attachment security and toddler exploration and sociability. Child Development, 82, 1381–1389. doi: 10.1111/j.1467–8624.2011.01638.x Surbey, M. K. (1987). Anorexia nervosa, amenorrhea, and adaptation. Ethology and Sociobiology, 8, 47–61. doi: 10.1016/0162– 3095(87)90018–5

Wakefield, J. C. (2011). Darwin, functional explanation, and the philosophy of psychiatry. In P. R. Adriaens & A. De Block (Eds.), Maladapting minds: Philosophy, psychiatry, and evolutionary theory (pp. 141–172). Oxford, UK: Oxford University Press. Watson, D. (2005). Rethinking the mood and anxiety disorders: A quantitative hierarchical model for DSM-V. Journal of Abnormal Psychology, 114, 522–536. doi: 10.1037/0021–843X.114.4.522 Watson, P. J., & Andrews, P. W. (2002). Towards a revised evolutionary adaptationist analysis of depression: The social navigation

58

Evolutionary Foundations of Developmental Psychopathology

hypothesis. Journal of Affective Disorders, 72, 1–14. doi: 10.1016/ S0165–0327(01)00459–1 Waynforth, D. (2012). Life-history theory, chronic childhood illness and the timing of first reproduction in a British birth cohort. Proceedings of the Royal Society B, 279, 2998–3002. doi: 10.1098/rspb .2012.0220 West, S. A., Griffin, A. S., & Gardner, A. (2007). Evolutionary explanations for cooperation. Current Biology, 17, R661–R672. doi: 10.1016/j.cub.2007.06.004 West-Eberhard, M. J. (2003). Developmental plasticity and evolution. New York, NY: Oxford University Press. Westen, D., & Harnden-Fischer, J. (2001). Personality profiles in eating disorders: Rethinking the distinction between axis I and axis II. American Journal of Psychiatry, 158, 547–562. doi:10.1176 /appi.ajp. 158.4.547 Wilkins, J. F., & Haig, D. (2003). What good is genomic imprinting: the function of parent-specific gene expression. Nature Reviews Genetics, 4, 359–368. doi:10.1038/nrg1062

Williams, G. C. (1966). Adaptation and natural selection. Princeton, NJ: Princeton University Press. Wilson, D. S., & Yoshimura, J. (1994). On the coexistence of specialists and generalists. American Naturalist, 144, 692–707. Wilson, M., Daly, M., & Pound, N. (2002). An evolutionary psychological perspective on the modulation of competitive confrontation and risk-taking. In D. W. Pfaff, A. P. Arnold, A. M. Etgen, S. E. Fahrbach, & R. T. Rubin (Eds.), Hormones, brain and behavior (pp. 381–408). San Diego, CA: Academic Press. Wolf, M., van Doorn, G. S., & Weissing, F. J. (2008). Evolutionary emergence of responsive and unresponsive personalities. Proceedings of the National Academy of Sciences USA, 105, 15825–15830. doi: 10.1073/pnas.0805473105 Woody, E. Z., & Szechtman, H. (2011). Adaptation to potential threat: The evolution, neurobiology, and psychopathology of the security motivation system. Neuroscience and Biobehavioral Reviews, 35, 1019–1033. doi: 10.1016/j.neubiorev.2010.08.003

CHAPTER 2

Differential Susceptibility to Environmental Influences JAY BELSKY and MICHAEL PLUESS

DIFFERENTIAL SUSCEPTIBILITY TO ENVIRONMENTAL INFLUENCES 59 DIATHESIS–STRESS 61 Developmental Psychopathology Foundations 61 BEYOND DIATHESIS–STRESS 63 Evolutionary-Developmental Theories of Differential Susceptibility 63 Toward an Integrated Differential Susceptibility Paradigm 66 METHODOLOGICAL CONSIDERATIONS IN EVALUATING DIFFERENTIAL SUSCEPTIBILITY 67 The Importance of Securing Adequate Environmental Variance 68 Ecological, Cultural, and Racial-Ethnic Dimensions of Differential Susceptibility 68 Statistical Criteria for Evaluating Differential Susceptibility 69 BEHAVIORAL MARKERS OF DIFFERENTIAL SUSCEPTIBILITY 69 Negative Emotionality and Difficult Temperament as Plasticity Markers 70 Comment 74 PHYSIOLOGICAL MARKERS OF DIFFERENTIAL SUSCEPTIBILITY 75 GENETIC MARKERS OF DIFFERENTIAL SUSCEPTIBILITY 77 Dopamine Receptor D4 Gene (DRD4) 77 Serotonin Transporter Gene (5-HTT) 78 Monoamine Oxidase A Gene (MAOA) 80 Serotonin Receptor 2A Gene (HTR2A) 81 Tryptophan Hydroxylase 1 Gene (TPH1) 81

Dopamine Receptor D2 Gene (DRD2) 82 Additional Plasticity Genes? 82 Polygenetic Plasticity 83 GxE Mechanisms 84 EXPERIMENTAL EVALUATION OF VARIATION IN DEVELOPMENTAL PLASTICITY 85 Negative Emotionality and Physiological Reactivity 85 Genetics 86 REPEATED MEASUREMENTS 87 UNKNOWNS IN THE DIFFERENTIAL SUSCEPTIBILITY EQUATION 88 Same Individuals, Different Plasticity Markers? 88 Categorical or Dimensional Scaling of Plasticity? 88 Domain Specific or Domain General? 88 Origins of Plasticity: Nature or Nurture (or Both)? 89 Population Variation in Plasticity? 90 Gender Differences in Plasticity? 91 Competitive Evaluation of Models of Person–Environment Interaction 91 Variation in Environmental Cue Reliability 92 Parent–Child Conflict of Interest 92 Family Dynamics 92 Timing of Susceptibility 93 For Better and For Worse—or Just for Better? 93 FUTURE DIRECTIONS IN RESEARCH ON DIFFERENTIAL SUSCEPTIBILITY 94 GENERAL CONCLUSION 96 REFERENCES 96

DIFFERENTIAL SUSCEPTIBILITY TO ENVIRONMENTAL INFLUENCES

including attachment theory, social-learning theory and life-course sociology. Moreover, it guides a diverse array of research, including, for example, work chronicling links between prenatal maternal stress and temperament in infancy (Huizink, de Medina, Mulder, Visser, & Buitelaar, 2002), between maternal sensitivity in the first year and attachment security thereafter (Bakermans-Kranenburg, van IJzendoorn, & Juffer, 2003), between quality of child care in the preschool years and vocabulary scores at age 10 (Belsky, Vandell, et al., 2007), between parental divorce/separation in early adolescence and school grades in the later teenage years (Lansford et al., 2006), and

Central to much thinking about human development is the presumption that experiences in the early years of life— and even thereafter—influence individual differences in later development. Indeed, such developmental plasticity is a central tenant of most theories of human development, Jay Belsky’s work on this chapter was made possible with the support of the endowment for the Robert M. and Natalie Dorn Professorship that he holds. 59

60

Differential Susceptibility

between early family environment and depression symptomatology in early adulthood (Taylor et al., 2006). In this chapter we provide a detailed introduction to theory and evidence pertaining to differential susceptibility (DS) (Belsky, 1997b, 2005; Belsky & Pluess, 2009a, 2013a), a relatively new conceptual framework stipulating that individuals differ substantially in their degree of developmental plasticity, with some being generally more and some generally less plastic. Importantly, this differential susceptibility hypothesis is based on evolutionary reasoning about developmental plasticity. What many developmentalists have failed to appreciate regarding developmental plasticity is that such plasticity carries costs (DeWitt, Sih, & Wilson, 1998; Frankenhuis & Del Giudice, 2012; Schlichting & Pigliucci, 1998; Sih, 2011). First there is the cost associated with increased phenotypic complexity that may be required for an organism to be able to develop in diverse ways, depending on contextual conditions. Some of these costs derive from the fact that as complexity increases, so does the risk of something going wrong, in the same way that a machine with more moving parts has more ways of breaking down than one with fewer. Consider as well the following and related costs of plasticity enumerated by DeWitt and associates (1998): energetic costs of the sensory and regulatory mechanisms of plasticity; the production cost of environmentally inducible structures and processes; the cost of acquiring information about the environment, including the energy expended and potential risks incurred; and the imprecision in manufacturing a phenotype based on environmental experience. Collectively, these would seem to imply a trade-off between the benefits of being able to adjust development in accord with prevailing or anticipated contextual conditions—and thus being able to thrive in more than a single ecological niche—and the costs of a breakdown in the developmental machinery and processes needed to do so. Also important to appreciate with respect to the costs of plasticity is that modifying development early in life in response to developmental experiences and environmental exposures may result in a “mismatch” between the developed organism and the environment in which it finds itself later in life (Nederhof & Schmidt, 2012; Sih, 2011). In this way, seemingly adaptive developmental plasticity, selected to allow the organism to be programmed on the basis of early experience to fit the later environment, may, in some cases, actually undermine developmental opportunity and eventual fitness rather than enhance it. The potentially high costs associated with developmental plasticity as well as the potential risk of mismatch between programmed phenotype and future environment

would seem to provide grounds to expect individual differences in the degree of developmental plasticity within a species, the fundamental focus of this chapter. After all, individuals whose development is more canalized and less programmable by environmental exposures would be less likely to experience a mismatch resulting from developmental plasticity. Of course, such highly canalized specialists with more or less fixed characteristics could still end up mismatched to their environment, but for reasons different from that of the more plastic, programmable generalist who may develop in a manner guided by experiences had earlier in life. In sum, what these observations imply is that being developmentally malleable should not be regarded, as it typically is by developmental scholars, as an unmitigated good for the developing organism and thus as something that should be pronounced in all individuals. In some cases it would appear wiser not to be plastic or to be less malleable—manifesting a more canalized pattern of development—than others and thus less susceptible to environmental regulation. What this suggests, as now noted repeatedly, is that we should expect individual differences in developmental plasticity. In other word, developmental plasticity should be conceptualized as a phenotype in its own right, something that will vary across individuals—for a variety of reasons (Belsky & Pluess, 2009b). As it turns out, this notion of individual differences in developmental plasticity has long been appreciated by students of both normal development and psychopathology, even if not always expressed in such terms. Indeed, the notion of individual differences in plasticity is central to the model of human development and environmental action that dominates so much thinking and research. For example, the diathesis–stress framework presumes that not all individuals are equally vulnerable to the negative effects of adverse environmental influences (Gottesman & Shields, 1967; Monroe & Simons, 1991; Zuckerman, 1999). In recent years, an alternative framework has been advanced, that of differential susceptibility to environmental influence (Belsky, 1997b, 2005; Belsky, Bakermans-Kranenburg, & van IJzendoorn, 2007; Belsky & Pluess, 2009a; Ellis, Boyce, Belsky, Bakermans-Kranenburg, & van IJzendoorn, 2011), which is the primary subject of this chapter. We thus begin by considering the diathesis–stress framework before outlining that of differential susceptibility. We then discuss some methodological issues pertinent to the investigation of differential susceptibility before proceeding to review evidence consistent with the latter, highlighting organismic factors (i.e., temperament, physiology, genotype) that moderate responsiveness to environmental influences. Because

Diathesis–Stress

As already noted, behavioral scientists and developmental psychopathologists have long appreciated that individuals vary in their susceptibility to the influence of environmental factors and forces. It is this presumption, after all, that has long motivated interest among many in treatment–aptitude interactions, including personalized medicine today (Shoham & Insel, 2011). But in the realm of both normal and psychopathological functioning and development, it is the diathesis–stress model of environmental action (Gottesman & Shields, 1967; Monroe & Simons, 1991; Zuckerman, 1999) that most clearly highlights the issue at hand—individual differences in responsiveness to the environment. Of especial importance is that the diathesis–stress framework is principally, even exclusively, concerned with variation in response to contextual adversity. More specifically, diathesis–stress thinking stipulates that some individuals are disproportionately, if not exclusively, likely to succumb to the negative effects of some contextual stressor (e.g., poverty, negative life events). This heightened vulnerability to adversity may be due to organismic characteristics of the individual, such as genetic makeup, personality or temperament, which will be of principal concern here, as well as family factors and processes or extrafamilial conditions. Not unrelatedly, developmental psychopathologists have embraced the concept of resilience, which is really the inverse of vulnerability (Garmezy, Masten, & Tellegen, 1984; Masten, Best, & Garmezy, 1990; Rutter, 1987; Werner & Smith, 1982). Empirical research on risk and resilience first emerged in the 1970s (e.g., Garmezy, 1974; Rutter, 1979; Werner & Smith, 1977) and has come to exert a major influence on the study of development and psychopathology (for reviews, see Luthar, Cicchetti, &

Diathesis–Stress Model positive

Developmental Psychopathology Foundations

Outcome

DIATHESIS–STRESS

Becker, 2000; Masten, 2007; Rutter, 2012). Indeed, one of the most important and valuable insights that emerged over the past decades has been that individuals differ in the degree to which they are affected by contextual adversity. Whereas some vulnerable children and adults are likely to develop problematically in response to negative contextual conditions—by becoming emotionally and behaviorally dysregulated or disordered—others are not and thus regarded as resilient. Like vulnerability, resilience may be a function of a variety of factors, including organismic (e.g., intelligence, humor, genes) and contextual ones (e.g., supportive friend or teacher (Garmezy, 1974)) (Figure 2.1). To be appreciated, however, is that resilience from this perspective is understood as a dynamic process rather than an inherent characteristic of an individual (Cicchetti & Rogosch, 2012; Rutter, 2012). Prior to the emergence of theorizing about adaptive functioning in the face of significant adversity (i.e., resilience), investigations of risk and psychopathology generally portrayed the developmental process as somewhat deterministic, resulting in negative or maladaptive outcomes (Cicchetti & Rogosch, 2012). Ultimately, the concept of resilience had its origins in the observation that “there is huge heterogeneity in response to all manner of environmental hazards: physical and psychosocial”

Resilience

negative

all the work reviewed through this point is observational in character, a subsequent section focuses on experimental manipulations of the environment that illuminate whether intervention effects are moderated by plasticity factors. Thereafter, we call attention to studies that actually consider repeated-measurement involving how the very same individuals behave when exposed to both negative and positive contextual conditions. Finally, we raise a host of issues under the guise of unknowns in the differential susceptibility equation to guide future research before drawing some general conclusions.

61

ility

rab

lne Vu

negative

positive environment/experience

Figure 2.1 Graphical illustration of diathesis–stress. Vulnerability (i.e., diathesis) describes the propensity to respond negatively to adverse experiences, as a function of individual characteristics, whereas resilience reflects protective resistance from the same negative influences. No differences are predicted in response to positive influences. Source: Adapted from M. J. Bakermans-Kranenburg & M. H. van IJzendoorn, Gene–environment interaction of the dopamine D4 receptor (DRD4) and observed maternal insensitivity predicting externalizing behavior in preschoolers. Developmental Psychobiology, 48(5), 2006, Figure 1.

62

Differential Susceptibility

(Rutter, 2012, p. 336). But it needs to be appreciated given what is to follow in the discussion of differential susceptibility that “resilience is defined in terms of a better outcome than that seen in other individuals from a similarly adverse background . . . there is no requirement of superior functioning in relation to the non-deprived population as a whole” (Rutter, 2012, p. 342). As we will see, this feature of the resilient individual distinguishes diathesis–stress thinking from that of differential susceptibility. Intriguingly, similar views were advanced by ecological scientists and biologists (Hanson & Gottesman, 2012), starting with Holling’s (1973) work on ecological systems, with resilience defined as the capacity of a system to absorb disturbance and reorganize while undergoing change to retain essentially the same function, structure, identity, and feedbacks (Walker, Hollin, Carpenter, & Kinzig, 2004). Developmental geneticists captured a similar notion using the term canalization (Gottesman, 1974; Waddington, 1942). A Psychological Challenge It is not surprising that those who think predominately in diathesis–stress and resilience terms when considering human development—given their specific research interest in disordered, troubled and problematic functioning or the absence thereof—have rarely wondered about what might be regarded as the other half of the person–context interaction equation: How do individual differences, including with regard to organismic vulnerability factors, play out when circumstances are not adverse but benign, supportive or even enriching (Pluess & Belsky, 2013)? This critical question is central to this chapter because the differential susceptibility paradigm posits that those putatively vulnerable to adversity are also likely to be more responsive to positive and growth-facilitating developmental experiences and environmental exposures (Belsky, 1997b, 2005; Belsky, Bakermans-Kranenburg, et al., 2007; Belsky & Pluess, 2009a; Ellis, Boyce, et al., 2011). From this perspective, questions can be raised as to whether one should even speak of vulnerability factors (alone). By the same token, from the differential susceptibility perspective, those resilient and thus protected or buffered from the negative effects of adversity are also less likely—or entirely unlikely—to reap the rewards of a benign, supportive or enriching environment. This alternative analysis raises the prospect that so-called vulnerable individuals are not as unlucky or unfortunate as they might appear, nor are the so-called resilient as fortunate as they would seem to be according to traditional developmental psychopathological thinking.

An Evolutionary Challenge An evolutionary perspective also raises issues about the prevailing developmental-psychopathological analysis of dysfunctional or maladaptive outcomes within settings of adversity. In particular, it contends that both stressful and supportive environments have been part of human experience throughout our evolutionary history, and that developmental systems shaped by natural selection respond adaptively to both kinds of contexts. Thus, when people encounter stressful environments, this does not so much disturb their development—as widely assumed—as direct or regulate it toward strategies that are adaptive under stressful conditions—even if those strategies are currently harmful vis-à-vis the long-term welfare of the individual or society as a whole (Ellis, Boyce, et al., 2011; Hinde & Stevenson-Hinde, 1990; Main & Solomon, 1990). Consider the extensive experimental work conducted by Michael Meaney and colleagues showing that putatively low-quality maternal care in the rat (i.e., low levels of maternal licking and grooming) alters pups’ stress physiology and brain morphology. Although such changes seem disadvantageous (i.e., higher corticosterone levels, shorter dendritic branch lengths, and lower spine density in hippocampal neurons), they actually enhance learning and memory processes under stressful conditions (Champagne et al., 2008). Moreover, such physiological and morphological changes mediate the effects of maternal behavior on central features of defensive and reproductive strategies—behavior under threat, open-field exploration, pubertal development, sexual behavior, and parenting (Cameron, Fish, & Meaney, 2008; Cameron et al., 2005)—in ways consistent with evolutionary models of adaptive reproductive strategies (Belsky, Steinberg, & Draper, 1991; Chisholm, 1993). In the rodent model, then, enhanced learning under stressful conditions, increased fearful and defensive behaviors, accelerated sexual maturation, increased sexual behavior, and reduced parental investment in offspring apparently represent strategic—that is, functional—ways of developing when suffering parental neglect. In this context, neglect itself can be regarded as a mechanism through which rat mothers guide or program their offspring’s development toward optimal survival and reproductive strategies under conditions of adversity. It would seem mistaken, therefore, to view diminished licking and grooming as inherently negative or the development induced by such care as disturbed. From an evolutionary perspective, the care provided by the putatively neglectful parents may be appropriate preparation of their offspring for the

Beyond Diathesis–Stress

patterns of development in response to both stressful and supportive environmental conditions (within the range encountered over human evolution) constitute adaptive variation (Ellis, Boyce, et al., 2011).

BEYOND DIATHESIS–STRESS Evolutionary-Developmental Theories of Differential Susceptibility In addition to shaping species-typical developmental responses to diverse environmental conditions, natural selection has also maintained variation—adaptive individual differences—in neurobiological susceptibility to the environment. Two different evolutionary accounts of such variation in human development have emerged in recent years (Ellis, Boyce, et al., 2011): DS (Belsky, 1997b, 2005; Belsky, Bakermans-Kranenburg, et al., 2007); and biological sensitivity to context (BSC; Boyce & Ellis, 2005; Boyce et al., 1995; Ellis, Essex, & Boyce, 2005). With a central focus on person–environment interactions, both models highlight the role of organismic characteristics of individuals in moderating the effects of both stressful and supportive environmental conditions on human development (Figure 2.2). Thus, both presume that there

Outcome

positive

Differential Susceptibility Model

Low susceptibility

negative

ecological conditions into which they are likely to mature (Ellis, Boyce, et al., 2011). It is important to note that optimal adaptation (in the evolutionary sense) to challenging environments is not without real consequences and costs. Harsh environments often harm or kill children, and the fact that children developmentally adapt to such rearing conditions (reviewed in Ellis, Figueredo, Brumbach, & Schlomer, 2009; Pollak, 2008) does not imply that such conditions either promote child well-being or should be accepted as unmodifiable facts of life (i.e., David Hume’s naturalistic fallacy). There can be no doubt that high-stress environments that are dangerous and lack essential resources, compared with low-stress environments that are safe and well-resourced, cause substantial distress and eventually undermine fitness. In other words, developmental adaptations to high-stress environments simply enable individuals to make the best of a bad situation (i.e., to mitigate the inevitable fitness costs), even though the best may still constitute a high-risk strategy that jeopardizes the person’s health, well-being and survival (e.g., Mulvihill, 2005; Shonkoff, Boyce, & McEwen, 2009). It is also important to mention here the possibility of genuinely novel environments that are beyond the normative range of conditions encountered over human evolution. Exposures to such challenging yet (evolutionarily) unprecedented conditions can be expected to induce pathological development that does not reflect evolutionarily adaptive strategies (Ellis, Boyce, et al., 2011). In sum, an evolutionary-developmental perspective emphasizes conditional adaptation: “evolved mechanisms that detect and respond to specific features of childhood environments, features that have proven reliable over evolutionary time in predicting the nature of the social and physical world into which children will mature, and entrain developmental pathways that reliably matched those features during a species’ natural selective history” (Boyce & Ellis, 2005, p. 290; for a comprehensive treatment of conditional adaptation, see West-Eberhard, 2003; for applications to human development, see Belsky et al., 1991, 2000; Chisholm, 1996; Ellis, 2004). From within such a perspective, the highly susceptible child who responds to a dangerous environment by developing insecure attachments, adopting an opportunistic interpersonal orientation, and sustaining an early sexual debut is no less functional than the context-sensitive child who responds to a well-resourced and supportive social environment by developing the opposing characteristics and orientations. Children have evolved to function competently—to survive and ultimately reproduce—in a variety of contexts, and the default assumption should be that alternative

63

ti ep

bil

it y

sc

h Hig

su

negative

positive environment/experience

Figure 2.2 Graphical illustration of differential susceptibility. High susceptibility is characterized by increased responsivity to both negative and positive experiences, as a function of individual characteristics, in contrast to low susceptibility, which reflects a general psychological inertia to environmental influences independent of quality. Source: Adapted from M. J. Bakermans-Kranenburg & M. H. van IJzendoorn, Gene–environment interaction of the dopamine D4 receptor (DRD4) and observed maternal insensitivity predicting externalizing behavior in preschoolers. Developmental Psychobiology, 48(5), 2006, Figure 1.

64

Differential Susceptibility

are individual differences in sensitivity and responsiveness to environmental influence but, importantly, not in the manner of diathesis–stress. Both theories incorporate the traditional diathesis–stress view while extending it, by making the critical observation that those individuals most likely to be adversely affected (according to conventional mental health standards) by negative environmental conditions are also most likely to benefit from supportive ones. Focusing on the role of the person in moderating environmental effects on development, DS and BSC theories follow in the tradition of Bronfenbrenner’s (1993) person–process–context model, which posits that parenting and other environmental factors may vary in their developmental influence as a function of the characteristics of the child. The theories also converge with Wachs and Gandour’s (1983) organismic-specificity hypothesis, which posits differential reactivity of children to similar rearing experiences as a function of differences in children’s cognitive, behavioral, or emotional characteristics. Before proceeding with a review and comparison of DS and BSC theories, it is important to note that these perspectives share much in common with Aron and Aron’s (1997; Aron, Aron, & Davies, 2005) theory of sensory-processing sensitivity (SPS). The main difference is that DS and BSC began with a focus on child-developmental processes, whereas SPS started with a focus on cognitive processes in adults (in terms of variation in depth of processing of sensory information, with more sensitive individuals processing information more thoroughly before acting). Although some SPS studies have included retrospective assessments of childhood experiences, SPS did not originate from a developmental perspective. Given that longitudinal work on the role of SPS in regulating child (or adult) development has yet to be conducted, we consider it premature to endeavor to integrating SPS theory with BSC and DS at this time. Differential Susceptibility Theory The question of why childhood experiences should influence later development—one rarely posed by developmentalists concerned principally with how experience influences development—was the origin of DS theory, along with challenges posed by behavior-genetic critiques by Rowe (2000) of Belsky et al.’s (1991) evolutionary theory of socialization (Belsky, 2000). From an evolutionary perspective, developmental mechanisms that use earlier experiences to guide later development should only evolve in recurring contexts in which the future is tolerably related to the past (Pigliucci, 2001), at least within generations.

Only then could there be reliable fitness payoffs in using experiences in childhood to regulate adolescent and adult development (i.e., conditional adaptation). The fact that the future is inherently uncertain, however, meant that conditional adaptation was theoretically problematic (i.e., risk for potential mismatch). This realization led Belsky (1997a, 1997b, 2000, 2005) to propose that, as a form of bet-hedging against an uncertain future, natural selection has maintained genes for both conditional and alternative developmental strategies. Whereas conditional strategies are shaped by environmental factors to better fit the organism to the future environment, alternative strategies, in a manner consistent with much behavior-genetic thinking, are largely fixed and less subject to environmental regulation (Rowe, Vazsonyi, & Figueredo, 1997). Because the future is and always has been uncertain, no parent could ever know for certain what rearing strategies, whether consciously or unconsciously implemented, would prove most successful in terms of promoting the child’s reproductive fitness and thus the parent’s inclusive fitness. This suggested to Belsky (1997a, 1997b, 2000, 2005) that—especially within families—children should vary in their susceptibility to the rearing environment, broadly construed. Indeed, while it would make sense for parents to produce some children who pursued alternative strategies, perhaps entirely impervious to socialization efforts, and who would thrive in particular contexts that fit their proclivities, it would also make sense for them to bear some conditional strategists capable of fitting and thriving in a variety of niches depending upon the rearing conditions encountered while growing up. Belsky (2005) subsequently observed not only that parents would be (unconsciously) hedging their bets by diversifying their progeny’s susceptibility to rearing influence but also that the same would be true of children themselves. This was due to the fact that, just like parents and children, siblings share 50% of the same genetic alleles. Thus, if one child benefited from parental influence, so would the other, less susceptible sibling—but indirectly, via shared genes. By the same token, if one child’s development was undermined, inadvertently, by parental influence, the less susceptible child would be protected, thereby providing an indirect, inclusive-fitness benefit to the child whose susceptibility proved counterproductive in individual terms. From the perspective of both parent and child, then, differential susceptibility to rearing and perhaps other environmental factors and processes was considered to be evolutionarily advantageous. Although Belsky’s theorizing stipulated that children should vary in their susceptibility to environmental

Beyond Diathesis–Stress

influence, it did not specify what might distinguish those children who were more susceptible from those who were less. Indeed, the theory lacked any explanatory mechanism, something we will see distinguished it substantially from BSC theorizing which was stimulated by mechanistic thinking. Early attempts to identify potential susceptibility factors or markers called attention, somewhat surprisingly, to negative emotionality or difficult temperament early in life (Belsky, 1997a, 2005; Belsky, Hsieh, & Crnic, 1998), whereas gene × environment interaction work, as well as theory and research on physiological reactivity by Boyce and Ellis (Boyce et al., 1995, 2005), called attention to physiological and genetic markers of variation in susceptibility (e.g., Bakermans-Kranenburg & van IJzendoorn, 2007; Belsky & Pluess, 2009a). The differential susceptibility hypothesis, while not directly influenced by Plomin and Daniel’s (1987) important insights about nonshared environmental effects, was eminently consistent with it. Indeed, in some respects the differential susceptibility perspective offered an explanation as to why nonshared, within-family effects should be the rule, as they have proven to be in behavior-genetic research, rather than the exception, with shared environmental effects proving so modest (Plomin & Daniels, 1987; Reiss, Neiderhiser, Hetherington, & Plomin, 2000; Turkheimer & Waldron, 2000). Because children within families vary in their susceptibility to rearing influences, they should be differentially affected by exposure to the very same developmental experience. Differential susceptibility thinking could also account for why environmental effects have proven both variable and generally modest across studies—perhaps because samples have varied, inadvertently, in the proportion of more and less susceptible individuals which they include, and because such individuals have not been distinguished in estimations of average rearing effects (Belsky, Hsieh & Crnic, 1998). Biological Sensitivity to Context Theory Whereas DS thinking emerged initially as a totally theoretical account of human development, not stimulated by any empirical findings, just the reverse proved true of BSC thinking. Indeed, the concept of BSC has its early roots in Boyce and colleagues’ (1995) two studies of naturally occurring environmental adversities and stress reactivity as predictors of respiratory illnesses in 3–5-year-old children. Results revealed, first, that children showing low cardiovascular or immune reactivity to stressors had approximately equal rates of respiratory illnesses in both low and high adversity settings. Second, and consistent with the prevailing diathesis–stress model, highly biologically reactive

65

children exposed to high adversity child care settings or home environments had substantially higher illness incidences than all other groups of children. The third—and unexpected—finding was that highly reactive children living in lower adversity conditions (i.e., more supportive child care or family settings) had the lowest illness rates, significantly lower than even low reactivity children in comparable settings. Such unanticipated results—in need of an explanation—led Boyce et al. (1995, p. 419) to suggest that “reactive children . . . with a heightened sensitivity to . . . the environment might . . . be expected to experience unusually poor outcomes in high-stress, unsupportive social conditions. The same children might flourish, on the other hand, under low-stress, nurturing, and predictable conditions. . . .” Such a conclusion would subsequently lead to the differential susceptibility view that some children are more responsive to the environment for better and for worse (Belsky, Bakermans-Kranenburg, et al., 2007). Of critical and distinguishing importance is that the initial Boyce et al. (1995) research, together with subsequent work (Boyce & Ellis, 2005), identified a physiological mechanism of environmental susceptibility—autonomic, adrenocortical, or immune reactivity to psychosocial stressors—and proposed that psychobiologic reactivity moderated the effects of early environmental exposures on physical and mental health outcomes in a bivalent manner. More reactive children displayed heightened sensitivity to both positive and negative environmental influences and thus were given the shorthand designation of orchid children, signifying their special susceptibility to both highly stressful and highly nurturing environments. Children low in reactivity were designated as dandelion children, reflecting their relative ability to function adequately in species-typical circumstances of all varieties (Boyce & Ellis, 2005). Mechanistic analysis of BSC led Boyce and Ellis (2005) to highlight the significance of the stress response system, but not just as a long-appreciated means of regulating behavior in the face of adversity, but also in response to resources and support in the ambient environment (e.g., positive social opportunities, cooperative information). This dual function signified the need to re-conceptualize stress reactivity more broadly as biological sensitivity to context, which Boyce and Ellis (2005) defined as neurobiological susceptibility to both cost-inflicting and benefit-conferring features of the environment and operationalized as a physiological property indexed by heightened reactivity in one or more of the stress response systems. Depending on levels of nurturance and support versus harshness and unpredictability in their

66

Differential Susceptibility

developmental environments, highly reactive children experience either the best or the worst of psychiatric and biomedical outcomes within the populations from which they are drawn (Boyce, 1996; Boyce et al., 1995, 2006; Bubier, Drabick, & Breiner, 2009; Ellis, Shirtcliff, Boyce, Deardorff, & Essex, 2011; Essex, Armstrong, Burk, Goldsmith, & Boyce, 2011; Obradovic, Bush, & Boyce, 2011; Obradovic, Bush, Stamperdahl, Adler, & Boyce, 2010; Quas, Bauer, & Boyce, 2004). BSC theory therefore posits that individual differences in the magnitude of biological stress responses function to regulate openness or susceptibility to environmental influences, ranging from harmful to protective. Given past evidence that early trauma can increase stress reactivity and newer evidence that high reactivity can enhance developmental functioning in highly supportive settings, Boyce and Ellis (2005) postulated a curvilinear, U-shaped relation between levels of early support-adversity and the magnitude of biological response dispositions. Specifically, Boyce and Ellis hypothesized that: (1) exposure to acutely stressful childhood environments up-regulates BSC, increasing the capacity and tendency of individuals to detect and respond to environmental dangers and threats; (2) exposure to especially supportive childhood environments also up-regulates BSC, increasing susceptibility to social resources and support; and (3) by contrast, and typical of the majority of children, exposure to childhood environments that are not extreme in either direction down-regulates BSC, buffering individuals against the chronic stressors encountered in a world that is neither highly threatening nor consistently safe. Toward an Integrated Differential Susceptibility Paradigm Even though DS and BSC theories emerged independently and differ in important respects, they share much in common (Ellis, Boyce, et al., 2011). First, both theories, based as they are on an evolutionary analysis of human development, advance the claim that individuals differ systematically in their susceptibility to environmental influences (i.e., developmental plasticity) and seek to explain the nature of such individual differences. As will become apparent in the next major section of this chapter, this begins with the recognition that susceptibility to the environment is instantiated in the multiple genetic polymorphisms, physiological mechanisms, and behavioral phenotypes that operate as susceptibility factors, moderating the influence of environmental exposures on developmental and life outcomes. Several of the physiological mechanisms may be considered endophenotypes (Gottesman & Gould, 2003)

constituting links between genes and behavior, whereby (single or multiple) genetic markers of differential susceptibility operate through the same neurobiological processes in which behavioral indicators of differential susceptibility are grounded. Second, both DS and BSC theories presume that individuals should differ in their neurobiological susceptibility to environmental influence, and that such differential susceptibility underlies many reliable person–environment interactions regulating human development and functioning. Indeed, both theories embrace, implicitly if not explicitly, Bronfenbrenner’s (1979) dictum that when it comes to human development, the main effects are in the interactions, be those gene–environment, temperament– parenting, stress reactivity–family stress, or other person– environment processes. More specifically, central to both DS and BSC theories is the assumption that environmental influences on developmental and life outcomes are moderated by neurobiological susceptibility to the environment. Both theories thus conceptualize variation in neurobiological susceptibility as a central mechanism in the regulation of alternative patterns of human development. Third, neither theory presumes that differential susceptibility is restricted to any one developmental period or even just to childhood; both appreciate that at all developmental stages during the life span, individuals may differ in the extent to which they are affected by both supportive and challenging environments. Along these lines, research documenting differential susceptibility in children, adolescents, and adults is considered later in this chapter. Nonetheless, both DS and BSC theories were originally developed to explain neurobiological susceptibility in childhood, as clearly reflected in the preceding summaries of these perspectives. Finally, a core assertion of both DS and BSC theories is that individuals differ in neurobiological susceptibility to environmental contexts that are both positive in character (i.e., afford resources and support that potentially enhance fitness) and negative in character (i.e., embody stressors and adversities that potentially undermine fitness). In other words, and in contrast to widely embraced diathesis–stress models, it is not simply that some individuals are more susceptible to the negative effects of adversity, making them vulnerable. Rather, DS and BSC theories regard those disproportionately vulnerable to adversity as also disproportionately likely to benefit from supportive and enriching environments. Such individuals are thus affected by the environment in a manner that can be characterized as for better and for worse (Belsky, Bakermans-Kranenburg, et al., 2007). Another way of

Methodological Considerations in Evaluating Differential Susceptibility

describing or even conceptualizing the individual differences in question is in terms of reaction norms (Manuck, 2010). Whereas some individuals have a wide range of reaction in terms of their developmental functioning, depending on the environments they encounter, others have a much narrower range of reaction, responding less markedly—if at all—to positive and negative life experiences. It is critical to appreciate that differential susceptibility to positive and negative environments has different implications when viewed from developmental-psychopathology and evolutionary perspectives (Ellis, Boyce, et al., 2011). In the developmental-psychopathology framework, heightened neurobiological susceptibility increases the ability and tendency of individuals to experience good outcomes in positive environments (i.e., for better, as defined by dominant Western values like secure attachment, happiness, high self-esteem, emotion regulation, educational and professional success, stable marriage) and bad outcomes in negative environments (i.e., for worse, as defined by that same value system such as insecure attachment, substance abuse, conduct problems, depression, school failure, teenage pregnancy). By contrast, according to the evolutionary perspectives central to both DS and BSC theories, heightened neurobiological susceptibility to the environment functions to direct or regulate development in ways that, over human evolution, recurrently matched individuals to both positive and negative environments, thereby promoting reproductive fitness. In positive environments, this translates into adjusting development to optimize reproductively relevant processes and behaviors such as growth, status, fertility, and offspring quality. This form of conditional adaptation would typically be considered for better in a developmental psychopathology framework. In negative environments, however, this translates into making the best of a bad situation, often resulting in developmental outcomes that are typically regarded as nonoptimal in Western culture. A key difference, then, between the evolutionary and developmental-psychopathology perspectives is that evolutionary models conceptualize conditional adaptation to negative environments as a functional and strategic output of evolved developmental systems shaped by natural selection in the service of fitness goals. Consequently, even though susceptible individuals in negative environments may be especially vulnerable to poor mental health (as defined by dominant Western values), for example, they may still be acting in ways that promote—or once promoted—status and reproductive success in dangerous environments (e.g., gang membership in bad neighborhoods (see Palmer & Tilley, 1995), advantage taking, sexual promiscuity, limited parental investment).

67

Beyond these points of agreement, it is worth highlighting that both DS and BSC theories derive from evolutionary analyses of human development which, even if not identical, are grounded in the view that natural selection maintains alternative patterns of development (phenotypic variation) in the context of multiniche environments (e.g., Hinde & Stevenson-Hinde, 1990; Penke, Denissen, & Miller, 2007). Both theories posit that variation in neurobiological susceptibility to the environment has been adaptively structured and functions as a central mechanism in regulating alternative developmental pathways to match—as best possible—different environmental niches (Ellis, Boyce, et al., 2011). Fundamental to this view is the assumption that optimal developmental strategies vary as a function of the physical, economic, and social parameters of an individual’s specific environment. Levels of neurobiological susceptibility that promote success in some environments may therefore lead to failure in others. It is this kind of environmental heterogeneity—multiniche environments in which a trait’s effect on fitness varies across time or space—that provides the ecological basis for the maintenance of adaptive phenotypic variation (whether through balancing selection, conditional adaptation, bet hedging, or a combination thereof; see Ellis, Jackson, & Boyce, 2006; Ellis, Shirtcliff, Boyce, & Essex, 2009; Penke et al., 2007). In sum, the DS and BSC models largely converge on an integrated theory of neurobiological susceptibility to the environment. Taken together, these perspectives shed new light on the potency—and impotency—of a broad range of environmental contexts, from highly positive and enriching to dangerous and corrosive, to shape the gamut of developmental outcomes. Having delineated the theoretical framework of differential susceptibility while contrasting it to that of diathesis–stress, attention now turns to empirical evidence, seemingly consistent with differential susceptibility. Before reviewing relevant research, however, some methodological comments are regarding the empirical evaluation of differential susceptibility.

METHODOLOGICAL CONSIDERATIONS IN EVALUATING DIFFERENTIAL SUSCEPTIBILITY Three distinct methodological issues relating to differential susceptibility are of concern here: measurement of the environment; cultural and racial variation; statistical evaluation.

68

Differential Susceptibility

The Importance of Securing Adequate Environmental Variance Because the essence of the differential susceptibility model is that individuals who display high neurobiological susceptibility to the environment are not only disproportionately affected by negative contexts but also respond more favorably to positive environmental influences (Figure 2.2), it is imperative that tests of this hypothesis secure adequate variance in environmental conditions (Belsky, Bakermans-Kranenburg, et al., 2007; Ellis et al., 2005). This is because targeting risky environments may obscure the potential benefits of exposure to positive contexts for susceptible individuals, while a narrow focus on only positive environments or outcomes (though less often seen in practice) may have the opposite effect. Thus, a broad range of environmental qualities is a minimum condition to reveal differential susceptibility—a condition that is often not met in studies of developmental psychopathology and much of the research to be cited below. The importance of careful measurement of the environment (and the outcome) cannot be overestimated (Ellis, Boyce, et al., 2011). Some two decades ago Wachs and Plomin (1991) identified what they called Plomin’s paradox: if interaction effects—which are central to the evaluation of differential susceptibility—are ubiquitous in nature, why are they so difficult to detect in behavioral studies (see, McClelland & Judd, 1993)? One of the explanations could have to do with unstable or unreliable measures of the environment. When the error components of the genetic and environmental parts of the GxE equation strongly diverge, testing for moderation is at risk for both type 1 and type 2 errors. Therefore, in studies on differential susceptibility, more than usual care should be taken to assess the environment (and behavioral outcome) reliably and validly by extended behavioral observations in various situations, the use of multiple informants, and aggregation of data across settings and measures. Many of these conditions are not met in the research to be reviewed. Tests of differential susceptibility involve variation in organismic characteristics, environmental factors, and developmental outcomes (Figure 2.2). These variance components may be highly interdependent and context-dependent. Indeed, child effects and by extension person–environment-interaction effects are highly dependent on variance in the environment (Ellis, Boyce, et al., 2011). Quite importantly, person–environment interactions are much more likely to emerge in the range of average expectable environments than at environmental extremes (Cicchetti & Valentino, 2006; Hartmann, 1958),

where the power of context to shape human development may restrict the range of phenotypic variation. In twin studies, for example, physical growth (e.g., height) and IQ have been shown to be highly heritable, but the majority of children growing up in institutions show delays in growth and diminished IQ that become exponentially greater with longer institutionalization (e.g., Rutter & English Romanian Adoptees, 1998; van IJzendoorn, Luijk, & Juffer, 2008). Drastically improved care through adoption, foster care, or interventions within institutions results in massive catch up for virtually all children (Marian J. Bakermans-Kranenburg, van IJzendoorn, & Juffer, 2008; Nelson et al., 2007), leaving little room for interaction effects between environment and child characteristics. Ecological, Cultural, and Racial-Ethnic Dimensions of Differential Susceptibility Further, cross-cultural research raises the issue of the cultural specificity of the environment, the susceptibility/ plasticity factor, and the developmental outcome that together constitute the differential susceptibility equation (Ellis, Boyce, et al., 2011). Cross-cultural studies of differential susceptibility have documented that the meaning and context of specific parental behaviors, as well as the value placed on specific developmental outcomes, may vary between different cultural groups and across different ecological niches (Deater-Deckard, Pinkerton, & Scarr, 1996; Hinde & Stevenson-Hinde, 1990; Scheper-Hughes, 1992). Moreover, and with regard to race/ethnicity, it is not just familial factors and developmental outcomes that may be different, but neurobiological susceptibility as well. Attempting to replicate and extend Caspi and associates’ (2002) GxE findings (on child maltreatment, the MAOA genotype, and antisocial behavior), Widom and Brzustowicz (2006) replicated the original GxE result and detected evidence of differential susceptibility (Belsky & Pluess, 2009a). However, they also discovered that the GxE finding only applied to Caucasians, not to African Americans. (For examples pertaining to dopamine-related genes, see Bakermans-Kranenburg & van IJzendoorn, 2011). The explanation for this racial difference may be that the genetic effects in question are dependent on race: Short alleles of 5-HTTLPR are associated with the production of higher levels of serotonergic function in the central nervous system of African American participants but lower levels of serotonergic function among European American participants (Williams et al., 2003). Moreover, the distributions of genotypes differs substantially among the various parts of the world, showing for instance lower prevalence of

Behavioral Markers of Differential Susceptibility

the DRD4–7-repeat allele on the African continent and more carriers of the 5HTT short allele in Asia compared with Europe and North America (Chen, Burton, Greenberger, & Dmitrieva, 1999; Gelernter, Cubells, Kidd, Pakstis, & Kidd, 1999). Ethnically homogeneous samples are thus preferred in GxE investigations but not restricted to Caucasian samples in Western countries. The generalizability of GxE interaction effects reflecting differential susceptibility to populations of different cultures and races is not self-evident but should each time be empirically established. Statistical Criteria for Evaluating Differential Susceptibility The statistical test of differential susceptibility first proposed by Belsky et al. (2007) consists of a series of five consecutive steps. In short, the first always concerns the application of conventional statistical criteria for evaluating moderation (Dearing & Hamilton, 2006). The second step refers to various shapes moderator effects can take, not all of which are indicative of differential susceptibility (see Belsky, Bakermans-Kranenburg, et al., 2007, for a figure delineating various interaction effects). Interactions with regression lines that do not cross (sometimes referred to as removable interactions) do not document differential susceptibility, although they are not necessarily incompatible with aspects of differential susceptibility (depending on the range of environments covered). Differential susceptibility is more conclusively shown when the moderation reflects a crossover interaction that covers both the positive and the negative aspects of the environment. The slope for individuals high in neurobiological susceptibility should be significantly different from zero and at the same time significantly steeper than the slope for the individuals low in neurobiological susceptibility. If both slopes are significantly different from zero but in opposite directions (i.e., look like an X), contrastive effects rather than differential susceptibility effects are indicated. The next step requires that there be no association between the moderator (i.e., the susceptibility factor) and the environment (see Kraemer, Stice, Kazdin, Offord, & Kupfer, 2001). In other words, differential susceptibility should be distinguished from GxE or temperament– environment correlations (which may reflect rearing experiences evoked by child characteristics) and from dual-risk models (Figure 2.1). Belsky et al. (1998), for example, tested the independence of negative emotionality and parenting as a critical step in their investigation of differential susceptibility. Had these factors been correlated, then the

69

evidence would not have shown that the predictive power of parenting was greater for highly negative infants. Instead, it would have indicated either that highly negative infants evoke negative parenting or that negative parenting fosters infant negativity. Similarly, Caspi and Moffitt (2006) determined that boys’ MAOA genotype did not elicit maltreatment. Correlations between the environment and the moderator may be dealt with by partialling out the variance in the moderator explained by the environmental factor prior to testing for differential susceptibility as described already. Pluess and Belsky (2010) used this approach to overcome the problem in their investigation of differential susceptibility to parenting. A further step concerns the association between moderator (i.e., susceptibility factor) and outcome, which has to be zero to reflect differential susceptibility. If the susceptibility factor and the outcome are related, dual risk (or gain, when positive factors are involved) may be a more plausible model. Along these lines, if an intervention proves successful in counter acting the negative effects of a risk gene (or other risk factor) that is correlated with the outcome (e.g., 5-HTTLPR and youth risk behavior initiation, as in Brody, Beach, Philibert, Chen, & Murry, 2009), the experiment tests a protective factor model more than a differential susceptibility model in which the polymorphism should be a risk as well as a susceptibility gene. As a final step, Belsky et al. (2007) suggested testing the specificity of the effect by replacing the susceptibility factor (i.e., moderator) and outcomes (see Caspi, Harrington, et al., 2003, for an example of such replacement establishing discriminant validity in the GxE tradition). Because we are still in the early stages of testing differential susceptibility, not all of the empirical work to be considered below actually meets all the statistical criteria just enumerated. Moreover, as we will highlight in the last major section of the chapter on unknowns in the differential susceptibility equation, new statistical approaches have been proposed recently which afford competitive evaluation of differential susceptibility and other models of environmental action, including diathesis–stress. Now, however, attention turns to research chronicling differential susceptibility, organized in terms of first psychological, then physiological and finally genetic susceptibility factors.

BEHAVIORAL MARKERS OF DIFFERENTIAL SUSCEPTIBILITY The first evidence of differential susceptibility to be considered comes from research showing that temperamental and

70

Differential Susceptibility

emotional characteristics of (mostly very young) children moderate the effect of developmental experience on behavioral development. Some of this work highlights the role of early negative emotionality/difficult temperament vis-à-vis parenting effects, some of it the moderated influence of child-care experience, and some of it individual differences in plasticity beyond the early childhood years. Each set of evidence is considered in turn. Of note is that unless otherwise indicated the work to be cited was conducted without regard to the differential susceptibility hypothesis and, if anything, informed by diathesis–stress thinking. It is highlighted, nevertheless, due to evidence generated of differential susceptibility, whether appreciated by the investigators or not. Negative Emotionality and Difficult Temperament as Plasticity Markers Some of the earliest and most suggestive evidence of differential susceptibility to environmental influences emerged in research on temperament–parenting interaction, a long-standing focus of inquiry, typically conducted from a dual-risk/diathesis–stress perspective (Rothbart & Bates, 2006). In fact, after advancing the differential susceptibility hypothesis that, it will be recalled, included no claims regarding proximate factors or mechanisms that would make some children more susceptible than others, Belsky sought to identify some in existing and continually emerging developmental research. Indeed, well before any GxE research on humans was reported, he called attention to infant and toddler negative emotionality and difficult temperament as potential differential susceptibility factors (for review, see Belsky, 2005). Whether operationalized in terms of difficult temperament, irritability, fearfulness, or inhibition, cross sectional and longitudinal studies by Kochanska (1993); Belsky, Hsieh, and Crnic (1998); and Feldman, Greenbaum, and Yirmiya (1999) showed, for example, that diverse measures of rearing of infants and toddlers (e.g., discipline, interactional synchrony, positive and negative parenting) accounted for substantially more variance in self-control, externalizing problems and inhibition in the case of more negatively emotional infants/toddlers than other children. (All but the Kochanska, 1993, study were explicitly designed to test differential susceptibility.) But it was not just in research on very young children that such shorter and longer term rearing effects emerged. Morris and associates (2002) found, for example, that harsh and hostile mothering also proved to be a stronger predictor of teacher-reported externalizing problems in first and

second grade when children scored high rather than low on irritable distress. Most of the research on children reviewed by Belsky (2005) showed, following documentation of significant temperament–parenting interactions, that greater variance in a variety of developmental outcomes could be explained by rearing experiences in the case of more negatively emotional children, statistical analyses in the studies reviewed often did not afford determination of whether this result was itself a function of a for-better-and-for-worse parenting effect. In consequence, it remained undetermined whether individual differences in plasticity—or just vulnerability—were responsible for the repeatedly detected finding that more variance was explained in one group’s functioning than in another’s by the environmental factor investigated. Fortunately, a growing number of studies provide substantial empirical evidence that rearing and other environmental effects do not just account for more variance in the functioning of more negatively emotional children—or even that such individuals are simply more vulnerable to negative experiences as the dual-risk/diathesis–stress model would suggest—but that they are indeed differentially susceptible to environmental experiences in a for-better-and-for-worse manner. Van Aken, Junger, Verhoeven and Dekovic (2007) found, for example, that 16–19-month-old boys (N = 115) with difficult temperament showed the smallest increase six months later in externalizing problems scores when reared by highly sensitive mothers who only infrequently used negative control, but the largest increase when highly insensitive mothers relied heavily on negative control. These striking parenting effects simply did not obtain in the case of other children. In a series of investigations Kochanska, Aksan, and Joy (2007) evaluated whether child temperament moderated parenting effects on positive developmental outcomes. In one study children’s fearfulness, maternal power assertion, and mother–child positive relations were assessed behaviorally when children were 22 and 33 months old and children’s moral self was measured using a puppet interview at 56 months (N = 74). Although no parenting effects emerged in the case of children who, as toddlers, scored low in fear, those who were highly fearful evinced a greater moral sense if their mothers (at 22 months only) relied little on power assertion to regulate their behavior, but a limited one if their mothers relied heavily on power to control earlier child behavior. The fact that child fearfulness was itself significantly and negatively related to maternal power assertion raises some questions about how much confidence to place in this study when it comes

Behavioral Markers of Differential Susceptibility

to inferring individual differences in plasticity. Recall that in circumstances such as these in which the putative susceptibility factor is itself related to the environmental factor, thereby raising the prospect of evocative person effects, a better approach would be to partial the effect of the susceptibility factor from the environmental one before testing interactions between the two. This would statistically eliminate the potential evocative effect that confounds interpretation of differential susceptibility. In a second study (N = 100 families), this time focused on father’s reliance on power assertion (15 months), children’s fearfulness (7 and 15 months) and their rule compatible conduct (38 months), Kochanska et al. (2007) once again documented evidence of for-better-and-for-worse parenting effects: Whereas high versus low power assertiveness made no apparent difference for children scoring low in fearfulness (at 7 and 15 months), children who had been highly fearful infants proved less obedient than all others when fathers’ power assertion was high, yet more obedient than all others when fathers’ power assertion was low. And this time child fearfulness was not associated with paternal power assertion or rule-compatible conduct, hence fulfilling important criteria for inferring differential susceptibility. Drawing on data of the large-scale longitudinal National Institute of Child Health and Human Development (NICHD) study of Early Child Care and Youth Development (NICHD Early Child Care Research Network, 2005) and focusing on maternally reported difficult temperament at one and six months (composited), Bradley and Corwyn (2008) specifically tested and discerned (repeated) evidence of differential susceptibility. More specifically, these investigators evaluated effects of three measures of parenting quality (sensitivity, harshness, productive activity), each composited and averaged across multiple measurement occasions from infancy through first grade, on teacher-reported behavior problems in first grade (N = 929). Results showed that children with more difficult temperaments had more behavior problems in first grade than all other children if they experienced low-quality parenting, but fewer problems than all other children if they experienced high-quality parenting; the anticipated effect of parenting quality was weaker in the case for children with intermediate levels of difficult temperament and weaker still in the case of children scoring very low on difficult temperament (i.e., easy temperament). Such findings suggest that rather than conceptualizing some children, categorically, as malleable and others as not, it might be more appropriate to conceptualize and measure individual differences in plasticity dimensionally, in terms

71

of a plasticity gradient (Belsky, 2000; see also Warren & Simmens, 2005, for similar results using the same data). Stright, Gallagher, and Kelley (2008) were able to extend work explicitly testing differential susceptibility to positive developmental outcomes by also drawing on data from the NICHD study of Early Child Care (NICHD Early Child Care Research Network, 2005). Once again a temperament–parenting interaction emerged, this time between difficult temperament (at 6 months) and parenting style (composited across 6, 15, 24, 36, and 54 months) in predicting teacher-rated academic competence, social skills, teacher–child relationships, and peer-status at first grade. Predictive power proved greater for infants with more difficult temperaments than for infants with less difficult temperaments. Although all interactions were of a cross-over nature and in line with a for-better-and-for-worse parenting effect for only some children, not all criteria for differential susceptibility were met; of special significance, as already noted, is that the temperament susceptibility factor itself predicted the parenting predictor (as well as at least one outcome measured). Pluess and Belsky (2010) overcame this problem in their analysis of differential susceptibility while drawing on the same NICHD study. After finding that parenting quality prior to school entry predicted reading, math, picture vocabulary, social competence and academic work habits in the fifth grade more strongly for children with difficult temperament than for those with easy temperaments—and in a for-better-and-for-for worse manner—they reran their analysis using the method proposed earlier for discounting discerned evocative effects of a putative susceptibility factor (i.e., temperament) on the environmental predictor (i.e., parenting quality). When the composite measure of parenting quality was statistically adjusted to control for the effect of 6-month, mother-reported difficult temperament, differential susceptibility findings remained virtually unchanged. Since reviewing the just-cited evidence some years ago (Belsky & Pluess, 2009a), even more evidence of this differential susceptibility related, negative emotionality/ difficult temperament moderational effect has emerged in the case of children. This has been revealed in research linking maternal empathy (Pitzer, Jennen-Steinmetz, Esser, Schmidt, & Laucht, 2011) and anger (Poehlmann et al., 2012) with externalizing problems, mutual responsiveness observed in the mother–child dyad with effortful control (Kim & Kochanska, 2012), intrusive maternal behavior (Conway & Stifter, 2012), and poverty (Raver, Blair, & Willoughby, 2013) with executive functioning; and sensitive

72

Differential Susceptibility

parenting with social, emotional and cognitive-academic development (Roisman et al., 2012). Moreover, other new research suggests that “emotion regulation in part explains why negative reactivity can be associated with poor outcomes in supportive contexts and beneficial outcomes in supportive contexts” (Ursache, Blair, Stifter, & Voegtline, 2013, p. 129). More specifically, after discovering that executive functioning at 4 years of age was highest among children who were highly negatively reactive (to fear-evoking stimuli) and high in regulating such reactivity at 15 months of age, but lowest among 4-year-olds who were similarly high in negativity but low in regulation as toddlers, Ursache and associates (2013) tested—and found support for—the proposition that the former children had experience especially positive parenting. Considered together, such data suggest that by promoting emotional regulation, supportive rearing sets the stage for highly negatively reactive children to become especially competent. Although discussion through this point has focused upon the influential role of parenting, it is important to note that it is not just susceptibility to the effects of parenting that negativity seems to moderate in a for-better-and-for-worse manner, as documented in recent work examining the influence of teacher-child conflict on change in symptomology during the primary-school years (Essex et al., 2011). We would be remiss, however, if we failed to call attention to studies showing it was less rather than more negatively emotional or reactive children who proved highly responsive to environmental experiences in a manner consistent with differential susceptibility (Essex et al., 2011, Figure 2, bottom; Schudlich, White, Fleischhauer, & Fitzgerald, 2011) and to other evidence showing that sometimes difficult temperament operates only in a manner consistent with the diathesis–stress rather than differential susceptibility framework (Kochanska & Kim, 2013; Roisman et al., 2012; Yaman, Mesman, van IJzendoorn, & Bakermans-Kranenburg, 2010). The fact that the research cited reveals that diverse indices of negative emotionality function as plasticity factors (e.g., fear, inhibition, distress, difficult temperament, negative affect) raises several issues that future research should address. First, is there a particular component or feature of negativity that is principally responsible for the moderational effects detected (Buss, 2011)? Or might more progress be made with regard to this issue by embracing the integrated, evolutionary, hawk-dove model of temperament (Korte, Koolhaas, Wingfield, & McEwen, 2005) that Sturge-Apple, Davies, Martin, Cicchetti, and Hentges (2012) recently found to be empirically

useful. Whereas hawk-like strategies are characterized by approach, dominant-negative affect (e.g., anger), and activity, dove-like ones involving avoidance, inhibition, and vulnerable affect (e.g., fear) seem more consistent with heightened plasticity according to existing research. We can also wonder about the extent to which the moderational effect of negative emotionality depends on the origins of this plasticity factor, given evidence that negativity is heritable (e.g., Buss & Plomin, 1984; Rhee et al., 2012), but also shaped by prenatal and postnatal experience (Belsky & Pluess, 2009b; Davis, Glynn, Waffarn, & Sandman, 2011; Glynn & Sandman, 2011; for review, see Pluess & Belsky, 2011)—though sometimes more so in the case of some genotypes than others (Ivorra et al., 2010; Pluess et al., 2011). Thus, what may be particularly important is distinguishing negative emotionality, which reflects an experience-induced failure to self-regulate and a general hypersensitivity to the environment. Beyond Parenting: Child-Care Quality All of the rearing data considered through this point pertains to parenting. But as children, especially in the contemporary Western world, are routinely cared for by alternative caregivers in child-care settings, the question arises as to whether quality of child care differentially affects children’s development as a function of their phenotype. In perhaps the earliest pertinent study, Volling and Feagans (1995) detected a relevant and thus noteworthy interaction between children’s social fear (i.e., negative emotionality?), as rated by mothers, and the observed quality of center-based child care in the prediction of observed nonsocial activity (i.e., solitary play, onlooker behavior) a year later when children were 14–48 months of age (N = 36). The highly fearful children manifested both the most and least nonsocial activity, depending on the quality of child care, whereas no such environmental effect emerged in the case of the low-fear children. Given Volling and Feagan’s (1995) small sample size, perhaps more convincing evidence that differential susceptibility characterizes some effects of child care comes from a recent analysis of data from the aforementioned NICHD Study of Early Child which investigated both negative and positive developmental outcomes (Pluess & Belsky, 2009). In this work on samples ranging from 761 to 915 children that explicitly tested differential susceptibility, the observed quality of care (averaged across measurements at 6, 15, 24, 36, and 54 months) differentially predicted behavior problems and social competence rated by caregivers in the year before school entry and by kindergarten teachers. Not

Behavioral Markers of Differential Susceptibility

only did children with difficult temperament (6 months) have more behavior problems when reared in low-quality environments and fewer problems when quality was high compared with children with easy temperaments, but also the regression lines (i.e., slopes) proved significant only for the children who scored high on difficult temperament as infants. Similar results emerged when Pluess and Belsky (2010) extended their research to determine if, after imputing missing data and thereby increasing their sample size to 1,364, differential susceptibility to the effects of good and poor quality child care in the first 4.5 years of life extended to teacher-reported behavior problems and teacher–child conflicts when children were 10–11 years of age. Beyond the Early Years: Middle Childhood Although the Pluess and Belsky (2009, 2010) work indicates that differential susceptibility effects pertaining to early parenting and child-care experience and involving phenotypic susceptibility factors extend beyond the early childhood years, the question arises as to whether rearing and related experiences in later childhood operate in a similar manner. Evidence that they might emerges in research on conflict in families with 7–10-year-olds. Using structural equation modeling with nested-group comparisons, Ramos, Wright Guerin, Gottfried, Bathurst, and Oliver (2005) found that greater conflict predicted more behavior problems (N = 108), but principally in the case of children evaluated as having difficult temperaments when 3–5 years of age and not in the case of those scoring low in difficulty. Unfortunately, the manner in which data are presented precludes determination of whether a true for-better-and-for-worse rather than just dual-risk pattern of results was responsible for the findings reported. Perhaps somewhat more compelling differential susceptibility evidence comes from Lengua’s (2008) temperament–parenting interaction study that sought to explain change in internalizing and externalizing problems using a community sample (N = 188) of 8–12-year-old boys and girls. Children’s reports of their mothers’ parenting style (i.e., rejection–acceptance, inconsistent discipline) predicted change over a 1-year period in mother-reported internalizing and externalizing problems, but differentially as a function of temperament. No doubt in line with expectations developed through this point, results revealed that children highly prone to negative emotion in the form of frustration increased in externalizing problems over time when mothers were rejecting, but decreased when mothers manifested little rejection, with no such apparent effects of rejection evident in the case of children scoring low on frustration. It needs to be noted, however, that other

73

findings from this inquiry, even when consistent with the differential susceptibility hypothesis, clearly seem to be at odds with what other cited investigations have found. For example, Lengua (2008) also reported that it was children who scored high on frustration who were not affected by inconsistent discipline with respect to the development of internalizing problems and children who scored low on frustration who were. Although another interaction indicated that inconsistent discipline did not affect the development of internalizing problems in low-anxious boys but did so in the case of high-anxious ones, thus being in line with previously cited evidence, other reported results proved inconsistent with expectations. Most notable was the fact that low levels of inconsistent parenting predicted increases in internalizing problems among high-anxious boys, with high levels of such putatively problematic parenting predicting decreases. Other perhaps surprising middle-childhood findings emerge in Leve, Kim, and Pears (2005) work on behaviorproblem trajectories from ages 5–17 years (N = 373). After controlling for marital adjustment, parental depression and income (all at age 5), evidence indicated, consistent with expectations, that highly impulsive 5-year-old girls manifest both the most and least externalizing problems across the developmental period under study (parental report) and that this depended on whether they experienced, respectively, high or low levels of harsh punishment at age 5 (parental interview). No such rearing effect emerged in the case of those rated low on impulsivity. At the same time, however, girls rated high on fear/shyness proved less susceptible than those rated low in terms of the influence of harsh punishment on externalizing problems. In summary, even though evidence in line with the differential susceptibility hypothesis emerged in all cases just considered, it should be clear that the substantive nature of differential effects, including who proves most susceptible, are by no means always nor entirely consistent across—or even within—studies. Given claims by Harris (1998) that when it comes to influences on development, peers are actually more important than parents, recent research in middle childhood chronicling differential susceptibility-like effects takes on special importance. In a noteworthy inquiry focused on peer effects, Mezulis, Hyde, and Abramson (2006) observed that 11-year-olds (N = 289) reporting more negative life events involving age mates showed greater cognitive vulnerability to depression when 9 and 11 but somewhat less negative cognitive style when experiencing fewer life events compared with children low in the temperamental trait of withdrawal negativity.

74

Differential Susceptibility

Also of special interest is work that treats physiological traits, rather than just behavioral functioning, as developmental outcomes shaped by developmental experience. Gilissen, Bakermans-Kranenburg, van IJzendoorn, and van der Veer (2008) detected a significant interaction between child fearfulness (parent report) and parent–child relationship quality (observed during interaction or based on a story-completion task administered to children) in the prediction of skin conductance level (SCL) reactivity in response to a fear-inducing film clip. This cross sectional investigation involving 78 4-year-olds and 92 7-year-olds specifically testing differential susceptibility found that more fearful children manifest lower and higher SCL reactivity than all low-fear children, depending on whether their parent–child relationships were, respectively, secure or insecure. Of indisputable interest is that SCL reactivity—a marker for the activity of the sympathetic nervous system—has been found to moderate effects of the environment in a differential susceptibility manner, as will be reported in the next major subsection dealing with physiological susceptibility factors. Important to mention, the just summarized Dutch findings are consistent with the Boyce and Ellis (2005) claim that plasticity may itself be a function of environmental influence (Belsky & Pluess, 2009b). Beyond Middle Childhood: Adulthood Intriguingly, there is also suggestive evidence of the legacy or even operation of differential susceptibility in early adulthood. Of importance in this regard are a series of cross sectional studies testing hypothesized interactions between sensory-processing sensitivity, a personality characteristic measureable by means of the Highly Sensitive Person scale (see Aron & Aron, 1997) and various environmental factors in predicting adult shyness and negative affectivity (Aron et al., 2005). According to Aron and Aron (1997), about 20% of the population are characterized by a high-sensitive personality which encompasses a sensitive nervous system, awareness of subtleties in surroundings, and a tendency to be more easily overwhelmed when in a highly stimulating environment. One study of especial relevance showed that a problematic (and retrospectively reported) childrearing history predicted high levels of (self-reported) shyness and negative affectivity in a sample of 213 undergraduate students, whereas its absence predicted low levels of these same dependent constructs; but this relation obtained principally in the case of students scoring high on sensory-processing sensitivity, resulting in significant differences between regression lines (i.e., slopes) for high- and low-sensitivity groups. Also of

note is research showing that marital quality proves more strongly related to sensitive mothering—and in a for-better-and-for-worse manner—in the case of mothers who score high on negative affect (Jessee et al., 2010). Considered together, the findings just summarized make it clear that it is not just children who prove differentially susceptible to environmental effects. Comment The repeatedly discerned moderational effect of negative emotionality/difficult temperament in the case of parenting, child care quality, and other environmental experiences raises the question of why this should be the case. This would seem especially important to address in view of the fact that even though Belsky (1997a, 2005) theorized that children should vary in their susceptibility to environmental influences (i.e., developmental plasticity) for evolutionary-biological reasons, his differential susceptibility hypothesis did not stipulate that more negatively emotional children or those with difficult temperament would prove especially malleable; this was an empirical observation (Belsky, 2005). One possible reason why those high in negative emotionality, operationalized as it has been in a variety of ways, may prove most susceptible to environmental influence is because a negatively emotional/difficult temperament reflects a highly sensitive nervous system on which experience registers especially strongly, irrespective of whether the experience is positive and growth-promoting or negative and undermining of well-being, broadly conceived (see also Aron & Aron, 1997). Whatever the mechanisms involved in making more negatively emotional children seemingly more malleable—in an often for-better-and-for-worse manner, consistent with differential susceptibility thinking—it would be mistaken to conclude that this is the most important psychological marker of plasticity. Even though this could turn out to be the case, it may well be an artifact of the disproportionate attention that investigators guided by diathesis–stress/dual-risk perspective pay to individual risk factors that interact with contextual adversity in producing problematic functioning (Belsky & Pluess, 2009b). If this is so, then it certainly behooves the field to consider other potential behavioral markers of plasticity/malleability rather than reify one. Nevertheless, the evidence considered through this point indisputably suggests that differential susceptibility is operative in not just the opening years of life and, most critically, that regarding negatively emotional or difficult infants and children as

Physiological Markers of Differential Susceptibility

at risk and vulnerable may represent a fundamental mischaracterization of their more general highly malleable/ plastic nature. PHYSIOLOGICAL MARKERS OF DIFFERENTIAL SUSCEPTIBILITY Recall that central to Boyce and Ellis’ (2005) biologicalsensitivity-to-context theorizing is the claim that children who are highly physiologically reactive to stress manifest the most developmental plasticity. Given that many such children probably begin life as highly negative infants or ones with difficult temperaments, it seems likely that many of the very same children Belsky (1997a, 2005) first called attention to in this regard are being identified by different means. In any event, what Boyce and Ellis (2005) viewpoint highlights is the fact that physiological characteristics, not just the psychological ones considered in the preceding section, might moderate environmental influences, functioning thereby as plasticity markers. In this section evidence consistent with the claim is considered after first providing a brief summary of the two separate physiological systems with specific functions—the autonomous nervous system and the neuroendocrine system. The so-called fight-or-flight response to stress is primarily controlled by the autonomous nervous system (ANS), which is further divided into the sympathetic (SNS) and the parasympathetic nervous system (PNS). The SNS controls those activities that are mobilizing during stress and anxiety (e.g., acceleration of heart rate, increased blood pressure, enhanced blood flow to the skeletal muscles, decreased blood flow to the internal organs and extremities, sweating). Physiologically opposing activities under PNS control serve the basic functions of rest, repair, and relaxation of the body and restoration of energy stores (e.g., decreases in heart rate and blood pressure, stimulation of the digestive system, sexual arousal, sleep). The neuroendocrine response to stress is primarily controlled by the hypothalamus-pituitary-adrenal axis (HPA). Corticotropin releasing hormone (CRH)—which is released from the hypothalamus in response to stress—activates the secretion of adrenocorticotropic hormone (ACTH) from the pituitary gland, which then causes the adrenal cortex to release cortisol into the general bloodstream. Finally, cortisol leads to a large number of diverse physiological and metabolic changes to prepare the organism for optimal functioning under stressful conditions (e.g., increase of blood pressure and blood sugar, breakdown of lipids and proteins, mobilization of amino acids, reduction of immune responses).

75

In the earliest pertinent investigation of physiological reactivity of which we are aware that detects differential susceptibility-like effects, Gannon, Banks, Shelton and Luchetta (1989) studied 50 undergraduates on whom a range of SNS markers of physiological reactivity were obtained (before and after a math-problems’ stress test; plasticity factor). These students also reported on daily hassles (environmental factor), as well as common physical symptoms and depression. Compared to individuals showing low reactivity of blood volume pulse amplitude, high-reactive students reported both few physical symptoms when experiencing few daily hassles and many when experiencing many hassles. Also consistent with differential susceptibility thinking, those students showing slow heart rate recovery after the stress test reported fewer depressive symptoms when experiencing fewer daily hassles and more symptoms when experiencing more daily hassles compared with individuals with a fast recovery. Findings in line with those just presented, but evident at much younger ages, emerged in Boyce and associates (1995) aforementioned test of the hypothesis that mean arterial blood pressure reactivity to a stress test at ages 3–5 would interact with a composite measure of child-care quality (measured across a 2-year period) in predicting frequency of respiratory illness during the six months following the physiological-reactivity assessment. Specifically, children with higher blood pressure reactivity exhibited higher rates of respiratory illness than other children when growing up in stressful rearing contexts, yet under low-stress conditions such high-reactive children had a significantly lower incidence of respiratory illnesses than other children. Reactivity-moderated effects of environmental experiences are also evident when skin conductance level (SCL) reactivity serves as the index of physiological functioning. This is perhaps noteworthy in view of the fact that SCL is controlled solely by the SNS, in contrast to the other cardiovascular-reactivity measures that are generally innervated and controlled by both SNS and PNS. Thus, El-Sheikh, Keller, and Erath (2007b) investigated associations between SCL reactivity (assessed during a star-tracing problem solving task), marital conflict (parent report), and change (from age 9–11.5 years) in adjustment problems (parent report). Compared with girls with low SCL reactivity, highly reactive girls showed the largest increase in internalizing problems if from highly conflicted homes, but the smallest increase when marital conflict was low in their families. A significant cross-over interaction consistent with differential susceptibility also emerged for boys with respect to change in externalizing problems but, rather surprisingly, it was those scoring low in physiological

76

Differential Susceptibility

reactivity who appeared (exclusively) susceptible to the adverse effect of marital conflict. The same research team has also used vagal tone (indexed by respiratory sinus arrhythmia [RSA]) and vagal suppression (during exposure to an audio recording of a male-female verbal conflict) to investigate whether and how PNS measures moderate effects of marital conflict on child adjustment in middle childhood (El-Sheikh, Harger, & Whitson, 2001). Compared to children with high vagal tone (who were not seemingly affected by marital conflict), those scoring low in vagal tone proved less anxious when growing up in families with little marital conflict, but more anxious when residing in high-conflict homes. Similar cross-over-interaction results emerged with respect to vagal suppression, but for boys only. In a cross sectional study of more than 300 5-years-olds, Obradovic, Bush, Stamperdahl, Adler, and Boyce (2010) reported yet more data chronicling the role of RSA in moderating environmental effects, along with some pertaining to cortisol reactivity (both assessed during a stress test). In this research, a composite index of childhood adversity (based on parental reports of financial stress, parenting overload, marital conflict, negative/anger expressiveness, maternal depression, and harsh and restrictive parenting) proved predictive of composite well-being measures (based on parent, teacher, and child reports); but this proved true more so in the case of children with a more reactive parasympathetic nervous system (in contrast to El-Sheikh, Erath, & Keller, 2007; El-Sheikh et al., 2001). More specifically, children with high RSA reactivity were rated as more prosocial under low-adversity conditions and less prosocial under high-adversity conditions compared with children with low-RSA reactivity. High RSA reactivity children also scored higher on school engagement under low-adversity conditions and lower under high-adversity conditions compared with children with low-RSA reactivity. Despite the fact that multiple PNS investigations provide evidence in line with the differential susceptibility hypothesis, seeming to highlight individual differences in plasticity, not just vulnerability, work involving the neuroendocrine system appears to provide comparable evidence. In the just summarized work by Obradovic et al. (2010), children with high-cortisol reactivity were rated as more prosocial under low adversity and less prosocial under high adversity relative to children with low-cortisol reactivity. It is difficult to be sure that this apparent imbalance in evidence across the autonomous nervous system and the neuroendocrine system is due to the two stress reactivity systems playing fundamentally different roles vis-à-vis environmental influences; the alternative

possibility is that one has just received more empirical attention as a moderator of environmental effects. This would seem likely given that most developmentalists measuring cortisol reactivity in studies of environmental effects treat it as an outcome to be explained rather than as a moderator of environmental influences on development (Fernald, Burke, & Gunnar, 2008; Gunnar & Quevedo, 2007). Beyond the work just cited and summarized by Belsky and Pluess (2009a), ever more research highlights the potential role of physiological reactivity as a plasticity factor. Consider in this regard recent evidence indicating that more physiologically reactive children are more susceptible, in a for-better-and-for-worse fashion, to the effect of actual marital conflict (Obradovic et al., 2011) and simulated interparental aggression (Davies, Sturge-Apple, & Cicchetti, 2011), on externalizing problems; of family adversity on school achievement (Obradovic et al., 2010); of chronic family aggression on adolescent anti-social behavior and symptoms of posttraumatic stress disorder (Saxbe, Margolin, Shapiro, & Baucom, 2012); of peer victimization on self-reported depressive symptoms (Rudolph, Troop-Gordon, & Granger, 2011); and of teacher–child conflict on change in symptom severity (Essex et al., 2011). Work showing similar moderational effects of physiological reactivity with respect to family influences on female pubertal development by Ellis and associates (2011) is especially noteworthy, given the roots of this work in psychosocial acceleration theory (Belsky et al., 1991) and the way its findings proved consistent with the differential susceptibility-related revision of the theory (Belsky, 2000). After all, Belsky (2000) postulated that some females would be alternative strategists whose pubertal development proved unrelated to their rearing experiences, just as Ellis and associates (2011) found for girls low in physiological reactivity, whereas others would be conditional strategists susceptible to the accelerating and delaying effects of unsupportive and supportive rearing, respectively, just as Ellis et al. (2011) observed in the case of highly reactive girls. Once again, though, we would be remiss if we did not highlight some evidence inconsistent with the BSC thinking, such as that from the aforementioned Obradovic et al. (2011) investigation showing that it was young children with low—rather than high—RSA reactivity who seemed, in some analyses, to benefit from growing up under supportive family conditions characterized by little marital strife. As it turns out, it may not just be high levels of physiological reactivity that demarcates more malleable individuals. In studying the effects of change over time

Genetic Markers of Differential Susceptibility

in parent depressive symptoms on change in children’s internalizing problems, Laurent and associates (2013) recently observed plasticity-like processes when evaluating the moderating effect of children’s evening cortisol levels. Thus, children with higher evening levels of cortisol increased in internalizing problems over time if fathers’ depressive symptoms increased, but decreased if they had declined; no father depression effect of any kind was discernible in the case of children scoring low in evening cortisol. The fact that the study was focused on adoptive parents and their adopted children means, of course, that the physiologically moderated environmental effects discerned in this inquiry could not be the result of genes shared between parent and child. One of the key questions that still remains to be answered regarding BSC concerns its life span implications. More specifically, do those individuals who are induced to be highly physiologically reactive early in life— and thus developmentally plastic—remain that way over time? It seems hard to imagine that a child induced by supportive rearing to be highly physiologically reactive, in accordance with the theory (Boyce & Ellis, 2005), would remain highly reactive as the years go by. But, of course, this is an empirical question—one that has implications for our understanding of the role physiological reactivity does or does not play as a plasticity factor as the individual develops. GENETIC MARKERS OF DIFFERENTIAL SUSCEPTIBILITY Whereas almost all the evidence cited through this point derives from studies of children, gene–environment (GxE) interaction findings consistent with the differential susceptibility hypothesis often derive from research with adults. This is especially true of psychiatric research focused upon pathological outcomes (e.g., depression, antisocial behavior). The fact that most of this work has been guided by traditional diathesis–stress thinking—at least until recently—means that on many occasions evidence that those carrying a putative risk allele actually function better than others when not exposed to the risk condition being studied (e.g., negative life events) frequently goes unnoticed or at least is not discussed in primary reports. Uher and McGuffin (2008) called attention to such differential susceptibility-like findings, even if not in such terms, in their review of GxE research on life events, the serotonin-transporter gene and depression. Here we discuss findings seemingly reflective of differential susceptibility across a diverse array of candidate genes.

77

Dopamine Receptor D4 Gene (DRD4) A number of genes are known to play a role in regulating the dopaminergic system, which is engaged in attentional, motivational, and reward mechanisms have figured importantly in GxE research. One of these—which probably has revealed more evidence of genetic plasticity than any other polymorphism—is the dopamine receptor D4 (DRD4) gene. Variants of the DRD4 differ by the number of 48-base pair tandem repeats in exon III, ranging from 2 to 11. The 7-repeat variant has been identified in psychiatric-genetic studies as a vulnerability factor due to its links to ADHD (Faraone, Doyle, Mick, & Biederman, 2001), high novelty seeking behavior (Kluger, Siegfried, & Ebstein, 2002), and low dopamine reception efficiency (Robbins & Everitt, 1999), among other correlates. A number of observational studies indicate that children carrying this putative risk allele are not only more adversely affected by poorer quality parenting than other children; they also benefit more than others from good-quality rearing. The first two investigations to be considered may be regarded as particularly important because a good environment was not just operationalized as the absence of adversity—as is all too often the case in GxE investigations—but in terms of high-quality parenting. In a longitudinal investigation of 47 infants, maternal sensitivity observed when children were 10 months predicted externalizing problems reported by mother more than two years later, but only for children carrying the 7-repeat DRD4 allele (Bakermans-Kranenburg & van IJzendoorn, 2006). Moreover, although children with the 7-repeat DRD4 allele displayed the most externalizing behavior of all children when mothers were judged insensitive, they also manifested the least externalizing behavior when mothers were highly sensitive (but see, for contradictory results, Propper, Willoughby, Halpern, Carbone, & Cox, 2007). A cross sectional study of sensation seeking in 45 18–21-month-old children generated results in line with those just summarized. Toddlers carrying the 7-repeat allele were rated by parents as showing, compared with children without the 7-repeat allele, less sensation seeking behavior when parenting quality was high but more when it was low (Sheese, Voelker, Rothbart, & Posner, 2007). Whereas parenting proved significantly associated with sensation seeking in the 7-repeat individuals, it did not in other children. The same team of Dutch of investigators whose work examining effects of maternal sensitivity on toddler’s externalizing problems also found evidence that the DRD4 7-repeat allele moderated the effect of a maternal

78

Differential Susceptibility

psychological condition, unresolved loss or trauma (as measured by means of the Adult Attachment Interview), on a known early developmental marker of psychological disturbance later in life, infant attachment disorganization (Carlson, 1998). But, importantly, this proved true only in the case of infants carrying the 7-repeat allele (van IJzendoorn & Bakermans-Kranenburg, 2006). Such infants manifest both the most and least disorganized attachment behavior when stressed depending on whether their mothers had or had not experienced unresolved loss or trauma in their own lives. In a study with a focus rather different from all those previously considered, Seeger, Schloss, Schmidt, Ruter-Jungfleisch, and Henn (2004) evaluated whether the season of the year in which a child was born interacted with the dopamine DRD4 polymorphism in predicting hyperkinetic conduct disorder (ADHD). Employing a cross sectional design involving 64 children with the disorder and 163 healthy controls (mean age 11–12, +/–3 years), they found that it did—and in ways consistent with what is known about photoperiod exposure during pregnancy. When comparing patients with controls, children with one copy of the DRD4 7-repeat allele born in autumn and winter (i.e., long photoperiod during pregnancy) had a 5.4-fold decreased relative risk for hyperkinetic conduct disorder, whereas children with the same genotype born in spring and summer (i.e., short photoperiod) had a 2.8-fold increased relative risk for hyperkinetic conduct disorder. Neither season of birth nor the presence of DRD4 7-repeat allele increased the risk of hyperkinetic conduct disorder on their own. Beyond these DRD4-related studies first reviewed by Belsky and Pluess (2009a), even more recent work has emerged indicating the 7R allele functions as a plasticity factor. Consider in this regard work showing heightened—or exclusive—susceptibility of individuals carrying the 7-repeat allele when the environmental predictor and developmental outcome were, respectively, maternal positivity and prosocial behavior (Knafo, Israel, & Ebstein, 2011); early nonfamilial childcare and social competence (Belsky & Pluess, 2013b); contextual stress and support and adolescent negative arousal (Beach et al., 2012); childhood adversity and young adult persistent alcohol dependence (Park, Sher, Todorov, & Heath, 2011); and newborn risk status (i.e., , gestational age, birth weight for gestational age, length of stay in NICU) and observed maternal sensitivity (Fortuna et al., 2011). To be noted, however, is that at least one investigatory team finds that it is those without the 7-repeat allele who prove most responsive to childhood adversities—in a diathesis–stress

manner (Das, Cherbuin, Tan, Anstey, & Easteal, 2011). Nevertheless, a recent meta-analysis of GxE research involving dopamine-related genes revealed that children eight and younger respond to positive and negative developmental experiences and environmental exposures in a manner consistent with differential susceptibility (Bakermans-Kranenburg & van IJzendoorn, 2011). Serotonin Transporter Gene (5-HTT) The serotonin-transporter-linked polymorphic region (5-HTTLPR) is another genetic polymorphism that has figured importantly in psychiatric-genetic research. The 5-HTTLPR is a degenerate repeat polymorphic region in SLC6A4, the gene that codes for the serotonin transporter. Most research focuses on two variants—those carrying at least one short allele (s/s, s/l) and those homozygous for the long allele (l/l)—though more variants than these have been identified (Nakamura, Ueno, Sano, & Tanabe, 2000). The short allele has generally been associated with reduced expression of the serotonin transporter molecule—which is involved in the reuptake of serotonin from the synaptic cleft—and thus considered to be related to depression, either directly or in the face of adversity. Caspi and associates (2003a) were the first to show that the 5-HTTLPR moderates effects of stressful life events during early adulthood on depressive symptoms and on probability of suicide ideation/attempts and of major depression episode at age 26 years. Individuals with two s alleles proved most adversely affected whereas effects on l/l genotypes were weaker or entirely absent. Of special significance, however, is that carriers of the s/s allele scored best on the outcomes just mentioned when stressful life events were absent—though just as was true among low-MAOA activity individuals in Caspi et al. (2002)—not by very much. Multiple research groups have attempted to replicate Caspi et al.’s findings (2003) of increased vulnerability to depression in response to stressful life events for individuals with one or more copies of the s allele, with many succeeding (see below), even if not all (Surtees et al., 2006). Of note is that the data presented in quite a number of these studies indicates that individuals carrying short alleles (s/s, s/l) did not just function most poorly when exposed to many stressors, but best—showing least problems—when encountering few or none. While calling explicit attention to such a pattern of results, Taylor and associates’ (2006) reported that young adults homozygous for short alleles (s/s) manifested greater depressive symptomatology than individuals with other allelic variants when exposed to

Genetic Markers of Differential Susceptibility

early adversity (i.e., problematic childrearing history), as well as many recent negative life events, yet the fewest symptoms when they experienced a supportive early environment or recent positive experiences, that is, not just the absence of adversity. A similar pattern of environmental effects emerged in still other investigations of stressful life events and depression, including one targeting depressed patients, healthy controls and experiences during the six-month prior to study enrollment (Zalsman et al., 2006), and another of a sizeable community sample (N = 567) and life events up to two years prior to the assessment of depression (Lazary et al., 2008). The same for-better-and-for-worse pattern of results are evident—and noted—in Brummett et al.’s (2008) investigation of more than 200 adults (mean age 58 years) who differed in whether or not they served as caregiver of a relative with Alzheimer’s disease and in Eley et al.’s (2004) research on adolescent girls who were and were not exposed to risky family environments. In fact, secondary analysis of the latter findings revealed that the for-better environmental effect size exceeded the routinely evaluated for-worse size (Belsky et al., 2009). Although all the work just cited, with the exception of Caspi et al. (2003), was cross sectional in design, Wilhelm and associates’ (2006) longitudinal data also shows that individuals with the s/s genotype had the lowest probability of lifetime major depression if exposed to no adverse life events in a five-year study period, but the highest probability when reporting two, three, or more adverse life events compared with other genotypes. The effect of 5-HTTLPR in moderating environmental influences in a manner consistent with differential susceptibility is not restricted to depression and its symptoms. It also emerges, perhaps unsurprisingly, in studies of anxiety and ADHD. Gunthert et al. (2007) documented the former result in longitudinal research on 350 college students. At study entry and a year later, participants reported anxiety and negative events daily for 30 days. Genotyping distinguished three alleles, but the LG allele was grouped with s alleles due to its functional equivalence vis-à-vis promoter activity. Individuals homozygous for short alleles (including s/LG and LG /LG ) reported more anxiety in the evening when daily-event stress was high compared with individuals with different genotypes, but also less anxiety than other genotypes when experiencing little daily-event stress; this pattern of results proved consistent across measurement occasions. In a second study focused on undergraduate students (N = 247) and anxiety (Stein, Schork, & Gelernter, 2008), but this time concerned with (retrospectively reported)

79

emotional abuse in childhood, a GxE interaction once again emerged. The significantly steeper abuse–anxiety slope in the case of students homozygous for short alleles relative to those with one or more long alleles indicated that s/s individuals scored highest in anxiety sensitivity when exposed to abuse and lowest when not exposed. Moving on to consider ADHD (in childhood and adulthood), Retz and associates (2008) focused on the moderated effects of an adverse childhood environment in their study of 184 male delinquents who averaged 34 years of age. Using a retrospective assessment of childhood ADHD, as well as of early adversity, but a clinical interview to assess functioning in adulthood, these investigators detected a crossover interaction with respect to the persistence of ADHD over time. Compared with l/l genotypes, individuals with s alleles had more and less persistent ADHD, depending on whether or not, respectively, they experienced an adverse early environment. Yet another differential susceptibility-relevant finding involving 5-HTTLPR comes from Manuck, Flory, Ferrell, and Muldoon (2004) test of a GxE interaction involving socioeconomic status in the prediction of central nervous serotonergic responsivity; the sample included 139 adults ranging in age from 26 to 60. Central serotonergic responsivity was measured indirectly by means of the fenfluramine challenge test. Fenfluramine increases serotonergic neurotransmission by release of serotonin stores and reuptake inhibition. Such stimulation of hypothalamic serotonin receptors promotes as well the pituitary release of the hormone prolactin. This relative release in circulating prolactin concentration provides an index of the serotonergic responsivity in the HPA axis. Consistent with all the findings summarized above pertaining to depression, anxiety and persistent ADHD, s/s individuals manifest the most and least serotonergic responsivity, depending on whether they were, respectively, low or high SES. In the short time since the aforementioned studies involving 5-HTTLPR were reviewed by Belsky and Pluess (2009a), yet more evidence that short alleles operate as plasticity factors has emerged. Consider in this regard, evidence showing for-better-and-for-worse results in the case of those carrying one or more 5-HTTLPR short alleles when the rearing predictor and child outcome were, respectively, maternal responsiveness and moral internalization (Kochanska, Kim, Barry, & Philibert, 2011), child maltreatment and antisocial behavior (Cicchetti, Rogosch, & Thibodeau, 2012), and supportive parenting and positive affect (Hankin et al., 2011). Differential susceptibility-related findings also emerged (among male African American adolescents) when perceived racial

80

Differential Susceptibility

discrimination was used to predict conduct problems (Brody et al., 2011); when life events were used to predict neuroticism (Pluess, Belsky, Way, & Taylor, 2010) and life satisfaction of young adults (Kuepper et al., 2012); and when retrospectively reported childhood adversity was used to explain aspects of impulsivity among college students (e.g., pervasive influence of feelings, feelings trigger action; Carver, Johnson, Joormann, Kim, & Nam, 2011). Just to be clear, other 5-HTTLPR-related findings prove more consistent with the diathesis–stress than differential susceptibility framework (e.g., Bakermans-Kranenburg, Dobrova-Krol, & van IJzendoorn, 2012; Brody et al., 2013; Pluess et al., 2011), sometimes even in GxE studies which also provide evidence of differential susceptibility (e.g., Kochanska et al., 2011). Moreover, in apparently rare instances it is—surprisingly—those with long rather than short alleles who prove most susceptible to environmental influences, even when differential susceptibility-related findings emerge (e.g., Cicchetti, Rogosch, & Oshri, 2011; Sulik et al., 2012). Especially noteworthy, therefore, are the results of a recent meta-analysis of GxE findings pertaining to children under 18 years of age showing that short-allele carriers are more susceptible to the effects of both positive and negative developmental experiences and environmental exposures, at least in the case of Caucasians (van IJzendoorn, Belsky, & Bakermans-Kranenburg, 2012). Monoamine Oxidase A Gene (MAOA) The neurotransmitter-metabolizing enzyme monoamino oxidase A or MAOA gene is located on the X chromosome. It encodes the MAOA enzyme, which metabolizes neurotransmitters such as norepinephrine, serotonin, and dopamine, rendering them inactive. Two sets of evidence—one linking the low-activity MAOA allele to antisocial behavior and another linking abuse and neglect in childhood to the same developmental outcome—led Caspi and associates (2002) to hypothesize that inconsistency in findings in both literatures could be a result of the fact that maltreatment effects are moderated by genotype. This is exactly what they discovered in their groundbreaking and widely cited GxE research carried out on a New Zealand birth cohort followed into young adulthood. More specifically, it was principally young men—females were not studied—with one form of the gene (i.e., that associated with low MAOA activity) who proved more violence prone when subject to child maltreatment. For those children with the high-MAOA-activity allele, a substantially smaller effect of child maltreatment emerged.

Although most interpret these findings, not unreasonably, in diathesis–stress terms, few seem to have noticed that those most vulnerable to the adverse effects of maltreatment actually scored lowest in antisocial behavior when not exposed to maltreatment, suggesting perhaps greater plasticity rather than just greater vulnerability to adversity in the case of those with the low-activity MAOA allele. This interpretation would seem substantiated by results of a significant number of efforts to replicate the Caspi et al. (2002) findings. For example, Kim-Cohen et al. (2006) studied 975 boys to determine whether the MAOA polymorphism moderated effects of mother-reported physical abuse in early childhood on later mental health problems. At age 7 years boys with the low-MAOA-activity variant were rated by mothers and teachers as having more mental health problems—and specifically ADHD symptoms—if they had been victims of abuse, but fewer problems if they had not, compared with boys with the high-MAOA-activity genotype. In a longitudinal study of 514 adolescent twin boys aged 8 to 17 years, Foley et al (2004) found that childhood adversity—based on parent and child report—predicted three-month history of conduct disorder (DSM-III) differently for children with the low- and high-activity MAOA allele. Once again, boys with the low-MAOA-activity allele were more likely to be diagnosed with conduct disorder if exposed to higher levels of childhood adversity and less likely if exposed to lower levels of adversity, compared with boys with the high-MAOA-activity allele. Similar results emerged in Nilsson and associates’ (2006) cross sectional investigation of 81 adolescent boys when the predictor was psychosocial risk, operationalized in terms of maltreatment experience and living arrangement. Only boys with the low-MAOA-activity allele were affected by such risk, such that those with a history of adversity engaged in more criminal behavior (composite of vandalism, violence, stealing) and those lacking this history engaged in less. Three additional studies extend the Caspi et al. (2002) findings: One was a prospective investigation of 631 male and female and white and black victims of (court-substantiated) child abuse and neglect, along with a comparison group matched on age, sex, race/ethnicity, and social class background (Widom & Brzustowicz, 2006); the second a retrospective study of 235 adult psychiatric outpatients and healthy controls who reported on trauma experienced in childhood and physical aggression in adulthood (Frazzetto et al., 2007); and the third a cross sectional retrospective study with an American Indian sample of 291 adult women, 50% of whom had a history of childhood sexual abuse (Ducci et al., 2008). White (but not black)

Genetic Markers of Differential Susceptibility

males and females with the low-MAOA-activity allele in the longitudinal study manifest the most violent and antisocial behavior during adolescence and across their lifetimes when measured around 40 years of age, if they had been maltreated. But those with the same alleles engaged in the least such behavior (at both times of measurement) if they had not been victims of abuse. In the second study, men (only) with the low-MAOA-activity variant reported more physical aggression if they experienced one or more (retrospectively reported) objective traumatic events while growing up (e.g., death of mother, severe physical handicap of sibling) and less physical aggression if there was no history of trauma, compared with high-MAOA-activity men (for whom trauma proved unrelated to aggression). In the third study, women homozygous for the low MAOA activity variant had the highest count of antisocial personality disorder symptoms when reporting childhood sexual abuse and the lowest count when having no history of sexual abuse, compared with women homozygous for the high-MAOA-activity genotype. The fact of the matter is that since the just-cited studies were reviewed by Belsky and Pluess (2009a), no additional GxE evidence has emerged suggesting that MAOA should be regarded as a plasticity rather than vulnerability factor. The same is true for the remaining genes to be considered in this section, raising questions about the confidence that can be placed in genes other than perhaps DRD4 and 5-HTTLPR as plasticity genes. To be appreciated, of course, is that none of these other genes have received the empirical attention in the GxE literature that DRD4 and 5-HTTLPR have, which could account, at least in part, as to why no new differential susceptibility-related findings have emerged for these other genes that Belsky and Pluess (2009a) highlighted in their review of relevant research. Serotonin Receptor 2A Gene (HTR2A) Irrespective of these necessary and cautionary comments, evidence consistent with differential susceptibility thinking emerges from GxE studies of another serotonergic system related genetic polymorphism: the single nucleotide polymorphism (SNP) rs6313, located on the serotonin receptor gene (HTR2A), which comes in two forms—the C and the T allele. Whereas some research reveals an association between the C allele and depression (e.g., see Du, Bakish, Lapierre, Ravindran, & Hrdina, 2000), others find the T allele to confer depression risk (Eley et al., 2004). Recent work by Jokela and associates (Jokela, Keltikangas-Jarvinen, et al., 2007a; Jokela, Lehtimaki, & Keltikangas-Jarvinen, 2007b, c)—drawing from data

81

of a longitudinal population-based study in Finland, the Cardiovascular Risk in Young Finns Study, that will be cited frequently in coming paragraphs—suggests that individuals carrying the T allele are generally more affected by environmental factors than others in a for-better-and-for-worse manner. When nurturance reported by mother was averaged across baseline—when 820–1212 study participants were 3–18 years old—and again 3 years later, offspring with at least one T allele scored highest and lowest on self-reported depression some 2 decades after baseline, depending on whether they experienced, respectively, more or less supportive care (Jokela, Keltikangas-Jarvinen, et al., 2007a). Similar results emerged when the predictor was family SES and the outcome to be explained was self-reported harm avoidance; those carrying one or more T alleles scored highest on harm avoidance if they grew up in low-SES households but lowest if they grew up in high-SES ones (Jokela, Lehtimaki, et al., 2007b). This result led the investigators to suggest that this allele might function as an opportunity allele, not just a risk gene. Yet more support for this differential susceptibility interpretation comes from a third Finnish report focused on whether adults resided in remote rural, rural, suburban, or urban areas. T/T individuals showed the most and least depressive symptoms of all study participants depending on whether they resided, respectively, in remote rural or urban areas (Jokela et al., 2007b). In sum, these Finnish findings are consistent with a differential susceptibility interpretation. Tryptophan Hydroxylase 1 Gene (TPH1) As it turns out, the same research group identified similar interaction effects involving yet another SNP on a serotonergic system related gene: rs1800532 located in the tryptophan hydroxylase 1 gene (Jokela, Räikkönen, Lehtimäki, Rontu, & Keltikangas-Järvinen, 2007); this gene codes for a rate-limiting enzyme in the biosynthesis of serotonin. Of the two variants A and C, the A allele has been associated with low serotonin levels (Jönsson et al., 1997) and suicidal behavior/depression, though findings are not consistent (e.g., see Bellivier, Chaste, & Malafosse, 2004; Lalovic & Turecki, 2002). Jokela, Räikkönen, et al. (2007) detected a moderating effect of TPH1 on the association between social support and depressive symptoms, again using data from the Cardiovascular Risk in Young Finns Study. Depressive symptoms and social support were assessed when participants were between 20 and 35 years old, with the

82

Differential Susceptibility

former measurements taken again four year later. A GxE interaction emerged irrespective of when the depression outcome was assessed. A/A individuals experienced the most and least depression compared with all other individuals, depending on whether they experienced, respectively, low or high social support. This proved to be even more so the case when change in depressive symptoms was the outcome to be explained. When the Finnish researchers turned their attention to the effect of a hostile childhood environment (based on maternal report) on harm avoidance in adulthood, yet more differential susceptibility like findings emerged involving TPH. Keltikangas-Jarvinen and associates (2007b) found that female participants carrying the A/A genotype reported more adulthood harm avoidance if raised in a hostile environment but less if raised in a nonhostile one relative to those with one or two A alleles. In fact, no apparent effect of childhood environment was detected in the case of women carrying C alleles (or in men). Dopamine Receptor D2 Gene (DRD2) Another polymorphism in the dopaminergic system that has been a focus of GXE research is located on the DRD2 gene that encodes the D2 subtype of the dopamine receptor. The Taq1A (A1) SNP (rs1800497)—a C to T substitution located in a noncoding region of the DRD2 locus—is thought to affect dopamine receptor D2 availability in postmortem striatal samples (Thompson et al., 1997). The A1 allele has been associated with low dopamine density and lower mean relative glucose metabolic rate in dopaminergic regions in the human brain (Noble, Gottschalk, Fallon, Ritchie, & Wu, 1997), high novelty seeking (Suhara et al., 2001), and a number of substance use disorders, particularly alcoholism (Bowirrat & Oscar-Berman, 2005). Some study findings suggest that the DRD2 polymorphism moderates environmental influences in differential susceptibility terms. For example, Mills-Koonce and associates (2007) detected an interaction between DRD2 and sensitive mothering observed when children were 6 and 12 months of age in the prediction of children’s psychological well-being. Infants with the A1 allele reared by more and less sensitive mothers had, respectively, fewer and more affective problems at 3 years of age than agemates with other genotypes. Indeed, rearing experience proved unrelated to well-being among these latter children. A differential susceptibility like effect also emerged in what appears to be the first ever report of GxE in humans, one involving an interaction between the DRD2 gene and family stress in the prediction of visuospatial ability

(i.e., line orientation test) and 2 years later, P300 amplitude in ERP research (Berman & Noble, 1997). In this cross sectional study of 10–13-year-old boys, those carrying the A1 allele scored lower than those lacking this allele under high stress conditions, but higher under low stress conditions (even if the differential was more substantial under higher stress conditions). Focusing as well on a physiological outcome, Propper et al. (2008) found DRD2 to moderate effects of 6-month maternal sensitivity on, 6 months later, respiratory sinus arrhythmia (RSA) reactivity. Children carrying the A1 allele had lower RSA change at 12 months when experiencing low maternal sensitivity and higher RSA change when experiencing high maternal sensitivity compared with children without the A1 allele. Turning to cognitive outcomes, Keltikangas-Järvinen et al (2007b) found DRD2 to moderate the association between birth weight, presumed to reflect quality of the uterine environment to which the fetus has been exposed, and educational achievement in the Cardiovascular Risk in Young Finns Study, though only in the case of males in this large-sample study of 1,512 adults (27 to 34 years old). Whereas men with the A1 allele achieved less academically if born of low birth weight but more if born of high birth weight, no such birth-condition-achievement association emerged in the case of other men. (Individuals with birth weight > 5000g were excluded from the analysis.) Somewhat surprisingly given the DRD2 work already reviewed, other findings from the Cardiovascular Risk in Young Finn’s study suggest that it is not always individuals carrying the A1 genotype who appear disproportionately susceptible to environmental conditions. When it came to predicting depressive symptoms in 1,636 24–39-year-olds, Elovainio et al. (2007) found that it was young adults lacking the A1 allele who had more depressive symptoms if they experienced many negative life events and fewer depressive symptoms if they reported no such events across the preceding 9-year period compared with their counterparts carrying this allele. Additional Plasticity Genes? Emerging evidence has also begun to suggest that a polymorphism in the brain-derived-neurotrophic factor gene (BDNF) may function a plasticity gene (e.g., Chen, Li, & McGue, 2012; Gunnar et al., 2012; Juhasz et al., 2011; Mata, Thompson, & Gotlib, 2010; Suzuki et al., 2011) and a polymorphisms in the oxytocin receptor gene (OXTR; Johansson et al., 2012; Poulin, Holman, & Buffone, 2012; Sturge-Apple, Cicchetti, Davies, & Suor, 2012),

Genetic Markers of Differential Susceptibility

and the FKBP5 gene (Bevilacqua et al., 2012; Xie et al., 2010; Zimmermann et al., 2011). There are perhaps less frequent indications of this in the case of a SNP in the catechol-o-methyl transferase gene (COMT; Laucht et al., 2012; Nijmeijer et al., 2010). It will be worth monitoring these polymorphism—and others—as more work emerges to see whether additional evidence proves consistent with differential susceptibility. In fact, given that virtually all the polymorphisms just considered have been studied in GxE perspective as a result of the original interests of psychiatric geneticists in simple genotype-phenotype linkages, we cannot but wonder whether an entirely different set of polymorphisms reflecting sensitivity to the environment would have emerged had the phenotype of interest been not some psychiatric disorder but developmental plasticity itself. It therefore behooves investigators to expand the list of genetic suspects beyond those thought to be related to disturbances in functioning by thinking biologically about genes that might be related to physiological processes instantiating plasticity. An excellent recent example of this approach is to be found in Grazioplene and associates (2013) work showing that a polymorphisms located on CHRNA4 interacts with childhood maltreatment to predict features of personality in a differential susceptibility-related manner. Of especial significance is that this polymorphism was selected for investigation because it is involved in the production of the neurotransmitter acetylcholine, a component of the cholinergic system that is strongly involved in neural plasticity and learning. A similarly noteworthy GxE study focused on a GABRA2 polymorphism on the presumption that it enhances responsiveness to the environment due to the role it plays in regulating neuronal excitability. Simons and associates (2013) observed that it functioned just as one would expect of a plasticity gene when investigating the influence of parenting experienced at age 10–12 on hostility toward romantic partners in early adulthood. Polygenetic Plasticity Most GxE research, like that just considered, has focused on one or another polymorphism. In recent years, however, work has emerged which we believe should be encouraged, focusing on multiple polymorphisms and thus reflecting the operation of epistatic (i.e., GxG) interactions (e.g., Beaver, Sak, Vaske, & Nilsson, 2010; Conner, Hellemann, Ritchie, & Noble, 2010), as well as GxGxE ones. One can distinguish polygenetic GxE research in terms of the basis used for creating multigene composites. One strategy that seems fundamentally problematic to us involves identifying genes

83

that show main effects and then compositing only these to then test an interaction with some environmental parameter (e.g., Docherty, Kovas, & Plomin, 2011). What leads us to question the wisdom of this approach is that genes—or any other factors for that matter—that do not exert main effects could still be involved in interactions with the environment, just as some recent GxE research reveals (e.g., Cicchetti & Rogosch, 2012; Cicchetti, Rogosch, & Oshri, 2011; Pluess et al., 2011). Another approach is to composite genes for a secondary, follow-up analysis that have been found in a first round of inquiry to generate significant GxE interactions (e.g., Brody, Chen, & Beach, 2013; Simons et al., 2013; Sonuga-Barke et al., 2009). When Cicchetti and Rogosch (2012, Figure 2) applied this approach using four different polymorphisms, they found that as the number of sensitivity-to-the-environment genotypes increased (i.e., S/S of 5-HTTLPR, zero copies of CRH1 TAT haplotype, the T/T genotype of DRD4–521C/T, and A carrier of OXTR), so did the degree to which maltreated and nonmaltreated low-income children differed on a composite measure of resilient functioning—in a for-better-and-for-worse manner. A third approach which has now been used successfully a number of times to chronicle differential susceptibility involves compositing a set of genes selected on an a-priori basis before evaluating GxE (e.g., Brody et al., 2013). Consider in this regard evidence indicating that two-gene composites moderate links between sexual abuse and adolescent depression/anxiety and somatic symptoms (Cicchetti, Rogosch, & Sturge-Apple, 2007), between perceived racial discrimination and risk-related cognitions reflecting a fast versus slow life-history strategy (Gibbons et al., 2012, Figure 2), between contextual stress/support and aggression in young adulthood (Simons et al., 2011) and between social class and postpartum depression (Mitchell et al., 2011); that a three-gene composite moderates the relation between a hostile-demoralizing community and family environment and aggression in early adulthood (Simons et al., 2011); and that a five-gene composite moderates the relation between parenting and adolescent self-control (Belsky & Beaver, 2011), as well as between marital stability experienced across the first 15 years of life and income earned in young adulthood (Wickrama & O’Neal, 2013). No matter how informative the approach to compositing polymorphisms on a-priori grounds has proven to be when it comes to detecting differential susceptibility, we suspect that the wave of the future will be the system-level genetic approach. This involves combining

84

Differential Susceptibility

genes considered to collectively influence a particular biological process or pathway, such as the dopaminergic or serotonergic system (e.g., Brody et al., 2013; Conner et al., 2010; Nikolova, Ferrell, Manuck, & Hariri, 2011) or neurological endophenotypes for increased susceptibility, such as amygdala volume (Yap et al., 2008), hippocampus volume (Whittle et al., 2011) and ventral striatum reactivity (Nikolova, Bogdan, Brigidi, & Hariri, 2012), to list just a few possibilities. Further insight could be gained, conceivably, by even more refined analysis of such subsystems, including those reflecting synthesis, degradation/transport, receptor and modulation (Chen et al., 2011). GxE Mechanisms There are likely multiple explanations as to why—in terms of mechanisms—the GxE findings we have highlighted and the ones we have not prove consistent with differential susceptibility and thus Belsky and Pluess’s (2013a) claim that developmental plasticity should be regarded as a phenotype in its own right. When considering mechanisms, however, we must be sensitive to the fact that this term means very different things to different scholars. For a brain scientist, mechanism may refer to brain structure or function, whereas for an endocrinologically oriented investigator it may refer to a hormonal phenomenon, for a geneticist to an epigenetic cascade, and for a cognitive neuroscientist an attentional process, to name just some of the possibilities. Thus, what for a cognitive scientist may qualify as a proximate mechanism (e.g., attentional bias) may well be regarded by a geneticist focused on the epigenome as a distal predictor. Developmental and behavioral scientists should not be arrogant, then, when it comes to stipulating what does and does not qualify as a mechanism. It ultimately depends on the level—or levels—of analysis one is working at. In what follows we call attention, based on recent empirical evidence, to some possible differential susceptibilityrelated mechanisms—at different levels of analysis. But rather than emphasizing—as most genotype–phenotype and even GxE reports do—findings linking particular genes directly to particular psychological or biological processes, we highlight instead research showing that such processes are themselves responsive to environmental exposures in a differential susceptibility-related manner. When it comes to cognitive processes, the following have been found to be moderated by 5-HTTLPR and an environmental factor in a for-better-and-for-worse manner: positive and negative attentional biases (Fox, Zougkou, Ridgewell, & Garner, 2011; see also Pergamin-Hight,

Bakermans-Kranenburg, van IJzendoorn, & Bar-Haim, 2012); reward sensitivity (Roiser, Rogers, Cook, & Sahakian, 2006); and accurate processing of emotional faces (Jacobs et al., 2011). The same is true of inattention in the case of DRD4 (Berry, Deater-Deckard, McCartney, Wang, & Petrill, 2013); working-memory accuracy and reaction time in the case of BDNF (Gatt et al., 2009); and rumination in the case of a two-gene composite (5-HTTLPR, BDNF: Clasen, Wells, Knopik, McGeary, & Beevers, 2011) and street code reflecting high value placed on being tough and aggressive, standing up for one’s rights and being ready to fight in the case of a three-gene one (5-HTTLPR, DRD4, MAOA: Simons et al., 2012). When it comes to physiological processes, 5-HTTLPR and an environmental factor have been found to interact in a differential susceptibility related manner in predicting cortisol reactivity (Way & Taylor, 2010), with the same being true of BDNF when the dependent variable is heart rate (Gatt et al., 2009) or evening salivary cortisol (Vinberg et al., 2009). Turning to brain measurements, Alexander and colleagues (2012) present evidence that 5-HTTLPR interacts with stressful life events in a for-better-and-for-worse manner to predict left amygdala reactivity to fearful faces, while Gatt et al. (2009) document much the same when using BDNF to forecast grey matter volume in both the hippocampus and the amygdala. Additionally, Klucken and associates (2013) report that BDNF interacts with positive and negative feedback in a for-better-and-for-worse manner during appetitive conditioning when predicting hemodynamic response in the amygdala. Also worth considering is evidence that individuals homozygous for the Iso allele of the rs5522 SNP located on the mineralocorticoid receptor gene (MR) manifest the most and least right amygdala reactivity, relative to Val carriers, when they (retrospectively) reported experiencing, respectively, high and low levels of emotional neglect in childhood (Bogdan, Williamson, & Hariri, 2012). But it may not be just adversity to which the amygdala of certain individuals proves especially sensitive to. After all, an extensive meta-analysis showed that the amygdala responds to both negative and—even more strongly—to positive stimuli (Sergerie, Chochol, & Armony, 2008). This suggests to us, as does the work cited in the preceding paragraph, that amygdala reactivity might be one of several central nervous mechanism by which differential susceptibility operates. This would certainly be consistent with the hypothesis, proposed by the differential susceptibility framework (Belsky & Pluess, 2009a) as well as the concept of sensory-processing sensitivity (Aron & Aron, 1997; Aron, Aron, & Jagiellowicz, 2012), that heightened

Experimental Evaluation of Variation in Developmental Plasticity

susceptibility is the function of a more sensitive central nervous system on which experiences register more easily and deeply. According to this neurosensitivity hypothesis (Pluess & Belsky, 2013), specific gene variants (e.g., 5-HTTLPR short allele, DRD4 7-repeat) contribute to the increased sensitivity and responsivity of specific brain regions that then manifests itself in increased negative emotionality and physiological reactivity (Pluess, Stevens, & Belsky, 2013), in part because highly sensitive individuals are more easily aroused. What this brief summary of diverse findings related to putative plasticity genes should make clear is that a variety of plausible mechanisms—at multiple levels of analysis—may help to account for the GxE findings noted or referred to earlier. Not only is more such research required, but the big challenge going forward will surely be to put Humpty Dumpty together, even if not again, so that a multilevel system of causation can be illuminated (Hyde, Bogdan, & Hariri, 2011). Needless to say, this represents a huge agenda, if not the holy grail of inquiry.

EXPERIMENTAL EVALUATION OF VARIATION IN DEVELOPMENTAL PLASTICITY All the research cited through this point has been observational and correlational, even if often longitudinal, in character. One of the fundamental risks of interpreting evidence of such work, as noted earlier when considering methodological issues, is that what appears to be a function of person–environment interaction could actually reflect person–environment correlation. In other words, rather than such interaction effects reflecting the moderation of environmental influence by some person characteristic, they could reflect the fact that person characteristics and environments are correlated, perhaps due to the shared heritability of each, the fact that certain people tend to be exposed to certain experiences or that individuals with certain attributes evoke or provoke particular developmental experiences. The common strategy employed in observational work for discounting the possibility that organism-environment correlation is masquerading as organism–environment interaction is to insure that the temperamental, physiological, or genetic moderator under study is not correlated with the contextual predictor under consideration. However laudable such efforts are, they are not without limits. For example, just because the moderating polymorphism under investigation (e.g., DRD4) proves unrelated to the contextual predictor of interest (e.g., sensitive parenting) does not mean that this is true of all

85

other unmeasured or unexamined organismic factors (e.g., BDNF). Ultimately, the fact that environments have not been systematically manipulated in most plasticity-related research means that one cannot be certain, even in the case of the putatively malleable, that truly causal environmental effects are being detected (van IJzendoorn & Bakermans-Kranenburg, 2012). One way around this problem, pioneered by Dutch investigators in the case of genetic plasticity factors (Bakermans-Kranenburg, van IJzendoorn, Mesman, Alink, & Juffer, 2008b; Bakermans-Kranenburg, van IJzendoorn, Pijlman, Mesman, & Juffer, 2008c), involves conducting intervention experiments in which individuals are randomly assigned to a treatment condition. Because it is unethical to assign individuals, especially children, to conditions of adversity, this approach has only been used to evaluate differential response—as a function of some hypothesized plasticity factor—to an experience presumed to promote positive functioning. Such a design is of course limited, as it can only evaluate the for-better side of developmental plasticity (Pluess & Belsky, 2013). Nevertheless, it represents an ideal way to determine whether those considered to be resilient to adversity on the basis of observational research also prove relatively unresponsive to supportive conditions and whether those considered vulnerable to adversity disproportionately benefit from environmental enrichment. As it turns out, there is ever growing experimental evidence that this is indeed the case. Negative Emotionality and Physiological Reactivity With regard to the plasticity factor of negative emotionality, consider the re-analysis of data from the Infant Health and Development Program (1990), a well-known, early intervention which involved the random assignment of poor, low-birth infants and their families to treatment or control condition, putatively generating positive, across-the-board program effects. Blair (2002) tested Belsky’s (1997a, 1997b) proposition that an enriched rearing experience (involving educational day care in the second and third year of life, combined with home visiting and parent support over the child’s first three years) would differentially impact children with varying temperaments. As predicted, infants who were highly negatively emotional and assigned to the early-intervention group scored substantially lower on externalizing problems at 3 years of age than did similarly tempered infants randomly assigned to the control group; no such treatment effect occurred in the case of other infants. Especially intriguing given the fact that virtually all research considered through this

86

Differential Susceptibility

point has focused on differential susceptibility vis-à-vis social and emotional functioning, is that exactly the same results emerged when the outcome in question was severely impaired cognitive functioning. More specifically, highly negative infants assigned to the experimental intervention were five times less likely to score at or below 75 on an IQ test at age three than their negatively emotional counterparts assigned to the control condition; no such experimental effect occurred in the case of infants scoring low on negative emotionality. Not inconsistent with these findings are those showing that (1) it was infants who scored relatively low on irritability as newborns who failed to benefit from an otherwise security-promoting intervention (Cassidy, Woodhouse, Sherman, Stupica, & Lejuez, 2011) and (2) infants who showed few, if any, mild perinatal adversities—known to be related to limited negative emotionality—who failed to benefit from computer-based instruction otherwise found to promote preschoolers’ phonemic awareness and early literacy (van der Kooy-Hofland, van der Kooy, Bus, van IJzendoorn, & Bonsel, 2012). In other words, only the putatively vulnerable—those manifesting or likely to manifest high levels of negativity—experienced developmental enhancement as a function of the interventions cited. Similar results emerge among older children. Consider in this regard the results of Scott and O’Connor’s (2012) parenting intervention that generated the most positive change in conduct among emotionally dysregulated children (i.e., loses temper, angry, touchy). Van de Wiel and associates (2004) presented related results when the plasticity factor was physiological reactivity, finding that an intervention for children with disruptive behavior disorder proved effective, but only for those displaying high cortisol-related, stress reactivity. Genetics Recent genetically informed evaluations of intervention programs also indicate that alleles presumed to place individuals at risk in the face of adversity (e.g., 5-HTTLPR short allele, DRD4 7-repeat) or to promote resilience (i.e., no 5-HTTLPR short allele/no DRD4 7-repeat) are associated with them being, respectfully, susceptible or not to the benefits of intervention (but see Cicchetti, Rogosch, & Toth, 2011, for counter evidence). Consider first experimental intervention research designed to enhance parenting which chronicles a moderating effect of the 7-repeat allele on parenting. When Bakermans-Kranenburg, van IJzendorn, Pijlman, Mesman, and Juffer (2008) looked at change over time in

parenting—from before to well after a video-feedback parenting intervention was provided on a random basis to 157 mothers of 1–3-year-olds who scored high on externalizing problems—they found that the intervention succeeded in promoting more sensitive parenting and positive discipline. Moreover, this intervention effect translated into improvements in child behavior, but only for those children carrying the DRD4 7-repeat allele. The same was true when, at post-treatment follow up, stress reactivity was measured by means of change in salivary cortisol before and after the administration of an experimental stressor (i.e., area under the curve) (Bakermans-Kranenburg, van IJzendoorn, Mesman, et al., 2008). Indeed, DRD4 7-repeat children in the experimental group not only showed the least physiological stress reactivity of all children, but the most if their mothers had been assigned to the control group. Of note, too, is that the most pronounced reduction in children’s problem behavior occurred when two conditions obtained: (1) the parenting intervention substantially improved mother’s use of positive discipline techniques; and (2) the child carried the DRD4 7-repeat allele. When only the first condition was met, children’s behavior did not change. One cannot help but wonder why some mothers benefited more from the experimental treatment than others in terms of improved parenting and thus whether an untested GxGxE interaction was responsible for the problem-behavior findings discerned. Could it be that mothers with certain susceptibility genes—including perhaps the 5-HTTLPR short allele—proved most responsive to the intervention and it was this, in combination with their child’s genetic susceptibility to rearing, that generated the results described? This question raises a more general one: when parenting interventions prove effective in changing child behavior, does a small subset of parent–child dyads carry the overall treatment effect—and, specifically, ones comprised of a parent and a child who are both highly malleable for genetic reasons? As no intervention investigation has considered parent as well as child genotype, this possibility remains to be evaluated. Further evidence of DRD4 in moderating intervention effects can be found in Kegel, Bus, and van IJzendoorn’s (2011) work showing that it was DRD4 7-repeat carriers who benefited from specially designed computer games promoting phonemic awareness and, thereby, early literacy in their RCT. Other RCT results point in the same direction with regard to DRD4 7-repeat, including research on African-American teens in which substance use was the outcome examined (Beach, Brody, Lei, & Philibert, 2010; Brody et al., 2013, 2014)

Repeated Measurements

Also of importance is intervention evidence highlighting the moderating role of 5-HTTLPR. Consider in this regard Drury and associates’ (2012) data showing that it was only children growing up in Romanian orphanages who carried short alleles who benefited from being randomly assigned to high-quality foster care—in terms of reductions in the display of indiscriminant friendliness. Eley and associates (2012) also documented intervention benefits restricted to short-allele carriers, but their design included only treated children (i.e., not RCT). Attention must be called, however, to another non-RCT intervention study which found that it was individuals suffering from Posttraumatic Stress Disorder who carried short alleles who benefited the least from cognitive behavior therapy (Bryant et al., 2010). Last to be considered is research examining genetically moderated intervention effects using multigene composites rather than single candidate genes. Consider in this regard the Drury et al. (2012) findings showing that even though BDNF—all by itself—did not operate as a plasticity factor when it came to distinguishing those who did and did not benefit from the foster-care intervention, the already-noted moderating effect of 5-HTTLPR was amplified if a child carried BDNF Met alleles along with short 5-HTTLPR alleles. In other words, the more putative plasticity alleles children carried, the more their indiscriminant friendliness declined over time when assigned to foster care and the more it increased if they remained institutionalized. In light of these results and other research from nonintervention studies focused on BDNF suggesting that BDNF Met alleles are associated with susceptibility to environmental effects in a positive and negative manner (e.g., Chen et al., 2012; Gunnar et al., 2012; Juhasz et al., 2011; Mata et al., 2010; Suzuki et al., 2011), it is somewhat surprising to discover that in a non-RCT treatment study it was individuals carrying BDNF Val rather than Met alleles who benefitted the most from exposure therapy when it came to reducing the symptoms of posttraumatic stress disorder (Felmingham, Dobson-Stone, Schofield, Quirk, & Bryant, 2013). In an intervention study that involved a RCT, Brody, Beach and Chen (2013) confirmed their prediction that the more GABAergic and dopaminergic plasticity-related gene variants African American teens carried, the more protected they were from increasing their alcohol use over time when enrolled in a whole-family prevention program. These findings along with those of Drury et al. (2012) once again call attention to the benefits of moving beyond single polymorphisms when it comes to operationalizing the plasticity phenotype. Moreover, the Drury et al. (2012)

87

results indicate that even if a single gene may not by itself moderate an intervention (or other environmental) effect, in could still play a role in determining the degree to which an individual benefits. These are insights future investigators—and interventionists—should keep in mind when seeking to discover what works for whom.

REPEATED MEASUREMENTS All the differential susceptibility-related work cited herein, as well as in our earlier reviews (Belsky, 2005; Belsky & Pluess, 2009a; Ellis, Boyce, et al., 2011), essentially presumes that children and adults who share the same characteristics, be they temperamental, physiological or genetic plasticity factors, would function in a manner opposite to what was observed were they exposed to contrasting conditions. For example, we have presumed that children carrying the DRD4 7-repeat and showing high levels of externalizing problems in the face of insensitive parenting would score low on such problems if they were raised by skilled caregivers, and this is certainly what some of the just summarized research on genetically moderated intervention effects suggests. But the truth is that there is almost no research validating this assumption, in part because it is unethical to provide a child growing up under for-better, supportive conditions with for-worse, adverse ones just to test for differential susceptibility. One of the few studies that systematically addressed the repeated-measurement issue under consideration, even if not in such differential susceptibility-related terms or with children, examined the probability of taking a risk in a gambling-decision-making game when the chances of winning were high (i.e., for-better condition) and when they were low (i.e., for-worse condition) (Roiser et al., 2006). Results revealed the more 5-HTTLPR short alleles an individual carried, the more likely the person would make a bet (i.e., take a risk) when the chance of winning was high (i.e., for-better condition) and the less likely such a decision would be made when the probability of winning was low (i.e., for-worse condition). In addition to constructing repeated-measurement experiments of the kind just described—though not necessarily having to do with gambling decision making—future researchers would be well advised to take advantage of experiments in nature that could serve this purpose. Consider in this regard the imaginative work of Verschoor and Markus (2011) who took advantage of days in which undergraduates did and did not have examinations to see how those homozygous for long and short 5-HTTLPR

88

Differential Susceptibility

alleles functioned, finding that that the latter experienced the most tension, negative mood and perceived stress on exam days (i.e., for-worse condition), yet the least on nonexam days (i.e., for-better condition). Consider as well Schoebi and associates’ (2012) marital research indicating that spouses homozygous for 5-HTTLPR short alleles proved more sensitive to their partner’s positive and negative emotion during marital interactions, seeming to reflect exposure to for-better and for-worse conditions.

UNKNOWNS IN THE DIFFERENTIAL SUSCEPTIBILITY EQUATION As much as anything else, the central thesis of this chapter, namely, that individuals differ in their developmental plasticity and thus susceptibility to environmental influence, raises many issues in need of further investigation, many of which have already been highlighted. Here we call attention to additional ones. Same Individuals, Different Plasticity Markers? Is the research conducted by investigators often working in different fields—psychology, physiology, genetics— actually identifying the same more-and-less-susceptible individuals by different means? Consider in this regard that the very children who score high in negative emotionality, physiological stress reactivity and who have 5-HTTPLR short alleles could often be one and the same (5-HTTLPR and cortisol reactivity: Gotlib, Joormann, Minor, & Hallmayer, 2008; 5-HTTLPR and infant negative emotionality: Holmboe, Nemoda, Fearon, Sasvari-Szekely, & Johnson, 2011). Some recent evidence raises the possibility, however, that empirically related plasticity factors may not be interchangeable. Even though recent work indicates that DRD4 and negative emotionality early in life are related (Holmboe et al., 2011; Ivorra et al., 2010), Belsky and Pluess (2013b) observed that the moderating effect of DRD4 does not account for a parallel moderating effect of negative emotionality, at least in a study of effects of child care on children’s social functioning. Categorical or Dimensional Scaling of Plasticity? Another issue that merits consideration in future work is whether individual differences in plasticity are best conceptualized in typological or dimensional terms. Adopting evolutionary terminology pertaining to reproductive strategy, we can ask whether there exist plastic and fixed

strategists who are and are not, respectively, susceptible to environmental experiences, thereby following conditional and alternative pathways of development (Belsky, 2000). Or does it make more sense to think in terms of a plasticity gradient, as mentioned earlier, with individuals varying in degree of susceptibility to environmental influences? But perhaps in the same way that light is best conceived as both a wave and a particle, depending upon the circumstance, the choice between a typological and dimensional conceptualization and parameterization of plasticity is a false one, as the approach which proves best will vary across conceptual purposes and empirical inquiries. Domain Specific or Domain General? Also unknown is whether it makes the most sense to regard more and less plasticity as a global, macro, trait-like characteristic of individuals or consider it in more domain-specific terms? Are some people simply more malleable than others across the board, almost irrespective of the environmental factor and aspect of functioning under consideration, or are people a complex mosaic of components that are more and less susceptible to environmental influence, thus making them both more and less malleable relative to others? Whereas the latter conceptualization might make more intuitive sense, two new sets of evidence underscore the domain generality of plasticity. The first involves the nature and effects of two dramatically different and already cited Dutch interventions which used different approaches to influence children, one indirectly and the other directly, and focused on different developmental outcomes, yet revealed that children sharing the same plasticity factor were the primary beneficiaries of each. More specifically, even though one intervention promoted sensitive parenting via video feedback to reduce toddler’s externalizing behavior (Bakermans-Kranenburg, van IJzendoorn, Pijlman, et al., 2008) and cortisol-related stress reactivity (Bakermans-Kranenburg, van IJzendoorn, Mesman, et al., 2008) and the other relied on a computerized instructional program to promote preschooler’s phonemic awareness and, thereby, early literacy (Kegel et al., 2011), it was children carrying the DRD4–7-repeat allele who disproportionately or exclusively benefited in both cases. Then there is the longitudinal data evaluated by Roisman and associates (2012) showing that infants with difficult temperaments measured in the first six months of life proved more susceptible to the positive and negative effects, respectively, of sensitive and insensitive mothering experienced during their first three years of life when

Unknowns in the Differential Susceptibility Equation

it came to predicting teacher-reported social competence, behavior problems, and academic skills across the primary-school years. Also, this was so even with different teachers rating the children each and every academic year. Moreover, even though a diathesis–stress rather than differential susceptibility model fit the data best when it came to predicting performance on objective tests of academic performance through 15 years of age, it once again proved to be children who as infants had difficult temperaments who were disproportionately susceptible to the shorter and longer term effects of early mothering. Despite such evidence, it still seems premature to embrace the broad domain-general interpretation of developmental plasticity. Nevertheless, the results just summarized caution against its premature dismissal. Ultimately, further work is called for before any firm conclusions can be drawn. Our own suspicion is that there will be variation across individuals in that some will prove highly susceptible to many developmental experiences and environmental exposures (i.e., extremely plastic) and that some will prove susceptible to virtually none (i.e., extremely fixed), but that most will prove to be intermediate between these extremes. One might even envision a bell curve with regard to the degree of plasticity that individuals manifest. Origins of Plasticity: Nature or Nurture (or Both)? A fundamental issue that should be raised for future research pertains to whether, or at least the extent to which, plasticity should be regarded as principally a function of nature or nurture. Certainly the GxE evidence calls attention to heritable individual differences in plasticity, as well as to the fact that so-called vulnerability genes or risk alleles might in many cases be better conceptualized as plasticity genes (Belsky et al., 2009). After all and with regard to the latter point, why would natural selection, for example, maintain much less select genes that only functioned to foster depression in the face of negative life events or antisocial behavior in the face of child maltreatment? Were these perhaps down-side costs of selecting and preserving genes that engendered benefit in the face of supportive contextual conditions, or even operated as adaptations when also functioning in a diathesis–stress manner, it would seem to make more sense for them to be selected. But just because GxE studies are replete with evidence, often unnoticed, of differential susceptibility should not lead to the presumption that plasticity is only function of genetics. Central to Boyce and Ellis’s (2005) thinking, it will be recalled, is the role of extremely supportive and

89

unsupportive environments in fostering physiological reactivity and, thereby, developmental plasticity. Especially notable, in fact, is recent research on the putatively adverse effects on the developing child of maternal stress during pregnancy. This is because so-called fetal programming appears to influence several of the very susceptibility factors already considered that seem to be markers of differential susceptibility (Belsky & Pluess, 2009b; Pluess & Belsky, 2011). Consider in this regard research showing (1) that maternal stress during pregnancy predicts difficult temperament at 3 months of age (Huizink et al., 2002) and emotional reactivity to novelty in 4-month-olds (Möhler, Parzer, Brunner, Wiebel, & Resch, 2006); (2) that prenatal maternal depression and elevated cortisol levels in late pregnancy predict negative reactivity at age 2 (Davis et al., 2007); and (3) that maternal prenatal anxiety predicts awakening cortisol in 10-year-olds (O’Connor et al., 2005). On one hand, such data suggest that very early experience—in the womb—may shape plasticity, not just genetics, as the outcomes just mentioned are among the very child characteristics found in work cited herein to demarcate heightened susceptibility to environmental influences. Just as importantly, this re-interpretation of putatively negative effects of prenatal stress raises fundamental questions about the disturbance-focused perspective that pervades virtually all research and theory on prenatal programming: Is it the case that prenatal stressors compromise later development, as prevailing thinking presumes, or do these prenatal experiences promote plasticity—and thus the organism’s openness to future experiential input, be it positive or negative in character? That is, is there prenatal programming of postnatal programming (Pluess & Belsky, 2011)? Oberlander et al.’s (2008) epigenetic findings showing that maternal depressed mood in pregnancy predicts increased methylation of the human glucocorticoid receptor gene (NR3C1, measured in neonatal cord blood), which itself forecasts elevated cortisol stress reactivity at age three months, illuminates at least one biological mechanism that may be central to such fetal programming of postnatal plasticity. Recall in this regard that cortisol reactivity may well demarcate heightened susceptibility to rearing influences (Boyce & Ellis, 2005). Before concluding on the basis of fetal programming research that plasticity may be a function of experience as much as a function of genetics, we should not lose sight of the fact that GxE may characterize the fetal programming process (Gluckman & Hanson, 2005). This raises the possibility that some fetuses may be more susceptible to fetal programming than others, for genetic reasons.

90

Differential Susceptibility

If they are, it would suggest that plasticity is a function not just of nature or nurture, but of their interaction (Belsky & Pluess, 2009b). That is, some individuals may be more likely than others to be affected by experience, most notably perhaps fetal experience, in ways that subsequently affect whether or to what degree they will be influenced by the postnatal world they encounter. Initial evidence to this effect was found in recent GxE work based on data from a large Dutch cohort study in which maternal anxiety during pregnancy significantly predicted negative emotionality—a putative susceptibility factor—in infants carrying one or more copies of the 5-HTTLPR short allele, but not in those homozygous for the long allele (Pluess et al., 2011). These results provide the first empirical evidence consistent with the aforementioned hypothesis that individuals with certain genotypes who are also exposed to specific prenatal environments may be more likely than those not carrying such alleles or not so exposed to prove susceptible to postnatal environmental influences (Pluess et al., 2011). However, it has to be noted that whether infants characterized by high negative emotionality where also more affected by subsequent postnatal contextual factors, consistent with a differential susceptibility pattern, has not yet been determined within this sample. This would be necessary to conclusively test whether genetic factors moderate the effects of prenatal programming on postnatal plasticity (Pluess & Belsky, 2011). Not to be ignored in this discussion of the determinants of developmental plasticity is evidence that physiological and behavioral susceptibility factors are also

1

Genotype

shaped by postnatal experiences (e.g., Heim et al., 2000; Kaplan, Evans, & Monk, 2008). Figure 2.3 represents a schematic model outlining the multiple means and pathways by which developmental plasticity is likely regulated: (1) genetic contributions to susceptibility; (2) genetic contributions mediated by behavioral and physiological susceptibility factors; (3) prenatal and (4) postnatal environmental effects on behavioral and physiological susceptibility factors; (5) GxE interactions involving the prenatal and (6) postnatal environment; and (7) interactions between the prenatal and postnatal environment in shaping susceptibility factors. Population Variation in Plasticity? Do populations differ in the degree to which children are malleable? If we consider for a moment the fact that selection for plasticity perhaps only pays off in fitness terms, especially given its potential costs, if what happens at one point in time is systematically related to what happens at a later point in time, the possibility emerges that human populations may vary in the degree to which children are malleable. After all, in some ecological niches past and future could have been more systematically related than in others. Future research can thus address whether in ecological niches in which the present and future are (or have been) more related, the payoff for plasticity could well be greater, with greater selection for plasticity in human populations. Three observations seem noteworthy in this context. First, the DRD4 7-repeat allele not only recently

Susceptibility Factors Genes (e.g., short allele of the 5-HTTLPR)

GxE

2 Programming

ExE Postnatal

Postnatal

3

Eplgenetic Mechanisms

Environment

Prenatal

Prenatal

Physiology (e.g., cortisol stress reactivity)

Individual Degree of Developmental Plasticity

Behavior (e.g., infant temperament)

Programming

Figure 2.3 Nature, nurture, and differential susceptibility: a process model. 1 = genetic contribution to general susceptibility partially mediated by susceptibility factors (nature); 2 = prenatal environment shapes susceptibility factors (nurture); 3 = postnatal environment shapes susceptibility factors (nurture). In addition, genotype interacts with both prenatal and postnatal environment and prenatal and postnatal environments interact with each other to shape susceptibility factors. GxE = Gene–environment interaction; ExE = environment–environment interaction. Source: Pluess, M., Stevens, S., & Belsky, J. (2013). Differential Susceptibility: Developmental and Evolutionary Mechanisms of Gene-Environment Interactions. In M. Legerstee, D. W. Haley & M. H. Bornstein (Eds.), The Infant Mind: Origins of the Social Brain (pp. 77–96). New York: Guilford.

Unknowns in the Differential Susceptibility Equation

emerged in human populations (∼40k years ago), but varies substantially across them, having an extremely low incidence in Asia yet a high frequency in the Americas (Ding et al., 2002); intriguingly, the reverse is true of another possible plasticity gene, the s/s genotype of 5-HTTLPR (Kim et al., 2007). Second, one wild bird population shows evidence that selection favoring individuals who are highly plastic with regard to the timing of reproduction has intensified over the past three decades, perhaps in response to climate change causing a mismatch between the breeding times of the birds and their caterpillar prey (Nussey, Postma, Gienapp, & Visser, 2005). Finally, Suomi (2006) observed that only two species of primates fill diverse ecological niches around the world, humans and rhesus macaques, and that what distinguishes these two weed species from all other primates is the presence of a polymorphism in the serotonin transporter gene (5-HTTLPR). In the same way, then, that not all primates seem to share the 5-HTTLPR, populations may differ generally in the presence of plasticity alleles, as already noted, due to variation in historical continuity of contextual conditions over time. Gender Differences in Plasticity? Originally, the differential susceptibility hypothesis did not have anything to say about the prospect of gender differences in developmental plasticity (Belsky, 1997b, 2005; Belsky & Pluess, 2009b). Therefore, it is not surprising that most research pertaining to differential susceptibility do not consider gender other than as a covariate to control for gender differences in the dependent variables under investigation. Nevertheless, it seems potentially—even if not certainly—noteworthy that some research on differential susceptibility has revealed gender differences in developmental plasticity. Consider in this regard one study involving a behavioral susceptibility factor, infant negativity, which detected differential susceptibility related effects of father involvement on prosocial behavior, but only in the case of girls (Ramchandani, van IJzendoorn, & Bakermans-Kranenburg, 2010). Consider next research focused on physiological reactivity as a susceptibility factor, with El-Sheikh, Keller, et al. (2007) reporting that skin conductance reactivity of children moderated the effects of marital conflict on externalizing problems, again in a differential susceptibility related manner, but this time only in the case of boys. Several GxE studies involving the 5-HTTLPR have also detected the for-better-and-for-worse pattern of environmental influence, but only for females (Brummett et al.,

91

2008; Eley et al., 2004; Hammen, Brennan, Keenan-Miller, Hazel, & Najman, 2010). Nevertheless, other work which has formally tested three-way interactions involving gender, 5-HTTLPR and an environmental factor has failed to detect sex-differentiated GxE patterns taking the form of differential susceptibility (Pluess et al., 2010, 2011). Finally, GxE-related gender differences emerged in work that composited putative plasticity genes, finding that it was only adolescent boys whose self-control proved related to the parenting they experienced in a for-better-and-for-worse manner (Belsky & Beaver, 2011). In summary, although gender differences in the operation of differential susceptibility emerge in a number of investigations, one is hard pressed to discern order in the empirical chaos. Perhaps if the issue is addressed more systematically and consistently—and theory as to why such should be expected emerges to motivate such inquiry—future research will tell a different story. Competitive Evaluation of Models of Person–Environment Interaction Much of the evidence cited in our earlier reviews of differential susceptibility-related evidence (Belsky et al., 2009; Belsky & Pluess, 2009a), as well as herein, was selected for consideration based on eyeball tests of graphed interactions between an environmental predictor and developmental outcome moderated by a plasticity factor. The limits of this approach to testing for differential susceptibility led Belsky et al. (2007) to propose explicit empirical criteria which we summarized earlier, while calling attention to the fact that much of the research cited in this chapter does not meet such seemingly rigorous evidentiary standards. Indeed, Kochanska and associates (2011) extended this effort to establish standards for distinguishing differential susceptibility from diathesis stress using the regions-of-significance approach (Preacher, Curran, & Bauer, 2006; Takane & Cramer, 1975) to evaluate, separately, the for-better and for-worse sides of the differential susceptibility hypothesis upon documenting the required cross-over interaction. Roisman and associates (2012) recently offered an even more demanding approach for evaluating differential susceptibility by providing two new quantitative metrics in combination with the regions-of-significance analysis: the proportion of interaction index, which represents the proportion of the total area of an interaction plot that is uniquely attributable to differential susceptibility; and the proportion affected index, which quantifies the proportion of people in the sample who fall above the crossover point

92

Differential Susceptibility

(i.e., proportion of sample in the for-better condition). In addition, Roisman et al. (2012) recommend testing for nonlinear relationships between predictor and outcome and correction for multiple comparisons. All the approaches offered to date to evaluate differential susceptibility have involved conducting essentially exploratory tests of organism–environment interactions as a first step before further interrogating the data at hand. Widaman and associates (Belsky, Pluess, & Widaman, 2013; Widaman et al., 2012) recently advanced a strategy which sidesteps such approaches, moving directly to test competing theoretical models. In a first step the cross-over point between high- and low-susceptible individuals is estimated with the help of a re-parameterized regression equation (Widaman et al., 2012). If the estimated cross-over point lies within the observed range of the predictor, four additional re-parameterized regression models are applied, two constrained to reflect a weak and a strong version of diathesis–stress and two to reflect a weak- and a strong-version of differential susceptibility. Fit indices of the four models are then compared to evaluate whether the data fits better with a diathesis–stress or differential susceptibility model (Belsky et al., 2013). The strong version of differential susceptibility presupposes that there are those who are susceptible to both positive and negative contextual conditions and others who are not, whereas the weak version presupposes that while many may be susceptible to environmental influences to some extent, some are more affected than others. The time has thus come to move beyond exploratory tests of organism–environment interaction in regression and ANOVA designs, and more formally evaluate in a competitive manner the a-priori hypotheses central to diathesis–stress and differential susceptibility models of environmental action.

the environment, but the latter deferring commitment and choosing to keep sampling the environment until a better estimate of the environment can be established. Whether this is actually the case or not is an issue that should be pursued empirically in the future. Indeed, Frankenhuis and Panchanathan’s (2011) intriguing ideas lead us to wonder whether it is when cue reliability is low that individuals call upon the epigenome to infer from markings passed down across multiple generations what past environments were like and thus what is the best developmental pathway to pursue. Such a process would be consistent with Kuzawa’s (2005) notion of intergenerational phenotypic inertia that highlights the fact that it is not just the experiences of the immediately preceding generation that developing organisms are sensitive to, but even ones preceding that. Parent–Child Conflict of Interest Like all-too-many developmental scientists lacking foundational knowledge about evolutionary theory, virtually all our thinking and writing about differential susceptibility to date has presumed that malleable children will have their development regulated by their rearing experiences. But as Trivers (1974) pointed out more than four decades ago, the fact that parents and children share, on average, 50% of their genes means that their biological interests are not one and the same and thus that they often experience a conflict of interest (see also Schlomer, Del Giudice, & Ellis, 2011). In discussing the influence of prenatal stress on development, Del Giudice (2012) recently pointed out that it is because of this fact that children, even highly malleable ones, may not simply take instructions from their parents. What are the implications of this observation for understanding developmental plasticity? Family Dynamics

Variation in Environmental Cue Reliability Whether children develop in a generally supportive or unsupportive environment, it is likely that the reliability of environmental cues indicative of the valence of the rearing milieu will vary. One can imagine in this regard two different hostile environments, one in which conflict is constant and thus highly predictable and another in which conflict is episodic and unpredictable. Frankenhuis and Panchanathan (2011) have insightfully observed that highly plastic individuals growing up in these two environments may differ in terms of when they commit to a developmental strategy, with the former deciding to specialize early, having high confidence about the state of

However intrigued we have become with theory and evidence pertaining to differential susceptibility and the notion that there exist individual differences in developmental plasticity, one issue has always caused us concern: We have never heard a parent characterize one of their children as more susceptible to their influence than another, which we regard as distinct from their saying that their children are temperamentally different, with perhaps one being more sensitive to challenging or stressful situations than others or even being more difficult to care for. Why is that? Evolutionary reasoning would seem to imply that parents should detect for whom their socialization efforts, be they implicit or explicit, pay off and for whom they

Unknowns in the Differential Susceptibility Equation

do not and so adjust those efforts accordingly. Yet we are aware of no evidence suggesting, much less indicating this to be the case. Indeed, should not more malleable siblings seek to alert their parents to the fact that they are more responsive to rearing inputs to obtain more attention and other resources? Might it be the case, however, that such children do engage in such efforts, but that they do not prove effective because less malleable children employ tactics to camouflage in some way this characteristic of theirs? Perhaps it is such children who prove most likely to claim it’s not fair within families when siblings are perceived as being treated better as a way of throwing their parents off track. At the very least, it would be interesting to study parental awareness or perception of variation in susceptibility across siblings within a family. Timing of Susceptibility It is widely assumed by many developmentalists that it is during the early years of life when human development is most susceptible to environmental influences, for better and for worse (e.g., Blair & Raver, 2012; Ganzel & Morris, 2011), as plasticity is presumed to be greatest when biological systems are being laid down. But even if this is true in general, does it mean that it is true in the case of each and every individual? It certainly seems conceivable that at times in the ancestral past that individuals who were especially susceptible to environmental influence at some later point or points in the life span might have experienced an adaptive advantage, leading perhaps to the selection of genes for later rather than earlier plasticity. If this is so, it implies that variation should exist in terms of when children and even adults are especially developmentally plastic. And if this is the case, does it not imply that those who appear not to be especially malleable may simply not be so at the developmental point in time when developmental experiences and environmental exposures are measured? In fact, might it be the case that some are especially susceptible to contextual regulation early, others later, others at all times and still others more or less never? For Better and For Worse—or Just for Better? One important unknown in the differential susceptibility equation is whether some individuals could be especially susceptible to just adversity, some to just environmental support and enrichment, some to both, and some to neither. Of note in this regard is that whereas the English language has terms to characterize those highly susceptible to both positive and negative conditions

93

(i.e., plastic/malleable) and highly susceptible to adversity (i.e., vulnerable), it is difficult to find a term that would characterize those disproportionately responsive to supportive conditions only. Indeed, the only term that Belsky and Pluess (2009b) could identify—in tongue-and-cheek fashion—to characterize those disproportionately likely to benefit from positive experiences and exposures was lucky. And this was after asking speakers of diverse languages, including French, German, Italian, Chinese, Czech, Spanish, Korean, and Polish, whether there was a word or term in their native tongue that captured the for-better side of differential susceptibility. The fact that there did not seem to be such a term in any of these languages raised the intriguing possibility that one reason variability in response to positive—as opposed to negative—experiences went unnoticed, or at least unheralded, for so long was that we simply lack terminology to direct attention to it. Recently, Manuck and associates (2011a; Sweitzer et al., 2013) introduced the term vantage sensitivity to characterize the for-better side of differential susceptibility and more generally variability in response to positive experiences. Vantage is short for advantage, but in addition to implying benefit, gain or profit, it is also defined as a position, condition, or opportunity that is likely to provide superiority or an advantage (HMCO, 2000). In Manuck’s own words (Manuck, 2011b), vantage bespeaks a position conferring advantage, benefit, or gain without bearing the singularity of a particular advantage. We embrace and promote the term vantage sensitivity to describe the notion that some individuals are more sensitive and positively responsive to the environmental advantages to which they are exposed to, but not especially susceptible to adverse experiences, as is presumed in the case of differential susceptibility. These advantages may take the form of security of attachment derived from sensitive parenting, academic achievement resulting from high-quality child care, prosocial behavior due to supportive friendship networks, and life satisfaction stemming from positive life events, as well as sense of efficacy following psychotherapy, to name just a few possibilities. In hopes of stimulating work which would distinguish vantage sensitivity from differential susceptibility, Pluess and Belsky (2013) proposed the following concepts to characterize variability in response to positive experiences: (1) vantage sensitivity reflects the general proclivity of an individual to benefit from positive and presumptively well-being– and competence-promoting features of the environment, just as vulnerability depicts the tendency to succumb to negative effects of adversity in the diathesis–stress framework; (2) the degree

94

Differential Susceptibility

of vantage sensitivity is a function of the presence of vantage-sensitivity factors (i.e., promotive factors), just as vulnerability/risk factors increase vulnerability to negative effects of adversity in the diathesis–stress framework; (3) vantage resistance describes the failure to benefit from positive influences, just as resilience characterizes resistance to negative effects of adversity in the diathesis–stress framework; and (4) the degree of vantage resistance is a function of the presence of vantage-resistance factors or absence of vantage-sensitivity ones, just as protective factors increase resilience to negative effects of adversity in the diathesis–stress framework. In summary: vantage-sensitivity factors increase vantage sensitivity to the beneficial effects of positive experiences and exposures, whereas vantage-resistance factors diminish or even completely eliminate positive response to the same supportive conditions (Figure 2.4 for graphical illustration). Having delineated the nature of vantage sensitivity, some evidence consistent with the notion merits consideration. Testing the a priori hypothesis that children characterized by a high sensitive personality (Aron & Aron, 1997) would show vantage sensitivity to psychological intervention, Pluess and Boniwell (2015) investigated variation in the anticipated positive effects of a school-based resilience-promoting program administered to a sample

positive

Vantage Sensitivity Model ty

ivi

sit

en es

tag

Outcome

n Va

negative

Vantage resistance

negative

positive

of 200 11-year-old girls in one of the most deprived areas in London, United Kingdom (Pluess, Boniwell, Hefferon, & Tunariu, submitted). The intervention led to a decrease of depression symptoms observable up to the 12 month follow-up assessment, but, consistent with vantage sensitivity, exclusively among children who scored in the upper quartile of the highly-sensitive-child questionnaire. All other children failed to benefit from the intervention, at least regarding changes in depression symptoms. An example of vantage sensitivity as a function of a physiological moderating factor emerged recently in Eisenberg et al.’s (2012) work of 213 toddlers and their families. Baseline RSA moderated effects of the home environment (based on a composite of maternal and paternal education, family income, and marital quality) in early childhood on repeatedly assessed child aggression. For children with high RSA activity, high environmental quality was associated with less aggression at 54 months, whereas children characterized by low RSA activity—displaying vantage resistance—did not benefit from high-quality environments at all. Eley et al.’s (2012) evaluation of whether 5-HTTLPR moderated effects of cognitive-behavioral therapy for anxiety disorders provides an example of genetic vantage sensitivity. Clinical diagnoses of anxiety disorders and symptom severity were assessed before and after treatment as well as 6 months after treatment ended. Although all children appeared to benefit from the treatment, the positive effect of the intervention was particularly pronounced in the case of those children carrying the short allele in this work with 6–13-year-old boys and girls (N = 359). More specifically, those homozygous for the 5-HTTLPR short allele showed a significantly greater reduction in symptom severity from pre-treatment to follow-up assessment, so much so, in fact, that they proved 20% more likely than others to be free of anxiety disorder at the 6-month follow-up assessment.

FUTURE DIRECTIONS IN RESEARCH ON DIFFERENTIAL SUSCEPTIBILITY

environment/experience

Figure 2.4 Graphical illustration of vantage sensitivity. Vantage sensitivity describes the propensity to respond favorably to positive experiences as a function of individual characteristics, whereas vantage resistance reflects the inability to benefit from supportive influences. No differences are predicted in response to negative influences. Source: Adapted from M. J. Bakermans-Kranenburg & M. H. van IJzendoorn, Gene–environment interaction of the dopamine D4 receptor (DRD4) and observed maternal insensitivity predicting externalizing behavior in preschoolers. Developmental Psychobiology, 48(5), 2006, Figure 1.

Here we highlight a number of issues that future research should prioritize to advance knowledge regarding individual differences in developmental plasticity. Following this, we draw some general conclusions. Differential susceptibility is a rather new concept with most empirical work having emerged in just the past few years. Although already an influential framework, particularly within the fields of developmental and clinical psychology, providing answers to some long-standing

Future Directions in Research on Differential Susceptibility

questions regarding individual differences in response to environmental influences, differential susceptibility reasoning actually appears to raise more new questions than answering old ones. 1. Mechanisms of differential susceptibility. Among all the unknowns in the differential susceptibility equation that have been discussed in this chapter as well as in previous publications (Belsky & Pluess, 2009a, 2009b, 2013a; Pluess & Belsky, 2013), that regarding mechanisms underlying individual differences in susceptibility seems especially important to apply differential susceptibility in practice, a thorough understanding of how susceptibility gets instantiated will be crucial. Future work should therefore endeavor to identify specific mechanisms and related processes across different levels of analysis from genetic and molecular, to neurological and hormonal, to behavioral. 2. Domain specificity. Are some individuals generally more and others less susceptible to all kinds of developmental experiences and environmental exposures, or are they more susceptible to certain influences compared with others? For example, might some be especially sensitive to peer influence and others to parental influence; or some especially susceptible to the effects of punishment and others to the effects of reward? Empirical investigation of this important question will require studies testing whether the same susceptibility factor increases plasticity of the same individual across different environmental influences and with respect to different phenotypic characteristics. 3. Nature and nurture of susceptibility. Is an individual’s degree of susceptibility biologically fixed or can it be influenced through intervention or particular experiences? As discussed earlier, susceptibility seems to have a genetic basis. The question of determinants of susceptibility is particularly important regarding potential intervention efforts to foster both resilience and vantage sensitivity. Future studies with large longitudinal prospective samples that include molecular genetic measures are required to investigate the genetic and environmental determinants of susceptibility as well as the interaction of the two. 4. Development of developmental plasticity. Closely linked to the question of whether susceptibility is a fixed or malleable trait is a developmental question: Does susceptibility evolve during development or is it established early in life, perhaps even by the time of birth? Future research should seek to illuminate, for example, whether an individuals’ propensity for vulnerability to

95

environmental adversity may increase in response to accumulated negative experiences (i.e., diathesis–stress) and whether susceptibility to positive experiences (i.e., vantage sensitivity) may increase over time as a result of exposure to positive influences. Evidence of such would prove consistent with the notion of upward spiral dynamics (Fredrickson & Joiner, 2002). 5. Susceptibility to what? There appear to be three general models of person–environment interaction today: diathesis–stress (i.e., variability in response to exclusively negative experiences, Monroe & Simons, 1991), vantage sensitivity (i.e., variability in response to exclusively positive experiences, Pluess & Belsky, 2013), and differential susceptibility (i.e., variability in response to both negative and positive experiences, Belsky & Pluess, 2009a), with the latter reflecting the combination of both diathesis–stress and vantage sensitivity as a function of the same moderating variable. Empirical evidence for each of these models suggests that whereas some individuals may be disproportionately susceptible to negative experiences others may be disproportionately susceptible to positive environmental conditions and still others may be disproportionately susceptible to both—or to neither. Future research should investigate whether these four different types of environmental sensitivity exist in the population (Pluess, 2015). 6. Susceptibility versus vulnerability versus vantage sensitivity factors. Susceptibility, Vulnerability and Vantage Sensitivity have often been associated with the same psychological, physiological and genetic characteristics. Future research should aim to determine whether there are, in fact, specific factors for vulnerability, vantage sensitivity, and differential susceptibility and associated specific mechanisms and related processes. 7. Sensitive periods. Is an individual’s susceptibility constant across development or are there sensitive periods where susceptibility to environmental influences is more pronounced compared with other developmental periods? Furthermore, do different susceptibility factors increasing susceptibility to environmental influences during different developmental periods? Addressing these questions requires studies testing interactions between putative susceptibility factors and environmental influences within the same person repeatedly across different developmental periods. 8. Measures of susceptibility. To apply differential susceptibility to psychological and educational services, it will be important to develop measures of susceptibility that are both reliable and easily applicable. Measures of

96

Differential Susceptibility

physiological reactivity or genetic susceptibility are currently not suited for that purpose. Future efforts should be directed toward the development of rating scales and observational indices of susceptibility. Although some existing self-report measures have been found to predict differential susceptibility as well as vantage sensitivity (for example, sensory-processing sensitivity, Aron & Aron, 1997; Pluess & Boniwell, 2015) more specific measures—validated rigorously in empirical studies—are required. 9. Ethical considerations. Give the potential application of differential susceptibility reasoning in clinical, educational, child care and other settings, it will be important to consider ethical challenges before applying differential susceptibility thinking to these services. Besides the mere identification of ethical challenges (e.g., systematic screening for high and low susceptibility, stigma of being found high or low in susceptibility, treating individuals differently given their susceptibility) it will be important to explore solutions for dealing with these potential problematic issues (e.g., offering alternative treatments rather than no treatment to low susceptible individuals). GENERAL CONCLUSION One of the most striking features of the empirical work reviewed in this chapter is how diverse the evidence base for differential susceptibility is, involving a myriad of developmental experiences and environmental exposures (e.g., parenting, child-care quality, life events, rural vs. urban residence, birth season), numerous phenotypic outcomes from different developmental periods (e.g., disorganized infant attachment, externalizing problems in childhood, antisocial behavior in adolescence, depression throughout adulthood), and distinct susceptibility factors (e.g., temperamental and other behavioral attributes of children, physiological and genetic characteristics). Beyond highlighting empirical evidence for differential susceptibility, we have sought to emphasize the utility of thinking about development from an evolutionary perspective, without which the differential susceptibility framework seems unlikely to have emerged. Indeed, what is striking is how simple a concept differential susceptibility appears to be. It thus seems likely that failure to appreciate that there are individual differences in developmental plasticity, whether general or specific in nature, has been due, at least in part, by the general lack of an evolutionary orientation among many students of human development and related disciplines.

Beyond documenting the utility of an evolutionary perspective, one great benefit of differential susceptibility reasoning is its generativity—in stimulating new thinking and thus new avenues of research about the human development process. We thus feel confident that differential susceptibility thinking will continue to inspire, stimulate, and guide empirical and theoretical research efforts in human development and look forward to the scientific progress this should occasion.

REFERENCES Alexander, N., Klucken, T., Koppe, G., Osinsky, R., Walter, B., Vaitl, D., . . . Hennig, J. (2012). Interaction of the serotonin transporter-linked polymorphic region and environmental adversity: Increased amygdala-hypothalamus connectivity as a potential mechanism linking neural and endocrine hyperreactivity. Biological Psychiatry, 72(1), 49–56. doi: 10.1016/j.biopsych.2012.01.030 Aron, E. N., & Aron, A. (1997). Sensory-processing sensitivity and its relation to introversion and emotionality. Journal of Personality and Social Psychology, 73(2), 345–368. Aron, E. N., Aron, A., & Davies, K. M. (2005). Adult shyness: The interaction of temperamental sensitivity and an adverse childhood environment. Personality and Social Psychology Bulletin, 31(2), 181–197. Aron, E. N., Aron, A., & Jagiellowicz, J. (2012). Sensory processing sensitivity: A review in the light of the evolution of biological responsivity. Personality and Social Psychology Review, 16(3), 262–282. doi: 10.1177/1088868311434213 Bakermans-Kranenburg, M. J., Dobrova-Krol, N., & van IJzendoorn, M. (2012). Impact of institutional care on attachment disorganization and insecurity of Ukrainian preschoolers: Protective effect of the long variant of the serotonin transporter gene (5HTT). International Journal of Behavioral Development, 36(1), 11–18. doi: 10.1177/0165025411406858 Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2006). Gene–environment interaction of the dopamine D4 receptor (DRD4) and observed maternal insensitivity predicting externalizing behavior in preschoolers. Developmental Psychobiology, 48(5), 406–409. Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2007). Research review: genetic vulnerability or differential susceptibility in child development: The case of attachment. Journal of Child Psychology and Psychiatry, 48(12), 1160–1173. Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to rearing environment depending on dopamine-related genes: New evidence and a meta-analysis. Development and Psychopathology, 23(1), 39–52. doi: 10.1017 /s0954579410000635 Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & Juffer, F. (2003). Less is more: Meta-analyses of sensitivity and attachment interventions in early childhood. Psychological Bulletin, 129(2), 195–215. Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & Juffer, F. (2008a). Earlier is better: A meta-analysis of 70 years of intervention improving cognitive development in institutionalized children. Monographs of the Society for Research in Child Development, 73(3), 279–293. doi: 10.1111/j.1540–5834.2008.00498.x Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., Mesman, J., Alink, L. R., & Juffer, F. (2008b). Effects of an attachment-based intervention on daily cortisol moderated by dopamine receptor D4: A

References randomized control trial on 1- to 3-year-olds screened for externalizing behavior. Development and Psychopathology, 20(3), 805–820. Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., Pijlman, F. T., Mesman, J., & Juffer, F. (2008c). Experimental evidence for differential susceptibility: Dopamine D4 receptor polymorphism (DRD4 VNTR) moderates intervention effects on toddlers’ externalizing behavior in a randomized controlled trial. Developmental Psychology, 44(1), 293–300. Beach, S. R. H., Brody, G. H., Lei, M.-K., & Philibert, R. A. (2010). Differential susceptibility to parenting among African American youths: Testing the DRD4 hypothesis. Journal of Family Psychology, 24(5), 513–521. doi: 10.1037/a0020835 Beach, S. R. H., Lei, M.-H., Brody, G. H., Simons, R. L., Cutrona, C., & Philibert, R. A. (2012). Genetic moderation of contextual effects on negative arousal and parenting in African-American parents. Journal of Family Psychology, 26(1), 46–55. doi: 10.1037/a0026236 Beaver, K. M., Sak, A., Vaske, J., & Nilsson, J. (2010). Genetic risk, parent–child relations, and antisocial phenotypes in a sample of African-American males. Psychiatry Research, 175(1–2), 160–164. doi: 10.1016/j.psychres.2009.01.024 Bellivier, F., Chaste, P., & Malafosse, A. (2004). Association between the TPH gene A218C polymorphism and suicidal behavior: A meta-analysis. American Journal of Medical Genetics B Neuropsychiatric Genetics, 124B(1), 87–91. Belsky, J. (1997a). Theory testing, effect-size evaluation, and differential susceptibility to rearing influence: the case of mothering and attachment. Child Development, 68(4), 598–600. Belsky, J. (1997b). Variation in susceptibility to rearing influences: An evolutionary argument. Psychological Inquiry, 8, 182–186. Belsky, J. (2000). Conditional and alternative reproductive strategies: Individual differences in susceptibility to rearing experience. In J. Rodgers, D. Rowe, & W. Miller (Eds.), Genetic influences on human fertility and sexuality: Theoretical and empirical contributions from the biological and behavioral sciences (pp. 127–146). Boston, MA: Kluwer. Belsky, J. (2005). Differential susceptibility to rearing influences: An evolutionary hypothesis and some evidence. In B. Ellis & D. Bjorklund (Eds.), Origins of the social mind: Evolutionary psychology and child development (pp. 139–163). New York, NY: Guildford. Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2007a). For better and for worse: Differential susceptibility to environmental influences. Current Directions in Psychological Science, 16(6), 300–304. Belsky, J., & Beaver, K.M. (2011). Cumulative-genetic plasticity, parenting and adolescent self-control/regulation. Journal of Child Psychology and Psychiatry, 52(5), 619–626. Belsky, J., Hsieh, K. H., & Crnic, K. (1998). Mothering, fathering, and infant negativity as antecedents of boys’ externalizing problems and inhibition at age 3 years: Differential susceptibility to rearing experience? Development and Psychopathology, 10(2), 301–319. Belsky, J., Jonassaint, C., Pluess, M., Stanton, M., Brummett, B., & Williams, R. (2009). Vulnerability genes or plasticity genes? Molecular Psychiatry, 14, 746–754. Belsky, J., & Pluess, M. (2009a). Beyond diathesis–stress: Differential susceptibility to environmental influences. Psychological Bulletin, 135(6), 885–908. Belsky, J., & Pluess, M. (2009b). The nature (and nurture?) of plasticity in early human development. Perspectives on Psychological Science, 4(4), 345–351. Belsky, J., & Pluess, M. (2013a). Beyond risk, resilience, and dysregulation: Phenotypic plasticity and human development. Development and Psychopathology, 25(4 Pt 2), 1243–1261. doi: 10.1017/S095457941300059X

97

Belsky, J., & Pluess, M. (2013b). Genetic moderation of early child-care effects on social functioning across childhood: A developmental analysis. Child Development, 84(4), 1209–1225. doi: 10.1111/cdev.12058 Belsky, J., Pluess, M., & Widaman, K. F. (2013). Confirmatory and competitive evaluation of alternative gene–environment interaction hypotheses. Journal of Child Psychology and Psychiatry and Allied Disciplines, 54(10), 1135–1143. doi: 10.1111/jcpp. 12075 Belsky, J., Steinberg, L., & Draper, P. (1991). Childhood experience, interpersonal development, and reproductive strategy: An evolutionary theory of socialization. Child Development, 62(4), 647–670. Belsky, J., Vandell, D. L., Burchinal, M., Clarke-Stewart, K. A., McCartney, K., & Owen, M. T. (2007). Are there long-term effects of early child care? Child Development, 78(2), 681–701. Berman, S. M., & Noble, E. P. (1997). The D2 dopamine receptor (DRD2) gene and family stress: Interactive effects on cognitive functions in children. Behavior Genetics, 27(1), 33–43. Berry, D., Deater-Deckard, K., McCartney, K., Wang, Z., & Petrill, S. A. (2013). Gene–environment interaction between dopamine receptor D4 7-repeat polymorphism and early maternal sensitivity predicts inattention trajectories across middle childhood. Development and Psychopathology, 25(2), 291–306. doi: 10.1017/s095457941200106x Bevilacqua, L., Carli, V., Sarchiapone, M., George, D. K., Goldman, D., Roy, A., & Enoch, M.-A. (2012). Interaction between FKBP5 and childhood trauma and risk of aggressive behavior. Archives of General Psychiatry, 69(1), 62–70. Blair, C. (2002). Early intervention for low birth weight, preterm infants: the role of negative emotionality in the specification of effects. Development and Psychopathology, 14(2), 311–332. Blair, C., & Raver, C. C. (2012). Individual development and evolution: Experiential canalization of self-regulation. Developmental Psychology, 48(3), 647–657. doi: 10.1037/a0026472 Bogdan, R., Williamson, D. E., & Hariri, A. R. (2012). Mineralocorticoid receptor Iso/Val (rs5522) genotype moderates the association between previous childhood emotional neglect and amygdala reactivity. American Journal of Psychiatry, 169(5), 515–522. doi: 10.1176/appi.ajp. 2011.11060855 Bowirrat, A., & Oscar-Berman, M. (2005). Relationship between dopaminergic neurotransmission, alcoholism, and reward deficiency syndrome. American Journal of Medical Genetics B Neuropsychiatric Genetics, 132B(1), 29–37. Boyce, W. T. (1996). Biobehavioral reactivity and injuries in children and adolescents. In M. H. Bornstein & J. Genevro (Eds.), Child development and behavioral pediatrics: Toward understanding children and health (pp. 35–58). Mahwah, NJ: Erlbaum. Boyce, W. T., Chesney, M., Alkon, A., Tschann, J. M., Adams, S., Chesterman, B., . . . Wara, D. (1995). Psychobiologic reactivity to stress and childhood respiratory illnesses: results of two prospective studies. Psychosomatic Medicine, 57(5), 411–422. Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary-developmental theory of the origins and functions of stress reactivity. Development and Psychopathology, 17(2), 271–301. Boyce, W. T., Essex, M. J., Alkon, A., Goldsmith, H. H., Kraemer, H. C., & Kupfer, D. J. (2006). Early father involvement moderates biobehavioral susceptibility to mental health problems in middle childhood. Journal of the American Academy of Child and Adolescent Psychiatry, 45(12), 1510–1520. Bradley, R. H., & Corwyn, R. F. (2008). Infant temperament, parenting, and externalizing behavior in first grade: a test of the differential susceptibility hypothesis. Journal of Child Psychology and Psychiatry, 49(2), 124–131. Brody, G. H., Beach, S. R., Philibert, R. A., Chen, Y. F., & Murry, V. M. (2009). Prevention effects moderate the association of 5-HTTLPR

98

Differential Susceptibility

and youth risk behavior initiation: Gene × environment hypotheses tested via a randomized prevention design. Child Development, 80(3), 645–661. Brody, G. H., Beach, S. R. H., Chen, Y.-F., Obasi, E., Philibert, R. A., Kogan, S. M., & Simons, R. L. (2011). Perceived discrimination, serotonin transporter linked polymorphic region status, and the development of conduct problems. Development and Psychopathology, 23(2), 617–627. doi: 10.1017/s0954579411000046 Brody, G. H., Chen, Y. F., & Beach, S. R. H. (2013). Differential susceptibility to prevention: GABAergic, dopaminergic, and multilocus effects. Journal of Child Psychology and Psychiatry, 54(8), 863–871. doi: 10.1111/Jcpp. 12042 Brody, G. H., Chen, Y. F., Beach, S. R., Kogan, S. M., Yu, T., Diclemente, R. J., . . . Philibert, R. A. (2014). Differential sensitivity to prevention programming: A dopaminergic polymorphism-enhanced prevention effect on protective parenting and adolescent substance use. Health Psychology, 33(2), 182–191. doi: 10.1037/a0031253 Brody, G. H., Yu, T., Chen, Y. F., Kogan, S. M., Evans, G. W., Windle, M., . . . Philibert, R. A. (2012). Supportive family environments, genes that confer sensitivity, and allostatic load among rural African American emerging adults: A prospective analysis. Journal of Family Psychology, 27(1), 22. doi: 10.1037/a0027829 Bronfenbrenner, U. (1979). The ecology of human development. Cambridge, MA: Harvard University Press. Bronfenbrenner, U. (1993). The ecology of cognitive development. In R. Wozniak & K. Fischer (Eds.), Scientific environments (pp. 3–44). Hillsdale, NJ: Erlbaum. Brummett, B. H., Boyle, S. H., Siegler, I. C., Kuhn, C. M., Ashley-Koch, A., Jonassaint, C. R., . . . Williams, R. B. (2008). Effects of environmental stress and gender on associations among symptoms of depression and the serotonin transporter gene linked polymorphic region (5-HTTLPR). Behavior Genetics, 38(1), 34–43. Bryant, R. A., Felmingham, K. L., Falconer, E. M., Benito, L. P., Dobson-Stone, C., Pierce, K. D., & Schofield, P. R. (2010). Preliminary evidence of the short allele of the serotonin transporter gene predicting poor response to cognitive behavior therapy in posttraumatic stress disorder. Biological Psychiatry, 67(12), 1217–1219. doi: 10.1016/j.biopsych.2010.03.016 Bubier, J. L., Drabick, D. A. G., & Breiner, T. (2009). Autonomic functioning moderates the relations between contextual factors and externalizing behaviors among inner-city children. Journal of Family Psychology, 23(4), 500–510. doi: 10.1037/a0015555 Buss, A., & Plomin, R. (1984). Temperament: Early developing personality traits. Hillsdale, NJ: Erlbaum. Buss, K. A. (2011). Which fearful toddlers should we worry about? Context, fear regulation, and anxiety risk. Developmental Psychology, 47(3), 804–819. doi: 10.1037/a0023227 Cameron, N. M., Champagne, F. A., Parent, C., Fish, E. W., Ozaki-Kuroda, K., & Meaney, M. J. (2005). The programming of individual differences in defensive responses and reproductive strategies in the rat through variations in maternal care. Neuroscience and Biobehavioral Reviews, 29(4–5), 843–865. Cameron, N. M., Fish, E. W., & Meaney, M. J. (2008). Maternal influences on the sexual behavior and reproductive success of the female rat. Hormones and Behavior, 54(1), 178–184. Carlson, E. A. (1998). A prospective longitudinal study of attachment disorganization/disorientation. Child Development, 69(4), 1107–1128. Carver, C. S., Johnson, S. L., Joormann, J., Kim, Y., & Nam, J. Y. (2011). Serotonin transporter polymorphism interacts with childhood adversity to predict aspects of impulsivity. Psychological Science, 22(5), 589–595. doi: 10.1177/0956797611404085

Caspi, A., Harrington, H., Milne, B., Amell, J. W., Theodore, R. F., & Moffitt, T. E. (2003a). Children’s behavioral styles at age 3 are linked to their adult personality traits at age 26. Journal of Personality, 71(4), 495–513. Caspi, A., McClay, J., Moffitt, T. E., Mill, J., Martin, J., Craig, I. W., . . . Poulton, R. (2002). Role of genotype in the cycle of violence in maltreated children. Science, 297(5582), 851–854. Caspi, A., & Moffitt, T. E. (2006). Gene–environment interactions in psychiatry: Joining forces with neuroscience. Nature Reviews: Neuroscience, 7(7), 583–590. Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . Poulton, R. (2003b). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301(5631), 386–389. Cassidy, J., Woodhouse, S. S., Sherman, L. J., Stupica, B., & Lejuez, C. W. (2011). Enhancing infant attachment security: An examination of treatment efficacy and differential susceptibility. Development and Psychopathology, 23(1), 131–148. doi: 10.1017/s0954579410000696 Champagne, D. L., Bagot, R. C., van Hasselt, F., Ramakers, G., Meaney, M. J., de Kloet, E. R., . . . Krugers, H. (2008). Maternal care and hippocampal plasticity: Evidence for experience-dependent structural plasticity, altered synaptic functioning, and differential responsiveness to glucocorticoids and stress. Journal of Neuroscience, 28(23), 6037–6045. doi: 10.1523/jneurosci.0526–08.2008 Chen, C., Moyzis, R., Stern, H., He, Q., Li, H., Li, J., . . . Dong, Q. (2011). Contributions of dopamine-related genes and environmental factors to highly sensitive personality: A multi-step neuronal system-level approach. PloS One, 6(7), e21636. doi: 10.1371/journal.pone.0021636 Chen, C. S., Burton, M., Greenberger, E., & Dmitrieva, J. (1999). Population migration and the variation of dopamine D4 receptor (DRD4) allele frequencies around the globe. Evolution and Human Behavior, 20(5), 309–324. doi: 10.1016/s1090–5138(99)00015-x Chen, J., Li, X., & McGue, M. (2012). Interacting effect of BDNF Val66Met polymorphism and stressful life events on adolescent depression. Genes, Brain, and Behavior, 11(8), 958–965. doi: 10.1111/j.1601–183X.2012.00843.x Chisholm, J. S. (1993). Death, hope, and sex—Life-history theory and the development of reproductive strategies. Current Anthropology, 34(1), 1–24. doi: 10.1086/204131 Chisholm, J. S. (1996). The evolutionary ecology of attachment organization. Human Nature—An Interdisciplinary Biosocial Perspective, 7(1), 1–37. Cicchetti, D., & Rogosch, F. A. (2012). Gene × environment interaction and resilience: Effects of child maltreatment and serotonin, corticotropin releasing hormone, dopamine, and oxytocin genes. Development and Psychopathology, 24(2), 411–427. doi: 10.1017/s0954579412000077 Cicchetti, D., Rogosch, F. A., & Oshri, A. (2011). Interactive effects of corticotropin releasing hormone receptor 1, serotonin transporter linked polymorphic region, and child maltreatment on diurnal cortisol regulation and internalizing symptomatology. Development and Psychopathology, 23(4), 1125–1138. doi: 10.1017/s0954579411000599 Cicchetti, D., Rogosch, F. A., & Sturge-Apple, M. L. (2007). Interactions of child maltreatment and serotonin transporter and monoamine oxidase A polymorphisms: Depressive symptomatology among adolescents from low socioeconomic status backgrounds. Development and Psychopathology, 19(4), 1161–1180. Cicchetti, D., Rogosch, F. A., & Thibodeau, E. L. (2012). The effects of child maltreatment on early signs of antisocial behavior: Genetic moderation by tryptophan hydroxylase, serotonin transporter, and monoamine oxidase A genes. Development and Psychopathology, 24(3), 907–928.

References Cicchetti, D., Rogosch, F. A., & Toth, S. L. (2011). The effects of child maltreatment and polymorphisms of the serotonin transporter and dopamine D4 receptor genes on infant attachment and intervention efficacy. Development and Psychopathology, 23(2), 357–372. doi: 10.1017/s0954579411000113 Cicchetti, D., & Valentino, K. (2006). An ecological transactional perspective on child maltreatment: Failure of the average expectable environment and its influence upon child development. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology: Risk, disorder, and adaptation (2nd ed., Vol. 3, pp. 129–201). New York, NY: Wiley. Clasen, P. C., Wells, T. T., Knopik, V. S., McGeary, J. E., & Beevers, C. G. (2011). 5-HTTLPR and BDNF Val66Met polymorphisms moderate effects of stress on rumination. Genes, Brain, and Behavior, 10(7), 740–746. doi: 10.1111/j.1601–183X.2011.00715.x Conner, B. T., Hellemann, G. S., Ritchie, T. L., & Noble, E. P. (2010). Genetic, personality, and environmental predictors of drug use in adolescents. Journal of Substance Abuse Treatment, 38(2), 178–190. doi: 10.1016/j.jsat.2009.07.004 Conway, A., & Stifter, C. A. (2012). Longitudinal antecedents of executive function in preschoolers. Child Development, 83(3), 1022–1036. doi: 10.1111/j.1467–8624.2012.01756.x Das, D., Cherbuin, N., Tan, X., Anstey, K. J., & Easteal, S. (2011). DRD4-exonIII-VNTR moderates the effect of childhood adversities on emotional resilience in young-adults. PloS One, 6(5). doi: 10.1371/journal.pone.0020177 Davies, P. T., Sturge-Apple, M. L., & Cicchetti, D. (2011). Interparental aggression and children’s adrenocortical reactivity: Testing an evolutionary model of allostatic load. Development and Psychopathology, 23(3), 801–814. doi: 10.1017/s0954579411000319 Davis, E. P., Glynn, L. M., Schetter, C. D., Hobel, C., Chicz-Demet, A., & Sandman, C. A. (2007). Prenatal exposure to maternal depression and cortisol influences infant temperament. Journal of the American Academy of Child and Adolescent Psychiatry, 46(6), 737–746. Davis, E. P., Glynn, L. M., Waffarn, F., & Sandman, C. A. (2011). Prenatal maternal stress programs infant stress regulation. Journal of Child Psychology and Psychiatry, 52(2), 119–129. doi: 10.1111 /j.1469–7610.2010.02314.x Dearing, E., & Hamilton, L. C. (2006). Contemporary advances and classic advice for analyzing mediating and moderating variables. Monographs of the Society for Research in Child Development, 71, 88–104. Deater-Deckard, K., Pinkerton, R., & Scarr, S. (1996). Child care quality and children’s behavioral adjustment: A four-year longitudinal study. Journal of Child Psychology and Psychiatry, 37(8), 937–948. Del Giudice, M. (2012). Fetal programming by maternal stress: Insights from a conflict perspective. Psychoneuroendocrinology, 37(10), 1614–1629. doi: 10.1016/j.psyneuen.2012.05.014 DeWitt, T. J., Sih, A., & Wilson, D. S. (1998). Costs and limits of phenotypic plasticity. Trends Ecological Evolution, 13(2), 77–81. doi: 10.1016/s0169–5347(97)01274–3 Ding, Yl, Chi, H., Grady, D. L., Morishima, A., Kidd, J., Kidd, K. K., . . . Moyzis, R. K. (2002). Evidence of positive selection acting at the human dopamine DF gene locus. PNAS, 99, 309–314. Docherty, S. J., Kovas, Y., & Plomin, R. (2011). Gene–environment interaction in the etiology of mathematical ability using SNP sets. Behavior Genetics, 41(1), 141–154. doi: 10.1007/s10519–010–9405–6 Drury, S. S., Gleason, M. M., Theall, K. P., Smyke, A. T., Nelson, C. A., Fox, N. A., & Zeanah, C. H. (2012). Genetic sensitivity to the caregiving context: The influence of 5httlpr and BDNF val66met on indiscriminate social behavior. Physiology and Behavior, 106(5), 728–735. doi: 10.1016/j.physbeh.2011.11.014

99

Du, L., Bakish, D., Lapierre, Y. D., Ravindran, A. V., & Hrdina, P. D. (2000). Association of polymorphism of serotonin 2A receptor gene with suicidal ideation in major depressive disorder. American Journal of Medical Genetics, 96(1), 56–60. Ducci, F., Enoch, M. A., Hodgkinson, C., Xu, K., Catena, M., Robin, R. W., & Goldman, D. (2008). Interaction between a functional MAOA locus and childhood sexual abuse predicts alcoholism and antisocial personality disorder in adult women. Molecular Psychiatry, 13(3), 334–347. Eisenberg, N., Sulik, M. J., Spinrad, T. L., Edwards, A., Eggum, N. D., Liew, J., . . . Hart, D. (2012). Differential susceptibility and the early development of aggression: Interactive effects of respiratory sinus arrhythmia and environmental quality. Developmental Psychology, 48(3), 755–768. doi: 10.1037/a0026518 El-Sheikh, M., Erath, S. A., & Keller, P. S. (2007a). Children’s sleep and adjustment: the moderating role of vagal regulation. Journal of Sleep Research, 16(4), 396–405. El-Sheikh, M., Harger, J., & Whitson, S. M. (2001). Exposure to interparental conflict and children’s adjustment and physical health: the moderating role of vagal tone. Child Development, 72(6), 1617–1636. El-Sheikh, M., Keller, P. S., & Erath, S. A. (2007b). Marital conflict and risk for child maladjustment over time: skin conductance level reactivity as a vulnerability factor. Journal of Abnormal Child Psychology, 35(5), 715–727. Eley, T. C., Hudson, J. L., Creswell, C., Tropeano, M., Lester, K. J., Cooper, P., . . . Collier, D. A. (2012). Therapygenetics: The 5HTTLPR and response to psychological therapy. Molecular Psychiatry, 17, 236–241. doi: 10.1038/mp. 2011.132 Eley, T. C., Sugden, K., Corsico, A., Gregory, A. M., Sham, P., McGuffin, P., . . . Craig, I. W. (2004). Gene–environment interaction analysis of serotonin system markers with adolescent depression. Molecular Psychiatry, 9(10), 908–915. Ellis, B. J. (2004). Timing of pubertal maturation in girls: An integrated life history approach. Psychological Bulletin, 130(6), 920–958. Ellis, B. J., Boyce, W. T., Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to the environment: An evolutionary–neurodevelopmental theory. Development and Psychopathology, 23(1), 7–28. doi: 10.1017/s0954579410000611 Ellis, B. J., Essex, M. J., & Boyce, W. T. (2005). Biological sensitivity to context: II. Empirical explorations of an evolutionary-developmental theory. Development and Psychopathology, 17(2), 303–328. Ellis, B. J., Figueredo, A. J., Brumbach, B. H., & Schlomer, G. L. (2009a). Fundamental dimensions of environmental risk. Human Nature-an Interdisciplinary Biosocial Perspective, 20(2), 204–268. doi: 10.1007/s12110–009–9063–7 Ellis, B. J., Jackson, J. J., & Boyce, W. T. (2006). The stress response systems: Universality and adaptive individual differences. Developmental Review, 26(2), 175–212. Ellis, B. J., Shirtcliff, E. A., Boyce, W. T., Deardorff, J., & Essex, M. J. (2011). Quality of early family relationships and the timing and tempo of puberty: Effects depend on biological sensitivity to context. Development and Psychopathology, 23(1), 85–99. doi: 10.1017/s0954579410000660 Ellis, B. J., Shirtcliff, E. A., Boyce, W. T., & Essex, M. J. (2009b). Biological sensitivity to context: Stress reactivity moderates the effect of parental supportiveness on pubertal maturation. SRCD Presentation. Paper presented at the Biennial Meeting of the Society for Research in Child Development, Denver, Colorado. Elovainio, M., Jokela, M., Kivimaki, M., Pulkki-Raback, L., Lehtimaki, T., Airla, N., & Keltikangas-Jarvinen, L. (2007). Genetic variants in the DRD2 gene moderate the relationship between stressful life events

100

Differential Susceptibility

and depressive symptoms in adults: cardiovascular risk in young Finns study. Psychosomatic Medicine, 69(5), 391–395. Essex, M. J., Armstrong, J. M., Burk, L. R., Goldsmith, H. H., & Boyce, W. T. (2011). Biological sensitivity to context moderates the effects of the early teacher–child relationship on the development of mental health by adolescence. Development and Psychopathology, 23(1), 149–161. doi: 10.1017/s0954579410000702 Faraone, S. V., Doyle, A. E., Mick, E., & Biederman, J. (2001). Meta-analysis of the association between the 7-repeat allele of the dopamine D(4) receptor gene and attention deficit hyperactivity disorder. American Journal of Psychiatry, 158(7), 1052–1057. Feldman, R., Greenbaum, C. W., & Yirmiya, N. (1999). Mother-infant affect synchrony as an antecedent of the emergence of self-control. Developmental Psychology, 35(1), 223–231. Felmingham, K. L., Dobson-Stone, C., Schofield, P. R., Quirk, G. J., & Bryant, R. A. (2013). The brain-derived neurotrophic factor Val66Met polymorphism predicts response to exposure therapy in posttraumatic stress disorder. Biological Psychiatry, 73(11), 1059–1063. doi: 10.1016/j.biopsych.2012.10.033 Fernald, L. C., Burke, H. M., & Gunnar, M. R. (2008). Salivary cortisol levels in children of low-income women with high depressive symptomatology. Development and Psychopathology, 20(2), 423–436. Foley, D. L., Eaves, L. J., Wormley, B., Silberg, J. L., Maes, H. H., Kuhn, J., & Riley, B. (2004). Childhood adversity, monoamine oxidase a genotype, and risk for conduct disorder. Archives of General Psychiatry, 61(7), 738–744. Fortuna, Keren, van IJzendoorn, Marinus H., Mankuta, David, Kaitz, Marsha, Avinun, Reut, Ebstein, Richard P., & Knafo, Ariel. (2011). Differential genetic susceptibility to child risk at birth in predicting observed maternal behavior. PloS One, 6(5). doi: 10.1371/journal .pone.0019765 Fox, E., Zougkou, K., Ridgewell, A., & Garner, K. (2011). The serotonin transporter gene alters sensitivity to attention bias modification: Evidence for a plasticity gene. Biological Psychiatry, 70(11), 1049–1054. doi: 10.1016/j.biopsych.2011.07.004 Frankenhuis, W. E., & Del Giudice, M. (2012). When do adaptive developmental mechanisms yield maladaptive outcomes? Developmental Psychology, 48(3), 628–642. doi: 10.1037/a0025629 Frankenhuis, W. E., & Panchanathan, K. (2011). Individual differences in developmental plasticity may result from stochastic sampling. Perspectives on Psychological Science, 6(4), 336–347. doi: 10.1177/1745691611412602 Frazzetto, G., Di Lorenzo, G., Carola, V., Proietti, L., Sokolowska, E., Siracusano, A., . . . Troisi, A. (2007). Early trauma and increased risk for physical aggression during adulthood: The moderating role of MAOA genotype. PloS One, 2(5), e486. Fredrickson, B. L., & Joiner, T. (2002). Positive emotions trigger upward spirals toward emotional well-being. Psychological Science, 13(2), 172–175. Gannon, L., Banks, J., Shelton, D., & Luchetta, T. (1989). The mediating effects of psychophysiological reactivity and recovery on the relationship between environmental stress and illness. Journal of Psychosomatic Research, 33(2), 167–175. Ganzel, B. L., & Morris, P. A. (2011). Allostasis and the developing human brain: Explicit consideration of implicit models. Development and Psychopathology, 23(4), 955–974. doi: 10.1017/s0954579411000447 Garmezy, N. (1974). The study of competence in children at risk for severe psychopathology. In E. J. Anthony & C. Koupernik (Eds.), The child in his family: Children at psychiatric risk (Vol. 3, pp. 77–97). New York, NY: Wiley.

Garmezy, N., Masten, A. S., & Tellegen, A. (1984). The study of stress and competence in children—A building block for developmental psychopathology. Child Development, 55(1), 97–111. doi: 10.1111/j.1467–8624.1984.tb00276.x Gatt, J. M., Nemeroff, C. B., Dobson-Stone, C., Paul, R. H., Bryant, R. A., Schofield, P. R., . . . Williams, L. M. (2009). Interactions between BDNF Val66Met polymorphism and early life stress predict brain and arousal pathways to syndromal depression and anxiety. Molecular Psychiatry, 14(7), 681–695. Gelernter, J., Cubells, J. F., Kidd, J. R., Pakstis, A. J., & Kidd, K. K. (1999). Population studies of polymorphisms of the serotonin transporter protein gene. American Journal of Medical Genetics, 88(1), 61–66. Gibbons, F. X., Roberts, M. E., Gerrard, M., Li, Z., Beach, S. R. H., Simons, R. L., . . . Philibert, R. A. (2012). The impact of stress on the life history strategies of African American adolescents: Cognitions, genetic moderation, and the role of discrimination. Developmental Psychology, 48(3), 722–739. doi: 10.1037/a0026599 Gilissen, R., Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & van der Veer, R. (2008). Parent–child relationship, temperament, and physiological reactions to fear-inducing film clips: Further evidence for differential susceptibility. Journal of Experimental Child Psychology, 99(3), 182–195. Gluckman, P., & Hanson, M. (2005). The fetal matrix: Evolution, development and disease. Cambridge, UK: Cambridge University Press. Glynn, L. M., & Sandman, C. A. (2011). Prenatal origins of neurological development: A Critical period for fetus and mother. Current Directions in Psychological Science, 20(6), 384–389. doi: 10.1177/0963721411422056 Gotlib, I. H., Joormann, J., Minor, K. L., & Hallmayer, J. (2008). HPA axis reactivity: a mechanism underlying the associations among 5-HTTLPR, stress, and depression. Biological Psychiatry, 63(9), 847–851. Gottesman, I. I., & Gould, T. D. (2003). The endophenotype concept in psychiatry: Etymology and strategic intentions. American Journal of Psychiatry, 160(4), 636–645. Gottesman, I. I., & Shields, J. (1967). A polygenic theory of schizophrenia. Proceedings of the National Academy of Sciences of the United States of America, 58(1), 199–205. Gottesman, I. I. (1974). Developmental genetics and ontogenetic psychology. In A. D. Pick & H. L. Pick, Jr. (Eds.), Minnesota symposium on child psychology (Vol. 8, pp. 55–80). Minneapolis, MN: University of Minnesota Press. Grazioplene, R. G., DeYoung, C. G., Rogosch, F. A., & Cicchetti, D. (2013). A novel differential susceptibility gene: CHRNA4 and moderation of the effect of maltreatment on child personality. Journal of Child Psychology and Psychiatry, 54(8), 872–880. doi:10.1111/Jcpp. 12031 Gunnar, M., & Quevedo, K. (2007). The neurobiology of stress and development. Annual Review of Psychology, 58, 145–173. Gunnar, M. R., Wenner, J. A., Thomas, K. M., Glatt, C. E., McKenna, M. C., & Clark, A. G. (2012). The brain-derived neurotrophic factor Val66Met polymorphism moderates early deprivation effects on attention problems. Development and Psychopathology, 24(4), 1215–1223. doi: 10.1017/s095457941200065x Gunthert, K. C., Conner, T. S., Armeli, S., Tennen, H., Covault, J., & Kranzler, H. R. (2007). Serotonin transporter gene polymorphism (5-HTTLPR) and anxiety reactivity in daily life: A daily process approach to gene–environment interaction. Psychosomatic Medicine, 69(8), 762–768. Hammen, C., Brennan, P. A., Keenan-Miller, D., Hazel, N. A., & Najman, J. M. (2010). Chronic and acute stress, gender, and serotonin transporter gene–environment interactions predicting depression

References symptoms in youth. Journal of Child Psychology and Psychiatry and Allied Disciplines, 51(2), 180–187. Hankin, B. L., Nederhof, E., Oppenheimer, C. W., Jenness, J., Young, J. F., Abela, J. R. Z., . . . Oldehinkel, A. J. (2011). Differential susceptibility in youth: Evidence that 5-HTTLPR × positive parenting is associated with positive affect “for better and worse.” Translational Psychiatry, 1, e44. doi: 10.1038/tp. 2011.44 Hanson, D. R., & Gottesman, I. I. (2012). Biologically flavored perspectives on Garmezian resilience. Development and Psychopathology, 24(2), 363–369. doi: 10.1017/s0954579412000041 Harris, J. R. (1998). The nurture assumption: Why children turn out the way they do. New York, NY: Touchstone. Hartmann, H. (1958). Ego psychology and the problem of adaptation. New York, NY: International Universities Press. Heim, C., Newport, D. J., Heit, S., Graham, Y. P., Wilcox, M., Bonsall, R., . . . Nemeroff, C. B. (2000). Pituitary-adrenal and autonomic responses to stress in women after sexual and physical abuse in childhood. Journal of the American Medical Association, 284(5), 592–597. Hinde, R. A., & Stevenson-Hinde, J. (1990). Attachment—Biological, cultural and individual desiderata. Human Development, 33(1), 62–72. The American Heritage Dictionary of the English Language. (2000). (4th ed.). Boston, MA: Houghton Mifflin Harcourt. Holling, Crawford S. (1973). Resilience and stability of ecological systems. Annual review of ecology and systematics, 4, 1–23. Holmboe, K., Nemoda, Z., Fearon, R. M., Sasvari-Szekely, M., & Johnson, M. H. (2011). Dopamine D4 receptor and serotonin transporter gene effects on the longitudinal development of infant temperament. Genes, Brain, and Behavior, 10(5), 513–522. doi: 10.1111/j.1601–183X.2010.00669.x Huizink, A. C., de Medina, P. G., Mulder, E. J., Visser, G. H., & Buitelaar, J. K. (2002). Psychological measures of prenatal stress as predictors of infant temperament. Journal of the American Academy of Child and Adolescent Psychiatry, 41(9), 1078–1085. Hyde, L. W., Bogdan, R., & Hariri, A. R. (2011). Understanding risk for psychopathology through imaging gene–environment interactions. Trends in Cognitive Sciences, 15(9), 417–427. doi: 10.1016/j.tics.2011.07.001 Infant Health and Development Program. (1990). Enhancing the outcomes of low-birth-weight, premature infants. Journal of the American Medical Association, 263, 3035–3042. Ivorra, J. L., Sanjuan, J., Jover, M., Carot, J. M., Frutos, R., & Molto, M. D. (2010). Gene–environment interaction of child temperament. Journal of Developmental and Behavioral Pediatrics, 31(7), 545–554. doi: 10.1097/DBP.0b013e3181ee4072 Jacobs, R. H., Pine, D. S., Schoeny, M. E., Henry, D. B., Gollan, J. K., Moy, G., . . . Wakschlag, L. S. (2011). Maternal depressive history, teen 5HTTLPR genotype, and the processing of emotional faces: Exploring mechanisms of risk. Behaviour Research and Therapy, 49(1), 80–84. doi: 10.1016/j.brat.2010.10.004 Jessee, A., Mangelsdorf, S. C., Brown, G. L., Schoppe-Sullivan, S. J., Shigeto, A., & Wong, M. S. (2010). Parents’ differential susceptibility to the effects of marital quality on sensitivity across the first year. Infant Behavior & Development, 33(4), 442–452. doi: 10.1016/j.infbeh.2010.04.010 Johansson, A., Bergman, H., Corander, J., Waldman, I. D., Karrani, N., Salo, B., . . . Westberg, L. (2012). Alcohol and aggressive behavior in men—Moderating effects of oxytocin receptor gene (OXTR) polymorphisms. Genes Brain and Behavior, 11(2), 214–221. doi: 10.1111/j.1601–183X.2011.00744.x Jokela, M., Keltikangas-Jarvinen, L., Kivimaki, M., Puttonen, S., Elovainio, M., Rontu, R., & Lehtimaki, T. (2007a). Serotonin receptor 2A gene and the influence of childhood maternal nurturance on

101

adulthood depressive symptoms. Archives of General Psychiatry, 64(3), 356–360. Jokela, M., Lehtimaki, T., & Keltikangas-Jarvinen, L. (2007b). The influence of urban/rural residency on depressive symptoms is moderated by the serotonin receptor 2A gene. American Journal of Medical Genetics. Part B: Neuropsychiatric Genetics, 144B(7), 918–922. Jokela, M., Lehtimaki, T., & Keltikangas-Jarvinen, L. (2007c). The serotonin receptor 2A gene moderates the influence of parental socioeconomic status on adulthood harm avoidance. Behavior Genetics, 37(4), 567–574. Jokela, M., Räikkönen, K., Lehtimäki, T., Rontu, R., & Keltikangas-Järvinen, L. (2007d). Tryptophan hydroxylase 1 gene (TPH1) moderates the influence of social support on depressive symptoms in adults. Journal of Affective Disorders, 100(1–3), 191–197. Jönsson, E. G., Goldman, D., Spurlock, G., Gustavsson, J. P., Nielsen, D. A., Linnoila, M., . . . Sedvall, G. C. (1997). Tryptophan hydroxylase and catechol-O-methyltransferase gene polymorphisms: relationships to monoamine metabolite concentrations in CSF of healthy volunteers. European Archives of Psychiatry and Clinical Neuroscience, 247(6), 297–302. Juhasz, G., Dunham, J. S., McKie, S., Thomas, E., Downey, D., Chase, D., . . . Deakin, J. F. (2011). The CREB1-BDNF-NTRK2 pathway in depression: multiple gene-cognition-environment interactions. Biological Psychiatry, 69(8), 762–771. doi: 10.1016/j.biopsych.2010.11.019 Kaplan, L. A., Evans, L., & Monk, C. (2008). Effects of mothers’ prenatal psychiatric status and postnatal caregiving on infant biobehavioral regulation: can prenatal programming be modified? Early Human Development, 84(4), 249–256. Kegel, C. A. T., Bus, A. G., & van IJzendoorn, M. H. (2011). Differential susceptibility in early literacy instruction through computer games: The role of the dopamine D4 receptor gene (DRD4). Mind, Brain, And Education, 5(2), 71–78. doi: 10.1111/j.1751–228X.2011.01112.x Keltikangas-Jarvinen, L., Elovainio, M., Kivimaki, M., Raitakari, O. T., Viikari, J. S., & Lehtimaki, T. (2007a). Dopamine receptor D2 gene Taq1A (C32806T) polymorphism modifies the relationship between birth weight and educational attainment in adulthood: 21-year follow-up of the Cardiovascular Risk in Young Finns study. Pediatrics, 120(4), 756–761. Keltikangas-Jarvinen, L., Puttonen, S., Kivimaki, M., Elovainio, M., Rontu, R., & Lehtimaki, T. (2007b). Tryptophan hydroxylase 1 gene haplotypes modify the effect of a hostile childhood environment on adulthood harm avoidance. Genes, Brain, and Behaviour, 6(4), 305–313. Kim-Cohen, J., Caspi, A., Taylor, A., Williams, B., Newcombe, R., Craig, I. W., & Moffitt, T. E. (2006). MAOA, maltreatment, and gene–environment interaction predicting children’s mental health: New evidence and a meta-analysis. Molecular Psychiatry, 11(10), 903–913. Kim, J. M., Stewart, R., Kim, S. W., Yang, S. J., Shin, I. S., Kim, Y. H., & Yoon, J. S. (2007). Interactions between life stressors and susceptibility genes (5-HTTLPR and BDNF) on depression in Korean elders. Biological Psychiatry, 62(5), 423–428. Kim, S., & Kochanska, G. (2012). Child temperament moderates effects of parent–child mutuality on self-regulation: A relationship-based path for emotionally negative infants. Child Development, 83(4), 1275–1289. doi: 10.1111/j.1467–8624.2012.01778.x Klucken, T., Wehrum, S., Schweckendiek, J., Merz, C. J., Hennig, J., Vaitl, D., & Stark, R. (2013). The 5-HTTLPR polymorphism is associated with altered hemodynamic responses during appetitive conditioning. Human Brain Mapping, 34(10), 2549–2560. doi: 10.1002/hbm.22085

102

Differential Susceptibility

Kluger, A. N., Siegfried, Z., & Ebstein, R. P. (2002). A meta-analysis of the association between DRD4 polymorphism and novelty seeking. Molecular Psychiatry, 7(7), 712–717.

Lengua, L. J. (2008). Anxiousness, frustration, and effortful control as moderators of the relation between parenting and adjustment in middle-childhood. Social Development, 17(3), 554–577.

Knafo, A., Israel, S., & Ebstein, R. P. (2011). Heritability of children’s prosocial behavior and differential susceptibility to parenting by variation in the dopamine receptor D4 gene. Development and Psychopathology, 23(1), 53–67. doi: 10.1017/s0954579410000647

Leve, L. D., Kim, H. K., & Pears, K. C. (2005). Childhood temperament and family environment as predictors of internalizing and externalizing trajectories from ages 5 to 17. Journal of Abnormal Child Psychology, 33(5), 505–520.

Kochanska, G., Aksan, N., & Joy, M. E. (2007). Children’s fearfulness as a moderator of parenting in early socialization: Two longitudinal studies. Developmental Psychology, 43(1), 222–237.

Luthar, S. S., Cicchetti, D., & Becker, B. (2000). The construct of resilience: a critical evaluation and guidelines for future work. Child Development, 71(3), 543–562.

Kochanska, G., & Kim, S. (2013). Difficult temperament moderates links between maternal responsiveness and children’s compliance and behavior problems in low-income families. Journal of Child Psychology and Psychiatry and Allied Disciplines, 54(3), 323–332. doi: 10.1111/jcpp. 12002

Main, M., & Solomon, J. (1990). Procedures for identifying infants as disorganized/disoriented during the Ainsworth Strange Situation. In M. T. Greenberg, D. Cicchetti, & E. M. Cummings (Eds.), Attachment in the preschool years: Theory, research, and intervention (pp. 121–160). Chicago, IL: University of Chicago Press.

Kochanska, G., Kim, S., Barry, R. A., & Philibert, R. A. (2011). Children’s genotypes interact with maternal responsive care in predicting children’s competence: Diathesis–stress or differential susceptibility? Development and Psychopathology, 23, 605–616. Kochanska, G. (1993). Toward a synthesis of parental socialization and child temperament in early development of conscience. Child Development, 64(2), 325–347. Korte, S. M., Koolhaas, J. M., Wingfield, J. C., & McEwen, B. S. (2005). The Darwinian concept of stress: Benefits of allostasis and costs of allostatic load and the trade-offs in health and disease. Neuroscience and Biobehavioral Reviews, 29(1), 3–38. doi: 10.1016/j.neubiorev.2004.08.009 Kraemer, H. C., Stice, E., Kazdin, A., Offord, D., & Kupfer, D. (2001). How do risk factors work together? Mediators, moderators, and independent, overlapping, and proxy risk factors. American Journal of Psychiatry, 158(6), 848–856. doi: 10.1176/appi.ajp. 158.6.848 Kuepper, Y., Wielpuetz, C., Alexander, N., Mueller, E., Grant, P., & Hennig, J. (2012). 5-HTTLPR S-allele: A genetic plasticity factor regarding the effects of life events on personality? Genes, Brain, and Behavior, 11(6), 643–650. doi: 10.1111/j.1601–183X.2012.00783.x Kuzawa, C. W. (2005). Fetal origins of developmental plasticity: Are fetal cues reliable predictors of future nutritional environments? American Journal of Human Biology, 17(1), 5–21. doi: 10.1002/ajhb.20091 Lalovic, A., & Turecki, G. (2002). Meta-analysis of the association between tryptophan hydroxylase and suicidal behavior. American Journal of Medical Genetics, 114(5), 533–540. Lansford, J. E., Malone, P. S., Castellino, D. R., Dodge, K. A., Pettit, G. S., & Bates, J. E. (2006). Trajectories of internalizing, externalizing, and grades for children who have and have not experienced their parents’ divorce or separation. Journal of Family Psychology, 20(2), 292–301. Laucht, M., Blomeyer, D., Buchmann, A. F., Treutlein, J., Schmidt, M. H., Esser, G., . . . Banaschewski, T. (2012). Catechol-O-methyltransferase Val158Met genotype, parenting practices and adolescent alcohol use: Testing the differential susceptibility hypothesis. Journal of Child Psychology and Psychiatry, 53(4), 351–359. doi: 10.1111/j.1469–7610.2011.02408.x Laurent, H. K., Leve, L. D., Neiderhiser, J. M., Natsuaki, M. N., Shaw, D. S., Fisher, P. A., . . . Reiss, D. (2013). Effects of parental depressive symptoms on child adjustment moderated by hypothalamic pituitary adrenal: Within- and between-family risk. Child Development, 84, 528–542. doi: 10.1111/j.1467–8624.2012.01859.x Lazary, J., Lazary, A., Gonda, X., Benko, A., Molnar, E., Juhasz, G., & Bagdy, G. (2008). New evidence for the association of the serotonin transporter gene (SLC6A4) haplotypes, threatening life events, and depressive phenotype. Biological Psychiatry, 64(6), 498–504.

Manuck, S. B. (2010). The reaction norm in gene × environment interaction. Molecular Psychiatry, 15(9), 881–882. doi: 10.1038/mp. 2009.139 Manuck, S. B. (2011a). Delay discounting covaries with childhood socioeconomic status as a function of genetic variation in the dopamine D4 receptor (DRD4). Paper presented at the Society for Research in Child Development, Montreal, Canada. Manuck, S. B. (2011b, 01.18.2011). [Personal Communication: Email]. Manuck, S. B., Flory, J. D., Ferrell, R. E., & Muldoon, M. F. (2004). Socio-economic status covaries with central nervous system serotonergic responsivity as a function of allelic variation in the serotonin transporter gene-linked polymorphic region. Psychoneuroendocrinology, 29(5), 651–668. Masten, A. S. (2007). Resilience in developing systems: progress and promise as the fourth wave rises. Development and Psychopathology, 19(3), 921–930. doi: 10.1017/s0954579407000442 Masten, A. S., Best, K., & Garmezy, N. (1990). Resilience and development: Contributions from the study of children who overcome adversity. Development and Psychopathology, 2, 425–444. Mata, J., Thompson, R. J., & Gotlib, I. H. (2010). BDNF genotype moderates the relation between physical activity and depressive symptoms. Health Psychology, 29(2), 130–133. doi: 10.1037/a0017261 McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties of detecting interactions and moderator effects. Psychological Bulletin, 114(2), 376–390. Mezulis, A. H., Hyde, J. S., & Abramson, L. Y. (2006). The developmental origins of cognitive vulnerability to depression: Temperament, parenting, and negative life events in childhood as contributors to negative cognitive style. Developmental Psychology, 42(6), 1012–1025. Mills-Koonce, W. R., Propper, C. B., Gariepy, J. L., Blair, C., Garrett-Peters, P., & Cox, M. J. (2007). Bidirectional genetic and environmental influences on mother and child behavior: The family system as the unit of analyses. Development and Psychopathology, 19(4), 1073–1087. Mitchell, C., Notterman, D., Brooks-Gunn, J., Hobcraft, J., Garfinkel, I., Jaeger, K., . . . McLanahan, S. (2011). Role of mother’s genes and environment in postpartum depression. Proceedings of the National Academy of Sciences of the United States of America, 108(20), 8189–8193. doi: 10.1073/pnas.1014129108 Möhler, E., Parzer, P., Brunner, R., Wiebel, A., & Resch, F. (2006). Emotional stress in pregnancy predicts human infant reactivity. Early Human Development, 82(11), 731–737. Monroe, S. M., & Simons, A. D. (1991). Diathesis–stress theories in the context of life stress research: Implications for the depressive disorders. Psychological Bulletin, 110(3), 406–425. Morris, A. S., Silk, J. S., Steinberg, L., Sessa, F. M., Avenevoli, S., & Essex, M. J. (2002). Temperamental vulnerability and negative parenting as

References interacting of child adjustment. Journal of Marriage and Family, 64(2), 461–471. Mulvihill, D. (2005). The health impact of childhood trauma: An interdisciplinary review, 1997–2003. Issues in Comprehensive Pediatric Nursing, 28(2), 115–136. doi: 10.1080/01460860590950890 Nakamura, M., Ueno, S., Sano, A., & Tanabe, H. (2000). The human serotonin transporter gene linked polymorphism (5-HTTLPR) shows ten novel allelic variants. Molecular Psychiatry, 5(1), 32–38. Nederhof, E., & Schmidt, M. V. (2012). Mismatch or cumulative stress: Toward an integrated hypothesis of programming effects. Physiology and Behavior, 106(5), 691–700. doi: 10.1016/j.physbeh .2011.12.008 Nelson, III, C. A., Zeanah, C. H., Fox, N. A., Marshall, P. J., Smyke, A. T., & Guthrie, D. (2007). Cognitive recovery in socially deprived young children: The Bucharest early intervention project. Science, 318(5858), 1937–1940. doi: 10.1126/science.1143921 National Institute of Child Health and Human Development (NICHD). Early Child Care Research Network. (2005). Child care and child development: Results of the NICHD study of early child care and youth development. New York, NY: Guilford Press. Nijmeijer, J. S., Hartman, C. A., Rommelse, N. N. J., Altink, M. E., Buschgens, C. J. M., Fliers, E. A., . . . Hoekstra, P. J. (2010). Perinatal risk factors interacting with catechol O-methyltransferase and the serotonin transporter gene predict ASD symptoms in children with ADHD. Journal of Child Psychology and Psychiatry, 51(11), 1242–1250. doi: 10.1111/j.1469–7610.2010.02277.x Nikolova, Y. S., Bogdan, R., Brigidi, B. D., & Hariri, A. R. (2012). Ventral striatum reactivity to reward and recent life stress interact to predict positive affect. Biological Psychiatry, 72(2), 157–163. doi: 10.1016/j.biopsych.2012.03.014 Nikolova, Y. S., Ferrell, R. E., Manuck, S. B., & Hariri, S. R. (2011). Multilocus genetic profile for dopamine signaling predicts ventral striatum reactivity. Neuropsychopharmacology, 36(9), 1940–1947. doi: 10.1038/npp. 2011.82 Nilsson, K. W., Sjoberg, R. L., Damberg, M., Leppert, J., Ohrvik, J., Alm, P. O., . . . Oreland, L. (2006). Role of monoamine oxidase A genotype and psychosocial factors in male adolescent criminal activity. Biological Psychiatry, 59(2), 121–127. Noble, E. P., Gottschalk, L. A., Fallon, J. H., Ritchie, T. L., & Wu, J. C. (1997). D2 dopamine receptor polymorphism and brain regional glucose metabolism. American Journal of Medical Genetics, 74(2), 162–166. Nussey, D. H., Postma, E., Gienapp, P., & Visser, M. E. (2005). Selection on heritable phenotypic plasticity in a wild bird population. Science, 310(5746), 304–306. O’Connor, T. G., Ben-Shlomo, Y., Heron, J., Golding, J., Adams, D., & Glover, V. (2005). Prenatal anxiety predicts individual differences in cortisol in pre-adolescent children. Biological Psychiatry, 58(3), 211–217. Oberlander, T. F., Weinberg, J., Papsdorf, M., Grunau, R., Misri, S., & Devlin, A. M. (2008). Prenatal exposure to maternal depression, neonatal methylation of human glucocorticoid receptor gene (NR3C1) and infant cortisol stress responses. Epigenetics, 3(2), 97–106. Obradovic, J., Bush, N. R., & Boyce, W. T. (2011). The interactive effect of marital conflict and stress reactivity on externalizing and internalizing symptoms: The role of laboratory stressors. Development and Psychopathology, 23(1), 101–114. doi: 10.1017/s0954579410000672 Obradovic, J., Bush, N. R., Stamperdahl, J., Adler, N. E., & Boyce, W. T. (2010). Biological sensitivity to context: The interactive effects of stress reactivity and family adversity on socio-emotional behavior and school readiness. Child Development, 81(1), 270–289.

103

Palmer, C. T., & Tilley, C. F. (1995). Sexual access to females as a motivation for joining gangs—An evolutionary approach. Journal of Sex Research, 32(3), 213–217. Park, A., Sher, K. J., Todorov, A. A., & Heath, A. C. (2011). Interaction between the DRD4 VNTR polymorphism and proximal and distal environments in alcohol dependence during emerging and young adulthood. Journal of Abnormal Psychology, 120(3), 585–595. doi: 10.1037/a0022648 Penke, L., Denissen, J. J. A., & Miller, G. F. (2007). The evolutionary genetics of personality. European Journal of Personality, 21(5), 549–587. doi: 10.1002/per.629 Pergamin-Hight, L., Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & Bar-Haim, Y. (2012). Variations in the promoter region of the serotonin transporter gene and biased attention for emotional information: A meta-analysis. Biological Psychiatry, 71(4), 373–379. doi: 10.1016/j.biopsych.2011.10.030 Pigliucci, M. (2001). Phenotypic plasticity: Beyond nature and nurture. Baltimore, MD: John Hopkins University Press. Pitzer, M., Jennen-Steinmetz, C., Esser, G., Schmidt, M. H., & Laucht, M. (2011). Differential susceptibility to environmental influences: The role of early temperament and parenting in the development of externalizing problems. Comprehensive Psychiatry, 52(6), 650–658. doi: 10.1016/j.comppsych.2010.10.017 Plomin, R., & Daniels, D. (1987). Why are children in the same family so different from one another? Behavioral and Brain Sciences, 10(1), 1–16. Pluess, M. (2015). Individual Differences in Environmental Sensitivity. Child Development Perspectives. Pluess, M., & Belsky, J. (2009). Differential susceptibility to rearing experience: The case of childcare. Journal of Child Psychology and Psychiatry and Allied Disciplines, 50(4), 396–404. Pluess, M., & Belsky, J. (2010). Differential susceptibility to parenting and quality child care. Developmental Psychology, 46(2), 379–390. Pluess, M., & Belsky, J. (2011). Prenatal programming of postnatal plasticity? Development and Psychopathology, 23(1), 29–38. Pluess, M., & Belsky, J. (2013). Vantage sensitivity: Individual differences in response to positive experiences. Psychological Bulletin, 139(4), 901–916. doi: 10.1037/a0030196 Pluess, M., Belsky, J., Way, B. M., & Taylor, S. E. (2010). 5-HTTLPR moderates effects of life events on neuroticism: Differential susceptibility to environmental influences. Progress in Neuro-Psychopharmacology and Biological Psychiatry, 34, 1070–1074. Pluess, M., & Boniwell, I. (2015). Sensory-Processing Sensitivity predicts treatment response to a school-based depression prevention program: Evidence of Vantage Sensitivity. Personality and Individual Differences, 82(0), 40–45. doi: 10.1016/j.paid.2015.03.011 Pluess, M., Boniwell, I., Hefferon, K., & Tunariu, A. (2015). Evaluation of a school-based resilience-promoting intervention in a high-risk population: An exploratory mixed-methods trial. Pluess, M., Stevens, S., & Belsky, J. (2013). Differential susceptibility: Developmental and evolutionary mechanisms of gene–environment interactions. In M. Legerstee, D. W. Haley, & M. H. Bornstein (Eds.), The infant mind: Origins of the social brain. New York, NY: Guilford. Pluess, M., Velders, F. P., Belsky, J., van IJzendoorn, M. H., BakermansKranenburg, M. J., Jaddoe, V. W., . . . Tiemeier, H. (2011). Serotonin transporter polymorphism moderates effects of prenatal maternal anxiety on infant negative emotionality. Biological Psychiatry, 69(6), 520–525. doi: 10.1016/j.biopsych.2010.10.006 Poehlmann, J., Hane, A., Burnson, C., Maleck, S., Hamburger, E., & Shah, P. E. (2012). Preterm infants who are prone to distress: differential effects of parenting on 36-month behavioral and cognitive

104

Differential Susceptibility

outcomes. Journal of Child Psychology and Psychiatry and Allied Disciplines, 53(10), 1018–1025. doi: 10.1111/j.1469-7610.2012.02564.x Pollak, S. D. (2008). Mechanisms linking early experience and the emergence of emotions: Illustrations from the study of maltreated children. Current Directions in Psychological Science, 17(6), 370–375. doi: 10.1111/j.1467–8721.2008.00608.x Poulin, M. J., Holman, E. A., & Buffone, A. (2012). The neurogenetics of nice: Receptor genes for oxytocin and vasopressin interact with threat to predict prosocial behavior. Psychological Science, 23(5), 446–452. Preacher, K. J., Curran, P. J., & Bauer, D. J. (2006). Computational tools for probing interactions in multiple linear regressions, multilevel modeling, and latent curve analysis. Journal of Educational and Behavioral Statistics, 31(4), 437–448. Propper, C., Mills-Koonce, W. R., Tucker Halpern, C., Hill-Soderlund, A. L., Carbone, M. A., Moore, G. A., . . . Cox, M. (2008). Gene–environment contributions to the development of infant vagal reactivity: The interaction of dopamine and maternal sensitivity. Child Development, 79(5), 1377–1394. Propper, C., Willoughby, M., Halpern, C. T., Carbone, M. A., & Cox, M. (2007). Parenting quality, DRD4, and the prediction of externalizing and internalizing behaviors in early childhood. Developmental Psychobiology, 49(6), 619–632. Quas, J. A., Bauer, A., & Boyce, W. T. (2004). Physiological reactivity, social support, and memory in early childhood. Child Development, 75(3), 797–814. doi: 10.1111/j.1467–8624.2004.00707.x Ramchandani, P. G., van IJzendoorn, M. H., & Bakermans-Kranenburg, M. J. (2010). Differential susceptibility to fathers’ care and involvement: The moderating effect of infant reactivity. Family Science, 1(2), 93–101. doi: 10.1080/19424621003599835 Ramos, M. C., Wright Guerin, D., Gottfried, A. W., Bathurst, K., & Oliver, P. H. (2005). Family conflicts and children’s behavior problems: The moderating role of child temperament. Structural Equation Modeling, 12(2), 278–298. Raver, C. C., Blair, C., & Willoughby, M. (2013). Poverty as a predictor of 4-year-olds’ executive function: new perspectives on models of differential susceptibility. Developmental Psychology, 49(2), 292–304. doi: 10.1037/a0028343 Reiss, D., Neiderhiser, J., Hetherington, E. M., & Plomin, R. (2000). The relationship code: Deciphering genetic and social patterns in adolescent development. Cambridge, MA: Harvard University Press. Retz, W., Freitag, C. M., Retz-Junginger, P., Wenzler, D., Schneider, M., Kissling, C., . . . Rosler, M. (2008). A functional serotonin transporter promoter gene polymorphism increases ADHD symptoms in delinquents: Interaction with adverse childhood environment. Psychiatry Research, 158(2), 123–131. Rhee, S. H., Corley, R. P., Friedman, N. P., Hewitt, J. K., Hink, L. K., Johnson, D. P., . . . Young, S. E. (2012). The etiology of observed negative emotionality from 14 to 24 months. Front Genet, 3, 9. doi: 10.3389/fgene.2012.00009 Robbins, T. W., & Everitt, B. J. (1999). Motivation and reward. In M. J. Zigmond, S. C. Bloom, S. C. Landis, J. L. Roberts, & L. R. Squire (Eds.), Fundamental neuroscience (pp. 1245–1260). San Diego, CA: Academic Press. Roiser, J. P., Rogers, R. D., Cook, L. J., & Sahakian, B. J. (2006). The effect of polymorphism at the serotonin transporter gene on decision-making, memory and executive function in ecstasy users and controls. Psychopharmacology, 188(2), 213–227. doi: 10.1007/s00213–006–0495-z Roisman, G. I., Newman, D. A., Fraley, R. C., Haltigan, J. D., Groh, A. M., & Haydon, K. C. (2012). Distinguishing differential susceptibility from diathesis–stress: Recommendations for evaluating

interaction effects. Development and Psychopathology, 24(2), 389–409. doi: 10.1017/S0954579412000065 Rothbart, M. K., & Bates, J. E. (2006). Temperament. In N. Eisenberg, W. Damon, & R. M. Lerner (Eds.), Handbook of child psychology: Social, emotional, and personality development (6th ed., Vol. 3, pp. 99–166). Hoboken, NJ: John Wiley & Sons Inc. Rowe, D. C. (2000). Environmental and genetic influences on pubertal development: Evolutionary life history traits? In J. L. Rodgers, D. C. Rowe, & W. B. Miller (Eds.), Genetic influences on human fertility and sexuality: Recent empirical and theoretical findings (pp. 147–168). Boston, MA: Kluwer. Rowe, D. C., Vazsonyi, A. T., & Figueredo, A. J. (1997). Mating-effort in adolescence: A conditional or alternative strategy. Personality and Individual Differences, 23, 105–115. Rudolph, K. D., Troop-Gordon, W., & Granger, D. A. (2011). Individual differences in biological stress responses moderate the contribution of early peer victimization to subsequent depressive symptoms. Psychopharmacology, 214(1), 209–219. doi: 10.1007/s00213–010– 1879–7 Rutter, M. (1979). Protective factors in children’s responses to stress and disadvantage. In M. W. Kent & J. E. Rolf (Eds.), Primary prevention of psychopathology (Vol. 3, pp. 49–74). Hanover, NH: University Press of New England. Rutter, M. (1987). Psychosocial resilience and protective mechanisms. American Journal of Orthopsychiatry, 57(3), 316–331. Rutter, M. (2012). Resilience as a dynamic concept. Development and Psychopathology, 24(2), 335–344. doi: 10.1017/s0954579412000028 Rutter, M., & English Romanian Adoptees, E. R. A. Study Team. (1998). Developmental catch-up, and deficit, following adoption after severe global early privation. Journal of Child Psychology and Psychiatry and Allied Disciplines, 39(4), 465–476. doi: 10.1017/s0021963098002236 Saxbe, D. E., Margolin, G., Shapiro, L. A. S., & Baucom, B. R. (2012). Does dampened physiological reactivity protect youth in aggressive family environments? Child Development, 83(3), 821–830. doi: 10.1111/j.1467–8624.2012.01752.x Scheper-Hughes, N. (1992). Death without weeping: The violence of everyday life in Brazil. Berkeley, CA: University of California Press. Schlichting, C. D., & Pigliucci, M. (1998). Phenotypic evolution: A reaction norm perspective. Sunderland, MA: Sinauer. Schlomer, G. L., Del Giudice, M., & Ellis, B. J. (2011). Parent–offspring conflict theory: An evolutionary framework for understanding conflict within human families. Psychological Review, 118(3), 496–521. doi: 10.1037/a0024043 Schoebi, D., Way, B. M., Karney, B. R., & Bradbury, T. N. (2012). Genetic moderation of sensitivity to positive and negative affect in marriage. Emotion, 12(2), 208–212. doi: 10.1037/a0026067 Schudlich, T. D. D. R., White, C. R., Fleischhauer, E. A., & Fitzgerald, K. A. (2011). Observed infant reactions during live interparental conflict. Journal of Marriage and Family, 73(1), 221–235. doi: 10.1111/j.1741–3737.2010.00800.x Scott, S., & O’Connor, T. G. (2012). An experimental test of differential susceptibility to parenting among emotionally-dysregulated children in a randomized controlled trial for oppositional behavior. Journal of Child Psychology and Psychiatry and Allied Disciplines, 53(11), 1184–1193. doi: 10.1111/j.1469-7610.2012.02586.x Seeger, G., Schloss, P., Schmidt, M. H., Ruter-Jungfleisch, A., & Henn, F. A. (2004). Gene–environment interaction in hyperkinetic conduct disorder (HD + CD) as indicated by season of birth variations in dopamine receptor (DRD4) gene polymorphism. Neuroscience Letters, 366(3), 282–286. Sergerie, K., Chochol, C., & Armony, J. L. (2008). The role of the amygdala in emotional processing: a quantitative meta-analysis of

References functional neuroimaging studies. Neuroscience and Biobehavioral Reviews, 32(4), 811–830. Sheese, B. E., Voelker, P. M., Rothbart, M. K., & Posner, M. I. (2007). Parenting quality interacts with genetic variation in dopamine receptor D4 to influence temperament in early childhood. Development and Psychopathology, 19, 1039–1046. Shoham, V., & Insel, T. R. (2011). Rebooting for whom? Portfolios, technology, and personalized intervention. Perspectives on Psychological Science, 6(5), 478–482. doi: 10.1177/1745691611418526 Shonkoff, J. P., Boyce, W. T., & McEwen, B. S. (2009). Neuroscience, molecular biology, and the childhood roots of health disparities building a new framework for health promotion and disease prevention. Journal of the American Medical Association, 301(21), 2252–2259. Sih, A. (2011). Effects of early stress on behavioral syndromes: An integrated adaptive perspective. Neuroscience and Biobehavioral Reviews, 35(7), 1452–1465. doi: 10.1016/j.neubiorev.2011.03.015 Simons, R. L., Lei, M. K., Beach, S. R., Brody, G. H., Philibert, R. A., & Gibbons, F. X. (2011). Social environmental variation, plasticity genes, and aggression: Evidence for the differential susceptibility hypothesis. American Sociological Review, 76(6), 833–912. doi: 10.1177/0003122411427580 Simons, R. L., Lei, M.-K., Stewart, E. A., Beach, S. R. H., Brody, G. H., Philibert, R. A., & Gibbons, F. X. (2012). Social adversity, genetic variation, street code, and aggression: A genetically informed model of violent behavior. Youth Violence and Juvenile Justice, 10(1), 3–24. doi: 10.1177/1541204011422087 Simons, R. L., Simons, L. G., Lei, M. K., Beach, S. R. H., Brody, G. H., Gibbons, F. X., & Philibert, R. A. (2013). Genetic moderation of the impact of parenting on hostility toward romantic partners. Journal of Marriage and Family, 75, 325–341. Sonuga-Barke, E. J., Oades, R. D., Psychogiou, L., Chen, W., Franke, B., Buitelaar, J., . . . Faraone, S. V. (2009). Dopamine and serotonin transporter genotypes moderate sensitivity to maternal expressed emotion: the case of conduct and emotional problems in attention deficit/hyperactivity disorder. Journal of Child Psychology and Psychiatry and Allied Disciplines, 50(9), 1052–1063. doi: 10.1111/j.1469-7610.2009.02095.x Stein, M. B., Schork, N. J., & Gelernter, J. (2008). Gene-by-environment (serotonin transporter and childhood maltreatment) interaction for anxiety sensitivity, an intermediate phenotype for anxiety disorders. Neuropsychopharmacology, 33(2), 312–319. Stright, A.D., Gallagher, K.C., & Kelley, K. (2008). Infant temperament moderates relations between maternal parenting in early childhood and children’s adjustment in first grade. Child Development, 79(1), 186–200. Sturge-Apple, M. L., Cicchetti, D., Davies, P. T., & Suor, J. H. (2012). Differential susceptibility in spillover between interparental conflict and maternal parenting practices: Evidence for OXTR and 5-HTT genes. Journal of Family Psychology, 26(3), 431–442. doi: 10.1037/a0028302 Sturge-Apple, M. L., Davies, P. T., Martin, M. J., Cicchetti, D., & Hentges, R. F. (2012). An examination of the impact of harsh parenting contexts on children’s adaptation within an evolutionary framework. Developmental Psychology, 48(3), 791–805. doi: 10.1037/a0026908 Suhara, T., Yasuno, F., Sudo, Y., Yamamoto, M., Inoue, M., Okubo, Y., & Suzuki, K. (2001). Dopamine D2 receptors in the insular cortex and the personality trait of novelty seeking. Neuroimage, 13(5), 891–895. Sulik, M. J., Eisenberg, N., Lemery-Chalfant, K., Spinrad, T. L., Silva, K. M., Eggum, N. D., . . . Verrelli, B. C. (2012). Interactions between serotonin transporter gene haplotypes and quality of mothers’ parenting predict the development of children’s noncompliance. Developmental Psychology, 48, 740–754. doi: 10.1037/a0025938

105

Suomi, S. J. (2006). Risk, resilience, and gene × environment interactions in rhesus monkeys. Annals of the New York Academy of Sciences, 1094, 52–62. Surtees, P. G., Wainwright, N. W., Willis-Owen, S. A., Luben, R., Day, N. E., & Flint, J. (2006). Social adversity, the serotonin transporter (5-HTTLPR) polymorphism and major depressive disorder. Biological Psychiatry, 59(3), 224–229. Suzuki, A., Matsumoto, Y., Shibuya, N., Sadahiro, R., Kamata, M., Goto, K., & Otani, K. (2011). The brain-derived neurotrophic factor Val66Met polymorphism modulates the effects of parental rearing on personality traits in healthy subjects. Genes Brain and Behavior, 10(4), 385–391. doi: 10.1111/j.1601–183X.2010.00673.x Sweitzer, M. M., Halder, I., Flory, J. D., Craig, A. E., Gianaros, P. J., Ferrell, R. E., & Manuck, S. B. (2013). Polymorphic variation in the dopamine D4 receptor predicts delay discounting as a function of childhood socioeconomic status: evidence for differential susceptibility. Social Cognitive and Affective Neuroscience, 8(5), 499–508. doi: 10.1093/scan/nss020 Takane, Y., & Cramer, E. M. (1975). Regions of significance in multiple-regression analysis. Multivariate Behavioral Research, 10(3), 373–383. doi: 10.1207/s15327906mbr1003_10 Taylor, S. E., Way, B. M., Welch, W. T., Hilmert, C. J., Lehman, B. J., & Eisenberger, N. I. (2006). Early family environment, current adversity, the serotonin transporter promoter polymorphism, and depressive symptomatology. Biological Psychiatry, 60(7), 671–676. Thompson, J., Thomas, N., Singleton, A., Piggott, M., Lloyd, S., Perry, E. K., . . . Court, J. A. (1997). D2 dopamine receptor gene (DRD2) Taq1 A polymorphism: reduced dopamine D2 receptor binding in the human striatum associated with the A1 allele. Pharmacogenetics, 7(6), 479–484. Trivers, R. L. (1974). Parent–offspring conflict. American Zoologist, 14(1), 249–264. Turkheimer, E., & Waldron, M. (2000). Nonshared environment: A theoretical, methodological, and quantitative review. Psychological Bulletin, 126(1), 78–108. doi: 10.1037//0033–2909.126.1.78 Uher, R., & McGuffin, P. (2008). The moderation by the serotonin transporter gene of environmental adversity in the aetiology of mental illness: Review and methodological analysis. Molecular Psychiatry, 13(2), 131–146. Ursache, A., Blair, C., Stifter, C., & Voegtline, K. (2013). Emotional reactivity and regulation in infancy interact to predict executive functioning in early childhood. Developmental Psychology, 49(1), 127–137. doi: 10.1037/a0027728 van Aken, C., Junger, M., Verhoeven, M., van Aken, M. A. G., & Dekovic, M. (2007). The interactive effects of temperament and maternal parenting on toddlers’ externalizing behaviours. Infant and Child Development, 16(5), 553–572. van de Wiel, N. M., van Goozen, S. H., Matthys, W., Snoek, H., & van Engeland, H. (2004). Cortisol and treatment effect in children with disruptive behavior disorders: A preliminary study. Journal of the American Academy of Child and Adolescent Psychiatry, 43(8), 1011–1018. van der Kooy-Hofland, V. A. C., van der Kooy, J., Bus, A. G., van IJzendoorn, M. H., & Bonsel, G. J. (2012). Differential susceptibility to early literacy intervention in children with mild perinatal adversities: Shortand long-term effects of a randomized control trial. Journal of Educational Psychology, 104(2), 337–349. doi: 10.1037/a0026984 van IJzendoorn, M. H., & Bakermans-Kranenburg, M. J. (2006). DRD4 7-repeat polymorphism moderates the association between maternal unresolved loss or trauma and infant disorganization. Attachment and Human Develeopment, 8(4), 291–307. van IJzendoorn, M. H., & Bakermans-Kranenburg, M. J. (2012). Differential susceptibility experiments: Going beyond correlational

106

Differential Susceptibility

van IJzendoorn, M. H., Belsky, J., & Bakermans-Kranenburg, M. J. (2012). Serotonin transporter genotype 5HTTLPR as a marker of differential susceptibility? A meta-analysis of child and adolescent gene-by-environment studies. Translational Psychiatry, 2, e147. doi: 10.1038/tp. 2012.73

Whittle, S., Yap, M. B. H., Sheeber, L., Dudgeon, P., Yuecel, M., Pantelis, C., . . . Allen, N. B. (2011). Hippocampal volume and sensitivity to maternal aggressive behavior: A prospective study of adolescent depressive symptoms. Development and Psychopathology, 23(1), 115–129. doi: 10.1017/s0954579410000684 Wickrama, K. A. S., & O’Neal, C. W. (2013). Family of origin, race/ethnicity and socioeconomic attainment: Genotype and intraindividual processes. Journal of Marriage and the Family, 75, 75–90.

van IJzendoorn, M. H., Luijk, M. P. C. M., & Juffer, F. (2008). IQ of children growing up in children’s homes—A meta-analysis on IQ delays in orphanages. Merrill-Palmer Quarterly-Journal of Developmental Psychology, 54(3), 341–366.

Widaman, K. F., Helm, J. L., Castro-Schilo, L., Pluess, M., Stallings, M. C., & Belsky, J. (2012). Distinguishing ordinal and disordinal interactions. Psychological Methods, 17(4), 615–622. doi: 10.1037/a0030003

Verschoor, E., & Markus, C. R. (2011). Affective and neuroendocrine stress reactivity to an academic examination: Influence of the 5-HTTLPR genotype and trait neuroticism. Biological Psychology, 87(3), 439–449. doi: 10.1016/j.biopsycho.2011.06.001

Widom, C. S., & Brzustowicz, L. M. (2006). MAOA and the “cycle of violence:” childhood abuse and neglect, MAOA genotype, and risk for violent and antisocial behavior. Biological Psychiatry, 60(7), 684–689. Wilhelm, K., Mitchell, P. B., Niven, H., Finch, A., Wedgwood, L., Scimone, A., . . . Schofield, P. R. (2006). Life events, first depression onset and the serotonin transporter gene. British Journal of Psychiatry, 188, 210–215. Williams, R. B., Marchuk, D. A., Gadde, K. M., Barefoot, J. C., Grichnik, K., Helms, M. J., . . . Siegler, I. C. (2003). Serotonin-related gene polymorphisms and central nervous system serotonin function. Neuropsychopharmacology, 28(3), 533–541. doi: 10.1038/sj.npp. 1300054

evidence: Comment on beyond mental health, differential susceptibility articles. Developmental Psychology, 48(3), 769–774. doi: 10.1037/a0027536

Vinberg, M., Trajkovska, V., Bennike, B., Knorr, U., Knudsen, G. M., & Kessing, L. V. (2009). The BDNF Val66Met polymorphism: Relation to familiar risk of affective disorder, BDNF levels and salivary cortisol. Psychoneuroendocrinology, 34(9), 1380–1389. Volling, B. L., & Feagans, L. V. (1995). Infant day care and children’s social competence. Infant Behavior & Development, 18(2), 177–188. Wachs, T. D., & Plomin, R. (1991). Conceptualization and measurement of organism-environment interaction synopsis. Washington, DC: American Psychological Association. Wachs, T. D., & Gandour, M. J. (1983). Temperament, environment, and six-month cognitive-intellectual development: A test of the organismic specificity hypothesis. International Journal of Behavioral Development, 6(2), 135–152.

Xie, P., Kranzler, H. R., Poling, J., Stein, M. B., Anton, R. F., Farrer, L. A., & Gelernter, J. (2010). Interaction of FKBP5 with childhood adversity on risk for post-traumatic stress disorder. Neuropsychopharmacology, 35(8), 1684–1692. doi: 10.1038/npp. 2010.37

Werner, E. E., & Smith, R. (1982). Vulnerable but invincible: A study of resilient children. New York, NY: McGraw-Hill.

Yaman, A., Mesman, J., van IJzendoorn, M. H., & BakermansKranenburg, M. J. (2010). Parenting and toddler aggression in second-generation immigrant families: The moderating role of child temperament. Journal of Family Psychology, 24(2), 208–211. doi: 10.1037/a0019100 Yap, M. B., Whittle, S., Yucel, M., Sheeber, L., Pantelis, C., Simmons, J. G., & Allen, N. B. (2008). Interaction of parenting experiences and brain structure in the prediction of depressive symptoms in adolescents. Archives of General Psychiatry, 65(12), 1377–1385. doi: 10.1001/archpsyc.65.12.1377 Zalsman, G., Huang, Y. Y., Oquendo, M. A., Burke, A. K., Hu, X. Z., Brent, D. A., . . . Mann, J. J. (2006). Association of a triallelic serotonin transporter gene promoter region (5-HTTLPR) polymorphism with stressful life events and severity of depression. American Journal of Psychiatry, 163(9), 1588–1593. Zimmermann, P., Brueckl, T., Nocon, A., Pfister, H., Binder, E. B., Uhr, M., . . . Ising, M. (2011). Interaction of FKBP5 gene variants and adverse life events in predicting depression onset: Results from a 10-year prospective community study. American Journal of Psychiatry, 168(10), 1107–1116. doi: 10.1176/appi.ajp. 2011.10111577

West-Eberhard, M. J. (2003). Developmental plasticity and evolution. New York, NY: Oxford University Press.

Zuckerman, M. (1999). Vulnerability to psychopathology: A biosocial model. Washington, DC: American Psychological Association.

Waddington, C. H. (1942). Canalization of development and the inheritance of acquired characteristics. Nature, 150, 563–565. Walker, B., Hollin, C. S., Carpenter, S. R., & Kinzig, A. (2004). Resilience, adaptability and transformability in social-ecological systems. Ecology and Society, 9(2). Warren, S. L., & Simmens, S. J. (2005). Predicting toddler anxiety/ depressive symptoms: Effects of caregiver sensitivity of temperamentally vulnerable children. Infant Mental Health Journal, 26(1), 40–55. Way, B. M., & Taylor, S. E. (2010). The serotonin transporter promoter polymorphism is associated with cortisol response to psychosocial stress. Biological Psychiatry, 67(5), 487–492. Werner, E. E., & Smith, R. (1977). Kauai’s children come of age. Honolulu, HI: University of Hawaii Press.

CHAPTER 3

Differential Sensitivity to Context: Implications for Developmental Psychopathology NICOLE R. BUSH and W. THOMAS BOYCE

INTRODUCTION 107 A History of the Nature–Nurture Culture Wars 107 Early Evidence for Biology–Context Interactions 108 SENSITIVITY TO CONTEXT WITHIN A DEVELOPMENTAL PSYCHOPATHOLOGY FRAMEWORK 110 NEUROBIOLOGICAL SENSITIVITY TO CONTEXT THEORIES AND EVIDENCE 110 Implications for Conditional Adaptation 113 Evidence of Differential Susceptibility Within Positive Environments 115 Cumulative Sensitivity to Environment? 117 CONCEPTUAL AND METHODOLOGICAL ISSUES FOR EXAMINATIONS OF DIFFERENTIAL NEUROBIOLOGICAL SUSCEPTIBILITY 119 Impediments to Discovery 119

GxE Debate 120 Is Reactivity or Susceptibility Maladaptive? 121 Evolutionary Thinking About Variation in Sensitivity 122 Developmental Timing and DNS 122 BIOLOGICAL PATHWAYS LINKING EARLY LIFE DIFFERENTIAL SUSCEPTIBILITY TO LATER PSYCHOPATHOLOGY 124 A Closer Consideration of Epigenetic Pathways 125 FUTURE DIRECTIONS 127 CONCLUSIONS 130 REFERENCES 130

INTRODUCTION

have occurred. And now, as we write this chapter, trainees are being steeped in a more-evolved formulation, representing yet another qualitative shift in understanding. While quickly paced paradigm shifts are somewhat commonplace in science, these nature–nurture debate shifts are particularly striking when one considers the centuries that passed with little progress in this argument before these more recent, rapid advancements. Students of philosophy will note that the nature versus nurture debate is centuries old. The debate rose to popular public awareness in the seventeenth century, when French philosopher René Descartes proposed that individual humans each possess certain innate ideas that underpin our approach to the world—a stance that was later strongly contested by British empiricists, exemplified by John Locke’s claim that humans were blank slates upon which experience marks its influence, fully shaping behavior. In 1874, Francis Galton, a second-cousin to Charles Darwin, presented a more integrated view of the influence of genetics and environment on a person’s development, stating: “[Nature and nurture are] a convenient jingle of words, for it separates under two distinct heads the innumerable

A History of the Nature–Nurture Culture Wars Modern technology is advancing and refining scientific knowledge at an unprecedented pace. In the three decades between the graduate academic training of this chapter’s authors, significant shifts in the nature vs. nurture debate

Preparation of this chapter was supported in part by funds provided by the UCSF Divisions of Child and Adolescent Psychiatry and Developmental Medicine to Dr. Bush. Dr. Bush is an Assistant Professor of Pediatrics and Psychiatry and the Associate Director of Research for the Division of Developmental-Behavioral Pediatrics at UCSF. Dr. Boyce is the Lisa and John Pritzker Distinguished Professor of Developmental and Behavioral Health in the Departments of Pediatrics and Psychiatry at UCSF; he is also a Senior Fellow and Codirector of the Child and Brain Development Program at the Canadian Institute for Advanced Research (CIFAR). We thank the Robert Wood Johnson Foundation Health and Society Scholars Program, the CIFAR, the John and Lisa Pritzker Family Foundation, and the National Institutes of Health for supporting our work. 107

108

Differential Sensitivity to Context: Implications for Developmental Psychopathology

elements of which personality is composed. Nature is all that a man brings with himself into the world; nurture is every influence that affects him after his birth” (Galton, 1874, p. 12). Although this latter approach considered both nature and nurture when describing development, both were presented as separate, unique entities with different time courses of effect. This type of dual conceptualization was carried forward into the study of disease over the next century. Few trains of thought were more integrative than the predominant dual conceptualization allowed, though historians note that Freud was one of the first to integrate contrasting conceptualizations of psychiatric disease as both hereditary and environmentally acquired through childhood experiences, which is one likely reason for the endurance of his thinking (Sulloway, 1992). However, by and large, the leading investigators and scholars within the field of developmental psychopathology were academically reared within a scientific generation marked by a confluence of two distinctive views of the origins of human disorders. Within this single generation, physicians, clinical and developmental psychologists, social workers, and laboratory investigators were steeped in the twin agendas of environmental and biological determinism. In the environmental determinism view, prominent in the scientific world of the 1960s and 1970s, disease and disorder were held to be products of contextual exposures and adversities. Human afflictions, it was believed, were due almost exclusively to the acute and chronic, cumulative influences of environmental agents of disease. Such agents included psychological stressors, impoverished living conditions, physical toxins, infectious pathogens, and inadequate or malevolent parenting. In the field of psychopathology, disease formulations were particularly influenced by behaviorist theories of B. F. Skinner and John Watson, who famously claimed, “Give me a dozen healthy infants . . . and my own specified world to bring them up in and I’ll guarantee to take any one at random and train him to become any type of specialist I might select—doctor, lawyer, artist, merchant-chief and yes, even beggar-man and thief, regardless of his talents, penchants, tendencies, abilities, vocations, and race of his ancestors (Watson, 1930 p. 82).” Within this nurture framework of behavior and disease, schizophrenia was misleadingly viewed as the product of psychological double-binds within dysfunctional family units, autism was regarded as the legacy of cold, distant mothers, and maternal overprotectiveness figured prominently in the presumed etiology of childhood anxiety disorders. Thus, prevention and treatment were taken to require alterations in the range of causative environmental exposures. Within a second canonical view, emerging at the pinnacle of the first by virtue of a revolution in molecular

biology, human disorder was thought principally to be the result of biological frailties built into the structure of the heritable genome. Individual genetic differences, occurring on average in one of every thousand nucleotide base pairs, became seen as likely biological substrates for many human disorders, and quantitative trait loci (QTLs) were thought to constitute the multiple gene systems that, in isolation, code for variation in complex behavioral traits (Plomin & Crabbe, 2000). Most if not all disease, it was posited by some, would one day become explicable within the frameworks of evolutionary biology and human genetics, and eagerly anticipated gene therapies would transform the treatment and prevention of disordered biology. Thus, Alzheimer disease would one day be accounted for by mutations in the gene coding for apolipoprotein E4; a polymorphism in the promoter region of the serotonin transporter gene would eventually elucidate anxiety disorders and phobias; and the long repeat allele of the DRD4 dopamine receptor gene would explain the etiology of ADHD and other externalizing behavior problems. Within 3 decades’ time, two opposing and mutually exclusive causal orthodoxies captured and held the high ground of scientific discourse on the genesis of human disease. Early Evidence for Biology–Context Interactions Vulnerabilities in both positions first became evident as three bodies of research findings emerged in the final years of the twentieth century. First, the nascent field of behavior genetics increasingly documented shared genetic and environmental accounts for variance in the incidence and severity of a broad array of behavioral and psychopathological disorders (Plomin, DeFries, McClearn, & McGuffin, 2001). Using heritability statistics derived from genetically informative research designs, behavioral geneticists became able to decompose phenotypic variance into genetic and environmental components. The genetic component was further subdivided into additive and nonadditive effects, and the environmental component was parsed into shared and nonshared influences (Goldsmith, Gottesman, & Lemery, 1997). While the neuroscience of schizophrenia provided compelling evidence for a heritable disorder of the brain, epidemiological studies revealed that, even among monozygotic twins, the concordance rate for a diagnosis of schizophrenia never exceeded 50%. Schizophrenia was demonstrably a biological disorder with heritable components, but as much as half of the variance in its rate of occurrence was attributable to environmental exposures. The science of behavior and development thus faced a Kierkegaardian dilemma (Kierkegaarad, 1986 (1843)), in which an uncompromising either–or became a more illuminating both–and. Schizophrenia, like

Introduction

virtually all forms of human behavior (Rutter et al., 1997), became understood as a product of both biological and contextual etiologies. A second set of findings revealed the previously unsuspected bidirectional influences of biology and context, each on the other. While it had become apparent that genes affected both normative and disordered human behavior, new evidence indicated: (1) the heritability of environmental experience, and (2) the potentially profound regulatory effects of social environmental exposures on the transcription of DNA. Discoveries of such bidirectional effects led to a recognition of genotype-environment correlations and to the categorization of such correlations into passive, reactive, and active forms (Plomin, DeFries, & Loehlin, 1977). Studies by Plomin and colleagues demonstrated heritable influences on parental behavior and the home environment (Braungart, Fulker, & Plomin, 1992; O’Connor, Hetherington, Reiss, & Plomin, 1995), and other work showed that even random misfortune, such as stressful events (Kendler, Neale, Kessler, Heath, & Eaves, 1993) and the loss of friends (Bergeman, Plomin, Pederson, McClearn, & Nesselroade, 1990), were partly attributable to genetic variation. Animal research by Meaney and colleagues demonstrated that brief maternal separations among rat pups alter maternal behavior, leading to down-regulation of the CRH system and diminished adrenocortical reactivity extending even into adult life (Anisman, Zaharia, Meaney, & Merali, 1998). Other studies in mice found that disruption of social hierarchies led to viral infection-related mortality, which was attributable to an over-expression of genes for cytokine proteins (Sheridan, Stark, Avitsur, & Padgett, 2000), and that caloric restriction prevented aging-related alterations in the expression of genes governing protein metabolism (Lee, Klopp, Weindruch, & Prolla, 1999). While both environmental and biological determinism assumed impenetrable divisions between contexts and genes, new evidence revealed a capacity for the genome to influence environmental experience and for social and physical environments to switch on and off the decoding of genetic material (Cacioppo, Berntson, Sheridan, & McClintock, 2000). Social experience became a heritable predisposition, and the transcription of genes became a process governable by the character of social experience. The third source of vulnerability in the twin dogmas of biological and environmental determinism at the turn of this century was the accelerating emergence of evidence for biology–context interactions in studies of disease etiology. Much of the work examining biological and contextual contributions to pathogenesis was initially preoccupied with partitioning variance into the two distinctive categories of causal factors. Other studies began to illuminate how the interplay between biology and context serves a

109

central etiologic role. Mednick, Gabrielli, and Hutchings (1987), for example, found that a child’s rearing by an adoptive parent with a criminal conviction did not increase the likelihood of criminality in the child, unless one of the child’s biological parents had also been convicted of a crime. Similarly, a polymorphism in the promoter region of the gene encoding corticotrophin releasing hormone (CRH) was shown to moderate the association between maternal stress and preterm birth (Wang et al., 2001). Self-perceived maternal stress was associated overall with pregnancies that were an average of 1.3 weeks shorter, but when a mother possessed a variant allele of the CRH gene, perceived stress was associated with a 2.9-week shortening. Another example was the finding that a polymorphism in the monoamine oxidase A (MAOA) gene promoter moderated the association between early life trauma and increased risk for antisocial behavior, such that the increased risk for individuals abused as children was higher for those carrying the genotype that confers lower levels of MAOA expression but those carrying the genotype that confers higher levels of MAOA expression were somewhat buffered from this maladaptive effect (Caspi et al., 2002). Even in so-called monogenic diseases, such as phenylketonuria, phenotypic expression of the disease depends upon dietary exposure to phenylalanine, suggesting environmental factors influence disease severity and course (Scriver & Waters, 1999). Common to each of these examples is a biogenetic vulnerability that amplifies the pathological effects of a social or physical environmental exposure. Taken together, those emerging lines of evidence constituted a convincing empirical threat to a monodeterministic view of pathogenesis, and a race to integrate biology and context into a coherent model ensued. Accumulating evidence from the past few decades has made real the current position that, while human morbidities will undoubtedly be identified that have their origins in solely genetic or environmental causes, variation in the onset and course of most human disorders will be most likely ultimately explained by interactions among biological and environmental forces (Rutter, 2006; Weatherall, 2001). As a field, we have moved past the simple partitioning of developmental variance into genetic and environmentally determined components and into the examination of more complex and potentially rewarding questions of genetic change and continuity, relations among dimensional symptoms and diagnosable psychopathology, and the interplay between biology and context (Meaney, 2010; Plomin & Crabbe, 2000). This chapter will outline some of the theories and empirical evidence for biology and context interactions in the etiology of healthy and pathological psychosocial development, with an eye toward highlighting

110

Differential Sensitivity to Context: Implications for Developmental Psychopathology

the conceptual and methodological controversies in this work and translational implications. We also describe the challenges awaiting a new generation of investigators and theorists poised to bring the nature–nurture discussion into its next evolutionary form. SENSITIVITY TO CONTEXT WITHIN A DEVELOPMENTAL PSYCHOPATHOLOGY FRAMEWORK Within the field of developmental psychopathology, as well as in the broader fields of medicine and biology, there is a strong focus on discovery of the mechanisms by which environmental forces get under the skin to affect health. Using the term biological embedding (Hertzman & Wiens, 1996) to describe the process whereby differential human experiences systematically affect health across the life cycle, Hertzman (1999, 2012) provided an articulation of the range of potential processes for these effects. Placing special emphasis on early development, he offered a hypothesis that systematic differences in early environment quality, including emotional and physical support and stimulation, will affect the neurochemistry and shaping of the central nervous system in ways that will adversely affect cognitive, social, and behavioral development. Although myriad biologic processes may be affected by social experience, some candidate systems are more likely to transduce social environmental factors into aspects of human biology with the capacity for embedding and influencing mental or physical health throughout the rest of the life course. Hertzman and Boyce (2010) identified four systems that are influenced by daily experience, respond to experience throughout an organism’s development, have meaningful impacts on health/learning/behavior, and are known to function differentially in response to variations in early experience. The four systems comprise: (1) the hypothalamic-pituitary-adrenal (HPA) axis and its expression of the glucocorticoid, cortisol; (2) the autonomic nervous system and its neurotransmitters, epinephrine and norepinephrine; (3) the prefrontal cortex, subserving memory, attention, and other executive functions; and (4) systems for social affiliation involving connections between the amygdala, locus coeruleus, and higher order cerebral connections, which are mediated by serotonin and other neurohormones. Other systems, such as the mesolimbic dopamine system, which mediates attentional processes, reward seeking, learning and behavioral engagement, and biological processes such as epigenetic modifications of neuroregulatory genes and telomerase activity are emerging as likely processes by which biological embedding occurs. Despite the plausibility of

biological embedding and the champions behind it, and that nationally representative studies have shown that adverse experiences early in life predict nearly 45% of childhood-onset and 30% of adult-onset psychopathology (Green et al., 2010), there is surprisingly little research on the early life effects of social environment on biology. Thus, the increasing attention to this gap in the empirical literature is merited. However, while understanding the mechanistic (i.e., mediation) processes by which an environmental exposure is linked to disordered development can be a powerful aid to elucidating pathogenesis and imagining novel interventions, understanding of mechanistic linkages, although perhaps necessary, is insufficient for a comprehensive understanding of causation. The quest for accurate determination of the mechanisms underpinning associations across early social contexts, stress, development, and psychopathology necessitates elucidation of the effects of modifiers (i.e., moderators) that reveal when, at what ages, or in what subgroups such associations hold (Baron & Kenny, 1986; Kraemer, Kiernan, Essex, & Kupfer, 2008; Kraemer, Stice, Kazdin, Offord, & Kupfer, 2001). NEUROBIOLOGICAL SENSITIVITY TO CONTEXT THEORIES AND EVIDENCE Contexts and experiences do not have universal influence across individual organisms, and the parsing of populations into subgroups of varying exposure susceptibility, through the discovery of moderator variables, can advance comprehension and the tractability of a given association between experience and pathology outcome. Indeed, the modern movement toward precision medicine (National Research Council, 2011) is, in part, motivated by the clinical advantage of identifying patients who are more susceptible to certain diseases, who will respond to treatments differently, and whose diseases or conditions may progress on a different course than others in the general population. Psychology has paid particular attention to understanding organismic factors (e.g., gender, IQ, and impulsivity) that place some individuals at greater risk for development of problems in adverse environments. This diathesis–stress or stress–vulnerability model (Figure 3.1) has been a prevailing view within medical and psychological frameworks since the 1960s, and emphasizes how individuals possessing biological vulnerabilities are most likely to develop functional impairments when exposed to particular stressors (for a review, see Ingram & Luxton, 2005). This framework is often conceptualized as a multicausal developmental model that proposes multiple risk factors across levels interact over the course of development, which has been remarkably useful in advancing

Neurobiological Sensitivity to Context Theories and Evidence

Level of Functioning

ity

Fixed or Resilient

iv sit en S ge nta Va e ibl pt e c us yS

all gic o l io ob ur e N ss tre S is es ath i D

Vantage Resistant

Positive

Negative Environment

Figure 3.1 Visual model of various theoretical concepts of sensitivity to environment.

understanding of individual variability in stress response and risk for pathology. Perhaps the most famous case of biologic sensitivity to adversity is the 2003 finding of Caspi and Moffitt’s group (Caspi, Sugden, Moffitt, Taylor, Craig, Harrington, McClay, et al., 2003), who pioneered longitudinal family gene–environment (GxE) research using their Dunedin study of roughly 1,000 individuals tracked over 40 years. Their 2003 paper linked experiences of childhood maltreatment with higher levels of adult depressive symptoms, but only among those possessing one or two copies of the short allele in the promoter region of the 5-HTT serotonin transporter gene—the gene that regulates transport of serotonin to the brain. They further found that possession of the longer allele was protective against those harmful effects. This first evidence demonstrating that one allele could lead to exacerbation or buffering of the stress–illness association was received with considerable scientific and media attention, and launched a gold rush of sorts into a quest for GxE interactions in mental health. This finding has been the source of considerable debate ever since, which we describe below, but nevertheless marked a major turning point in the conceptualization of and public awareness of the manner in which individual differences in a biological marker could confer varying sensitivity to harmful contexts. More recently, emphasis has shifted from the predominant diathesis–stress conceptualizations to a pursuit of understanding the individual difference factors that reliably allow some individuals to be anomalously resilient in the face of adversity (see Luthar & Lyman, in press; Masten, 2012; Rutter, 2012; Werner, 2012). In particular,

111

an individual’s stress reactivity appears to be an important individual difference variable that influences the manner in which an organism responds to environmental experiences (Bush & Boyce, 2014; Ellis, Boyce, Belsky, Bakermans-Kranenburg, & van IJzendoorn, 2011; Obradovi´c, Bush, Stamperdahl, Adler, & Boyce, 2010). Although heightened stress reactivity has traditionally been considered to be a factor that uniformly increases risk, burgeoning empirical evidence suggests instead a differential neurobiological susceptibility to environmental conditions among some individuals who show exaggerated sensitivity or permeability to both negative and positive environmental conditions (Belsky, 1997, 2005; Boyce & Ellis, 2005; Boyce et al., 1995; Ellis et al., 2011). Specifically, a theory known as biological sensitivity to context (BSC) suggested that higher levels of physiologic stress reactivity may promote adaptation in settings characterized by support, predictability, and protection, but exacerbate risk for maladaptive outcomes in stressful childhood environments characterized by adversity (Boyce, 1996, 2007; Boyce et al., 1995; Boyce & Ellis, 2005; Ellis, Essex, & Boyce, 2005b). Such individuals thus show either the least or most adaptive outcomes within the population, depending on the character of the proximal social contexts in which they are reared. BSC theory had clear parallels with Belsky’s differential susceptibility theory, which emphasized the manner in which behavioral reactivity made individuals particularly susceptible to both positive and negative experiences (Belsky, 1997, 2005). Although originally devised separately, through separate streams of thought and literature, the two theories shared considerable features and are now frequently presented together and referred to as differential neurobiological susceptibility (DNS; see Figure 3.1; Boyce, Obradovi´c, et al., 2012; Ellis et al., 2011; Pluess & Belsky, 2013), which is inclusive of multiple levels of endogenous susceptibility factors. Thus, in this chapter when referring to findings applicable to this integrated framework, we hereafter use this DNS term. A range of factors across multiple levels of biology have been thought to represent or subserve neurobiological sensitivity to social contexts, including genetic variation, chromatin modification, gene transcriptional control, ANS and HPA activation, and differences in brain structure and functions. The dominant lines of DNS research have tested differential susceptibility models by demonstrating interactive effects between various types of contextual stress (e.g., marital conflict, financial stress, parental psychopathology) and psychobiological reactivity in the prediction of health outcomes (e.g., Belsky & Pluess, 2009; Boyce et al., 1995; Obradovi´c et al., 2010). Such research finds support for the claim that associations between risky and

112

Differential Sensitivity to Context: Implications for Developmental Psychopathology

supportive environmental exposures and adaptation vary across different levels of stress reactivity. In one of our lab’s early empirical examinations of the DNS theory using the Peers and Wellness Study (PAWS) longitudinal community sample of 328 kindergarten children, we found support for DNS within two different biological systems and across multiple outcomes of adaptive functioning. Specifically, although children demonstrating high parasympathetic nervous system (PNS) reactivity (assessed via respiratory sinus arrhythmia) to a set of standardized challenges had the worst outcomes when they lived in families with high levels of family adversity, high parasympathetically reactive children demonstrated the highest levels of prosocial behaviors and school engagement and the lowest levels of externalizing behavior when they lived in contexts with low levels of family adversity. High parasympathetically reactive children also showed improvement in academic competence from fall to spring of the kindergarten year in the context of low adversity but declines in competence in the context of high adversity; whereas the inverse was true for low reactive children. In a similar fashion, children demonstrating high HPA axis reactivity (assessed via salivary cortisol) to the challenge protocol had the highest levels of prosocial behavior in the sample when they were in family contexts with low levels of adversity but the lowest levels of prosocial behavior when they were in family contexts with high adversity. The results of this study provided strong support for DNS in that the findings were relatively robust, particularly for parasympathetic reactivity, which emerged as a significant moderator of family environment across four domains of child adaptation—each pattern of which was consistent with BSC or DNS theory. Moreover, the DNS interactions were found for both positive and negative indices of adaptation, which are rarely examined simultaneously in empirical reports and allow for the test of the theory across a continuum of adjustment. These findings within a typically developing sample were also exciting because they suggested that, if stress reactivity moderated the effects of adversity on development in a community sample of children, who on average do not face extreme disadvantage or show clinical levels of problem behavior, the interactive effects are likely to be larger at the extreme ends of such distributions. Essex, Armstrong, Burk, Goldsmith, and Boyce (2011) tested DNS theory in a longitudinal sample of children to understand the roles of physiological and behavioral reactivity to the environment of classroom climate in the prediction of mental health symptoms across a 6-year period of childhood. Measuring both positive

(teacher–child closeness) and negative (teacher–child conflict) aspects of the classroom environment in first grade, they found that both autonomic (mean arterial blood pressure) and behavioral (objectively coded temperamental inhibition) reactivity moderated the impact of classroom climate on mental health problem severity in seventh grade, after adjusting for mental health problem severity in first grade. It is striking that all four biological sensitivity to context × teacher–child relationship interaction terms were statistically significant demonstrating DNS moderated associations of both adverse and supportive aspects of the early teacher–child relationship on the development of early adolescence symptom severity. Even within samples of children reared in poverty, wherein all study children are experiencing significant socioeconomic disadvantage, Conrad and colleagues found clear evidence that phenotypic reactivity can confer sensitivity to the risks and advantages conferred by other environmental factors such as objectively coded attachment security. In their study of infants born into disadvantaged conditions, infants with high baseline respiratory sinus arrhythmia (RSA) who were raised in contexts of high security had lower levels of problem behavior (actually below community norms), and infants with high baseline RSA who were raised in caregiving environments that fostered disorganization had the greatest number of problem behaviors at 17 months. Attachment security did not relate to problem behavior for infants with low RSA. These differences were also evident in the large effect sizes for these varying associations. These findings bolster evidence that poverty is not a uniform risk for problem behavior, particularly for children raised by nurturing caregivers; but they also reveal that children high in basal RSA are best able to take advantage of such opportunities and most at risk for harm if not provided with sensitive, responsive caregiving environments. Raver, Blair, and Willoughby’s (2013) work also demonstrates that even in low-income populations, variation in temperamental reactivity can confer sensitivity to variations in the stressors of poverty. Although chronicity of poverty shows a negative linear association with young children’s effortful control, their study found that it had no association for children low in reactivity. However, children high in reactivity demonstrated the highest levels of effortful control when they were not from families with chronic financial strain but the lowest levels of effortful control when they resided in the most chronically poor families—thus, demonstrating the for-better-or-for-worse effect even within a population with a restricted range of income.

Neurobiological Sensitivity to Context Theories and Evidence

Researchers using sophisticated modeling techniques on data from the Caucasian subsample of 1,030 adolescents and young adults from the nationally representative National Longitudinal Study of Adolescent Health have recently provided a compelling example of how sensitivity via a polymorphism in the serotonin transporter gene (5-HTTLPR) can convey DNS (Li, Berk, & Lee, 2013). Despite the widely demonstrated effects of high-quality family support on depression, they found that level of family support (cohesion, communication, and warmth) had little to no effect on depression scores in male individuals with the ll genotype. However, male s-carriers in families with high levels of support had the lowest levels of depression in the sample while s-carriers in families with the lowest levels of support had high levels of depression. This interaction was only marginally significant for females. Findings within this large representative sample, and across a broad range of environmental exposure (low to high quality of family support), have added support for DNS theories. Studies demonstrating this greater susceptibility of neurobiologically responsive children to both positive and negative aspects of their environments have now included a wide variety of the following: Stressors and adversities: peer victimization and aggression (Rudolph, Troop-Gordon, & Granger, 2010, 2011); parental depression and antisocial behavior (Cummings, El-Sheikh, Kouros, & Keller, 2007; Shannon, Beauchaine, Brenner, Neuhaus, & GatzkeKopp, 2007); marital conflict (El-Sheikh, 2005; El-Sheikh, Keller, & Erath, 2007); overall family distress (Obradovi´c et al., 2010); child maltreatment (Grazioplene, DeYoung, Rogosch, & Cicchetti, 2013) Positive environments: parental warmth (Ellis, McFadyenKetchum, Dodge, Pettit, & Bates, 1999); supportive child-care settings (Phillips et al., 2012); supportive interventions (Bakermans-Kranenburg, Van, Pijlman, Mesman, & Juffer, 2008) Defining behavioral and biological parameters: temperament (Lengua, 2008; Pluess & Belsky, 2009) and personality (Grazioplene et al., 2013); physiologic reactivity (Alkon et al., 2006; Boyce et al., 1995); structural and functional differences in brain circuitry (Whittle et al., 2011); gene polymorphisms (see Belsky & Pluess, 2009 for a review; Knafo, Israel, & Ebstein, 2011; Manuck, Craig, Flory, Halder, & Ferrell, 2011) Further, there is evidence for differential susceptibility across positive and negative types of environment as

113

described previously but also for differential susceptibility to experience across the variety of characters that play important roles in children’s lives, including mothering (Conradt, Measelle, & Ablow, 2013; Davies & Cicchetti, 2014), fathering (Boyce et al., 2006), quality of child-care settings (Pluess & Belsky, 2009), and teacher relationships (Essex et al., 2011). Most importantly, this body of literature reveals that highly susceptible children show bidirectional effects on outcomes in contrasting low- and high-stress settings, not simply an attenuation of negative effects in low-stress circumstances. Implications for Conditional Adaptation Within the Biological Sensitivity to Context Theory body of work, Boyce and Ellis (2005) also argued, in the second, less extensively explored component of the theory, for the importance of considering U-shaped associations, predicted from evolutionary principles, between the stressful versus supportive character of early rearing environments and the proportion of individuals evincing highly reactive phenotypes. The second of the 2005 papers advancing the BSC theory provided provisional evidence from two studies supporting the postulated U-shaped association between adversity and the prevalence of high reactivity phenotypes (Ellis, Essex, & Boyce, 2005a). The character of early contextual experiences likely plays a role, through conditional adaptations, in shaping the development of children’s physiology. Those raised in stimulating and nurturing contexts may disproportionately acquire heightened biological sensitivity as a means of maximizing the advantages of resources and opportunities therein. On the other hand, children reared in harsh, threatening environments might also develop greater biological sensitivity to enhance vigilance to threats and other hazards. In contrast, the majority of children, raised within species-typical environments falling within these two extremes, may acquire diminished, more normative biological sensitivity, as the environments to which they are exposed are neither highly nurturing nor highly threatening (see Ellis & Boyce, 2011; Ellis et al., 2011 for the full argument in this regard). More recently, the concept of differential susceptibility has been usefully elaborated in a more detailed evolutionary framework referred to as the adaptive calibration model (ACM; Del Giudice, Ellis, & Shirtcliff, 2011). The ACM is a theory of developmental programing that focuses on calibration of organismic stress response systems and life history strategies in response to local environmental conditions. The ACM framework emphasizes a model of highly variable inter-individual patterns of stress responsivity

114

Differential Sensitivity to Context: Implications for Developmental Psychopathology

across life stages, prototypical responsivity patterns, the emergence of sex differences, and the primacy of the adaptive nature of long-term changes in biobehavioral systems, which its authors suggest promotes cross-field theoretical and research integration. Boyce, Ellis, and colleagues thus suggest that highreactivity phenotypes will be most prevalent in the contexts of low and high adversity exposure. In accordance with this theory, a U-shaped association between adversity and physiologic reactivity could provide for diverging trajectories of subsequent mental health over time. For example, although high reactivity may be adaptive across very low and very high stress contexts in the short term, as high reactivity becomes established (i.e., canalized), subsequent exposures to adversity over time would predispose the highly reactive individual to disorders of mental health. However, only longitudinal studies of the interplay between environmental effects and stress reactivity, starting with prenatal development and within broadly variable social contexts, can provide credible empirical evidence for this hypothesis. Moreover, to date, research has rarely considered U-shaped associations despite the value to be gained from understanding with greater precision the shape of associations between contextual stressors and the biological systems that contribute to psychopathology. Three papers specifically explore curvilinear influences of early environmental stressors on the development of physiologic systems, specifically the HPA axis. Gunnar and colleagues (Gunnar, Frenn, Wewerka, & Van Ryzin, 2009) reported that children with moderate levels of early life adversity demonstrated lower cortisol reactivity to laboratory stressors than did children with either low or high levels of adversity, providing additional evidence for curvilinear associations between early life stress and HPA axis regulation. A second study, using a sample of young adults, found a curvilinear effect in the opposite direction from that proposed by BSC theory and found in Gunnar et al.’s study. Engert and colleagues (2010) found evidence for an inverted U-shaped relation between retrospective self-report of levels of maternal care received in childhood and young adults’ cortisol stress reactivity, such that stress-induced cortisol levels for low- and high-maternal care groups were lower than for those in medium-care group. Such divergent patterns found between studies may result from meaningful differences in developmental timing of physiological assessment, intensity of the stressors, and factors such as retrospective reporting of stressors vs. concurrent or prospective assessments. Both studies examine effects on reactivity to laboratory stressors, which may not generalize to daily levels of physiologic load that

are more relevant to allostatic load. A third paper, a review by Macri et al. (Macri, Zoratto, & Laviola, 2011), summarizes evidence from laboratory rodent studies and suggests that the link between neonatal stress exposures and adult phenotypic reactivity follows a U-shaped, curvilinear relation. The conclusions of these three papers support, at the very least, further consideration of curvilinear associations between early life stress and HPA axis regulation. Recent work from our laboratory provides some additional evidence. Using an ethnically diverse longitudinal sample of 338 kindergarten children, we examined the effects of cumulative contextual stressors on children’s developing HPA axis regulation (Bush, Obradovic, Adler, & Boyce, 2011). Chronic HPA axis regulation was assessed using cumulative, multiday measures of cortisol in both the fall and spring seasons of the kindergarten year. Hierarchical linear regression analyses revealed that contextual stressors related to ethnic minority status, SES, and family adversity each uniquely predicted children’s daily HPA activity and that some of those associations were curvilinear in conformation. Results showed that the quadratic, U-shaped influences of family SES and family adversity operate in different directions to predict children’s HPA axis regulation such that children from both the low and high SES families demonstrated higher levels of daily cortisol than their peers from middle-SES families and children from families with moderate adversity demonstrated higher daily cortisol than did those at either end of the adversity continuum. Results further suggested that these associations differed for white and ethnic minority children, with the opposing patterns of findings potentially reflecting different biological responses to environmental adversity by subgroup or different ranges of exposure to the stressors within subgroups. In total, this study revealed that early childhood experiences contribute to shifts in one of the principal neurobiological systems thought to contribute to allostatic load, and findings suggested that analyses of allostatic load and developmental theories accounting for its accrual would benefit from an inclusion of curvilinear associations in tested predictive models. A further consideration is that the preponderance of differential susceptibility research has examined the manner in which biology and context interact to predict psychopathology symptoms or diagnoses. More recently, we have investigated the manner in which “stress sensitive” alleles moderate the effects of early life adversity on stress physiology early in development—a phenotypic predictor of later development across a broad array of health outcomes. Exploring the established negative association between SES and chronic cortisol arousal in children,

Neurobiological Sensitivity to Context Theories and Evidence

Bush and colleagues found that both 5-HTTLPR and BDNF genotype moderated the effects of early life family SES on children’s chronic daily cortisol arousal (Bush, Guendelman, Adler, & Boyce, 2013a, 2013b), such that children with the l/l genotype of the serotonin transporter or who are carriers of the BDNF met allele had the highest chronic cortisol levels when growing up in low SES homes, but the lowest levels when growing up in high SES homes. Strikingly, for s carriers and Val/val genotypes, SES did not affect chronic daily cortisol levels. Beyond providing additional examples of DNS (and contributing to the small literature demonstrating that l/l genotypes are most sensitive to some environments (see, for example, Davies & Cicchetti, 2014), this work highlights how susceptibility via one biomarker (genes) can influence stress sensitivity in other systems (HPA axis arousal). Evidence of Differential Susceptibility Within Positive Environments As a result of the field’s focus on contextual risk for pathology over contexts that promote optimal development, the majority of studies reviewed above emphasize the negative impacts of adverse exposures or deprivation on children’s neurobiology. These studies tell us little about whether enriched environments produce optimized neurobiological development of the same magnitude as that negatively affected by adverse environments. A few studies, however, do point to enhanced sensitivity to positive environments. One excellent example is the work by Knafo and colleagues (Knafo et al., 2011) who used a sample of 168 twin pairs of Israeli preschool-aged children to parse genetic and environmental effects and investigate whether the 7-repeat allele of the dopamine receptor D4 (DRD4) would confer enhanced sensitivity to positive parenting. They found that positive parenting predicted maternal-rated prosocial behavior, but only for 7-repeat carriers. Similar findings occurred in a Dutch study in which 7-year-old children who were securely attached to their parents had higher rates of experimentally induced sharing assessed by willingness to donate their recently earned money to charity, but only for 7-repeat carriers (Bakermans-Kranenburg & van IJzendoorn, 2011). Additional evidence can be found from a study that found that the highest levels of social integration within objectively coded high-quality childcare settings was demonstrated by the most temperamentally reactive toddlers (Phillips et al., 2012). Together, these studies suggest that a variety of indices of sensitivity allow for increased sensitivity to enriched or optimal environments, yet many more studies in this area are needed to discern whether the sensitivity to both positive

115

and negative environments is similar in magnitude and frequency. Studies of differential susceptibility to adversity also do not help us ascertain whether extremely enriched contexts might counteract detrimental effects of adversity. However, a new literature is pointing to evidence for enhanced sensitivity to interventions, which represent positive environments designed to ameliorate negative effects of context on trajectories of mental health. The Bucharest Early Intervention Project, for example, is a novel study where institutionally reared Romanian children who had been abandoned in the first few years of life were randomized to either remain under institutional care or be placed in (relatively) enriched environments via in-home foster care placements (Nelson et al., 2007). It has provided compelling evidence for differential susceptibility to adversity and intervention effects (i.e., both negative and positive environments). Specifically, children with one of the more sensitive genotypes (in this study, the 5-HTTLPR s/s or BDNF Val66Met met-carriers) demonstrated the lowest levels of indiscriminate behavior in the enriched foster care environment and the highest levels in the continued institutionalization context (Drury et al., 2012). Children with either of the less sensitive alleles (in this study, 5-HTTLPR l-carriers or of BDNF val/val) demonstrated little difference in levels of indiscriminate behaviors over time and no group × genotype interaction. Strikingly, this pattern was amplified when GxGxE tests were conducted, revealing that the DNS effect was largest for children with both sensitive genotypes, relative to those possessing just one, and children with no plasticity alleles demonstrated no intervention effect on levels of indiscriminate behavior over time. Additional evidence can be found in more typically developing samples in the United States. In a randomized controlled trial of 169 families, only highly irritable infants showed positive benefit from an intervention designed to improve attachment security (Cassidy, Woodhouse, Sherman, Stupica, & Lejuez, 2011). Likewise, in a randomized controlled trial of 985 low birth weight infants from the Infant Health and Development Program, Blair (2002) found that a comprehensive compensatory education intervention had a substantially stronger effect on cognitive functioning and externalizing behavior outcomes for infants with greater temperamental negativity (Blair, 2002). In addition, preliminary analyses of a longitudinal evaluation of a clinical intervention for school-aged boys with disruptive behavior disorder found that only those with high cortisol reactivity prior to treatment showed decreases in parent-rated aggression and oppositional behavior after

116

Differential Sensitivity to Context: Implications for Developmental Psychopathology

treatment (van de Wiel, van Goozen, Matthys, Snoek, & van Engeland, 2004). Bakermans-Kranenburg and colleagues (2008) also found that DRD4 7-repeat carriers were more sensitive to an intervention designed to increase maternal positive discipline within families struggling with or at risk for child externalizing behavior problems; not only were the intervention effects on decreasing externalizing behavior strongest for 7-repeat carriers, but the association was also graded such that the effects were largest when mothers had better adherence to the goals of increased positive parenting. Particularly compelling is recent evidence from two randomized alcohol use prevention trials involving more than 900 African American youths presented by Brody and colleagues (Brody, Chen, & Beach, 2013). This group found that variation in dopaminergic (DRD2, DRD4, and ANKK1) and GABAergic (GABRG1 and GABRA2) genes predicted alcohol use across a 2-year period of adolescence such that genetically at-risk youth within the control condition displayed greater levels of alcohol consumption but genetically at risk youth assigned to the prevention condition showed the lowest levels of consumption. Strikingly, when adolescents were coded based on how many of the risk genes they possessed (a multilocus genetic score in this study) an interaction predicting longitudinal change in alcohol use revealed that youth carrying greater numbers of the risk genes showed the greatest increase in alcohol use if in the control condition but greater responsiveness to prevention (smaller increases in use) than carriers of zero or one risk gene. The large sample size and randomization increase confidence in these findings. Intervention Effects on Susceptibility In addition to thinking about differential susceptibility to interventions, it is important to consider how interventions might improve risk-associated phenotypic reactivity or sensitivity in a manner that affects later susceptibility to environment. Prospective research has found that parental responsivity during early childhood predicts adolescent cortisol reactivity to social stressors, even after adjusting for subjective appraisals of the stressor, other environmental risk factors, and current psychosocial functioning (Hackman et al., 2013). This points to interventions targeting parental behavior as potentially important for the shaping of neurobiological susceptibility to environment. Thus far, the limited empirical investigation of intensive interventions suggests that, although they may not be entirely reversible, some harmful effects of biological embedding of early adversity can be overcome.

For example, promise for recovery through intervention can be found in the recent results of primate rearing experiments. Although peer-reared monkeys exhibited higher cortisol levels in infancy and in response to stress, relative to mother-reared monkeys, 16 months after monkeys were relocated to a large social environment consisting of infants from all experimental rearing conditions, rearing groups were indistinguishable from one another physiologically and behaviorally (Dettmer, Novak, Suomi, & Meyer, 2012), suggesting transition to more normative environments can allow for recalibration of systems over time. The Bucharest Early Intervention Project has also shown that early intervention to correct a deeply impoverished early environment can improve brain structure and function as well as cognitive and emotional capabilities (Sheridan, Fox, Zeanah, McLaughlin, & Nelson, 2012), suggesting the potential for catch-up brain growth in human children, even following extreme deprivation. Further results point to behavioral improvements as well, indicating that children in the foster care intervention demonstrated higher levels of positive affect and attention at both 30 and 42 months of age, and improved cognitive outcomes at age 42 and 54 months, compared with children who remained institutionalized (Ghera et al., 2009; Nelson et al., 2007). Children who were transitioned to foster care at earlier ages demonstrated the most marked improvements, suggesting sensitive periods for brain development. Studies of foster care children living in the United States have similarly provided evidence that psychosocial interventions may indeed modify children’s disrupted biological systems, such as cortisol diurnal rhythms (Dozier et al., 2006; Fisher, Stoolmiller, Gunnar, & Burraston, 2007) and cortisol reactivity (Nelson & Spieker, 2013). For example, Nelson and Spieker found that an attachment promotion program with toddlers experiencing transitions in foster care placement led to increased cortisol reactivity (improvement in this context) in response to an experimental separation paradigm. There is also evidence that psychoeducation mixed with cognitive behavioral stabilizing group treatment can lead to decreases in levels of dorsal anterior cingulate cortex and left anterior insula activation to the Stroop task (thought to reflect selective attention and emotional regulation) for child abuse-related complex PTSD patients who previously demonstrated activation higher than that of matched controls (Thomaes et al., 2012). Together, these observed associations suggest some negative effects on neurobiology can be reversed or compensated for with early interventions, emphasizing the importance of both enrichment and adversity-reduction interventions in creating good

Neurobiological Sensitivity to Context Theories and Evidence

developmental environments for all children. Continued examination of intervention efforts with children exposed to early adversity and possessing stress reactivity phenotypes will broaden evidence for how contextual experience can shape the development of children’s physiology and which phenotypes might be most capable of reversing ill effects on biology. Vantage Sensitivity In contrast to the diathesis–stress framework’s emphasis on vulnerability to risk, Pluess and Belsky (2013) recently promoted a new construct of vantage sensitivity, which they argue can be distinguished from theoretically related concepts of differential-susceptibility and resilience, because it reflects variation in response to exclusively positive experiences as a function of individual endogenous characteristics (see Figure 3.1). They use this term to refer to the “bright side” of differential susceptibility in that some individuals are more sensitive and positively responsive to the environmental advantages to which they are exposed, be it security of attachment derived from sensitive parenting, prosocial behavior derived from supportive, healthy friendship networks, or academic achievement stemming from high-quality preschool or child care. This theoretical framework addresses some of the challenges that arise when attempting to examine certain combinations of individual characteristics and environments within a DNS framework. For example, from a DNS stance, a highly intelligent child might profit exceptionally from a high-quality education, but she may suffer disproportionately in a low-quality educational setting (perhaps due to being so smart that she disengages in the classroom and fails to learn); however a vantage sensitivity framework might allow for that child to have advantage in a good environment and no harm (unresponsive) to the adverse school environment, or even show protection from the adverse environment (increased resilience to adversity). In their review of recent empirical evidence for vantage sensitivity, they provide compelling examples featuring behavioral, physiological, and genetic factors as moderators of a wide range of positive experiences ranging from family environment and psychotherapy to educational intervention. We provide two illustrative examples here. First, in a large British cohort study, temperamentally reactive girls were more prosocial at age 6 if their fathers were highly involved in their care during early childhood, but nonreactive girls did not show any benefit of early father involvement on that outcome (Ramchandani, IJzendoorn, & Bakermans-Kranenburg, 2010). Second, Eisenberg and colleagues (Eisenberg et al.,

117

2012) found that children high in baseline RSA activity reared in high quality home environments displayed lower aggression, whereas children characterized by low RSA levels did not benefit from the high quality environments. Pluess and Belsky pointed out in their review that much of the empirical work to date does not allow for true determination of whether individual phenotypes make one differentially susceptible to positive environments alone, because most studies with evidence for vantage sensitivity are not positioned to test whether those individuals are also more responsive to negative contexts. This situation of leaning on available data to support portions of a theory is a challenge for differential susceptibility more generally and improved study designs with breadth of measurement of contextual predictors and adjustment outcomes across negative and positive ends of those spectrums will greatly advance the field. Cumulative Sensitivity to Environment? Although research has spanned biological systems and often tests for DNS effects across several sensitivity markers concurrently, only recently have scientists begun to examine how sensitivities might be additive to create cumulative risk or advantage to environmental experience. In particular, Belsky and colleagues (Belsky & Beaver, 2011) have made efforts to advance this body of work through multigenic studies within larger samples, to test whether there are additive effects of purported plasticity genes. In the Add Health nationally representative sample of 1586 adolescents, they examined whether maternal parenting quality (including involvement, disengagement and attachment) predicted concurrent adolescent attentional, emotional, and behavioral self-regulation, assessed by questionnaires. Children in the sample were assessed for five gene polymorphisms with known associations with behavior and mood, including genes affecting dopamine availability and processing (DAT1, DRD2, and DRD4), the serotonin transporter gene (5-HTTLPR), and the MAOA gene, to test for cumulative sensitivity to parenting. Findings revealed that possessing increasing numbers of plasticity variants was associated with increasing sensitivity to parenting quality, such that maternal parenting quality was not associated with attentional and behavioral self-regulation for those carrying zero or one sensitivity alleles but predicted more and less self-regulation under, respectively, supportive and unsupportive parenting conditions, for those carrying two or more alleles. The pattern of findings suggested an additive effect such that the strength of the interactive association was greater with an increasing

118

Differential Sensitivity to Context: Implications for Developmental Psychopathology

number of sensitivity alleles. However, this effect was only found for males, and was not significant for females in this sample. Of note, preliminary findings from our longitudinal sample of kindergarten children partially parallel this unique pattern of findings. Specifically, our findings (Allison, Bush, Adler, & Boyce, 2013) suggest that cumulative genetic sensitivity across several dopamine polymorphisms (DAT1, DRD2, and DRD4) predicts increasing severity of inattention and ADHD symptomatology later in the kindergarten school year, but only in males. However, this cumulative dopamine sensitivity variable did not significantly interact with family adversity to predict those symptoms. DAT1 alone, however, did interact with family adversity to predict inattention and ADHD symptoms in a manner consistent with the DNS theory. The variation in patterns of findings between the Belsky and Beaver analyses and those resulting from our preliminary analyses may be due to a variety of factors—a discussion of which highlights the complex challenge of establishing a body of evidence for cumulative sensitivity. First, we did not include two of the additional two sensitivity genes (5-HTTLPR and MAOA) used in their cumulative plasticity variable, as our interest was specifically in testing for cumulative effects within the dopamine pathway. Because studies rarely have exact replicates of variables for attempts at replications of findings, this sort of discrepancy will make it difficult to discern whether a specific genotype was driving the cumulative effect in a particular study and having multiple “risk” genotypes may just increase the odds that an individual possessed that particular variant. Second, we utilized different environmental variables (parenting vs. family adversity); although each are relevant exogenous influences on pathology, one measure may be more proximal for an individual child or more potent, and thus provide a better test of sensitivity to context. Third, our environmental variable had a limited positive range (lack of adversity does not necessarily imply a positive environment), and the parenting variable in the Belsky and Beaver study may have better reflected a broad range of experience in which to detect variability in response to environment. Fourth, there was an important difference in the ages assessed; the apparent cumulative sensitivity in adolescence may not manifest in a detectable manner by age 5, at least not in terms of ADHD symptomatology, which was relatively low in our community sample of young children. Although not framed as cumulative plasticity, new findings from a sample of 120 preschool-aged children also appear to reveal how polygenic effects may add together to

make individuals sensitive to environmental influences, for better or for worse. Pagliaccio and colleagues (2014) found crossover interactions such that children with greater numbers of genetic polymorphisms associated with risk of depression or higher cortisol had low hippocampal volume when they had histories relatively free of adversity but the highest hippocampal volumes when they had significant histories of adversity. The complex patterns of findings on children’s hippocampal volume and risk for adjustment in the literature to date prevent us from interpreting this finding in terms of which children are better off, yet the crossover interactions found in that study provide additional data for consideration in the quest for understanding DNS. Taking this cumulative plasticity approach is quite new, and additional work is needed to determine whether such conceptualizations and analyses will advance understanding of the etiology of developmental psychopathologies. Work such as that conducted by Brody and colleagues described above (Brody, Chen, et al., 2013), which revealed possession of greater numbers of susceptibility genes conferred increased risk for alcohol use across adolescence but greater sensitivity to a preventative intervention, is promising in this regard. Although theoretically and statistically distinguishable from cumulative additive effects, three-way interactions among biological sensitivities and context, such as the GxGxE differential susceptibility findings reviewed earlier in this chapter, may also reflect a cumulative sensitivity or plasticity. Another example of this GxGxE effect can be found in additional work published by Brody and colleagues (Brody, Yu, et al., 2013). Within a sample of 315 African American youth, they found that those who possessed both the short (s) allele of the 5-HTTLPR gene and an allele of DRD4 with 7 or more repeats were most sensitive to the effects of the family environments on biology; whereas family environment had no effect on those who possessed none or just one of those sensitivity alleles. Specifically, individuals who experienced low levels of family support in adolescence were more likely to demonstrate high levels of allostatic load (a compilation of measures of HPA axis activity, blood pressure, and other biological markers of health) if they were both s-carriers and 7-repeat carriers. Cicchetti and colleagues have also found GxGxE effects in allostatic load models within a large multiethnic sample of 10-year-old children. Their findings revealed that maltreatment and CRHR 1 (corticotropin releasing hormone receptor 1) genetic variation affect cortisol regulation but that maltreatment and the joint effects of CRHR 1 TAT haplotype and 5-HTTLPR moderated

Conceptual and Methodological Issues for Examinations of Differential Neurobiological Susceptibility

Impediments to Discovery Given the difficulties of discovering biology–context interactions (Wachs & Plomin, 1991), and that the challenging factors discussed in the above section apply not only to searches for replication of cumulative sensitivity effects, but to replication of interaction findings more generally, it is all the more remarkable that such interactions are being increasingly identified and characterized in studies of human disease. GxE interactions are assessed as the differential risk effects of an exposure among individuals with different genotypes, or as the differing effects of a genotype among those who are heterogeneous with respect to exposure (Kraemer et al., 2001). As noted by McClelland and Judd (1993), field studies tend to underestimate the magnitude of interaction effects due to the nonoptimal distributions of the component variables, as the power to detect GxE interactions depends on the distributions and frequencies of the genotypes and exposures in the sample (for a discussion and examples see Caspi, Hariri, Holmes, Uher, & Moffitt, 2010). Interactions accounting for as little as 1% of outcome variance in field research may conceal a true R2 many times that size that would be found using an optimal, experimental design. When demonstrated, biology–context interactions in human populations may disproportionately exist at the extremes of genetic and environmental variation (McGue & Bouchard, 1998). Mathematically, analyses of variance and related statistical tests often fail to identify nonadditive effects because they have much less power in tests of interactions than in tests of main effects (Wahlsten, 1990). Improved genetic research designs (Andrieu & Goldstein, 1998; Caspi et al., 2010; Dunn et al., 2011), along with statistical approaches that

ity

iv sit

Negative Adaptation/Functioning

CONCEPTUAL AND METHODOLOGICAL ISSUES FOR EXAMINATIONS OF DIFFERENTIAL NEUROBIOLOGICAL SUSCEPTIBILITY

are more sensitive to interaction effects (Kraemer et al., 2001) and can simultaneously consider multiple genetic markers within an entire targeted region (e.g., Belsky, Moffitt, & Caspi, 2013; Yu et al., 2012), are abetting the discovery of complex interactions between and among biologically and environmentally based factors. Single studies are likely to be capable of capturing only a portion of the underlying paradigmatic DNS interaction. As illustrated in Figure 3.2, which is adapted from a figure in our previous publication (Obradovi´c et al., 2010), the observational window (indicated by the area within the lightly shaded square) for DNS effects shifts from study to study, and even within a particular study depending on the focus of specific publications. The observational window can shift depending on a variety of factors, including: characteristics of the sample, such as intentionally oversampling impoverished children or unintentionally having greater success enrolling and retaining upper SES families, which greatly influences the range and qualities of environmental exposures; the outcomes of interest and whether they have a positive or negative valence (such as prosocial behavior and externalizing problems respectively) or more of a continuum such as with effortful control, or whether studies are examining early phenotypic risk factors for later

Positive Adaptation/Functioning

the effect of child maltreatment to predict internalizing symptoms (Cicchetti, Rogosch, & Oshri, 2011). Neither of these GxGxE studies presents findings that reveal the full spectrum of for-better-or-for-worse effects on adjustment due to their lack of differential susceptibility findings in better off family contexts or positive adjustment outcomes. However, they do reveal the ways multiple genes may simultaneously confer sensitivity to disadvantage via synergistic effects and point to the need for further multigene examination of sensitivity to context, particularly across positive and negative contexts and outcomes.

119

gh Hi

in

D

r iffe

en lS

tia

en

Low in Differential Sensitivity

Negative

Positive Environment

Figure 3.2 Conceptual model of the interactive effects between DNS and environmental influences, and the window of observation allowed by any one study design. Solid shaded gray area represents the window of observation by which any single study is capable of capturing a paradigmatic DNS interaction. The window can move in any direction, depending on study design and sample characteristics. Source: Figure adapted from previous work from our lab published in J. Obradovi´c, N. R. Bush, J. Stamperdahl, N. E. Adler, & W. T. Boyce, Biological sensitivity to context: The interactive effects of stress reactivity and family adversity on socioemotional behavior and school readiness, Child Development, 81(1), 270–289, 2010. Adapted with permission from the authors.

120

Differential Sensitivity to Context: Implications for Developmental Psychopathology

pathology or examining extreme psychopathology (which can affect the strength of associations or the variability in your outcomes); the measure of reactivity, because some indices may have greater inter-individual variability, and some indices may be quite difficult to ethically elicit reactivity in, (such as purely sympathetic preejection period response in young children) which restricts the ability to detect highly reactive children; the type of challenge or stressor used to elicit reactivity, since some individuals may be particularly reactive to some types and not others (for example, perhaps highly stressed in contexts of social threat but not in those evaluating math performance); developmental timing factors, such as ways children’s age or cognitive or biological development affect the capacity for studies to detect differences in phenotypic reactivity or adjustment outcome (e.g., some measures of physiological reactivity may not be appropriate early in development as the early development and functioning of those systems is not well understood); and a host of other possibilities. For example, the observational window may shift to the left in studies of highly disadvantaged child populations or to the right in intervention studies. The window may shift downwards and flatten in studies of children with severe behavioral problems but cover more of the y axis in community samples of typically developing children with a broad range of functioning. Similarly, the slopes of the two component lines may be determined by the researchers’ ability to capture the full range of stress reactivity responses, such as by truly activating the stress response system for reactive children or adequately sampling reactive and nonreactive genotypes, and thus distinguish between high and low reactive children. Failure to consider these factors in presentation of DNS findings, or in contexts when results fail to support DNS theory is an impediment to discovery, as it can lead to erroneous conclusions. It is critical, therefore, that developmental researchers examine and compare the interactive effects of stress reactivity and context across a range of outcomes and environmental influences and that the full scope of existing theories and literature be taken into account in the interpretation of such effects. GxE Debate As the cost of DNA testing has plummeted, an increasing number of tests of differential susceptibility are examining genetic polymorphisms as key biological plasticity factors. It is important to note that as enthusiasm for GxE studies in psychopathology has increased over the past decade, this has not occurred without controversy. The vast majority of GxE studies to date use the candidate gene approach,

which focuses on associations between allelic variation within prespecified genes of interest and phenotypes or disease states. Candidate genes are most often selected for study based on a priori knowledge of the gene’s biological functional impact on the phenotype or disease state being investigated. Strengths of these study designs often include: selection of both gene and environmental variables with plausible interactive biologic pathways; selection of an environmental variable with known association with the phenotype of interest; prospective and retrospective measurement of the environmental risk factor; and accepted and validated measures of the phenotype, based on DSM-IV (American Psychiatric Association, 2000; now DSM-V) criteria. Such an approach is thought to lead to accurate determination of true GxE effects. On the other hand, candidate gene analyses, examining either main effects or a specific polymorphism’s moderation of an environmental effect as in DNS models, have received criticism of late. GxE studies are notoriously difficult to replicate, and depending upon the methodology and comprehensiveness of an approach, meta-analyses have shown mixed evidence for GxE effects for even the most-studied candidate genes in psychopathology, such as the serotonin transporter polymorphism 5-HTTLPR (e.g., Karg et al., 2011; Risch et al., 2009). Such contrasting findings have advanced a robust debate among scholars, although studies with stronger methodological rigor appear more likely to replicate the original (Caspi et al., 2003) effects on depression. Additional rigorous work has extended these findings, suggesting that 5-HTTLPR confers sensitivity in the prediction of a variety of disease-related phenotypes such as hormone and inflammatory responses, brain wave patterns, and brain imaging results (Caspi et al., 2010). Recent years have led some from the field of genetics to argue that significant GxE results are likely to be the result of artifact or chance (see Rutter, Thapar, & Pickles, 2009), and that the smaller sample sizes required for these rich biological and environmental data analyses, inconsistency in which environments are selected for models, selective presentation of results, or publication bias may be inflating the appearance of a true effect. Proponents of this position often argue that genome-wide association studies (GWAS), which scan the entire genome for common genetic variation, are superior in that they avoid some of the bias problems within candidate-gene studies. Rather than beginning with a specific hypothesis about a candidate gene and psychopathology, as GxE researchers often do, whole-genome researchers typically begin with consideration of a particular disease, then examine the genomes of hundreds or thousands of people

Conceptual and Methodological Issues for Examinations of Differential Neurobiological Susceptibility

exhibiting that disease to determine whether they possess specific genes or gene variants absent in disease-free individuals (The Psychiatric GWAS Consortium Steering Committee, 2008). This atheoretical approach is viewed by some to be more scientific, and when no genes materialize in GWAS analyses it is asserted that no genetic association exists for that disorder. However, in the decade following the decoding of the human genome, the mass of large-scale GWAS have found many promising leads, yet no clarity about causes or cures for the most prominent, heritable psychopathologies facing our society. Although genome-wide approaches to identify loci involved in GxE interactions have recently begun to appear, they have yet to establish significant GxE interactions in common or well-documented sensitivity alleles (Aschard et al., 2012), leading to increased skepticism for the GxE framework. A crucial problem in this debate is that GWAS research is not able to test properly for DNS for a variety of reasons: 1. It is quite common that no significant main effects are found when true interactions exist, because crossover effects are masked due to the average level of outcome resulting from some individuals with the gene having very poor outcomes and others, also with the gene, having very good outcomes (see Rutter et al., 2009 for a discussion), 2. GxE effects of interest are not typically detectable in population-based whole-genome studies because such studies rarely examine environmental factors, and when they do it is often with one item, retrospective reporting, or other limiting methodologies, and 3. Very large surveys typically conducted for GWAS generally assess a particular categorical disease outcome, rather than a full continuum of positive and negative behaviors (such as emotion regulation), which is the hallmark of DNS (see Caspi et al., 2010 for a discussion of some of these and other issues). Even well-designed genome-wide environmental interaction (GWEI) studies face a broad range of methodological and statistical challenges (see Aschard et al., 2012, for a review) before they can begin to unravel the complexities of GxE effects in human disease and behavior. Advancement in this debate will require a variety of approaches from all sides, including increased appreciation for replication, and inclusion of rich environmental variables in genome-wide association studies. As is crucial in all scientific endeavors, multiple research strategies are needed, and it is likely that both the GWAS approach and the candidate GxE approach will be informative and ultimately result in a

121

synergistic or complementary understanding of these phenomena.

Is Reactivity or Susceptibility Maladaptive? The identification of factors contributing to the activation and dysregulation of physiological systems designed to maintain the allostatic balance has been thought of as essential for the design of prevention and intervention efforts (Lupien et al., 2006). Indeed, the term dysregulation is widespread in the stress–reactivity and regulation literature—often without reference values for normative or healthy levels, and without consideration of the adaptive function of those organismic systemic changes. This promotes confusion about the meaning of physiologic resting and reactivity levels, the sources for variability in these measures, and interpretation of variability or changes in these values. However, in recent years, scholars have begun to consider that the alterations in individual biological systems in response to environmental adversity, although sometimes seen as harmful or maladaptive, may actually be an appropriate physiologic coping response to the organism’s physical or social context. Thus, rather than being interpreted as a pathogenic response to negative experience, high reactivity would index heightened biological sensitivity to the quality of the environment, a characteristic with demonstrable adaptive features in certain rearing environments. Studies in rats and monkeys, for example, have revealed that negative neurobiological effects of low quality maternal care can actually contribute to adaptive physiologic and behavioral responses under future stressful conditions (Lyons, Parker, & Schatzberg, 2010; Oomen et al., 2010). Human children exhibiting muted or augmented reactivity in adverse contexts might be responding adaptively to chronic stressors, and although such physiological system modifications may do harm in the long run (e.g., allostatic load effects on pathology), they likely allow for improved functioning over the short term. Modifications such as enhanced amygdala reactivity to neutral stimuli (Tottenham & Sheridan, 2010) may also be a favorable response for children living in a high crime neighborhood, protecting one from physical harm albeit increasing risk for affective dysregulation. As such, variation in physiological outcomes sometimes found for minorities (e.g., Bush et al., 2011), relative to majority population levels, may be legitimately adaptive to contextual demands and might also reflect unique population characteristics that have other value for those populations (Garcia-Coll, Akerman, & Cicchetti, 2000).

122

Differential Sensitivity to Context: Implications for Developmental Psychopathology

Particularly when considering the evolutionary perspective of differential susceptibility, which we describe in more detail below (see also Belsky & Pluess, 2013), it may not be appropriate to think in terms of positive or negative outcomes, but rather alternative strategies by which organisms attempt to enhance their capacity for survival, given the ecological conditions in which they find themselves. Such a reactivity profile may work well in some setting but not others (e.g., crime-ridden neighborhood vs. classroom) and at some points in time but not others (e.g., successful in early childhood, but problematic as a child is expected to gain independence). Readers may also be interested in a new manuscript by Ellis and Del Giudice (2014) in which they outline the similarities and differences between the adaptive calibration model (ACM) and allostatic load models and articulate the value of long-term adaptive changes in biobehavioral systems, such as high reactivity phenotypes, that might otherwise be seen as dysregulated. Increased effort to discuss susceptibility factors and phenotypes with less deterministic wording and more context- and time-specific contingencies will likely prove to advance knowledge and prediction. Evolutionary Thinking About Variation in Sensitivity The origins, dimensions, and consequences of individual differences in phenotype are emerging as essential components in a full understanding of the biology of experience. Biology is teeming, of course, with both between- and within-species variation that bears convincing witness to the evolutionary uses of diversity, and elegant ethological and epigenetic work, such as that by Meaney (2010), McGowan and Szyf (2010), and Suomi (2006), reveals the adaptive benefits of phenotypic diversification. For example, the maternally and epigenetically regulated differentiation of rat pups’ adrenocortical responsivity produces a range of low- to high-reactivity phenotypes, each of which may maximize survival and fitness within particular early life and later life environments (Weaver et al., 2004). Similarly, neither the aggressively uninhibited nor the shy, neophobic phenotypes of young rhesus macaques can be warranted as normal or optimal; rather, each has adaptive value within specific social and physical contexts (Cirulli et al., 2009; Stevens, Leckman, Coplan, & Suomi, 2009). What is salient and important about phenotypic variants is their capacity for enhancing fitness within the diversity of species-typical environments encountered. Human variation in sensitivity to social environmental influence is but one illustrative case. Evolutionary arguments for the importance of these sensitivity polymorphisms or phenotypes suggest that the

DNS gene variants with the greatest empirical support thus far (e.g., s/s allele of SERT, the 7-repeat of DRD4, and the Met allele of BDNF) appear to have emerged (most likely through mutation) and then rapidly expanded through humankind only within the past 100 millennia, now existing in up to 1/3 of the population (most likely through natural selection). It is thought that, due to population dynamics, the adaptive edge of some variants only holds if the trait remains within a minority of the population, as they become nonadaptive when common. For example, although the high sensation seeking and risk taking found in 7R DRD4 carriers in many studies may have helped drive human expansion around the globe, these traits are expensive in evolutionary terms, increase risk for negative outcomes, and would likely lead to societal failure if all members shared this phenotype (Matthews & Butler, 2011; Wilson, 1998). Thus, populations of individuals possessing the fullest range of phenotypic variation, at both the individual and group levels, are likely to have the best odds of adapting to the constantly varying social, physical, and chemical environment of our world. However, it is also important to note that evolutionary frameworks emphasize adaptations that will improve fit in terms of survival and ability to successfully reproduce into the next generations. Other fields interested in understanding human experience and life course trajectories emphasize the impact of those adaptations on mental and physical health, which may have a different time course (e.g., occurring well beyond reproduction) or occurring and affecting well-being despite not significantly changing reproductive trajectories. Thus, although evolutionary frameworks are quite important to a deepening understanding of DNS, they are only part of the equation when considering the broad range of processes and outcomes related to DNS. Developmental Timing and DNS It is clear that the relevance and strength of these interactive associations depend very much upon the developmental time at which they are examined. Thus, examination of this DNS phenomena is enhanced when one considers the central role of time—evolutionary, historical, developmental, neurogenomic, and neurophysiological—in determining phenotypic variation and whether it is adaptive (Boyce, Sokolowski, & Robinson, 2012). Much of the biological sensitivity to current social contexts has resulted from response predispositions established, selectively and epigenetically, through adaptations to the temporally distant environments of early hominids (Dubos, 1965; Nesse & Young, 2000). Social disparities in health—products, in part, of the social, economic, and health policies of

Conceptual and Methodological Issues for Examinations of Differential Neurobiological Susceptibility

contemporary societies—wax and wane within historical time according to the era’s dominant sociopolitical philosophies (Beckfield & Krieger, 2009; Krieger, 2001). Developmental time is uneven in its potency, intensity of change, and accessibility to environmental influence. Thus, for example, critical or sensitive periods exist for the acquisition of language and the discrimination of speech sounds in human infants (Weikum, Oberlander, Hensch, & Werker, 2012), and exposure to music can change auditory preferences in young mice through changes in prelimbic and infralimbic medial prefrontal cortex during an early critical period (Yang, Lin, & Hensch, 2012). It also appears that early trauma and early SES have a greater effect on adult pathology than does later experience of those phenomena (see, for examples Adler, Bush, & Pantell, 2012; Heim, Plotsky, & Nemeroff, 2004; Kittleson et al., 2006), because of both the increased biological sensitivity to experience that exists early in development and because of the greater impact on trajectories of experience that early experiences can have (see Cicchetti & Toth, 2009; Gunnar & Vazquez, 2006; Meaney, 2010). Findings from animal research reveal evidence for the importance of neurogenomic time and neurophysiologic time, in that honey bees encode spatiotemporal mappings of foraging sites through differential gene expression signatures, allowing them to remember not only geographic locations of food but the circadian patterning of food availability (Naeger et al., 2011), and even fleeting social interactions between fish can trigger changes in neural firing and secretory action (Robinson, Fernald, & Clayton, 2008). Thus, at strikingly different levels of temporal resolution, time and timing appear to play crucial but not yet fully explored roles in guiding societal, organismic, and neurobiological responses to the conditions of early life (Boyce, Sokolowski, et al., 2012). Work within the multidisciplinary field of developmental psychopathology requires the use of multiple levels of analysis, considered simultaneously, and the outcomes under study generally result from transactional, interdependent co-determination among multiple levels of influence (Gottlieb & Halpern, 2002). Of course, human development is more than an interaction term between a biological and an environmental variable, as it is dynamic and ever-changing (we now know, for example, that epigenesis continues throughout the life course). As we have tried to illustrate throughout this chapter, biological sensitivity to context is not just a cause but is an outcome as well, shaped by previous social contextual experiences. This critical latter element is not yet fully appreciated, especially outside of the field of developmental psychopathology. The cumulative history of the person determines in part

123

the environmental contexts he or she finds him- or herself engaging with, reacting to, eliciting, or processing. Indeed, the organism reacts differently to the same environmental circumstance uniquely at different points in development, because it carries forward its history and the biological, cognitive and behavioral shaping that resulted from that history—thus engendering a slightly (or in some cases dramatically) different response to the same circumstances. Prenatal Exposures The prenatal environment is a relatively underexplored context and period that increasing evidence suggests plays a major role in DNS. As described already, susceptibility or plasticity can be shaped by the environment, and an organism’s first environmental exposure occurs at conception. Although nurture is often thought of as a postnatal experience, animal and human studies building upon Barker’s prenatal programming hypothesis (Barker & Sultan, 1995; Gluckman & Hanson, 2006) demonstrate that maternal stress hormones, nutritional cues, and other biological signals cross over the placental barrier to influence offspring phenotype, with the assumption that such information transfer is designed to support the optimal adaptation to the world into which the fetus will be born (for a review see Pluess & Belsky, 2011). For example, maternal psychosocial stress and stress hormone levels during pregnancy have been shown to predict infant birth weight (Worthman & Kuzara, 2005) (which has been linked to numerous neurodevelopmental outcomes, including difficult temperament (Pluess & Belsky, 2011)), child temperament (Davis et al., 2005) and behavioral recovery to a stressor (Davis, Glynn, Waffarn, & Sandman, 2011), HPA-axis regulation (Davis et al., 2011; Owen, Andrews, & Matthews, 2005) and stress-related gene methylation (Oberlander et al., 2008). Davis and colleagues (Davis et al., 2011) provide an excellent example of these effects in their finding that exposure to higher levels of maternal plasma levels of cortisol during the second and third trimesters of pregnancy predicted larger infant cortisol responses to the painful stress of a heel-stick procedure 24 hours after birth. Additionally, elevated levels of maternal cortisol early in pregnancy as well as prenatal maternal psychosocial stress throughout gestation predicted a slower rate of behavioral recovery to the heel-stick. Importantly, these findings could not be explained by socioeconomic status, child race or sex, or medical factors such as prenatal medical history, mode of delivery, or birth order. These prenatal programming effects appear to last across development. For example, a prospective, longitudinal study of healthy mother-child dyads found that higher maternal cortisol in early pregnancy was associated

124

Differential Sensitivity to Context: Implications for Developmental Psychopathology

with larger right amygdala volume in female offspring at age 7 and that this effect partially mediated the association between maternal stress during pregnancy and girls’ affective problems at age 7 (Buss et al., 2012). Additionally, these prenatal programming effects appear to last across development (e.g., Entringer, Kumsta, Hellharnmer, Wadhwa, & Wust, 2009). Such programmed physiological and behavioral phenotypes then moderate environmental effects on those offspring’s developmental trajectories. It is intriguing to note that a fetus’ own genetic sensitivity to environmental experience may make one fetus more sensitive to the uterine environment than another, adding a layer of DNS complexity to the understanding of the programming of DNS. Researchers have begun to hypothesize about individual difference characteristics that make some more or less susceptible to the prenatal environment. For example, in a first examination of EEG data within a DNS framework, Fortier and colleagues (Fortier et al., 2014) recently postulated that asymmetry in alpha-activation of the frontal lobe may be a marker of DNS. They highlight that frontal EEG alpha-activation asymmetry has a moderately heritable basis, is present in utero, and fairly stable across the life course. Although the increased risk for negative affectivity in those with greater activation of the right hemisphere (right frontal asymmetry; RFA) might be seen as the more sensitive phenotype, the authors posited that the variability in positive and negative outcomes across the literature for those with greater activation in the left hemisphere (left frontal asymmetry; LFA) was evidence for that phenotype being the more susceptible to context. Indeed, the authors found that LFA individuals were sensitive to prenatal environment (reflected in low vs. normal birth weight) whereas RFA individuals were not. Specifically, LFA individuals who were normal birth weight (presumably reflecting optimal intrauterine conditions for development) evidenced the best mental health outcomes a decade later, and those who were extremely low birth weight (presumably reflecting adverse intrauterine conditions) evidenced the highest attention problems and withdrawn behaviors. Prenatal conditions reflected in birth weight had no association with later life problems for RFA individuals. The patterns of interaction found in this study reflect DNS for better and for worse outcomes. New research is also attempting to span the prenatal programming of offspring reactivity literature with the DNS hypotheses. For example, a Swiss study examined the interplay between the prenatal environment, as indicated by maternal blood plasma cortisol levels in response to the Trier Social Stress test conducted late in pregnancy, and

neonates’ temperamental reactivity, as assessed by behavioral coding of newborns’ neurological and behavioral reactivity to stimulation 10–14 days after birth (Bolten et al., 2013). Consistent with DNS theories, newborns that were most reactive demonstrated the best emotion regulation during the stressful Still Face Paradigm, if they had gestated in mothers with lower HPA axis reactivity, but they had the lowest emotion regulation scores if they were born to mothers with high HPA axis reactivity during pregnancy. Although newborn reactivity has been linked to maternal HPA axis functioning during gestation, these results point to the possibility that variability in newborn reactivity (presumably stemming from reactivity developed during the fetal stage) may also serve as a susceptibility factor to maternal reactivity in utero (and presumably her reactivity phenotype postnatally), in the prediction of developing emotion regulation. Another research group has also explored similar associations and find that infant’s carrying the sensitive polymorphism of the 5-HTTLPR gene had higher levels of emotional dysregulation in early childhood if their mothers were depressed during gestation but the lowest levels of emotional dysregulation if their mothers had lower levels of prenatal depression (Babineau et al., 2015). These interaction findings held across multiple longitudinal time points and were robust to adjustments for postnatal factors such as maternal postnatal depression, but they are still limited by the shared reporter of maternal mood and offspring emotional regulation. Although both sets of these findings require replication and careful consideration of the timing of exposures and effects, they provide fodder for discussion of the complexity of the development and manifestation of DNS and highlight the opportunities for discovery when the prenatal period is examined.

BIOLOGICAL PATHWAYS LINKING EARLY LIFE DIFFERENTIAL SUSCEPTIBILITY TO LATER PSYCHOPATHOLOGY Throughout this chapter, we have focused on the accumulating evidence for the manner in which biological sensitivity can moderate the effects of context on developing physiology and behavior. These findings add to the growing body of literature supporting a differential susceptibility model of biology × environment interactions in developmental psychopathology. It remains to be seen whether the differential effects on behavior are mediated by individuals’ neurobiological changes, given the far less empirical evidence for biological processes linking early

Biological Pathways Linking Early Life Differential Susceptibility to Later Psychopathology

adversity with subsequent psychopathology (Pollak, 2005). In rat and nonhuman primate studies, adverse early life exposures, such as maternal separation or low quality maternal care, have been shown to result in subsequent HPA axis, autonomic, and immune system changes in response to stress, as well as marked behavioral changes in adulthood such as anxious behavior, anhedonia, social dysfunction, predisposition to alcohol abuse, decreased appetite and sleep disruptions (Dettmer et al., 2012; Liu et al., 1997; Plotsky & Meaney, 1993; Provencal et al., 2012; Sanchez, Ladd, & Plotsky, 2001). Although animal research lends itself to the rigorous study of causal environmental effects on biology, it has limitations in its ability to model human experiences of early life stress or developmental psychopathology. On the other hand, researchers studying children with mental health problems find it challenging to incorporate biological assessment into their studies and to follow individuals for the long periods of time required to assess these life course processes. Burghy and colleagues provide a particularly useful study that bridges several of the concepts reviewed above (Burghy et al., 2012). Their findings suggest that stress exposure during infancy increases adolescent depression and anxiety and decreases amygdala-vmPFC resting-state connectivity via cortisol elevations in childhood. In a parallel fashion, the Buss et al. work reviewed above (Buss et al., 2012) found that stress exposure during gestation, via maternal circulating cortisol levels, was associated with greater affective problems, and this association was mediated in part by amygdala volume. Both studies also reveal important gender differences for this vulnerability to dysregulated affect, as well as developmental sensitive periods for these effects. The relatively small sample sizes of youth in both studies merits caution for generalization of the findings, but the rich breadth of longitudinal data, strong theory-driven hypotheses, and sophisticated statistical modeling in those studies make them exemplars of the advances being made in this area. A Closer Consideration of Epigenetic Pathways The work just reviewed spans several biological systems and levels of inquiry into neurobiological susceptibility. However, given the tremendous increase in attention given to the field of epigenetics over the past decade, and the promise it holds for understanding human development and disease, we examine that literature in more detail here. In short, epigenetics (from the Greek root epi, meaning upon or over) has been defined as “the structural adaptation of chromosomal regions so as to register, signal

125

or perpetuate altered [gene] activity states” (Bird, 2007, p. 398). Epigenetic mechanisms change gene expression or activity through alterations in chromatin organization, but they do so without changing the genetic code of the DNA (Meaney, 2010). Thus, the epigenome can be viewed as a responsive overlay on the genome itself, and it possesses the potential to moderate the effects of genetic variation through up- and down-regulation of gene transcription. Before we review evidence for such processes, it may be helpful to review some of the basic biology involved. Chromatin configuration is controlled by a physicochemical process in which chemical tags (i.e., epigenetic marks, such as methyl, acetyl, phosphate, or ubiquitin groups) are placed on or removed from the DNA or histone proteins. DNA methylation is the most highly studied and best characterized epigenetic mark. Methylation involves a direct covalent, chemical modification of a cytosine base lying sequentially adjacent to a guanine base (thus a CpG dinucleotide), and is catalyzed by a group of enzymes called DNA methyltransferases (DNMTs). CpG methylation can occur during any stage of the cellular life cycle and, depending upon genomic context, can impede or foster gene expression or leave it unchanged (Klengel, Pape, Binder, & Mehta, 2014). Promoter DNA methylation (often in CpG islands, areas of the genome with comparatively high numbers of CpG dinucleotides) and gene body DNA methylation typically have contrasting associations with gene expression. High levels of promoter DNA methylation are frequently linked with diminished gene expression, but high levels of methylation in the gene body are more typically associated with increased gene expression (Jones, 1999; Kass, Landsberger, & Wolffe, 1997). DNA methylation is a relatively stable epigenetic tag, yet it has just recently become viewed as reversible and thus a potential mechanism of developmental plasticity (Wu & Zhang, 2014). Epigenetics and the Mediation of Differential Susceptibility to Environmental Influence Interactions All disorders of health and development can be seen as uniformly both genetic and environmental, in the sense that virtually all endpoints depend upon mutually interactive influences of both (Rothman & Greenland, 2005). The epigenetic marks and modifications controlling gene expression are gaining recognition as the molecular mechanisms underlying GxE interaction effects on developmental and mental health outcomes. Accordingly, these mechanisms constitute a tangible, physical nexus between environments and genes, between nurture and nature, between the exogenous and endogenous determinants of human development and psychological well-being.

126

Differential Sensitivity to Context: Implications for Developmental Psychopathology

Several recent leading edge discoveries support the perspective that epigenetic processes are molecular mechanisms for GxE at the population level. In a sample of 192 adults, Klengel and colleagues (Klengel et al., 2013) found that a functional polymorphism in the FK506 binding protein 5 (FKBP5) gene, an important functional, cytoplasmic regulator of the glucocorticoid receptor complex, interacted with childhood trauma exposure to predict the risk of developing stress-related psychiatric disorders in adulthood in a manner consistent with DNS. Although the interaction between FKBP5 and childhood maltreatment was known, their findings provide the first molecular evidence in humans that this type of GxE effect may be mediated via allele-specific, childhood trauma–dependent DNA demethylation in functional glucocorticoid response elements of FKBP5. Specifically, this research group found that demethylation was linked to increased stress-dependent gene transcription, followed by a long-term dysregulation of the stress hormone system and a global effect on the function of immune cells and brain areas associated with stress regulation. The authors suggest that, when occurring during developmentally sensitive periods, these epigenetic changes remain stable over time and further amplify the differential susceptibility to environment that the individual’s genetic makeup originally bestowed upon it—particularly through long-term changes in the stress hormone system regulation and alterations in neuronal circuits. In turn, the early life demethylation event results in a consequent augmentation of risk for PTSD across the life course. Although gene expression is likely controlled by a variety of molecular mechanisms, only one of which is epigenetic marks, this empirically driven intersection of the fields examining genetic and epigenetic variation offers one groundbreaking account of how and under what conditions GxE interactions arise. Work from the Holbrook lab (Teh et al., 2014) approached examination of this complex phenomena from the opposite direction. Their results demonstrate how neonatal human methylomes are affected by both unique intrauterine exposures and fixed genetic variation, resulting in GxE interactions during gestation. Specifically, within a sample of more than Singaporean newborns, the investigators found over 1,400 punctate genomic regions that were highly variable in methylation status across individuals— locations termed variably methylated regions (VMRs). Using birthweight as a proxy for environmental factors (e.g., nutrition, smoking, maternal depression, and maternal body mass index during gestation), they examined genetic and GxE origins of VMRs. Strikingly, GxE interactions were found to account for 75% of VMRs,

through a combination of intrauterine environmental signatures on the fetal epigenome and the physicochemical effects of sequence variation on a CpG site’s propensity for methylation. Of note, no VMRs were best accounted for by environmental linkages alone, independent of gene sequence variation. Allowing essentially the same inference as that made by Klengel and colleagues (2013), Teh et al.’s results reveal an increasingly convergent picture of how DNA sequence variation interacts with naturally occurring (and sometimes pernicious) variation in social environmental exposures at the molecular level, through chromatin modifications and other transcription regulatory processes, to produce complex developmental and health trajectories. Kobor and colleagues provide additional recent data that sheds light on these complex epigenetic processes. Their series of studies examining allelic and epigenetic variation within human populations demonstrates that human populations possess highly divergent DNA methylation profiles and that the sources of divergences between populations involve differences in allelic frequencies and complex GxG and GxE interactions (Fraser, Lam, Neumann, & Kobor, 2012). Examining over 14,000 genes in 180 different cell lines from European and African samples, they found population-level differences in DNA methylation near transcription start sites in over a third of genes. Further analyses indicated that these methylation differences were mostly attributable to differences in allele frequencies. The Kobor lab’s work augments the existing small evidence base for epigenetic dissimilarities between human populations (Krushkal, Tylavsky, & Thomas, 2011; Heyn et al., 2013; Moen et al., 2013) and points to the need for considering population heritage in epigenetic analyses but also for recognizing the manner in which population-level genetics drive variation in the epigenome. Considered together, these three recent studies and related observations suggest that considerable developmental variation, particularly when it places individuals at risk for developmental psychopathology, may be linked to interactions between genetic and environmental differences. Moreover, they add to the emerging evidence implicating epigenetic processes as molecular-level mechanisms by which GxE interactions may occur. In light of these findings, it seems that the increasingly apparent statistical interactions identified between allelic variants and risk-engendering early social environments may be rooted in and attributable to differences in DNA methylation, post-translational histone modifications, or other epigenetic marks, which contribute to inter-individual

Future Directions

variation in the transcription of genes linked to pathological phenotypes or endophenotypes. Variation in Epigenetic Marks as a Result of Differential Susceptibility to Context A related, important question is whether epigenetic modifications, acquired as a consequence of early environmental signaling, might also be linked to differentially susceptible phenotypes. For example, researchers have studied large, differentially methylated regions centered upon the NR3C1, GR gene in the hippocampi of both rats and humans experiencing substantially different levels or forms of early parental care (Suderman et al., 2012). Suderman and colleagues found that the methylation profiles of both species were extensively different in individuals receiving high vs. low level (rats) or abusive vs. nonabusive (humans) early parental care. They also found many between-species commonalities in the specific, differentially methylated sites. Examining stressors associated with socioeconomic environment within a sample of African American youth from working poor communities, Beach and colleagues (Beach et al., 2014) found that cumulative socioeconomic adversity and the S-allele of the 5-HTTLPR, serotonin transporter gene interactively predicted promoter region methylation within a group of more than 200 genes associated with depression. Youth who were S-allele carriers demonstrated either the highest or lowest levels of depression-related gene methylation, depending upon levels of exposure to poverty-related stress. Recently, researchers have argued that the increased susceptibility of infants to infectious agents of disease may be due to differential methylation of immune-regulating and other developmentally salient genes (Strunk, Jamieson, & Burgner, 2013). Finally, the Binder laboratory work (Klengel et al., 2013) demonstrated a differential susceptibility of individuals bearing the AG/AA risk allele of the FKBP5 gene in that such individuals have either higher or lower rates of adult PTSD, conditional upon childhood exposures to sexual or physical abuse. Further, their work demonstrated that the molecular process by which this epidemiologically observed interaction occurs is mediated through DNA demethylation in the glucocorticoid response elements of FKBP5. These observations are among the first to show how chromatin modification and epigenetic marks may constitute the actual molecular mechanisms for a differential susceptibility to environmental conditions, yet much remains to be elucidated. The proliferation of epigenetic publications and reports represents an epidemic of interest in this area, although

127

the tremendous initial enthusiasm with which the work was greeted by the scientific community has more recently been tempered by increasing awareness of the limiting realities that commonly characterize the rapid emergence of new fields and technologies. Not only is it inappropriate to expect epigenetic processes to explain all biological phenomena, but sometimes they also fail to reliably explain phenomena within systems or questions it is most reasonable to assume they should. Yet, despite the need to face these scientific realities, the research domain holds at least implicit promise for illuminating one of the oldest and deepest mysteries of human experience (i.e., how individual susceptibility and social conditions work together)—at the behavioral, physiologic, neural, cellular and molecular levels—to initiate and sustain disorders of development, behavior, and health.

FUTURE DIRECTIONS The body of work reviewed in this chapter is advancing at a rapid pace, and the opportunities for new research are vast and compelling. Although the developmental sequelae of early experience are well documented, current understanding of the individual-specific (moderated) manner in which those experiences get under the skin to affect psychopathology is quite limited, leaving much for the next generation to uncover. A recent, potentially transformative shift in conceptualization of developmental psychopathology and funding for discovery may provide stimulus for this work. The National Institute of Mental Health’s recent promotion of Research Domain Criteria (RDoC; Cuthbert & Insel, 2013; Insel et al., 2010) focuses on investigating the etiology and treatment of mental disorders as biological disorders that require advanced understanding of breakthroughs in genetics, physiology, and molecular, cellular and systems neuroscience. This movement away from diagnostic symptom criteria brings about additional reasons to increase attention to the range of ways variation in biology confers sensitivity to environment. As it is more likely that differential susceptibility to environments affects a broad range of physical and mental health phenotypes and outcomes (i.e., multifinality; Cicchetti & Rogosch, 1996), investigation of DNS is likely to benefit from this greater focus on intermediate phenotypes (heightened physiologic reactivity, patterns of epigenetic methylation, reactivity in regions of the brain, etc.) for a range of health outcomes, rather than predicting disease classifications.

128

Differential Sensitivity to Context: Implications for Developmental Psychopathology

As the fields reporting various forms of interactions between biologic predispositions and environments exponentially grow, what is now critically needed are programs of research examining the questions of how and by what mechanisms genes and early social contexts codetermine trajectories of behavioral and biological development. With respect to differences in complex behavior and its disorders, a focus on proximal, neurobiological processes must come strongly to the fore. With advances in neuroscience and human molecular genetics, next steps include identifying common variation in the genes and their expression that influence the functioning or availability of components in these stress–disease pathways. Illustrations of such approaches can be found in recent work within emerging imaging genomics (e.g., Hariri et al., 2005; Pezawas et al., 2005). A particularly groundbreaking example of progress in this realm that is specific to differential susceptibility research is the epigenetic work reported by Klengel and colleagues (Klengel et al., 2013) described earlier. Such sophisticated identification of molecular mechanisms for genotype-directed, long-term environmental reactivity is particularly informative, and further determination of how early in life such epigenetic modifications occur may elucidate the molecular origins of differential susceptibility and the deeper etiologies of developmental psychopathology. Future work must also advance theory. For example, there is the need for research to determine whether these behavioral, physiological, and genetic markers of susceptibility to context constitute the same phenomena expressed at different levels of scale or complexity or represent different forms of susceptibility (Obradovi´c & Boyce, 2009). To date, DNS studies typically focus on sensitivity in one level or system (albeit sometimes across various markers in that level), making it difficult to answer this question. Exceptions include work from our lab, such as that examining reactivity in both the HPA axis and within the parasympathetic nervous system to assess sensitivity to family adversity in the prediction of a variety of adjustment outcomes (Obradovi´c et al., 2010). Although the patterns of findings in that study were cohesive across the biological systems and various outcomes, DNS interaction findings were not exactly replicated across the two biological systems. Although this might be the result of measurement or sample variability, such discrepancies suggest unique roles for sensitivity in each system, or at least unique timing for the role of sensitivity to context. In parallel, Belsky and Pluess’s (2013) recent theory describing individuals as possessing plasticity phenotypes, rather than just system- or gene-specific plasticity, or acquiring cumulative plasticity,

are important ideas demanding empirical testing and expansion. A major impediment to uncovering the relations among adversity, biological sensitivities, and psychopathology is that the majority of behavioral and emotional pathologies are likely to result from constellations of aberrant biological systems, and multisystem studies and longitudinal analyses are rare. Considerable recent attention has been given to early adverse experience effects on the complex and varied neurobiology of depression (Bogdan & Hariri, 2012; Heim et al., 2004). For example, progressive effects on developing stress physiology might explain findings supporting the kindling hypothesis, in which an increase in spontaneous dysregulation or a lowering of the threshold for experiencing life stressors leads to easier triggering of depressive episodes (Monroe & Harkness, 2005). As precise, biologically plausible and multisystem models of psychopathology are developed (Pollak, 2005), our ability to seek and understand the manner in which early experience manifests as health and pathology will be greatly enhanced, as well. To this end, some work from our laboratory and those of colleagues is beginning to indicate ways we might use simultaneous measurement of reactivity across the HPA, sympathetic and parasympathetic systems to better understand developmental psychopathology (Alkon et al., 2003; Bauer, Quas, & Boyce, 2002; Quas et al., 2014). In particular, Quas et al. found six distinctive patterns of cross-system reactivity in a multicohort sample of 664 children with anticipatory and stress reactivity measures across the parasympathetic and sympathetic nervous systems and the hypothalamic-pituitary-adrenal axis. Although the majority of children in their samples displayed moderate reactivity indicated by an average level of cross-system reactivity, five other patterns also emerged: parasympathetic-specific reactivity, anticipatory arousal, multisystem reactivity, hypothalamic-pituitary-adrenal axis specific reactivity, and underarousal. Of note, groups meaningfully differed in socioeconomic status, family adversity, and age, suggesting important roles for developmental timing and contextual exposures in the shaping of these anticipatory and reactivity profiles. Surprisingly little attention has been given to delineating precisely how these systems interact with each other when confronted with stress, and the Quas et al. findings highlight the importance of attention to this issue. The lack of attention to date may be due to the difficulty and cost associated with cross system data collection, particularly in children; however, it is likely also the result of the dearth of theory addressing this issue. For example, although much

Future Directions

argument has been made for the importance of multisystem models, theoretical models that argue for the existence and importance of cross-system patterns of physiological reactivity to challenge or stress have yet to elaborate on how more than two biological systems operate in conjunction with one another. For example, the influential work of Berntson, Cacioppo, and colleagues (e.g., Berntson, Cacioppo, Quigley, & Fabro, 1994) does not comment on the role of HPA axis activation in their detailed account of autonomic nervous system indices, despite the importance of the HPA axis arousal and regulation in the activation of other stress-response systems. Similarly, the work by Bauer and colleagues (Bauer et al., 2002) did not consider the regulatory function of the parasympathetic system. The Adaptive Calibration Model (ADM) described above (Del Giudice et al., 2011), makes progress in this regard in that it theorizes about varying patterns of baseline arousal and stress-induced activation of both branches of the ANS as well as the HPA axis; however data examining the 3-system patterns they propose is lacking. A primary reason for the varied findings across DNS studies (those published and those left in the file drawer) may be the lack of attention to responses across multiple systems. Increased theoretical and empirical attention to these simultaneously occurring physiological phenomena is likely to clarify DNS and will certainly advance understanding of the development of psychopathology given the importance of both main and interactive effects of physiological arousal on pathology. Successful pursuit of advancement in the science of differential susceptibility within developmental psychopathology will surely also involve advanced methodology and more complex analyses. One example is that of Davies and Cicchetti (2014), who utilized structural equation modeling to test mediated moderation models to reveal that children’s angry reactivity to maternal negativity partly accounted for the greater susceptibility of homozygous L carriers of the 5-HTTLPR polymorphism to variations in maternal responsiveness. Additional sophisticated approaches include those incorporating the perspective of complex adaptive systems (Galea, Riddle, & Kaplan, 2010). Traditional epidemiological strategies for understanding the health effects of social environmental factors involve the ascertainment of such factors’ independent influences on a health end point through the use of multiple hierarchical regression models (Diez Roux, 2007). Although such an approach allows estimation of the isolated effects of single independent variables, it belies the reality that most human disorders are etiologically complex, with multiple interacting causes. Even detecting GxE interactions likely underserves the true

129

complexity of pathogenic processes, because allelic variation in a single gene likely interacts with polymorphisms in several other genes and multiple dimensions of the environment may also interactively influence outcomes. In such circumstances—circumstances that may eventually prove to predominate in disease causation—the use of more sophisticated models and analytical tools will be required to understand the multiply interactive networks of risk factors involved in the ontogeny of disordered development and health. If so, one such approach with increasingly demonstrable efficacy is the use of complex systems analysis, involving descriptive inventories of system components, nonlinear mathematical modeling, and the construction of agent-based models of causal networks (Diez Roux, 2007; Galea, Hall, & Kaplan, 2009). The apparent interaction between biological processes and contextual influences reviewed here points to the value of integrating biological measures into the design and evaluation of preventive interventions (Cicchetti & Gunnar, 2008). Moffitt (2013) elegantly articulated the need for and value of bringing stress–biology research and clinical intervention science together in the case of children exposed to violence, but the review and advocacy for the addition of a range of biomarkers to the study of violence-exposure interventions applies well across a variety of clinical contexts. Increasingly, the literature points to the potential for the inclusion of measures of HPA axis activation, inflammation, telomere length, gene expression, and epigenetic methylation. Given their increasing ease of collection and decreasing cost for assay, biological markers such as genotype and hair cortisol are beginning to prove useful for detecting who is most affected by interventions and whether those interventions affect biological systems related to later susceptibility. As costs go down for other neurobiological assessments, feasibility of using sophisticated biomarkers to advance our crafting and evaluation of interventions for psychopathology will improve, and the field will advance. Additionally, these endeavors, and the others described in this section should take account of the breadth of positive environments and buffering factors that operate across the life span, which will take us a step further in understanding healthy behavioral and emotional development. Just as use of genetic information has played a major role in certain aspects of precision medicine (National Research Council, 2011), determination of sensitivity to context in conjunction with advances in technology, methodology, and biological discovery will enable a level of personalization in the prediction and treatment of psychopathology that was not previously feasible or

130

Differential Sensitivity to Context: Implications for Developmental Psychopathology

practical. It is likely that failure to assess for differential susceptibility in most environmental intervention trials (e.g., parenting, family, classroom, neighborhood) has led to erroneous or blunted conclusions about the efficacy of certain approaches and a misestimation of treatment effects. Although it would be ideal to provide all children with optimal or enriched environments, practicality often demands targeted interventions to those at greatest risk and those most likely to be sensitive to intervention, as well as clarity about which types of contextual interventions are most beneficial for whom, which is made more feasible as we better understand differential susceptibility. On the other hand, there are circumstances in which understanding sensitivity to environmental factors mandates a public health intervention that improves well-being for the entire population. For example, the discovery that certain individuals possess genetic variants that make them particularly susceptible to the harmful effects of lead exposure (Astrin et al., 1987) led to improved federal guidelines and regulations for lead-safe practices, thereby reducing exposure and toxicity risk for everyone. One can imagine how identifying individuals who are particularly susceptible to toxic effects of a variety of chemical and social exposures might promote public attitudes and policies that would limit risk for these individuals and their fellow community members.

CONCLUSIONS A remarkable diversity of early, social environmental dimensions has been linked to important differences in mental and physical health, trajectories of development, and individual differences in behavior. As attention to differential susceptibility to those environmental dimensions grows, the capacity to place a finer, more exacting point on the specific kinds of environments that interact with particular variants of biological sensitivity will be important and likely illuminating. Indeed, studies examining the developmental biology of social experience show that we are on the cusp of providing deep mechanistic explications of the important, earlier insights of Waddington (1966), Gottlieb (1991) and others that organismic development is guided by the combined, interactive influences of biological predispositions and experiences. Conducting scientific inquiry related to interactions between individually varying sensitivity variables and the range of environments we are exposed to requires consideration of a broad range of theoretical and analytic possibilities for those patterns of association. Such openness improves our ability to determine the true nature

of these associations and better predict outcomes and optimal interventions. Rather than approaching these analyses with post hoc explanations for patterns of findings, researchers should endeavor to strongly consider the existing evidence base, evolutionary theories, and biological processes involved prior to developing hypotheses. However, if appropriate methods and statistics are used, post hoc analyses can disaggregate variability in responses to experiences and environments attributable to diathesis stress, differential susceptibility, vantage sensitivity, or some other form of host response singularity (e.g., Belsky, Pluess, & Widaman, 2013; Roisman et al., 2012). This approach has rarely been taken to date, but would greatly advance theories and evidence in this realm. In conclusion, the information presented in this chapter supports several main points. First, the data support reframing the conventional focus on individual risk and resilience factors toward a conceptualization of developmental psychopathology that considers whether certain neurobiological factors confer differential susceptibility to both positive and negative environmental experiences. Second, there is a remarkable convergence among theories and supporting evidence from fields (e.g., pediatrics, developmental psychology, evolutionary anthropology, biological psychology) and scientific laboratories originally unknown to each other (e.g., the compelling work of Klengel et al. 2013, was conducted by researchers largely unfamiliar with the DNS literature) toward this same conclusion, suggesting a shared scientific value of this reframing. Additionally, beyond aiding an understanding of etiology of psychopathology, a DNS framework expands the possibilities for understanding biological and environmental factors contributing to the full range of behaviors across the life course as well as individual difference variables that predict intervention success. In reviewing this literature, we have hoped to stimulate a new generation of research illuminating the manner in which early life contexts interact with children’s biological proclivities and biases to shape development and predict health and disorder. We hope that this work will support powerful insights into children’s typical and atypical development, highlight clear opportunities for the prevention of psychopathology, and play a formative role in newly individualized approaches to precision medicine. REFERENCES Adkins, R. M., Krushkal, J., Tylavsky, F. A., & Thomas, F. (2011). Racial differences in gene-specific DNA methylation levels are present at birth. Birth Defects Research Part A: Clinical and Molecular Teratology, 91(8), 728–736. doi: 10.1002/bdra.20770

References Adler, N., Bush, N. R., & Pantell, M. S. (2012). Rigor, vigor, and the study of health disparities. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Suppl 2), 17154–17159. doi: 10.1073/pnas.1121399109 Alkon, A., Goldstein, L., Smider, N., Essex, M., Boyce, W., & Kupfer, D. (2003). Developmental and contextual influences on autonomic reactivity in young children. Developmental Psychobiology, 42, 64–78. Alkon, A., Lippert, S., Vujan, N., Rodriquez, M. E., Boyce, W. T., & Eskenazi, B. (2006). The ontogeny of autonomic measures in 6- and 12-month-old infants. Developmental Psychobiology, 48(3), 197–208. doi: 10.1002/dev.20129 Allison, A. L., Bush, N., Adler, N. E., & Boyce, W. T. (2013). A multi-gene by environment perspective of ADHD symptomatology in young children. Dissertation, University of New Orleans, New Orleans. American Psychiatric Association (2000). Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition: DSM-IV-TR®. American Psychiatric Publishing. Andrieu, N., & Goldstein, A. M. (1998). Epidemiologic and genetic approaches in the study of gene–environment interaction: An overview of available methods. Epidemiologic Reviews, 20(2), 137–147. Anisman, H., Zaharia, M. D., Meaney, M. J., & Merali, Z. (1998). Do early-life events permanently alter behavioral and hormonal responses to stressors? International Journal of Developmental Neuroscience, 16(3–4), 149–164. Aschard, H., Lutz, S., Maus, B., Duell, E., Fingerlin, T., Chatterjee, N., . . . Steen, K. (2012). Challenges and opportunities in genome-wide environmental interaction (GWEI) studies. Human Genetics, 131(10), 1591–1613. doi: 10.1007/s00439–012–1192–0 Astrin, K. H., Bishop, D. F., Wetmur, J. G., Kaul, B., Davidow, B., & Desnick, R. J. (1987). delta-Aminolevulinic acid dehydratase isozymes and lead toxicity. Annals of the New York Academy of Sciences, 514(3), 23–29. Babineau, V., Gordon Green, C., Jolicoeur-Martineau, A., Minde, K., Sassi, R., St-Andre, M., . . . Wazana, A. (2015). Prenatal depression and 5-HTTLPR interact to predict dysregulation from 3 to 36 months—A differential susceptibility model. Journal of Child Psychology and Psychiatry, 56(1), 21–29. doi: 10.1111/jcpp. 12246 Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to rearing environment depending on dopamine-related genes: New evidence and a meta-analysis. Development and Psychopathology, 23(1), 39–52. doi: 10.1017 /s0954579410000635 Bakermans-Kranenburg, M. J., Van, I. M. H., Pijlman, F. T., Mesman, J., & Juffer, F. (2008). Experimental evidence for differential susceptibility: dopamine D4 receptor polymorphism (DRD4 VNTR) moderates intervention effects on toddlers’ externalizing behavior in a randomized controlled trial. Developmental Psychology, 44(1), 293–300. doi: 2007–19851–030 [pii]10.1037/0012–1649.44.1.293 Barker, D. J. P., & Sultan, H. Y. (1995). Fetal programming of human disease. In M. A. Hanson, J. A. D. Spencer, & J. H. Rodeck (Eds.), Growth (Vol. 3, pp. 255–276). Cambridge, UK: Cambridge University Press. Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173–1182. Bauer, A. M., Quas, J. A., & Boyce, W. T. (2002). Associations between physiological reactivity and children’s behavior: Advantages of a multisystem approach. Journal of Developmental and Behavioral Pediatrics, 23(2), 102–113. Beach, S. R., Brody, G. H., Lei, M. K., Kim, S., Cui, J., & Philibert, R. A. (2014). Is serotonin transporter genotype associated with epigenetic susceptibility or vulnerability? Examination

131

of the impact of socioeconomic status risk on African American youth. Development and Psychopathology, 26(2), 289–304. doi: 10.1017/s0954579413000990 Beckfield, J., & Krieger, N. (2009). Epi + demos + cracy: Linking political systems and priorities to the magnitude of health inequities—Evidence, gaps, and a research agenda. Epidemiologic Reviews, 31(1), 152–177. doi: 10.1093/epirev/mxp002 Belsky, D. W., Moffitt, T. E., & Caspi, A. (2013). Genetics in population health science: Strategies and opportunities. American Journal of Public Health, 103 Suppl 1, S73–S83. doi: 10.2105/ajph.2012.301139 Belsky, J. (1997). Variation in susceptibility to environmental influence: An evolutionary argument. Psychological Inquiry, 8(3), 182–186. Belsky, J. (2005). Differential susceptibility to rearing influence: An evolutionary hypothesis and some evidence. In B. Ellis & D. Bjorklund (Eds.), Origins of the social mind: Evolutionary psychology and child development (pp. 139–163). New York, NY: Guilford. Belsky, J., & Beaver, K. M. (2011). Cumulative-genetic plasticity, parenting and adolescent self-regulation. Journal of Child Psychology and Psychiatry, 52(5), 619–626. doi: 10.1111/j.1469–7610.2010.02327.x Belsky, J., & Pluess, M. (2009). Beyond diathesis stress: Differential susceptibility to environmental influences. Psychological Bulletin, 135(6), 885–908. doi: 2009–19763–005 [pii] 10.1037/a0017376 Belsky, J., & Pluess, M. (2013). Beyond risk, resilience and dysregulation: Phenotypic plasticity and human development. Development and Psychopathology, 25, 1243–1261. Belsky, J., Pluess, M., & Widaman, K. F. (2013). Confirmatory and competitive evaluation of alternative gene-environment interaction hypotheses. Journal of Child Psychology and Psychiatry, 54(10), 1135–1143. doi: 10.1111/jcpp. 12075 Bergeman, C. S., Plomin, R., Pederson, N. L., McClearn, G. E., & Nesselroade, J. R. (1990). Genetic and environmental influences on social support: The Swedish Adoption/Twin Study of Aging (SATSA). Journal of Gerontology, 45, 101–106. Berntson, G. G., Cacioppo, J. T., Quigley, K. S., & Fabro, V. T. (1994). Autonomic space and psychophysiological response. Psychophysiology, 31(1), 44–61. Bird, A. (2007). Perceptions of epigenetics. Nature, 447(7143), 396–398. doi: 10.1038/nature05913 Blair, C. (2002). Early intervention for low birth weight preterm infants: The role of negative emotionality in the specification of effects. Development and Psychopathology, 14, 311–332. Bogdan, R., & Hariri, A. R. (2012). Neural embedding of stress reactivity. Nature Neuroscience, 15(12), 1605–1607. Bolten, M., Nast, I., Skrundz, M., Stadler, C., Hellhammer, D. H., & Meinlschmidt, G. (2013). Prenatal programming of emotion regulation: Neonatal reactivity as a differential susceptibility factor moderating the outcome of prenatal cortisol levels. Journal of Psychosomatic Research, 75(4), 351–357. doi: 10.1016/j.jpsychores.2013.04.014 Boyce, W. T. (1996). Biobehavioral reactivity and injuries in children and adolescents. In M. H. Bornstein & J. Genevro (Eds.), Child development and behavioral pediatrics: Toward understanding children and health. Mahwah, NJ: Erlbaum: 35–58. Boyce, W. T. (2007). A biology of misfortune: Stress reactivity, social context, and the ontogeny of psychopathology in early life. In A. Masten (Ed.), Multilevel dynamics in developmental psychopathology: Pathways to the future, 34 45–82). Minneapolis, MN: University of Minnesota. Boyce, W. T., Chesney, M., Alkon, A., Tschann, J. M., Adams, S., Chesterman, B., . . . Wara, D. (1995). Psychobiologic reactivity to stress and childhood respiratory illnesses: Results of two prospective studies. Psychosomatic Medicine, 57(5), 411–422. Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary-developmental theory of the origins and

132

Differential Sensitivity to Context: Implications for Developmental Psychopathology

functions of stress reactivity. Development and Psychopathology, 17(2), 271–301. Boyce, W. T., Essex, M. J., Alkon, A., Goldsmith, H. H., Kraemer, H. C., & Kupfer, D. J. (2006). Early father involvement moderates biobehavioral susceptibility to mental health problems in middle childhood. Journal of the American Academy of Child & Adolescent Psychiatry, 45, 1510–1520. Boyce, W. T., Obradovi´c, J., Bush, N. R., Stamperdahl, J., Kim, Y. S., & Adler, N. (2012). Social stratification, classroom climate, and the behavioral adaptation of kindergarten children. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Supplement 2), 17168–17173. doi: 10.1073/pnas.1201730109 Boyce, W. T., Sokolowski, M. B., & Robinson, G. E. (2012). Toward a new biology of social adversity. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Supplement 2), 17143–17148. doi: 10.1073/pnas.1121264109 Braungart, J. M., Fulker, D. W., & Plomin, R. (1992). Genetic mediation of the home environment during infancy: A sibling adoption study of the HOME. Developmental Psychology, 28, 1048–1055. Brody, G. H., Chen, Y. F., & Beach, S. R. (2013). Differential susceptibility to prevention: GABAergic, dopaminergic, and multilocus effects. Journal of Child Psychology and Psychiatry, 54(8), 863–871. doi: 10.1111/jcpp. 12042 Brody, G. H., Yu, T., Chen, Y. F., Kogan, S. M., Evans, G. W., Windle, M., . . . Philibert, R. A. (2013). Supportive family environments, genes that confer sensitivity, and allostatic load among rural African American emerging adults: a prospective analysis. Journal of Family Psychology, 27(1), 22–29. doi: 10.1037/a0027829 Burghy, C. A., Stodola, D. E., Ruttle, P. L., Molloy, E. K., Armstrong, J. M., Oler, J. A., . . . Birn, R. M. (2012). Developmental pathways to amygdala-prefrontal function and internalizing symptoms in adolescence. Nature Neuroscience, 15(12), 1736–1741. doi: http://www.nature.com/neuro/journal/v15/n12/abs/nn.3257.html— supplementary-information Bush, N. R., & Boyce, W. T. (2014). The contributions of early experience to biological development and sensitivity to context. In M. Lewis & K. D. Rudolph (Eds.), Handbook of developmental psychopathology (3rd ed., pp. 287–309). New York, NY: Springer. Bush, N. R., Guendelman, M., Adler, N. E., & Boyce, W. T. (2013a). BDNF allelic variants moderate social disparities in children’s chronic cortisol expression. Manuscript submitted for publication. Bush, N. R., Guendelman, M., Adler, N. E., & Boyce, W. T. (2013b). Serotonin transporter allelic variants moderate social disparities in children’s chronic cortisol expression. Manuscript submitted for publication. Bush, N. R., Obradovic, J., Adler, N., & Boyce, W. T. (2011). Kindergarten stressors and cumulative adrenocortical activation: The “first straws” of allostatic load? Development and Psychopathology, 23(4), 1089–1106. doi: 10.1017/s0954579411000514 Buss, C., Davis, E. P., Shahbaba, B., Pruessner, J. C., Head, K., & Sandman, C. A. (2012). Maternal cortisol over the course of pregnancy and subsequent child amygdala and hippocampus volumes and affective problems. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(20), E1312–E1319. doi: 10.1073/pnas.1201295109 Cacioppo, J. T., Berntson, G. G., Sheridan, J. F., & McClintock, M. K. (2000). Multilevel integrative analyses of human behavior: Social neuroscience and the complementing nature of social and biological approaches. Psychological Bulletin, 126(6), 829–843. Caspi, A., Hariri, A. R., Holmes, A., Uher, R., & Moffitt, T. E. (2010). Genetic sensitivity to the environment: The case of the serotonin transporter gene and its implications for studying complex diseases

and traits. American Journal of Psychiatry, 167(5), 509–527. doi: 10.1176/appi.ajp. 2010.09101452 Caspi, A., McClay, J., Moffitt, T. E., Mill, J., Martin, J., Craig, I. W., . . . Poulton, R. (2002). Role of genotype in the cycle of violence in maltreated children. Science, 297(5582), 851–854. Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., McClay, J., Mill, J., Martin, J., Braithwaite, A., & Poulton, R. (2003). Influence of life stress on depression: moderation by a polymorphism in the 5-HTT gene. Science, 301(5631), 386–389. Cassidy, J., Woodhouse, S. S., Sherman, L. J., Stupica, B., & Lejuez, C. W. (2011). Enhancing infant attachment security: An examination of treatment efficacy and differential susceptibility. Development and Psychopathology, 23(1), 131–148. doi: doi:10.1017 /S0954579410000696 Cicchetti, D., & Gunnar, M. (2008). Integrating biological measures into the design and evaluation of preventative interventions. Development and Psychopathology, 20, 737–743. Cicchetti, D., & Rogosch, F. A. (1996). Equifinality and multifinality in developmental psychopathology. Development and Psychopathology, 8(4), 597–600. doi: doi:10.1017/S0954579400007318 Cicchetti, D., Rogosch, F. A., & Oshri, A. (2011). Interactive effects of corticotropin releasing hormone receptor 1, serotonin transporter linked polymorphic region, and child maltreatment on diurnal cortisol regulation and internalizing symptomatology. Development and Psychopathology, 23(4), 1125–1138. doi: 10.1017/s0954579411000599 Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50(1), 16–25. Cirulli, F., Francia, N., Berry, A., Aloe, L., Alleva, E., & Suomi, S. J. (2009). Early life stress as a risk factor for mental health: Role of neurotrophins from rodents to non-human primates. Neuroscience and Biobehavioral Reviews, 33(4), 573–585. doi: S0149–7634(08)00147–4 [pii]10.1016/j.neubiorev.2008.09.001 Conradt, E., Measelle, J., & Ablow, J. C. (2013). Poverty, problem behavior, and promise: differential susceptibility among infants reared in poverty. Psychological Science, 24(3), 235–242. doi: 10.1177 /0956797612457381 Cummings, E. M., El-Sheikh, M., Kouros, C. D., & Keller, P. S. (2007). Children’s skin conductance reactivity as a mechanism of risk in the context of parental depressive symptoms. Journal of Child Psychology and Psychiatry, 48(5), 436–445. doi: JCPP1713 [pii] 10.1111/j.1469–7610.2006.01713.x Cuthbert, B. N., & Insel, T. R. (2013). Toward the future of psychiatric diagnosis: The seven pillars of RDoC. BMC Medicine, 11, 126. doi: 10.1186/1741–7015–11–126 Davies, P. T., & Cicchetti, D. (2014). How and why does the 5-HTTLPR gene moderate associations between maternal unresponsiveness and children’s disruptive problems? Child Development, 85(2), 484–500. doi: 10.1111/cdev.12148 Davis, E. P., Glynn, L. M., Schetter, C. D., Hobel, C., Chicz-Demet, A., & Sandman, C. A. (2005). Corticotropin-releasing hormone during pregnancy is associated with infant temperament. Developmental Neuroscience, 27(5), 299–305. doi: 10.1159/000086709 Davis, E. P., Glynn, L. M., Waffarn, F., & Sandman, C. A. (2011). Prenatal maternal stress programs infant stress regulation. Journal of Child Psychology and Psychiatry, 52(2), 119–129. doi: 10.1111/j.1469–7610.2010.02314.x Del Giudice, M., Ellis, B. J., & Shirtcliff, E. A. (2011). The adaptive calibration model of stress responsivity. Neuroscience and Biobehavioral Reviews, 35(7), 1562–1592. Dettmer, A. M., Novak, M. A., Suomi, S. J., & Meyer, J. S. (2012). Physiological and behavioral adaptation to relocation stress in

References differentially reared rhesus monkeys: Hair cortisol as a biomarker for anxiety-related responses. Psychoneuroendocrinology, 37(2), 191–199. doi: 10.1016/j.psyneuen.2011.06.003 Diez Roux, A. V. (2007). Integrating social and biologic factors in health research: a systems view. Ann Epidemiol, 17(7), 569–574. doi: 10.1016/j.annepidem.2007.03.001 Dozier, M., Manni, M., Gordon, M. K., Peloso, E., Gunnar, M. R., Stovall-McClough, K. C., . . . Levine, S. (2006). Foster children’s diurnal production of cortisol: An exploratory study. Child Maltreatment, 11(2), 189–197. doi: 10.1177/1077559505285779 Drury, S. S., Gleason, M. M., Theall, K. P., Smyke, A. T., Nelson, C. A., Fox, N. A., & Zeanah, C. H. (2012). Genetic sensitivity to the caregiving context: The influence of 5-HTTLPR and BDNF val66met on indiscriminate social behavior. Physiology & Behavior, 106(5), 728–735. doi: 10.1016/j.physbeh.2011.11.014 Dubos, R. J. (1965). Man adapting. New Haven, CT: Yale University Press. Dunn, E. C., Uddin, M., Subramanian, S. V., Smoller, J. W., Galea, S., & Koenen, K. C. (2011). Research review: Gene–environment interaction research in youth depression—A systematic review with recommendations for future research. Journal of Child Psychology and Psychiatry, 52(12), 1223–1238. doi: 10.1111/j.1469–7610 .2011.02466.x Eisenberg, N., Sulik, M. J., Spinrad, T. L., Edwards, A., Eggum, N. D., Liew, J., . . . Hart, D. (2012). Differential susceptibility and the early development of aggression: interactive effects of respiratory sinus arrhythmia and environmental quality. Developmental Psychology, 48(3), 755–768. doi: 10.1037/a0026518 El-Sheikh, M. (2005). The role of emotional responses and physiological reactivity in the marital conflict-child functioning link. Journal of Child Psychology and Psychiatry, 46(11), 1191–1199. doi: JCPP418 [pii] 10.1111/j.1469–7610.2005.00418.x El-Sheikh, M., Keller, P. S., & Erath, S. A. (2007). Marital conflict and risk for child maladjustment over time: Skin conductance level reactivity as a vulnerability factor. Journal of Abnormal Child Psychology, 35(5), 715–727. doi: 10.1007/s10802–007–9127–2 Ellis, B. J., & Boyce, W. T. (2011). Differential susceptibility to the environment: Toward an understanding of sensitivity to developmental experiences and context. Development and Psychopathology, 23(1), 1–5. doi: 10.1017/s095457941000060x Ellis, B. J., Boyce, W. T., Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to the environment: An evolutionary-neurodevelopmental theory. Development and Psychopathology, 23(1), 7–28. doi: 10.1017/s0954579410000611 Ellis, B. J., & Del Giudice, M. (2014). Beyond allostatic load: Rethinking the role of stress in regulating human development. Development and Psychopathology, 26(1), 1–20. doi: doi:10.1017/S0954579413000849 Ellis, B. J., Essex, M. J., & Boyce, W. T. (2005a). Biological sensitivity to context: II. Empirical explorations of an evolutionary-developmental theory. Development and Psychopathology, 17(2), 303–328. Ellis, B. J., Essex, M. J., & Boyce, W. T. (2005b). Biological sensitivity to context: II. Empirical explorations of an evolutionary-developmental theory. Development and Psychopathology, 17, 303–328. Ellis, B. J., McFadyen-Ketchum, S., Dodge, K. A., Pettit, G. S., & Bates, J. E. (1999). Quality of early family relationships and individual differences in the timing of pubertal maturation in girls: A longitudinal test of an evolutionary model. Journal of Personality & Social Psychology, 77(2), 387–401. Engert, V., Efanov, S. I., Dedovic, K., Duchesne, A., Dagher, A., & Pruessner, J. C. (2010). Perceived early-life maternal care and the cortisol response to repeated psychosocial stress. Journal of Psychiatry & Neuroscience, 35(6), 370–377. doi: 10.1503/jpn.100022

133

Entringer, S., Kumsta, R., Hellharnmer, D. H., Wadhwa, P. D., & Wust, S. (2009). Prenatal exposure to maternal psychosocial stress and HPA axis regulation in young adults. Hormones and Behavior, 55(2), 292–298. doi: 10.1016/j.yhbeh.2008.11.006 Essex, M. J., Armstrong, J. M., Burk, L. R., Goldsmith, H. H., & Boyce, W. T. (2011). Biological sensitivity to context moderates the effects of the early teacher–child relationship on the development of mental health by adolescence. Development and Psychopathology, 23(1), 149–161. doi: 10.1017/S0954579410000702 Fisher, P. A., Stoolmiller, M., Gunnar, M. R., & Burraston, B. O. (2007). Effects of a therapeutic intervention for foster preschoolers on diurnal cortisol activity. Psychoneuroendocrinology, 32(8–10), 892–905. doi: 10.1016/j.psyneuen.2007.06.008 Fortier, P., Van Lieshout, R. J., Waxman, J. A., Boyle, M. H., Saigal, S., & Schmidt, L. A. (2014). Are orchids left and dandelions right? Frontal brain activation asymmetry and its sensitivity to developmental context. Psychological Science, 25(8), 1526–1533.doi: 10.1177/0956797614534267 Fraser, H. B., Lam, L. L., Neumann, S. M., & Kobor, M. S. (2012). Population-specificity of human DNA methylation. Genome Biology, 13(2), R8. doi: 10.1186/gb-2012–13–2-r8 Galea, S., Hall, C., & Kaplan, G. A. (2009). Social epidemiology and complex system dynamic modelling as applied to health behaviour and drug use research. International Journal of Drug Policy, 20(3), 209–216. Galea, S., Riddle, M., & Kaplan, G. A. (2010). Causal thinking and complex system approaches in epidemiology. International Journal of Epidemiology, 39(1), 97–106. doi: 10.1093/ije/dyp296 Galton, F. (1874). English men of science: Their nature and nurture. London, UK: MacMillan & Co. Garcia-Coll, C., Akerman, A., & Cicchetti, D. (2000). Cultural influences on developmental processes and outcomes: Implications for the study of development and psychopathology. Development and Psychopathology, 12, 333–356. Ghera, M. M., Marshall, P. J., Fox, N. A., Zeanah, C. H., Nelson, C. A., Smyke, A. T., & Guthrie, D. (2009). The effects of foster care intervention on socially deprived institutionalized children’s attention and positive affect: Results from the BEIP study. Journal of Child Psychology and Psychiatry, 50(3), 246–253. doi: 10.1111/j.1469–7610.2008.01954.x Gluckman, P. D., & Hanson, M. A. (Eds.). (2006). Developmental origins of health and disease. Cambridge, UK: Cambridge University Press. Goldsmith, H. H., Gottesman, I. I., & Lemery, K. S. (1997). Epigenetic approaches to developmental psychopathology. Development and Psychopathology, 9(2), 365–387. Gottlieb, G. (1991). Experiential canalization of behavioral development: Theory. Developmental Psychology, 27(1), 4–13. Gottlieb, G., & Halpern, C. T. (2002). A relational view of causality in normal and abnormal development. Development and Psychopathology, 14(3), 421–435. Grazioplene, R. G., DeYoung, C. G., Rogosch, F. A., & Cicchetti, D. (2013). A novel differential susceptibility gene: CHRNA4 and moderation of the effect of maltreatment on child personality. Journal of Child Psychology and Psychiatry, 54(8), 872–880. doi: 10.1111/jcpp. 12031 Green, J., McLaughlin, K. A., Berglund, P. A., Gruber, M., Sampson, N., Zaslavsky, A., & Kessler, R. (2010). Childhood adversities and adult psychiatric disorders in the national comorbidity survey replication i: Associations with first onset of dsm-iv disorders. Archives of General Psychiatry, 67(2), 113–123. doi: 10.1001/archgenpsychiatry.2009.186 Gunnar, M. R., Frenn, K., Wewerka, S. S., & Van Ryzin, M. J. (2009). Moderate versus severe early life stress: Associations with

134

Differential Sensitivity to Context: Implications for Developmental Psychopathology

stress reactivity and regulation in 10–12-year-old children. Psychoneuroendocrinology, 34(1), 62–75. doi: S0306–4530(08)00217–5 [pii] 10.1016/j.psyneuen.2008.08.013 Gunnar, M. R., & Vazquez, D. (2006). Stress neurobiology and developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental Psychopathology (Vol. 2, pp. 533–577). Hoboken, NJ: John Wiley & Sons. Hackman, D. A., Betancourt, L. M., Brodsky, N. L., Kobrin, L., Hurt, H., & Farah, M. J. (2013). Selective impact of early parental responsivity on adolescent stress reactivity. PLoS ONE, 8(3), e58250. doi: 10.1371/journal.pone.0058250 Hariri, A. R., Drabant, E. M., Munoz, K. E., Kolachana, B. S., Mattay, V. S., Egan, M. F., & Weinberger, D. R. (2005). A susceptibility gene for affective disorders and the response of the human amygdala. Archives of General Psychiatry, 62(2), 146–152. Heim, C., Plotsky, P. M., & Nemeroff, C. B. (2004). Importance of studying the contributions of early adverse experience to neurobiological findings in depression. Neuropsychopharmacology, 29(4), 641–648. doi: http://www.nature.com/npp/journal/v29/n4/suppinfo /1300397s1.html Hertzman, C. (1999). The biological embedding of early experience and its effects on health in adulthood. Annals of the New York Academy of Sciences, 896, 85–95. Hertzman, C. (2012). Putting the concept of biological embedding in historical perspective. Proceedings of the National Academy of Sciences of the United States of America, 109, 17160–17167. doi: 10.1073/pnas.1202203109 Hertzman, C., & Boyce, W. T. (2010). How experience gets under the skin to create gradients in developmental health. Annual Review of Public Health, 31, 329–347. doi: 10.1146/annurev.publhealth.012809.103538 Hertzman, C., & Wiens, M. (1996). Child development and long-term outcomes: A population health perspective and summary of successful interventions. Soc Sci Med, 43(7), 1083–1095. Heyn, H., Moran, S., Hernando-Herraez, I., Sayols, S., Gomez, A., Sandoval, J., . . . Esteller, M. (2013). DNA methylation contributes to natural human variation. Genome Research, 23(9), 1363–1372. doi: 10.1101/gr.154187.112 Ingram, R. E., & Luxton, D. D. (2005). Vulnerability-stress models. In B. L. Hankin & J. R. Z. Abela (Eds.), Development of psychopathology: A vulnerability stress perspective (pp. 32–46). Thousand Oaks, CA: Sage Publications Inc. Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., . . . Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167(7), 748–751. doi: 10.1176/appi.ajp. 2010.09091379 Jones, P. A. (1999). The DNA methylation paradox. Trends in Genetics, 15(1), 34–37. Karg, K., Burmeister, M., Shedden, K., & Sen, S. (2011). The serotonin transporter promoter variant (5-HTTLPR), stress, and depression metaanalysis revisited: Evidence of genetic moderation. Archives of General Psychiatry, 68, 444–454. Kass, S. U., Landsberger, N., & Wolffe, A. P. (1997). DNA methylation directs a time-dependent repression of transcription initiation. Current Biology, 7(3), 157–165. Kendler, K. S., Neale, M., Kessler, R., Heath, A., & Eaves, L. (1993). A twin study of recent life events and difficulties. Archives of General Psychiatry, 50, 789–796. Kierkegaarad, S. (1986). Either/or (Trans. S. L. Ross & G. L. Stengren). New York, NY: Harper & Row. (Original publication 1843.) Kittleson, M. M., Meoni, L. A., Wang, N. Y., Chu, A. Y., Ford, D. E., & Klag, M. J. (2006). Association of childhood socioeconomic status

with subsequent coronary heart disease in physicians. Archives of Internal Medicine, 166(21), 2356–2361. Klengel, T., Mehta, D., Anacker, C., Rex-Haffner, M., Pruessner, J. C., Pariante, C. M., . . . Binder, E. B. (2013). Allele-specific FKBP5 DNA demethylation mediates gene–childhood trauma interactions. Nature Neuroscience, 16(1), 33–41. doi: http://www.nature.com/neuro /journal/v16/n1/abs/nn.3275.html—supplementary-information Klengel, T., Pape, J., Binder, E. B., & Mehta, D. (2014). The role of DNA methylation in stress- related psychiatric disorders. Neuropharmacology, 80, 115–132. doi: 10.1016/j.neuropharm.2014.01.013 Knafo, A., Israel, S., & Ebstein, R. P. (2011). Heritability of children’s prosocial behavior and differential susceptibility to parenting by variation in the dopamine receptor D4 gene. Development and Psychopathology, 23(1), 53–67. doi: doi:10.1017/S0954579410000647 Kraemer, H. C., Kiernan, M., Essex, M., & Kupfer, D. J. (2008). How and why criteria defining moderators and mediators differ between the Baron & Kenny and MacArthur approaches. Health Psychology, 27(2 Suppl), S101–S108. doi: 2008–03488–002 [pii]10.1037 /0278–6133.27.2(Suppl.).S101 Kraemer, H. C., Stice, E., Kazdin, A., Offord, D., & Kupfer, D. (2001). How do risk factors work together? Mediators, moderators, independent, overlapping and proxy-risk factors. American Journal of Psychiatry, 158, 848–856. Krieger, N. (2001). Theories for social epidemiology in the 21st century: An ecosocial perspective. International Journal of Epidemiology, 30, 668–677. Lee, C.-K., Klopp, R. G., Weindruch, R., & Prolla, T. A. (1999). Gene expression profile of aging and its retardation by caloric restriction. Science, 285, 1390–1393. Lengua, L. J. (2008). Anxiousness, frustration, and effortful control as moderators of the relation between parenting and adjustment in middle-childhood. Social Development, 17(3), 554–577. doi: 10.1111/j.1467–9507.2007.00438.x Li, J. J., Berk, M. S., & Lee, S. S. (2013). Differential susceptibility in longitudinal models of gene–environment interaction for adolescent depression. Development and Psychopathology, 25(4 Pt 1), 991–1003. doi: 10.1017/s0954579413000321 Liu, D., Diorio, J., Tannenbaum, B., Caldji, C., Francis, D., Freedman, A., . . . Meaney, M. J. (1997). Maternal care, hippocampal glucocorticoid receptors, and hypothalamic-pituitary-adrenal responses to stress. Science, 277, 1659–1662. Lupien, S. J., Ouellet-Morin, I., Hupbach, A., Tu, M. T., Buss, C., Walker, D., . . . McEwen, B. (2006). Beyond the stress concept: Allostatic load—A developmental biological and cognitive perspective. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology: Developmental neuroscience (Vol. 2, pp. 578–628). Hoboken, NJ: John Wiley & Sons Inc. Luthar, S. S., & Lyman, E., & Crossman, E. J.. Resilience and positive psychology. In M. Lewis & K. Rudolph (Eds.), Handbook of developmental psychopathology (3rd ed.). Norwell, MA: Kluwer/ Academic Press: 125–142. Lyons, D. M., Parker, K. J., & Schatzberg, A. F. (2010). Animal models of early life stress: Implications for understanding resilience. Developmental Psychobiology, 52(5), 402–410. doi: 10.1002/dev.20429 Macri, S., Zoratto, F., & Laviola, G. (2011). Early-stress regulates resilience, vulnerability and experimental validity in laboratory rodents through mother–offspring hormonal transfer. Neuroscience and Biobehavioral Reviews, 35(7), 1534–1543. doi: 10.1016/j.neubiorev .2010.12.014 Manuck, S. B., Craig, A. E., Flory, J. D., Halder, I., & Ferrell, R. E. (2011). Reported early family environment covaries with menarcheal age as a function of polymorphic variation in estrogen

References receptor-𝛼. Development and Psychopathology, 23(1), 69–83. doi: doi:10.1017/S0954579410000659 Masten, A. (2012). Risk and resilience in development. In P. D. Zelazo (Ed.), Oxford handbook of developmental psychology. Oxford, UK: Oxford Press. Matthews, L. J., & Butler, P. M. (2011). Novelty-seeking DRD4 polymorphisms are associated with human migration distance out-of-Africa after controlling for neutral population gene structure. American Journal of Physical Anthropology, 145(3), 382–389. doi: 10.1002/ajpa.21507 McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties in detecting interactions and moderator effects. Psychological Bulletin, 114(2), 376–390. McGowan, P. O., & Szyf, M. (2010). The epigenetics of social adversity in early life: implications for mental health outcomes. Neurobiology of Disease, 39(1), 66–72. doi: 10.1016/j.nbd.2009.12.026 McGue, M., & Bouchard, T. J., Jr. (1998). Genetic and environmental influences on human behavioral differences. Annual Review of Neuroscience, 21(5), 1–24. Meaney, M. J. (2010). Epigenetics and the biological definition of gene × environment interactions. Child Development, 81(1), 41–79. doi: 10.1111/j.1467–8624.2009.01381.x Mednick, S. A., Gabrielli, W. F., & Hutchings, B. (1987). Genetic factors in the etiology of criminal behavior. In S. A. Mednick, W. F. Gabrielli, & B. Hutchings (Eds.), The causes of crime: New biological approaches (pp. 74–91). Cambridge, UK: Cambridge University Press. Moen, E. L., Zhang, X., Mu, W., Delaney, S. M., Wing, C., McQuade, J., . . . Zhang, W. (2013). Genome-wide variation of cytosine modifications between European and African populations and the implications for complex traits. Genetics, 194(4), 987–996. doi: 10.1534 /genetics.113.151381 Moffitt, T. E. (2013). Childhood exposure to violence and lifelong health: Clinical intervention science and stress-biology research join forces. Development and Psychopathology, 25(25th Anniversary Special Issue 4pt2), 1619–1634. doi: doi:10.1017/S0954579413000801 Monroe, S. M., & Harkness, K. L. (2005). Life stress, the “kindling” hypothesis, and the recurrence of depression: Considerations from a life stress perspective. Psychological Review, 112(2), 417–445. doi: 10.1192/bjp. 168.1.68 10.1016/s0165–0327(02)00112-x Naeger, N. L., Van Nest, B. N., Johnson, J. N., Boyd, S. D., Southey, B. R., Rodriguez-Zas, S. L., . . . Robinson, G. E. (2011). Neurogenomic signatures of spatiotemporal memories in time-trained forager honey bees. Journal of Experimental Biology, 214(Pt 6), 979–987. doi: 10.1242/jeb.053421 National Research Council. (2011). Toward precision medicine: Building a knowledge network for biomedical research and a new taxonomy of disease. Washington, D.C. National Academies Press. Nelson, C. A., Zeanah, C. H., Fox, N. A., Marshall, P. J., Smyke, A. T., & Guthrie, D. (2007). Cognitive recovery in socially deprived young children: The Bucharest early intervention project. Science, 318(5858), 1937–1940. doi: 10.1126/science.1143921 Nelson, E. M., & Spieker, S. J. (2013). Intervention effects on morning and stimulated cortisol responses among toddlers in foster care. Infant Mental Health Journal, 34(3), 211–221. doi: 10.1002/imhj.21382 Nesse, R. M., & Young, E. A. (2000). Evolutionary origins and functions of the stress response. In G. Fink (Ed.), Encyclopedia of Stress (Vol. 2, pp. 79–84). New York, NY: Academic Press. O’Connor, T. G., Hetherington, E. M., Reiss, D., & Plomin, R. (1995). A twin-sibling study of observed parent–adolescent interactions. Child Development, 66, 812–829. Oberlander, T. F., Weinberg, J., Papsdorf, M., Grunau, R., Misri, S., & Devlin, A. M. (2008). Prenatal exposure to maternal

135

depression, neonatal methylation of human glucocorticoid receptor gene (NR3C1) and infant cortisol stress response. Epigenetics, 3(2), 97–106. Obradovi´c, J., & Boyce, W. T. (2009). Individual differences in behavioral, physiological, and genetic sensitivities to contexts: Implications for development and adaptation. Developmental Neuroscience, 31(4), 300–308. Obradovi´c, J., Bush, N. R., Stamperdahl, J., Adler, N. E., & Boyce, W. T. (2010). Biological sensitivity to context: The interactive effects of stress reactivity and family adversity on socioemotional behavior and school readiness. Child Development, 81(1), 270–289. Oomen, C. A., Soeters, H., Audureau, N., Vermunt, L., van Hasselt, F. N., Manders, E. M. M., . . . Krugers, H. (2010). Severe early life stress hampers spatial learning and neurogenesis, but improves hippocampal synaptic plasticity and emotional learning under high-stress conditions in adulthood. Journal of Neuroscience, 30(19), 6635–6645. doi: 10.1523/jneurosci.0247–10.2010 Owen, D., Andrews, M. H., & Matthews, S. G. (2005). Maternal adversity, glucocorticoids and programming of neuroendocrine function and behaviour. Neuroscience Biobehavioral Review, 29(2), 209–226. Pagliaccio, D., Luby, J. L., Bogdan, R., Agrawal, A., Gaffrey, M. S., Belden, A. C., . . . Barch, D. M. (2014). Stress-system genes and life stress predict cortisol levels and amygdala and hippocampal volumes in children. Neuropsychopharmacology, 39(5), 1245–1253. doi: 10.1038/npp. 2013.327 Pezawas, L., Meyer-Lindenberg, A., Drabant, E. M., Verchinski, B. A., Munoz, K. E., Kolachana, B. S., . . . Weinberger, D. R. (2005). 5-HTTLPR polymorphism impacts human cingulate-amygdala interactions: A genetic susceptibility mechanism for depression. Nature Neuroscience, 8(6), 828–834. Phillips, D., Crowell, N. A., Sussman, A. L., Gunnar, M., Fox, N., Hane, A. A., & Bisgaier, J. (2012). Reactive temperament and sensitivity to context in childcare. Social Development, 21(3), 628–643. doi: 10.1111/j.1467–9507.2011.00649.x Plomin, R., & Crabbe, J. (2000). DNA. Psychological Bulletin, 126(6), 806–828. Plomin, R., DeFries, J. C., & Loehlin, J. C. (1977). Genotype–environment interaction and correlation in the analysis of human behavior. Psychological Bulletin, 84, 309–322. Plomin, R., DeFries, J. C., McClearn, G. E., & McGuffin, P. (2001). Behavioral genetics (4th ed.). New York, NY: Worth Publishers. Plotsky, P. M., & Meaney, M. J. (1993). Early, postnatal experience alters hypothalamic corticotropin-releasing factor (CRF) mRNA, median eminence CRF content and stress-induced release in adult rats. Brain Research Reviews, Molecular Brain Research, 18(3), 195–200. Pluess, M., & Belsky, J. (2009). Differential susceptibility to rearing experience: the case of childcare. Journal of Child Psychology and Psychiatry, 50(4), 396–404. doi: JCPP1992 [pii] 10.1111/j.1469–7610.2008.01992.x Pluess, M., & Belsky, J. (2011). Prenatal programming of postnatal plasticity? Development and Psychopathology, 23(1), 29–38. doi: 10.1017/S0954579410000623 Pluess, M., & Belsky, J. (2013). Vantage sensitivity: Individual differences in response to positive experiences. Psychological Bulletin, 139(4), 901–916. doi: 10.1037/a0030196 Pollak, S. D. (2005). Early adversity and mechanisms of plasticity: Integrating affective neuroscience with developmental approaches to psychopathology. Development and Psychopathology, 17(3), 735–752. doi: 10.1017/s0954579405050352 Provencal, N., Suderman, M. J., Guillemin, C., Massart, R., Ruggiero, A., Wang, D. S., . . . Szyf, M. (2012). The signature of maternal rearing in the methylome in rhesus macaque prefrontal cortex and T

136

Differential Sensitivity to Context: Implications for Developmental Psychopathology

cells. Journal of Neuroscience, 32(44), 15626–15642. doi: 10.1523 /jneurosci.1470–12.2012 Psychiatric GWAS Consortium Steering Committee. (2008). A framework for interpreting genome-wide association studies of psychiatric disorders. Molecular Psychiatry, 14, 10–17.

Shannon, K. E., Beauchaine, T. P., Brenner, S. L., Neuhaus, E., & Gatzke-Kopp, L. (2007). Familial and temperamental predictors of resilience in children at risk for conduct disorder and depression. Development and Psychopathology, 19(3), 701–727. doi: S0954579407000351 [pii] 10.1017/S0954579407000351

Quas, J. A., Yim, I. S., Oberlander, T. F., Nordstokke, D., Essex, M. J., Armstrong, J. M., . . . Boyce, W. T. (2014). The symphonic structure of childhood stress reactivity: Patterns of sympathetic, parasympathetic, and adrenocortical responses to psychological challenge. Development and Psychopathology, 26 (4pt1) pp 963–982. DOI: 10.1017/S0954579414000480

Sheridan, J. F., Stark, J. L., Avitsur, R., & Padgett, D. A. (2000). Social disruption, immunity, and susceptibility to viral infection. Role of glucocorticoid insensitivity and NGF. Annals of the New York Academy of Science, 917, 894–905.

Ramchandani, P. G., van IJzendoorn, M., & Bakermans-Kranenburg, M. J. (2010). Differential susceptibility to fathers’ care and involvement: The moderating effect of infant reactivity. Family Science, 1(2), 93–101. doi: 10.1080/19424621003599835 Raver, C. C., Blair, C., & Willoughby, M. (2013). Poverty as a predictor of 4-year-olds’ executive function: New perspectives on models of differential susceptibility. Developmental Psychology, 49(2), 292–304. doi: 10.1037/a0028343 Risch, N., Herrell, R., Lehner, T., Liang, K. Y., Eaves, L., Hoh, J., . . . Merikangas, K. R. (2009). Interaction between the serotonin transporter gene (5-HTTLPR), stressful life events, and risk of depression: A meta-analysis. Journal of the American Medical Association, 301(23), 2462–2471. Robinson, G. E., Fernald, R. D., & Clayton, D. F. (2008). Genes and social behavior. Science, 322(5903), 896–900. doi: 322/5903/896 [pii] 10.1126/science.1159277 Roisman, G. I., Newman, D. A., Fraley, R. C., Haltigan, J. D., Groh, A. M., & Haydon, K. C. (2012). Distinguishing differential susceptibility from diathesis–stress: Recommendations for evaluating interaction effects. Development and Psychopathology, 24(2), 389–409. doi: doi:10.1017/S0954579412000065 Rothman, K. J., & Greenland, S. (2005). Causation and causal inference in epidemiology. American Journal of Public Health, 95(Suppl 1), S144–S150. Rudolph, K. D., Troop-Gordon, W., & Granger, D. A. (2010). Peer victimization and aggression: Moderation by individual differences in salivary cortisol and alpha-amylase. Journal of Abnormal Child Psychology, 38(6), 843–856. doi: 10.1007/s10802–010–9412–3 Rudolph, K. D., Troop-Gordon, W., & Granger, D. A. (2011). Individual differences in biological stress responses moderate the contribution of early peer victimization to subsequent depressive symptoms. Psychopharmacology, 214(1), 209–219. doi: 10.1007/s00213–010–1879–7 Rutter, M. (2006). Genes and behavior: Nature–nurture interplay explained. Malden, MA: Blackwell Publishing. Rutter, M. (2012). Resilience as a dynamic concept. Development and Psychopathology, 24(2), 335–344. doi: 10.1017/s0954579412000028 Rutter, M., Dunn, J., Plomin, R., Simonoff, E., Pickles, A., Maughan, B., . . . Eaves, L. (1997). Integrating nature and nurture: Implications of person-environment correlations and interactions for developmental psychopathology. Development and Psychopathology, 9(2), 335–364. Rutter, M., Thapar, A., & Pickles, A. (2009). Gene–environment interactions: Biologically valid pathway or artifact? Archives of Genernal Psychiatry, 66(12), 1287–1289. Sanchez, M. M., Ladd, C. O., & Plotsky, P. M. (2001). Early adverse experience as a developmental risk factor for later psychopathology: Evidence from rodent and primate models. Development and Psychopathology, 13(3), 419–449. Scriver, C. R., & Waters, P. J. (1999). Monogenic traits are not simple: lessons from phenylketonuria. Trends in Genetics, 15(7), 267–272.

Sheridan, M. A., Fox, N. A., Zeanah, C. H., McLaughlin, K. A., & Nelson, C. A. (2012). Variation in neural development as a result of exposure to institutionalization early in childhood. PNAS Proceedings of the National Academy of Sciences, 109(32), 12927–12932. doi: 10.1073/pnas.1200041109 Stevens, H. E., Leckman, J. F., Coplan, J. D., & Suomi, S. J. (2009). Risk and resilience: early manipulation of macaque social experience and persistent behavioral and neurophysiological outcomes. Journal of Am Acad Child Adolesc Psychiatry, 48(2), 114–127. doi: 10.1097/CHI.0b013e318193064c Strunk, T., Jamieson, S. E., & Burgner, D. (2013). Genetic and epigenetic susceptibility to early life infection. Current Opinion in Infectious Disease, 26(3). doi: 10.1097/QCO.0b013e32835fb8d9 Suderman, M., McGowan, P. O., Sasaki, A., Huang, T. C. T., Hallett, M. T., Meaney, M. J., . . . Szyf, M. (2012). Conserved epigenetic sensitivity to early life experience in the rat and human hippocampus. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Suppl 2), 17266–17272. doi: 10.1073/pnas .1121260109 Sulloway, F. J. (1992). Freud, biologist of the mind: Beyond the psychoanlytic legend. Boston, MA: Harvard University Press. Suomi, S. J. (2006). Risk, resilience, and gene × environment interactions in rhesus monkeys. Annals New York Academy of Sciences, 1994, 52–62. doi: 10.1196/annals.1376.006 Teh, A. L., Pan, H., Chen, L., Ong, M. L., Dogra, S., Wong, J., . . . Holbrook, J. D. (2014). The effect of genotype and in utero environment on inter-individual variation in neonate DNA methylomes. Genome Research, 24, 1064–1074 doi: 10.1101/gr.171439.113 Thomaes, K., Dorrepaal, E., Draijer, N., de Ruiter, M. B., Elzinga, B. M., van Balkom, A. J., . . . Veltman, D. J. (2012). Treatment effects on insular and anterior cingulate cortex activation during classic and emotional Stroop interference in child abuse-related complex post-traumatic stress disorder. Psychological Medicine, 42(11), 2337–2349. doi: doi:10.1017/S0033291712000499 Tottenham, N., & Sheridan, M. A. (2010). A review of adversity, the amygdala and the hippocampus: a consideration of developmental timing. Frontiers in Human Neuroscience, 3, 1–18. doi: 10.3389/neuro.09.068.2009 van de Wiel, N. M. H., van Goozen, S. H. M., Matthys, W., Snoek, H., & van Engeland, H. (2004). Cortisol and Treatment effect in children with disruptive behavior disorders: A preliminary study. Journal of the American Academy of Child & Adolescent Psychiatry, 43(8), 1011–1018. doi: 10.1097/01.chi.0000126976.56955.43 Wachs, T. D., & Plomin, R. (Eds.). (1991). Conceptualization and measurement of organism–environment interaction. Washington, DC: American Psychological Association. Waddington, C. H. (1966). Principles of development and differentiation. New York, NY: MacMillan. Wahlsten, D. (1990). Insensitivity of the analysis of variance to heredity environment interaction. Behavioral & Brain Sciences, 13(1), 109–161.

References Wang, X., Zuckerman, B., Pearson, C., Kaufman, G., Chen, C., Wang, G., . . . Xu, X. (2001). Maternal stress, CRH genotype, and preterm birth. Paper presented at the Pediatric Academic Societies Annual Meeting, Baltimore, MD. Watson, J. B. (1930). Behaviorism (Revised edition). Chicago: University of Chicago Press. Weatherall, D. J. (2001). Phenotype–genotype relationships in monogenic disease: Lessons from the thalassaemias. Nature Reviews Genetics, 2(4), 245–255. Weaver, I. C., Cervoni, N., Champagne, F. A., D’Alessio, A. C., Sharma, S., Seckl, J. R., . . . Meaney, M. J. (2004). Epigenetic programming by maternal behavior. Nature Neuroscience, 7(8), 847–854. Weikum, W. M., Oberlander, T. F., Hensch, T. K., & Werker, J. F. (2012). Prenatal exposure to antidepressants and depressed maternal mood alter trajectory of infant speech perception. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Suppl 2), 17221–17227. doi: 10.1073/pnas.1121263109 Werner, E. E. (2012). Children and war: Risk, resilience, and recovery. Development and Psychopathology, 24(2), 553–558. doi: 10.1017 /s0954579412000156 Whittle, S., Yap, M. B. H., Sheeber, L., Dudgeon, P., Yücel, M., Pantelis, C., . . . Allen, N. B. (2011). Hippocampal volume and sensitivity

137

to maternal aggressive behavior: A prospective study of adolescent depressive symptoms. Development and Psychopathology, 23(1), 115–129. doi: 10.1017/S0954579410000684 Wilson, D. S. (1998). Adaptive individual differences within single populations. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 353(1366), 199–205. doi: 10.1098 /rstb.1998.0202 Worthman, C. M., & Kuzara, J. (2005). Life history and the early origins of health differentials. American Journal of Human Biology, 17(1), 95–112. Wu, H., & Zhang, Y. (2014). Reversing DNA methylation: Mechanisms, genomics, and biological functions. Cell, 156(1–2), 45–68. doi: 10.1016/j.cell.2013.12.019 Yang, E.-J., Lin, E. W., & Hensch, T. K. (2012). Critical period for acoustic preference in mice. PNAS Proceedings of the National Academy of Sciences of the United States of America, 109(Suppl 2), 17213–17220. doi: 10.1073/pnas.1200705109 Yu, K., Wacholder, S., Wheeler, W., Wang, Z., Caporaso, N., Landi, M. T., & Liang, F. (2012). A flexible Bayesian model for studying gene–environment interaction. PLoS Genetics, 8(1) e1002482. doi: 10.1371/journal.pgen.1002482

CHAPTER 4

Understanding Developmental Psychopathology: How Useful Are Evolutionary Perspectives? JAMES F. LECKMAN

INTRODUCTION 138 GOALS OF EVOLUTIONARY EXPLANATIONS 138 EVOLUTIONARY MECHANISMS THAT MAY ACCOUNT FOR THE EMERGENCE OF PSYCHOPATHOLOGY OVER THE COURSE OF DEVELOPMENT 140 Stress–Diathesis Models: Resilience and Allostasic Load 141 Differential Susceptibility and Biological Sensitivity to Context 141 Separation Challenges and Attachment Solutions 141 EVOLUTIONARY MECHANISMS THAT MAY ACCOUNT FOR THE PERSISTENCE OF DISCRETE FORMS OF PSYCHOPATHOLOGY 141 Failure of Conserved Patterns of Behavior to Develop Normally 141

Dysregulation of Conserved Behavioral Systems and Associated Mental States 147 Ancient Versus Current Environments and the Value of Diversity Within Populations 151 Co-optation of Neurobiological Systems Associated With Establishing Salience and Reward 152 An Evolutionary Arms Race: Infections and Autoimmunity 153 Other Evolutionary-Based Explanations 154 CONCLUSIONS AND CRITIQUE 154 Future Prospects 155 Clinical Implications 157 REFERENCES 157

INTRODUCTION

Perhaps most disappointing is its failure to provide a dynamic framework that would allow for a sensible dialectic with burgeoning areas of science including evolutionary biology, molecular genomics, environment-dependent epigenetic alterations of gene expression, systems and developmental neurosciences, neuroimmunology, ethology, child development, and cognitive and social neuroscience— to name a few. Despite these shortcomings there is an increasing recognition that there is a need to look beyond the proximal mechanisms that underlie specific disorders and to consider the distal evolutionary mechanisms that give rise to psychopathology (Stein & Neese, 2011). In this chapter the primary focus is on these emerging perspectives on how evolutionary processes may support the preservation of various forms of psychopathology.

What is the most fruitful perspective from which to view developmental psychopathology? At present a major vantage point of our field, as embodied in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5; American Psychiatric Association, 2013), is largely atheoretical and descriptive. Conceptually, it reifies psychopathological states as categorical entities that are modeled after discrete medical disorders (Kendler, 2012). To some extent this has been a serviceable and generative approach, well suited to today’s medical marketplace. Leaving aside the inelegant unspecified designations, DSM-5 also has a certain bureaucratic appeal with every disorder neatly in its place. The weaknesses of DSM-5 are also readily apparent. Many of the criteria are no more than a reflection of the conventional wisdom of a committee of experts with limited empirical justification. Being categorical in nature, it largely ignores the dimensional properties of many syndromes and their intimate relationship with environmental factors and the individual’s antecedent developmental history.

GOALS OF EVOLUTIONARY EXPLANATIONS The principal goal of an evolutionary perspective of psychopathology is to provide a coherent framework from which to view patterns of maladaptive behavior that are

138

Goals of Evolutionary Explanations

persistent in human populations (Leckman & Mayes, 1998). In such an evolutionary framework, the issue of persistence appears to be paradoxical given the editing power of natural selection. Darwin’s (1859) principle of natural selection posits (1) the existence of variation among individuals, (2) differential reproductive success for those individuals who exhibit traits that are useful in the struggle for life, and (3) differential inheritance of those factors that gave rise to the favorable traits. Why then would particular variations persist that place individuals at a reproductive disadvantage in the struggle for life? We also need to consider whether various forms of psychopathology are not truly maladaptive when viewed from an evolutionary perspective. As pointed out by Belsky and Pluess (2013), some forms of behavior that are generally considered to be psychopathological and maladaptive (e.g., impulsive, risky, antisocial) can be considered adaptive within certain conditions of adversity to the degree that they enhance an individual’s reproductive fitness. Before reviewing the proposed mechanisms, this chapter’s position should be made clear. A major thesis of this review is that the improbable cascade of evolutionary events that has led to the emergence of our species and our particular set of conserved behavioral and mental capacities has also left us vulnerable to certain forms of psychopathology. By conserved behavioral and mental capacities, we refer simply to adaptive capacities that have contributed to our evolutionary success as a species. On reflection, this set of capacities includes such behaviors (and associated mental states) as purposive reaching and holding, feeding, locomotion, grooming, communication (gestures, species-specific vocalizations and language), alarm systems activated by perceived threats such as separation from an attachment figure, courtship and pair bonding, reproductive and parental behaviors, the formation of social dominance hierarchies, and our capacity to be resilient and adapt and thrive in novel and challenging environments—among many others. Another fundamental aspect of this approach is to consider the developmental course of these conserved behavioral and mental capacities over the lifetime of the individual. Consideration of continuities and discontinuities between normality and psychopathology is a hallmark of developmental psychopathology (Cicchetti, 2006; Zigler & Glick, 1986). Consequently, in our examination of specific forms of psychopathology we will consider the developmental time course for the emergence of these conserved behavioral and mental capacities. In line with this developmental perspective, it is becoming increasingly clear that changes in the transcriptional

139

regulation of key developmentally expressed genes are major contributors to the evolution of our human species—especially the evolution of our brain and the emergence of brain-related disorders (Vallender, MekelBobrov, & Lahn, 2008; Varki, Geschwind, & Eichler, 2008). Indeed, variation in regulation of gene expression during development, rather than variation in protein sequence, is almost certainly the dominant factor in human brain evolution (King & Wilson, 1975). Such changes are thought to have led to the creation of new combinatorial expression patterns during development and ultimately to the formation of distinct neuronal circuits. In this context the field of evolutionary psychology and psychopathology needs to take into account our rapidly expanding knowledge of the human genome and the organ specific transcriptomes (Barbosa-Morais et al., 2012; ENCODE Project Consortium, 2012), and the complexity of the neurodevelopmental programs required for cellular proliferation and circuit formation within the human brain (Kang et al., 2012). Current estimates suggest that at least 80% of the human genome is functional. Our genome contains more than 20,000 protein-coding genes, more than 18,000 RNA genes as well as protein-bound regions, regions of histone modification and DNA methylation, and chromosome interacting regions (ENCODE Project Consortium, 2012). Many of these recently discovered regulatory elements are physically associated with one another and with expressed genes. A more complete understanding of their role should provide new insights into the specific mechanisms of gene regulation associated with neural development. A significant portion of these regulatory elements are likely to be associated with environmentally sensitive alterations in region-specific gene expression mediated by epigenetic mechanisms that take place during specific periods of development (Caldji, Hellstrom, Zhang, Diorio, & Meaney, 2011; Keverne, 2011; Pujadas & Feinberg, 2012). Another advance in understanding how species with similar repertoires of protein-coding genes differ so markedly in their biobehavioral repertoires has come from an in-depth analysis of organ transcriptomes from vertebrate species spanning ∼350 million years of evolution. These events likely have further contributed to the diversification of splicing and other transcriptomic changes that underlie the emergence of our species-specific bio-behavioral repertoire. For example, significant differences in alternative splicing complexity between vertebrate lineages has been documented with the highest complexity found in primates (Barbarosa-Morais et al.,

140

Understanding Developmental Psychopathology

2012). However, other investigators have found that much of our uniqueness as a species comes from variations in regulatory elements in mammalian aligned regions rather than from primate-specific regions (Ward & Kellis, 2012). The emerging picture of the intricate and dynamic complexity of the genetic programs associated with the development of the human brain is also worthy of note. Recent analyses of exon-level transcriptome and associated genotyping data, from multiple brain regions including several neocortical areas of developing and adult postmortem human brains, have found that over 80% of the more than 17,000 mostly protein-coding genes studied were expressed, and that 90% of these genes were differentially regulated at the whole-transcript or exon level across brain regions and/or time (Kang et al., 2012). Notably, many of the spatiotemporal differences detected were differently present prior to birth and included a number of genes that were sexually dimorphic in their patterns of expression (Kang et al., 2012). This study also identified in the developing and the adult human brain core gene co-expression networks associated with cellular proliferation, synapse and dendrite formation, and myelination. Moreover, because of functional redundancy of gene families in divergent biological processes, it is clear that the genetic and evolutionary origins of developmental disorders affecting behavior and emotions may be more systemic and polygenic in nature than is generally appreciated. Ongoing studies comparing genomes across species as well as comparing spatiotemporal differences of gene expression should provide a clearer understanding of the transcriptional foundations of human brain development and evolution of language and other uniquely human traits (Johnson et al., 2009). Future studies of human postmortem brain material that include recently discovered regulatory elements in the human genome should allow investigators to generate expression trajectories of genes commonly associated with specific mental disorders. Likewise, large-scale experimental assays in multiple individuals, cell types, and populations have the potential to identify the role of the regulatory elements that are key to both human evolution and disease. In an effort to integrate an interdisciplinary perspective on developmental psychopathology Killeen, Tannock, and Sagvolden (2012) have suggested that we consider Aristotle’s four causes to explain how things come to be (Hocutt, 1974). The formal cause describes the pattern or form which is necessary for us to recognize something as a particular something. At present, criteria contained

in the Diagnostic and Statistical Manual (DSM) and International Classification of Diseases (ICD) identify the specific requirements needed to diagnose mental disorders. The material cause is equivalent to the nature of the raw materials out of which something is made. Here it will be important to consider aspects of neural and somatic development occurring within specific environmental contexts—from the intrauterine environment to the social ecology of the caregivers and of the developing child (Rutter, 2008; Werner & Smith, 2001). The efficient cause identifies the triggers that cause something. Here again the individual’s interactions with multiple environments is key. Examples include the unfolding complex somatic and emotional environment associated with child birth leading to postpartum depression or the recreational use of certain drugs—leading to addiction. Last, but not least, is the final cause, or telos, which is defined as the purpose, end, aim, or goal of something, and here we return to evolutionary accounts. With this Aristotelian perspective in mind the focus of this chapter is to understand why the evolution of our species’ conserved behavioral and mental capacities has also left us vulnerable to certain forms of psychopathology. Although we focus on the final cause, it will also be important to consider the other three causes as well. EVOLUTIONARY MECHANISMS THAT MAY ACCOUNT FOR THE EMERGENCE OF PSYCHOPATHOLOGY OVER THE COURSE OF DEVELOPMENT As reviewed elsewhere in this volume (Chapter 1), evolutionary reasoning has played an increasingly prominent role in theoretical models that have been proposed to explain individual differences in behavior that emerge over the course of development. Several of these theoretical models were initially prompted by the limitations of biopsychosocial stress-diathesis models (Ingram & Luxton, 2005). While the concepts of resilience and adaptation remain useful concepts (Panter-Brick & Leckman, 2013; Chapter 6, Volume 4), other evolutionary-based theoretical models, including differential susceptibility and biological sensitivity to context, are gaining widespread use in both theoretical formulations and empirical studies of child development (Chapters 2 and 3, this volume). Likewise, efforts to integrate evolutionary science and neuroscience are being proposed that leave aside the prevailing taxonomy of mental disorders altogether and focus instead on the therapeutic importance of attachment solutions to the existential separation challenges that are inherent in life (Fricchione, 2011).

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

Stress–Diathesis Models: Resilience and Allostasic Load In the early 1970s, Garmezy and his team at the University of Minnesota were focused on the risk and protective factors associated with the development of schizophrenia (Garmezy & Streitman, 1974). In this context, the concept of resilience emerged to describe the dynamic process involving positive adaptations within the context of significant adversity (Luthar, Cicchetti, & Becker, 2000). While adaption and resilience can be readily seen from an evolutionary perspective (Lasker, 1969), the initial formulations of these constructs in the field of developmental psychology typically did not explicitly invoke evolution as being an important determinant (Rutter, 1979). By the time McEwen and his colleagues introduced the term allostasis in the 1990s to describe how organisms can adapt to environmental challenges and obtain stability through change, an evolutionary perspective was firmly in place (McEwen & Stellar, 1993).

Differential Susceptibility and Biological Sensitivity to Context The emergence of an evolutionary perspective led Belsky and colleagues to ask uncomfortable questions. For example, why would natural selection lead to the emergence of an individual that would respond to adversity by becoming putatively disturbed and dysregulated (Belsky, Steinberg, & Draper, 1991; Daly & Wilson, 1981)? Belsky et al.’s (1991) psychosocial acceleration theory was advanced to explain why behaviors such as early pubertal development, promiscuous sexual behavior, adolescent pregnancy, and limited parental investment in multiple offspring might make biological sense as a way to prepare for and insure the passage of genes from one generation to the next despite the likelihood of growing up in the context of environmental adversity. Subsequent work by Belsky, Boyce, and Ellis has led to several theoretical concepts including differential susceptibility and biological sensitivity to context (Belsky, 1997; Boyce & Ellis 2005). This evolutionary perspective has enriched the field of child development and led to a rich and expanding body of work that seeks to integrate scientific advances from a broad range of disciplines, from genetics and epigenetics to developmental neurobiology and neuroimaging to neuroimmunology. Although a comprehensive and critical exploration of this multifaceted field is well beyond the scope of this chapter, readers are referred to other chapters in this volume as well as to recent special issues of Development and Psychopathology (Ellis & Boyce, 2011) and

141

Developmental Psychology as well (Ellis & Bjoklund, 2012). There are also several other excellent reviews and editorials (Belsky & Pluess, 2013; Leckman, 2014; O’Conner, Moynihan, & Caserta, 2014). It is also becoming increasingly clear that early adversity can have enduring detrimental effects on physical as well as mental health over the life course (Danese et al., 2009) and that early interventions have the potential to alter this unfortunate, but evolutionarily understandable, trajectory (Campbell et al., 2014; Heckman, Moon, Pinto, Savelyev, & Yavitz, 2010). Separation Challenges and Attachment Solutions Fricchione (2011) provided one of the most detailed and compelling accounts of biological foundations of the evolutionary need for attachment solutions, from the earliest single cellular organisms that appear to our vertebrate and humanoid lineage leading to the emergence of our species and our incredibly complex and plastic biobehavioral systems. Fricchione’s conceptual framework relies heavily on attachment theory initially advanced by Bowlby (1969, 1973, 1980). In contrast to many evolutionary accounts that focus on the processes at work that lead to psychopathology, Frichonne relied on his evolutionary perspective to provide a theoretical foundation for his compassionate clinical care across the life span including the separation challenges that occur at the end of life.

EVOLUTIONARY MECHANISMS THAT MAY ACCOUNT FOR THE PERSISTENCE OF DISCRETE FORMS OF PSYCHOPATHOLOGY In this section five evolutionary-based mechanisms that may account for the persistence of some forms of psychopathology are presented. Table 4.1 presents a list of these mechanisms along with short descriptions and putative examples. Interested readers may also wish to consult anthologies and book-length reviews of evolutionary psychology and psychopathology (Barkow, Cosmides, & Tooby, 1995; Baron-Cohen, 1997; Brüne, 2008; Buss, 2011; Marks & Nesse, 1994; McGuire & Troisi, 1998; Tooby & Cosmides, 1990). Failure of Conserved Patterns of Behavior to Develop Normally If we depend on the normal expression of specific genes at the correct time and place in the developing brain, then we remain vulnerable to events that can disrupt the

142

Understanding Developmental Psychopathology

TABLE 4.1 Proposed Evolutionary Mechanisms Proposed evolutionary mechanism

Description

Examples

Failure of conserved neurobehavioral systems to develop normally. If we depend on the normal expression of specific genes at the correct time and place in the developing brain, then we remain vulnerable to events that can disrupt the development of these neural circuits.

Any failure of a conserved neurobehavioral system will set an individual apart. Although most of the proximal causes for such deficit states are accidental or stochastic, these events (mutations or untoward epigenetic events) occur early and constrain subsequent CNS development. Many of the mutations that have been identified involve regulatory regions of the human genome and not conserved protein-coding regions per se. These mutations often affect an entire cascade of neurobiological systems

Disorders associated with intellectual disability (e.g., fragile X, Prader-Willi syndrome, Angelman syndrome), learning disorders and communication disorders; autism spectrum disorder; and childhoodonset schizophrenia

Dysregulation of conserved neurobehavioral systems

Some conserved neurobehavioral systems are highly plastic and only appear in response to particular environment perturbations or at particular points in development. Some of these are behavioral alarm systems. Inappropriate or expressive activation in response to potential threats or failure to shut down such systems following loss may lead to psychopathology.

Mood and anxiety disorders, obsessivecompulsive disorder

Differential susceptibility

In response to chronic adversity during prior generations as well as during embryonic, fetal and postnatal life, individual differences may emerge that promote putative maladaptive behaviors (impulsivity, poor inhibitory control, and antisocial acts). However, these maladaptive behaviors may be adaptive in the context of chronic environmental adversity by enhancing an individual’s reproductive fitness despite the negative effects on the mental and physical health. The precise stochastic genetic, and epigenetic mechanisms that result in these individual differences and their impact of biobehavioral plasticity over the course of development have yet to be fully determined.

Increased susceptibility to develop behaviors associated with attention-deficit/ hyperactivity disorder (ADHD); conduct disorder; antisocial behavior; and substance use disorders

Environmental shifts and the value of diversity at the population level

Historical shifts in the environment away from the primeval environment of adaptation may make some individuals stand out by displaying previously adaptive behaviors that are now out of place. Alternatively, there may continue to be advantages in reproductive fitness at the group level for the presence of individuals in the group who are drawn to novel stimuli, adventurous, variable in their response to the same stimuli and willing to take risks despite the potential negative impact on their own reproductive fitness.

ADHD

Co-optation of conserved neurobiological systems

Conserved neurobehavioral systems are highly dependent on specific neurobiological substrates. Some of these pathways can be co-opted or hijacked by exogenous substances, leading to addiction, behavioral dysfunction, and impairment.

Substance use and addictive disorders

Evolutionary arms races

Some conserved biobehavioral systems may have evolved in relationship to other species (e.g., parasites, predators) that were important members of the primeval environment of adaptation.

Some forms of acute onset obsessivecompulsive disorder: pediatric autoimmune neuropsychiatric disorders associated with streptococcal infections (PANDAS)

Source: Adapted in part from J. F. Leckman & L. C. Mayes, Understanding developmental psychopathology: How useful are evolutionary perspectives? Journal of the American Academy of Child and Adolescent Psychiatry, 37, 1998, Table 1 (Leckman & Mayes, 1998).

development of these neural circuits which in turn can lead to the emergence of certain forms of psychopathology. In association with certain genetic mutations, characteristic phenotypes appear which reflect the loss or breakdown of conserved patterns of behavior. In some cases the reasons for persistence are clear; however, in many others the molecular mechanisms remain obscure. For many of these disorders there is a clear loss of reproductive fitness (Power et al., 2012). Four examples are briefly considered: fragile X syndrome, Prader-Willi and Angelman syndromes,

autism spectrum disorder (ASD), and childhood-onset schizophrenia (COS). Fragile X Syndrome Fragile X syndrome serves as a good example of a single-gene disorder with a characteristic profile of strengths and weakness that generally follow a specific course over development (Hagerman & Hagerman, 2002; Chapter 2, Volume 3). Typically, males with fragile X syndrome have a characteristic phenotype with both physical

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

stigmata (long face, large ears, hyperextensible joints, and large testes) and cognitive-behavioral features (mental impairment with sequential processing deficits, dysarthric speech, motoric hyperactivity, anxiety and unstable mood, and social difficulties). Fragile X syndrome is the single most common cause for heritable intellectual disability. In addition, it is also the most widely diagnosed form of intellectual disability that leads to ASD (Hatton et al., 2006). This X-linked disorder is caused by an expansion of a trinucleotide CGG repeat (>200) on the promoter region of the fragile X mental retardation 1 gene (FMR1). As a result, the promoter region often becomes methylated which leads to a deficiency or absence of the expression of the FMR1 protein (FMRP) (McLennan, Polussa, Tassone, & Hagerman, 2011). The transcription failure is usually due to the presence of an unstable region of DNA in which a trinucleotide repeat sequence, (CGG)n, is amplified from fewer than 54 repeats in typically developing individuals to more than 200 repeats in affected males (Fu et al., 1991). Individuals in these families with more than 52 but less than 200 repeats have a premutation. Curiously, a premutation expands to a full mutation only when it is transmitted by a female so that daughters of a male with a premutation have only a premutation and remain unaffected. The genetic anticipation (progressively more individuals in a kindred are affected in successive generations) associated with this mutation accounts in large part for the persistence of this mutation in the population. Similar DNA amplification abnormalities and associated patterns of genetic anticipation within large kindreds have also been documented in several other autosomal disorders including Huntington disease (Walker, 2007). These disorders provide a clear example of how a mutation in a regulatory region can have a major impact on neural development. Recent evidence suggests that the neuronal deficits in fragile X syndrome may occur in part due to the loss of translational constraint imposed by FMRP on transcripts for a number of synaptic proteins including neuroligin-3, neurexin-1, SHANK3, PTEN (Cowden syndrome), TSC2, and NF1 (neurofibromatosis type 1) (Darnell et al., 2011). Several of these genes including SHANK3 and NRXN1 have also been linked with ASD (State & Sestan, 2012). Also of note is that there are species-specific differences in the expression of FMR1 in the neocortex (Kwan et al., 2012) which in turn points to the potential limitations of animal model systems to explore the neuropathology of some neurodevelopmental disorders. From an evolutionary point of view this finding is consistent with the emerging knowledge that many of the genetic mutations responsible

143

for human diseases occur in regulatory regions of the genome that are not conserved across other mammalian or primate species (Ward & Kellis, 2012). Prader-Willi and Angelman Syndromes Prader-Willi syndrome in conjunction with Angelman syndrome perhaps represents the best example of genomic imprinting in humans. Genomic imprinting is a phenomenon by which certain genes are expressed in a parent-of-origin-specific manner. It is independent of the classical Mendelian inheritance. Genomic imprinting is an epigenetic process that involves DNA methylation and histone modifications to achieve monoallelic expression without altering the protein-coding sequence. These epigenetic marks are established in the germline and are maintained throughout the somatic cells of an organism. Forms of genomic imprinting have been demonstrated in mammals as well as insects and flowering plants. Prader-Willi syndrome is characterized by mild mental retardation, obesity, short stature, hypogonadism, and small hands and feet (Cassidy, Schwartz, Miller, & Discoll, 2012). Most commonly Prader-Willi syndrome is understood to involve the loss of a paternally imprinted stretch of five genes that encode polypeptides (MKRN3, MAGEL2, NECDIN, and SNURF-SNRPN) and a family of six paternal-only expressed small nucleolar RNA genes located on the proximal long arm of chromosome 15 (15q11.2-q13). There are three main molecular mechanisms that result in Prader-Willi syndrome: paternal deletion, maternal uniparental disomy and an imprinting defect. In 65–75% of the cases there is an interstitial microdeletion of the paternally inherited region on chromosome 15. In another 20–30% of cases, Prader-Willi is due to maternal uniparental disomy when the affected individual has two copies of chromosome 15, both of which are inherited from the mother. Maternal uniparental disomy has been shown to be associated with advanced maternal age (Gillessen-Kaesbach, Robinson, Lohmann, Kaya-Westerloh, Passargem, & Horsthemke, 1995). Finally, several forms of imprinting defects have been reported (Cassidy et al., 2012). Most of these defects arise from epigenetic causes while other individuals have a very small deletion (7.5 to >100 kb) in the promoter region of the SNRPN gene. In contrast to Prader-Willi syndrome, Angelman syndrome is characterized by a severe intellectual impairment as well as a speech impairment characterized by the absence or the minimal use of words, epilepsy, puppet-like ataxic movements, prognathism, tongue protrusion, paroxysms of laughter, abnormal sleep patterns, and motoric

144

Understanding Developmental Psychopathology

hyperactivity (Williams et al., 2006). Individuals with Angelman syndrome frequently exhibit socialization and communication deficits that meet the diagnostic criteria for ASD (Steffenburg, Gillberg, Steffenburg, & Kyllerman, 1996). Unlike Prader-Willi, Angelman’s syndrome is due to the loss of expression of two maternally imprinted genes (UBE3A and ATP10A). These two loci are immediately adjacent to the Prader-Willi region on chromosome 15. UBE3A encodes an E3 ubiquitin ligase that is expressed biallelically in most tissues but is maternally expressed in almost all neurons (Mabb, Judson, Zylka, & Philpot, 2011). One well-characterized substrate for UBE3A is the immediate early gene, Arc (activity-regulated cytoskeletonassociated protein). Arc is brain-specific and is rapidly upregulated in response to increases in neuronal activity. It is also required for learning and long-term memory (Greer et al., 2010). Angelman syndrome model mice, which carry a maternal Ube3a null mutation, recapitulate major features of Angelman syndrome in humans, including enhanced seizure susceptibility (Mabb et al., 2011). Interestingly, seizure or learning protocols induce abnormally elevated Arc expression in Ube3a knockout mice relative to wild-type controls (Greer et al., 2010), suggesting that UBE3A is required for Arc turnover in the brain during periods of increased synaptic activity. Excitatory neurotransmission onto neocortical pyramidal neurons has also been shown to be diminished (Wallace, Burette, Weinberg, & Philpot, 2012). These findings have resulted in an accelerating interest in UBE3A and its role in neural transmission. However, a number of critical gaps remain in our knowledge concerning UBE3A imprinting and its role in synaptic and circuit function. A deeper understanding of Prader-Willi and Angelman syndromes will likely shed light on why mammalian development is characterized by the nonequivalence of parental genomes. There are a few hints from the animal literature that genomic imprinting has allowed mammalian species to alter the dosage of particular genes and sculpt their patterns of expression using non-Mendelian mechanisms. For example, results from mouse chimera studies suggest that the accumulative effects of paternally derived imprinted genes differentially influence hypothalamic and septal structures but not the cerebral cortex. Conversely, the accumulative effect of maternally derived genes is evident in the cortex, striatum, and hippocampus, but not the hypothalamus (Keverne et al., 1996). Perhaps the most compelling argument is that monoallelic expression (either maternal or paternal) provides a clear source of stability amidst the temporal and spatial complexity that

characterizes the programming sequence associated with human CNS development (Keverne, 2011). Autism Spectrum Disorder Autistic behaviors have emerged as recognizable syndromes in the twentieth century based on the initial work of Kanner (1943), Asperger (1944), and Heller (1908). The current DSM-5 diagnostic criteria for ASD focus on two symptom dimensions: (1) persistent deficits in social communication and social interaction across contexts, not accounted for by general developmental delays; and (2) restricted, repetitive patterns of behavior, interests, or activities (American Psychiatric Association, 2013). At first glance ASD provides a clear example of a form of developmental psychopathology that involves a failure of conserved cognitive and emotional capacities to develop normally. Specifically, individuals with ASD are profoundly impaired in understanding the motives and mental states of other individuals. This impairment in their theory of mind extends into an understanding of one’s own mental states as well as the mental states of others (Baron-Cohen, Leslie, & Frith, 1985; Chapter 4, Volume 3). Yet there are perplexing issues. Why does ASD persist in the face of such a potentially negative effect on reproductive fitness? What is the evolutionary explanation for why these two dimensions (social communication and highly restrictive interests or repetitive behaviors) co-occur together? Intuitively these aspects of behavior do not seem to be closely interrelated. Some evolutionary explanations have approached this question. Most notably Baron-Cohen (2002) focused on the preponderance of males affected with ASD and proposed the extreme male brain theory of autism. This theory is based on the perspective that females on average have a stronger drive to empathize while males on average have a stronger drive to systemize. This theory is controversial, but it has led to an expanding body of empirical data focused on sex differences in brain structure and function and the potential role of fetal testosterone in influencing aspects of early brain development (Baron-Cohen, Lombardo, Auyeung, Ashwin, Chakrabarti, & Knickmeyer, 2011). More recently, Reser (2011) examined the hypothesis that some genes associated with ASD did in fact enhance reproductive fitness due to the adaptive benefits of being cognitively suited for solitary foraging. According to this theory, individuals with subclinical ASD included ecologically competent individuals that could have been adept at learning and implementing hunting and gathering skills on their own. During times when nutritional resources were sparse, human groups may have had to

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

periodically disband. In such circumstances individuals adept in solitary foraging were at an advantage. Reser also speculated that the most severe cases of ADS may be due to assortative mating. Evolutionary explanations also need to account for why children with ASD are much more commonly affected by a broad range of neurologic (e.g., epilepsy), neuropsychiatric, gastrointestinal and endocrine disorders (Chen, Chen, Liu, Huang, & Lin, 2009; Gurney, McPheeters, & Davis, 2006). The common neuropsychiatric and psychiatric comorbidities include stereotypic movement disorder, attention-deficit hyperactivity disorder (ADHD), obsessive-compulsive disorder (OCD), Tourette’s syndrome, and mood and anxiety disorders and sleep disorders (Gurney et al., 2006; Matson & Nebel-Schwalm, 2007). Ultimately, many of these questions will be answered by our rapidly expanding knowledge of the genetic and epigenetic factors that regulate the spatiotemporal development of the human brain. The neuropathogy associated with idiopathic ASD is variable although some structural and functional consistencies are emerging (Abrahams & Geschwind, 2008, 2010). Structurally, one well-replicated finding is that of accelerated postnatal growth in defined brain regions (Courchesne et al., 2010). There is also evidence of disrupted connectivity and laterality involving the frontal and temporal cortices, cerebellum, and amygdala (Eyler, Pierce, & Courchesne, 2012). Likewise, in response to viewing biological motion movements, children with ASD show differential patterns of activation compared with their unaffected siblings and typically developing controls (Kaiser et al., 2010). Biological motion refers to our evolutionarily conserved and robust visual sensitivity to the movements of other individuals (Blake & Shiffrar, 2007). The differential patterns of activation included prefrontal and temporal cortical regions as well as the amygdala and the fusiform gyrus (Kaiser et al., 2010). Intriguingly the unaffected siblings of the children with ASD also had unique patterns of functional activity in ventromedial prefrontal cortex and the right posterior temporal sulcus in response to biological motion. The authors speculate that since these regions have been previously implicated in aspects of social perception and social cognition (Adolphs, 1999), these regional activations may be compensatory in character such that these differential patterns might reflect a process through which brain function was altered over development to compensate for an increased genetic risk to develop ASD. In any case the social brain in its normal and dysfunctional states must be understood not in terms of specific structures but rather in terms

145

of their interaction in large-scale networks (Kennedy & Adolphs, 2012) Based on twin and family genetic studies, it is clear that ASD is highly heritable (Abrahams & Geschwind, 2008). A number of genetic mutations affecting single genetic loci have been described, including fragile X and Angelman syndromes. The number of single gene mutations associated with ASD phenotypes is increasing. One of the most well characterized disorders include Rett syndrome (Rett, 1966). Rett syndrome is one of the most common causes of ASD and intellectual disability in girls. It is characterized by early neurological regression that severely affects motor, cognitive and communication skills and is often associated with a seizure disorder. It is a monogenic X-linked dominant disorder related to mutation in MECP2, which encodes the methyl-CpG-binding protein MeCP2 (Smeets, Pelc, & Dan, 2012). Of interest MeCP2 deficiency in Rett syndrome causes epigenetic aberrations at the PWS/AS imprinting center that affects UBE3A expression (Makedonski, Abuhatzira, Kaufman, Razin, & Shemer, 2005). Among other single gene disorders, Novarino et al. (2012) recently identified inactivating mutations in the gene BCKDK (branched chain ketoacid dehydrogenase kinase) in consanguineous families with ASD, epilepsy, and intellectual disability. The encoded protein is responsible for phosphorylation-mediated inactivation of a branched-chain ketoacid dehydrogenase, causing individuals with homozygous mutations to display reductions in plasma branched-chain amino acids. Although the mechanism by which abnormal brain amino acid levels lead to autism, intellectual disability, and epilepsy is not fully understood, it likely involves the various biogenic amine pathways that underlie the dopaminergic, noradrenergic, serotonergic, and histaminergic pathways. It is also becoming clear that in addition to single gene disorders that impact multiple neurodevelopmental pathways, rare and de novo point mutations and submicroscopic variations in chromosomal structure contribute, but in isolation do not cause, a considerable number of ASD cases. To date more than 100 susceptibility genes have been identified (Betancur, 2011; State & Sestan, 2012). They include synaptic adhesion molecules and postsynaptic density proteins including FMR1, SHANK3, and NRXN1 as well as chromatin modifiers (MECP2, CHD8), and DNA binding proteins (POGZ), ion channels (SCN2A), microtubule-associated proteins (KATNAL2), neurotransmitter receptors (GRIN2B), and phosphorylation-regulated tyrosine kinases (DYRK1A) (Jamain et al., 2003; Kwan et al., 2012; Neale et al., 2012;

146

Understanding Developmental Psychopathology

O’Roak et al., 2011; Peça & Feng, 2012; Sanders et al., 2012; Sebat et al., 2007; Zachariah & Rastegar, 2012). In a recent exome sequencing of more than 340 children with ASD and one unaffected sibling Iossifov et al. (2012) estimated that there are at least 350 to 400 ASD susceptibility genes. In addition to rare and not so rare variants, it is also clear that common genetic variants likely account for substantial additive effects on ASD liability (Klei et al., 2012). Returning to the question—why does ASD persist in the face of such negative effects on reproductive fitness—we can now say with certainty that the normal expression of many of these susceptibility genes at the correct time and place in the developing brain are crucial to our capacity to understand the mental states of the other. However, it is also clear that unaffected siblings who share some of these susceptibility genes also have modestly reduced fecundity as well. This is particularly true of brothers (Power et al., 2012). This raises further questions about the persistence of the disorder and perhaps the possibility that the presence of individuals with certain autistic-like traits might benefit society as a whole. In addition, knowledge concerning the nature and timing of the relevant environmental insults that are associated with and increase the likelihood of ASD will also reveal further insights into the regulatory mechanisms that influence the emergence of our social and linguistic capacities over the course of development. For example, some of the strongest evidence indicates that advanced paternal age, especially for fathers over the age of 50 years; the immigrant status of the mother; prenatal nutrition; as well as pregnancy and birth complications, particularly gestational diabetes, can increase the risk of a child to develop ASD and related conditions including schizophrenia (Hamlyn, Duhig, McGrath, & Scott, 2013). The role of fetal levels of androgens (Lombardo et al., 2012) and events that can lead to activation of the fetal immune system early in life are also gaining increased attention (Crespi & Thiselton, 2011). Schizophrenia, Particularly Childhood-Onset Schizophrenia COS is variably defined by an onset of psychosis from age 12 to 18 years. Children with this condition have been identified and followed for decades (Asarnow, Tompson, & McGrath, 2004; Nicolson & Rapoport, 1999). The signs and symptoms of childhood/pediatric schizophrenia are nearly the same as adult-onset schizophrenia (Masi, Mucci, & Pari, 2006). Auditory hallucinations are the most frequent positive symptom, while visual and tactile hallucinations are rarer. The appearance of negative

symptoms including emotional, cognitive, and behavioral abnormalities is often seen years before the onset of psychosis. In many instances the child has a history of delays in cognitive, language and motor development (Alaghband-Rad et al., 1995; Hollis, 1995). Indeed, cognitive abilities in the borderline to the intellectually disabled range have been reported in up to 20% of children with schizophrenia and ASD (Sporn et al., 2004). A number of evolutionary theories have been advanced to account for the persistence of schizophrenia. For example, Brüne (2005) reviewed the available empirical evidence and concluded that, as in ASD, theory of mind is specifically impaired in schizophrenia and that many psychotic symptoms can be understood in light of a disturbed capacity in patients to recognize their own intentions, and to monitor others’ intentions. Similarly, Burns (2006) advanced the hypothesis that the evolution of our social brain and other higher order cognitive abilities is also responsible for the evolution of schizophrenia. The overlap of the natural histories of COS and ASD as well as the overlap of the evolutionary theories advanced to account for the persistence of ASD and schizophrenia provides the rationale for the inclusion of COS and schizophrenia more broadly in this section of the chapter. Structural brain imaging studies have documented an altered pace of neurodevelopmental trajectories in COS as well as in ASD. Specifically, in young adolescents with COS, there is increased velocity of gray matter thinning, which rectifies itself in parietal regions but partly persists in frontotemporal regions leading to the cortical deficit pattern similar to the one found in adult-onset schizophrenia (Gogtay, 2008; Rapoport & Gogtay, 2008). However, the trajectory of CNS development in COS differs from that seen in ASD. Genetically, there is clear evidence that many of the same genes convey vulnerability for both ASD and schizophrenia (Talkowski et al., 2012). For example, Lin, Hrabovsky, Pedrosa, Wang, Zheng, & Lachman (2012) using transcriptome sequencing identified 801 genes in differentiating neurons that were expressed in an allele-biased manner. These genes included a number of putative schizophrenia and ASD candidate genes including several involved in synapse formation and function (NRG1, NRG3, NLGN1, NLGN4, and NRXN1). Overall, there was a modest enrichment for schizophrenia and ASD candidate genes among those that showed evidence for allele-biased expression. This again points to the possibility of parent-of-origin effects and the possible role of genomic imprinting in the transmission of schizophrenia and ASD in some families.

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

Knowledge concerning the nature and timing of the relevant environmental insults that are associated with the increased risk of developing schizophrenia also overlap with those implicated in ASD. There is strong evidence which indicates that advanced paternal age increases the risk for both disorders (Miller et al., 2011). Curiously this meta-analysis also documented a modestly increased risk for schizophrenia among the offspring of younger fathers ( major depression) which likely contributes to the persistence of some of these traits in the population (Mathews & Reus, 2001). Most of these evolutionary hypotheses are not mutually exclusive. They also posit (or at least imply) that the neural mechanisms underlying these capacities can become dysregulated, leading to severe, chronic disabling and, at times, life-threatening symptoms and behaviors. The question then becomes how best to set the threshold for disorder and are there times when a “disorder” level impairment is normative and adaptive at a given developmental stage. While evolutionary theories may be useful in conceptualizing symptoms and disorders, population-based studies across age and gender groups and within diverse cultural and ethnic groups is a much more pragmatic and clinically useful approach. Given their universality, it is likely that fundamental aspects of human neural development are closely interrelated with these unpleasant emotions. The available neurobiological studies of rodents, primates and humans have identified key components of the emotional regulation and fear circuitry that include the amygdala and regions of the prefrontal cortex (Blackford & Pine, 2012; Canteras, Mota-Ortiz, & Motta, 2012; Kupfer, Frank, & Phillips,

2012). Studies of depressed youth also have implicated these same brain regions as well as areas of the brain involved in reward processing, particularly in a social context (Silk et al., 2012). Consistent with an evolutionary perspective, family, twin and adoption studies support an important role for genetic factors in the etiology of mood and anxiety symptoms and disorders (Hettema, Neale, & Kendler, 2001; Kendler, Gatz, Gardner, & Pedersen, 2006; Pawlby, Hay, Sharp, Waters, & O’Keane, 2009; Sullivan, Neale, & Kendler, 2000). These studies also consistently suggest that many of the genetic risk factors for mood and anxiety disorders are shared in common across these disorders (Kendler, Aggen, Knudsen, Roysamb, Neale, & Reichborn-Kjennerud, 2011; Kendler, Prescott, Myers, & Neale, 2003). However, efforts to identify specific genetic variants associated with an increased susceptibility to mood and anxiety disorders using large N genome-wide association studies (GWAS) have not been successful to date (Shyn et al., 2011). Despite this failure, it will be important to continue efforts to identify these susceptibility genes. Power et al. (2012) documented that siblings of patients with depression have significantly increased fecundity which more than compensated for the lower fecundity of affected individuals. This suggests that some of these susceptibility genes can increase an individual’s reproductive fitness. Efforts to identify gene by environment effects (GxE) with regard to mood and anxiety disorders have also been of great interest since the initial report by Caspi et al. (2003). They noted that a functional polymorphism in the promoter region of the serotonin transporter (5-HTT) gene (SLC6A4) moderated the influence of stressful life events on depression. In the initial report, individuals with one or two copies of the short allele of the 5-HTT promoter polymorphism exhibited more depressive symptoms, diagnosable depression, and suicidality in relation to stressful life events than individuals homozygous for the long allele. These findings have not been universally replicated, but the preponderance of the data from quantitative population genetics, neurobiological studies, and experimental studies in animals confirms the plausibility of their initial findings although the overall effect size may be modest (Caspi, Hariri, Holmes, Uher, & Moffitt, 2010). Two approaches of considerable promise for future research are the emerging data from population-based samples of twins and efforts to begin to understand the role of epigenetic modifications that occur as a result of environmental adversities. First, the most convincing data concerning GxE effects have come from recent efforts

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

to examine the impact of environmental experiences on symptoms of anxiety and depression across the life span using longitudinal data from eight American and European population-based twin registries (Kendler et al., 2011). By modeling data from monozygotic (MZ) twin pairs of widely varying ages this group of investigators were able to reject the hypothesis that genetic influences alone can account for the relative stability of anxiety and depression across the life span. Based on their analyses of the individual trajectories, they determined that levels of symptoms increasingly diverged between the MZ twins from childhood to late adulthood based on their environmental experiences. Only in late adulthood (∼70 years of age) did this divergence disappear. Efforts to delineate the multiple and diverse epigenetic and neurobiological mechanisms by which environmental experiences shape the character and continuity of mood and anxiety disorders is also an important area for future study. It is increasingly evident that environmental experiences can lead to epigenetically mediated alterations in gene expression that likely influence an individual’s vulnerability to depression and anxiety (Monk, Spicer, & Champagne, 2012). To date the strongest data from human and animal models concerns the epigenetic consequences of early adversity. For example, analysis of cord blood samples from infants born to depressed mothers indicated elevated levels of DNA methylation within the gene (NR3C1) that codes for the glucocorticoid receptor (Oberlander, Weinberg, Papsdorf, Grunau, Misri, & Devlin, 2008). At 3 months of age, these investigators assessed the infants’ stress responsivity by examining their salivary cortisol levels. Levels of NR3C1 methylation in fetal cord blood were found to predict infant cortisol response to stress, suggesting a functional consequence of this epigenetic variation. Similar epigenetic changes have also been reported for the gene encoding the serotonin transporter (SLC6A4) (Devlin, Brain, Austin, & Oberlander, 2010). Although the meaning of these epigenetic changes in blood is unclear with regard to brain function, as epigenetic alterations can be tissue specific and even specific to particular brain regions (Provençal et al., 2012), animal models have convincingly demonstrated that exposure to prenatal stress can induce similar epigenetic changes in the most relevant brain regions as well as increased depressive-like behavior in adult animals (Mueller & Bale, 2008). Viewed from an evolutionary perspective these data suggest that we and our mammalian ancestors have evolved molecular mechanisms that permit some degree of programming so that the offspring may be better prepared for the environmental context in which their rearing will take place.

149

Obsessive-Compulsive Disorder Dysregulation of otherwise adaptive behavioral patterns linked to specific environmental perturbations may also be readily evident in other forms of psychopathology. Although speculative, this list includes OCD. OCD has a lifetime prevalence of 2–3%, affecting all age groups across the globe (Ruscio, Stein Chiu, & Kessler, 2010). OCD is considered to be the fourth most common mental disorder and is frequently accompanied by family, social, school, and work dysfunctions (Ayuso-Mateos, 2006). Obsessive-compulsive (OC) symptoms, like symptoms of anxiety and depression are dimensional in character and evolutionary accounts for each of the symptom dimensions have been proposed (Evans & Leckman, 2006; Mataix-Cols, Rosario-Campos, & Leckman, 2005). Obsessions are ego-dystonic, repetitive, and intrusive thoughts or images that are not simply excessive worries about real-life problems (American Psychiatric Association, 2013; Leckman et al., 2010). Classic examples include intrusive worries about cleanliness or about harm coming to close family members. Compulsions are repetitive behaviors (e.g., hand washing, checking, ordering) or mental acts (e.g., praying, counting, repeating words silently) that the person feels driven to perform in response to an obsession. These rituals usually need to be performed according to certain rules. Often there is a heightened sensitivity in the home environment to things not being “just right” in the way they feel, look, or sound (Leckman, Walker, Goodman, Pauls, & Cohen, 1994). This awareness that something is not just right is commonly associated with a sense of dread that something terrible is about to happen. Other OC symptom dimensions include preoccupations about symmetry and exactness and compulsions of ordering and arranging, doing and redoing, and counting (Leckman et al., 1997). Here again we may see the presence of an ancient alarm system that may have served well our hunter–gatherer ancestors as they struggled to survive in the face of predators and unsanitary conditions (Brüne, 2008; Feygin, Swain, & Leckman, 2006; Marks & Nesse, 1994). Beginning in the second year of life, most normally developing children display a variety of intrusive thoughts, habits, and routines, some of which closely resemble the behaviors associated with OCD (Evans, Gray, & Leckman, 1999; Evans et al., 1997; Gesell, 1928; Leonard, Goldberger, Rapoport, Cheslow, & Swedo, 1990; Zohar & Felz, 2001). For example, in one large population-based study (n ∼ 1,500), we found that parents reported that over half of the 2-year-olds were very aware of minute details, such as imperfections in toys and clothes (Evans et al., 1997).

150

Understanding Developmental Psychopathology

Likewise, parents of 2-year-olds were also almost twice as likely to report that their child is very concerned with dirt, cleanliness, or neatness than were parents of 5-year-olds, and almost 7 times more likely than parents of children less than age 1. Finally, 63% of 2-year-olds were reported to arrange objects until they are just right compared with 4% of children less than 1 year of age and 41% of 5-year-olds. For children less than 4 years of age, repetitive behaviors were positively related to the child’s level of fearfulness in general as well as their fear of strangers (Evans et al., 1999). Interestingly, bedtime fears (fear of the dark, of monsters) were related to hoarding objects; fear of death was significantly associated with strong attachment to a specific object. For children older than age 4 years, a greater number of correlations emerged. Most notably, overall scores on the rituals inventory were related to bedtime fears. Just right behaviors were related to fears of contamination and fear of separation; and fear of death was positively related to repeats actions over and over, and arranges objects until they are just right (Evans et al., 1999). As noted by Evans and Leckman (2006), these findings echo many of the theories of Erikson (1968), Werner (1957), and Winnicott (1958a) that children’s rituals are associated with anxiety states brought on by the many fears and uncertainties that are common in childhood and that they bear some relationship to the presence or absence of attachment figures as a source of security and syncretic wholeness. These findings also emphasize the normal capacity that young children have to compare elements of the outside world with their existing mental representation to ensure that important aspects of their world are just right (Gesell, 1928; Piaget, 1962). This in turn recalls les sentiments d’incompletude (feelings of incompleteness) that Janet (1903) identified as the core problem for individuals with OCD at the end of the nineteenth century. The initiation of parenting is another developmental stage that is normally characterized by intense harm-avoidant preoccupations and a heightened sense of responsibility accompanied by compulsive attention to issues of safety and cleanliness characterized by anxious intrusive thoughts and harm avoidant behaviors (Kim, Mayes, Feldman, Leckman, & Swain, 2013; Leckman, Mayes, Feldman, Evans, King, & Cohen, 1999; Winnicott, 1958a). The normally occurring emergence of these symptoms may well account for the increased risk of onset or exacerbation of OCD seen late in pregnancy and during the early postpartum period (Abramowitz et al., 2010; McGuinness, Blissett, & Jones, 2011). Viewed in the context of the historically high rates of infant mortality

(Bideau, Desjardins, & Brignoli 1997) these observations are consistent with the hypothesis that the dysregulation of conserved behavioral patterns and associated mental states can contribute to the emergence of particular psychopathological outcomes. Like the other anxiety disorders, OCD has been conceptualized as an imbalance between subcortical and cortical circuits. A relatively large and growing number of neuroimaging studies have examined both functional and structural correlates of OCD in both children and adults (Blackford & Pine, 2012). Functional imaging studies have consistently reported hyperactivity in the orbitofrontal cortex, anterior cingulate cortex, and caudate nucleus of individuals with OCD. Although these findings and the results of structural studies are not sufficient to establish causality, a relatively compelling case can be made that an imbalance in the cortico-basal ganglia-thalamo-cortical loops that specifically involve the orbitofrontal and the anterior cingulate cortices play a causal role in this disorder (Maia, Cooney, & Peterson, 2008). Consistent with an evolutionary perspective, family, twin, and adoption studies have demonstrated that both genetic biological and environmental factors are important in the etiology of OCD (Pauls, 2010). Family-genetic studies published since the 1930s indicate that among first-degree relatives of OCD-affected children and adults there is a four- to tenfold OCD risk increase, respectively, as compared with relatives of controls. A review of twin studies concluded that obsessive-compulsive (OC) symptoms are heritable, with greater genetic influences in child-onset (45–65%), than in adult-onset OCD cases (27–47%) (Pauls, 2010). However, efforts to identify specific genetic variants associated with an increased susceptibility to OCD using large N GWAS have not been successful to date (Stewart et al., 2013). In conclusion, the available data support the plausibility of the hypothesis that many OC symptoms appear normally during the course of development and confer some degree of reproductive fitness. It is only when there is some dysregulation of evolutionarily conserved neurobiological pathways that extreme variants appear and a diagnosis of OCD is warranted. While future genetic research, e.g., next-generation sequencing strategies, the study of rare variants), longitudinal epidemiological studies, particularly population-based twin studies (Kendler et al., 2011), and the development of valid animal models (Chen et al., 2010; Yaddanapudi et al., 2010) will continue to elucidate the relevant etiological pathways, it is difficult to imagine how specific evolutionary hypotheses concerning OCD can

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

be formally tested. Consequently, these hypotheses are best viewed as strategically important in the conceptualization of the study of psychopathology and the integration of data from across multiple disciplines. Ancient Versus Current Environments and the Value of Diversity Within Populations Another evolutionary explanation focuses on the changing nature of our environments. This viewpoint posits that much of our human behavioral equipment can be best understood in relationship to our primeval environment where we evolved as a species. As Bowlby (1969) pointed out, the property of being adapted entails a reference both to a specified outcome and to a specified environment. Thus as modern societies shift away from their agrarian pasts, certain behavioral traits might become less adaptive. That said, things never stay the same and the multiple environments in which we dwell are constantly changing. Global warming and climate change are emerging as a major challenge to our species while the near universality of virtual worlds via social media and the internet are now part of everyday life even for preschool children. Indeed, our capacity as a species to explore and adapt to novel environments is one of our most distinguishing traits—compared with other species (Morán, 2008). Work in this area also highlights another aspect of evolutionary theory—namely, in the struggle for life it is the reproductive fitness within societal groups that is selected for rather than any particular individual. With this perspective in mind, let us consider ADHD. Attention-Deficit/Hyperactivity Disorder According to the DSM-5, ADHD is defined by a constellation of clinical features, including hyperactivity, inattention, and impulsivity (American Psychiatric Association, 2013). Children with ADHD are more likely to exhibit school behavior problems, academic underachievement, poor self-esteem, parent–child conflicts, and difficulties with peer acceptance (Abikoff, Gittelman-Klein, & Klein, 1977; Barkley, 1997; Erhardt & Hinshaw, 1994; Chapter 14, Volume 3). ADHD is a chronic disorder—as children age, become teenagers, and enter adulthood, they are more likely to engage in antisocial behaviors including substance abuse and criminal acts, as well as the persistence of attentional, familial, interpersonal, and occupational difficulties throughout their life span (Cherkasova, Sulla, Dalena, Pondé, & Hechtman, 2013; Klein & Mannuzza, 1991). These symptoms, and the compromised function they entail, should constitute strong evolutionary pressures

151

that perhaps should have eliminated ADHD from the population within a few score generations. Yet ADHD is one of the most common psychiatric disorders in children and adolescents, with a worldwide prevalence above 5% (Polanczyk, de Lima, Horta, Biederman, & Rohde, 2007). So why does ADHD persist in the face of such a potentially negative effect on reproductive fitness? Several evolutionary explanations have approached this question. For example, some have argued that in some instances the symptoms of ADHD are best viewed from the perspective of our ancestral environments that were unsafe, impoverished, and time critical. In such an environment there would be advantages to maintaining a response-ready versus a problem-solving attitude (Jensen et al., 1997; Urcelay & Dalley, 2012). Being distractible is perhaps better attuned to survival when every change in the environment could mean a threat or an opportunity. Acting on impulse may have left few opportunities for questioning later, but it has undeniable survival value in dangerous environments. Other potential advantages for ADHD-like symptoms include: alertness to novel stimuli; willingness to depart from local environments predisposed to high levels of prenatal and perinatal stressors, and potential positive founder effects upon discovery of novel, resource-rich environments. Among the many attempts to explain the persistence of ADHD, Killeen, Tannock, and Sagvolden (2012) provided the most current and comprehensive review as they consider the four causes of ADHD and point to the vast interdisciplinary literature from brain structure, electrophysiology, and neuropsychology to the role of neurotransmitters and genes. Specifically, interconnections between attention-regulating subsystems in temporal, parietal, and prefrontal cortices are likely involved in many cases of ADHD (Arnsten, Berridge, & McCracken, 2009). The prefrontal cortices, in particular, play a top-down inhibitory function by regulating emotions and behaviors, sustaining attention, and by screening out distracting stimuli. In this context, it appears that cortical development lags in children with ADHD compared with their peers (Shaw et al., 2007). Family, twin, and adoption studies also indicate that ADHD is highly heritable with heritability estimates consistently around 0.8 (Faraone & Mick, 2010). A large number of candidate genes have been and continue to be evaluated. Historically, much of the focus has been on genes involved in the synthesis and processing of biogenic amines. However, large scale GWAS have yet to identify convincingly any common genetic variants associated with increased susceptibility to ADHD (Neale et al.,

152

Understanding Developmental Psychopathology

2010). It appears likely that the genetic vulnerability to ADHD is mediated by multiple generalist genes each of small effect. In considering GxE, it is clear that perinatal risk factors such as birth trauma, anoxia, and prematurity can contribute to the risk of a child developing ADHD (Barry & Gill, 2007). It will also be important to examine how adverse environments characterized by low parental responsivity, high and critical expressed emotion, under stimulation, harsh or coercive disciplinary techniques, conflictual parental relationships, and low socio-economic status, can shift early brain development and neural systems more toward ADHD-like outcomes. Moving forward, the possibility that these risk factors operate through epigenetic mechanisms is an important set of molecular mechanisms to consider (Mill & Petronis, 2008). It is also worth noting that some individuals with ADHD are more likely to engage in risky sexual behaviors that may lead to their having more children with multiple partners (Flory, Molina, Pelham, Gnagy, & Smith, 2006; Hosain, Berenson, Tennen, Bauer, & Wu, 2012). These empirical observations are also consistent with the growing literature of differential susceptibility. As noted earlier in this chapter, in some adverse environments impulsivity, risk taking, and frank antisocial behavior may be adaptive to the degree that they enhance an individual’s reproductive fitness (Belsky & Pluess, 2013). The findings of an innovative computational modeling study undertaken by Williams and Taylor (2006) are also consistent with differential susceptibility. Briefly, they attempted to simulate the potential evolutionary benefits for groups when they have in their midst a minority of individuals who were highly variable, i.e., unpredictable, in their risk-taking approaches to secure food resources within changing environments. They initially constructed five groups of ten individuals. Nine of the individuals were programmed to be very predictable. Information about the availability of local foods was shared within each of the groups. Ninety per cent of the matings were programmed to occur outside the individual’s original group. In the simulation each of the offspring who had one of the five brittleness genes were randomly selected from the two parents, and in 10% of matings one of the five genes randomly mutated. Whenever the overall population exceeded 200, a group was randomly selected for elimination. Individuals started reproducing at age 10, and life span was limited to 50 years. When groups exceeded 20 individuals, they split into two groups. They then simulated the ability of two interventions to alter the reproductive fitness of the groups. These were the rate of environmental change and

the relationship between an individual’s predictability and rate of reproduction. They found that increased environmental variability produced a small, but definite and reversible reduction in population predictability. Given the parameters used in their simulations, they concluded that there is a class of tasks (group exploration tasks) in which unpredictable behavior by a minority of individuals optimizes results for the group. Characteristics of these tasks were risk taking, because its cost is borne mainly by the individual, and information sharing, because its benefits increase with group size. More importantly they found that a reproductive bias favoring the unpredictable individuals helped populations cope with rapid environmental change, without imposing major cost during periods of stability. In contrast, when they programmed in a reproductive bias favoring predictable individuals, they observed an opposite effect on reproductive fitness. Despite the obvious limitations of reducing ADHD to a single behavioral characteristic—unpredictability—the use of computational modeling may prove useful in evaluating other evolutionary hypotheses for ADHD and other forms of developmental psychopathology. This perspective is also consistent with the widely held view that the factors that shape human development do so in a probabilistic rather than a deterministic fashion (Belsky & Pluess, 2013). Co-optation of Neurobiological Systems Associated With Establishing Salience and Reward Perhaps the most compelling evolutionary explanation to be considered in this chapter relates to substance abuse and addictive disorders. Here it comes down to how our brains are built. If we rely on specific genetic pathways, neurotransmitters, and circuits to establish what is salient and rewarding in our world, then exogenous agents that affect these neurobiological systems may well have effects that are sufficiently maladaptive to be labeled a form of psychopathology. Substance Use Disorders and Addictive Disorders Drug abuse and addictive disorders including pathological gambling exact an enormous emotional, medical, and financial toll on societies across the globe in the form of the loss of an individual’s potential to contribute to society, loss of employment, crime, and family disintegration, in addition to medical complications in the case of an overdose (Anthony, Warner, & Kessler, 1994; Le Moal, 2010; Chapter 18 and 19 Volume 3). Considered together, alcohol, tobacco, and illicit drug use are implicated in

Evolutionary Mechanisms That May Account for the Persistence of Discrete Forms of Psychopathology

over 12% of mortality worldwide. Their use constitutes the leading cause of preventable death (World Health Organization, 2009a, 2009b). Addiction is a chronic, relapsing disorder that has been characterized by the impulsive and compulsive need to seek and consume the drug (craving); an inability to control intake leading to intoxication; and the emergence of a distressed emotional state, e.g., dysphoria, irritability, during periods of withdrawal when access to the drug is prevented (Koob & Volkow, 2010). Animal and human studies have identified widely distributed neural circuits that mediate each of the three stages of the addiction cycle. Remarkably, virtually all substance use disorders engage many of the same neural circuits. As reviewed by Koob and Volkow (2010), key elements of the reward and salience mesolimbic dopamine pathways are structures that once activated can lead to a cascade of neuroadaptations in the ventral striatum (i.e., a brain region that includes the nucleus accumbens core and shell and some nuclei of the olfactory tubercle) that receives projections from the ventral tegmental area and prefrontal cortex. Once addiction has been established, multiple interconnected brain regions are involved in craving including the orbitofrontal and other regions of the prefrontal cortices, as well as the insula, basolateral amygdala, and hippocampus. Although many individuals are exposed to nicotine, alcohol, and drugs of abuse, only a subset becomes addicted. Although the genetic contribution to risk for addiction is approximately 50%, the specific genetic variants have yet to be identified (Goldman, Oroszi, & Ducci, 2005; Kendler, Myers, & Prescott, 2007). One probable explanation for this failure is that the relevant sequences do not lie within the protein coding regions of these genes but rather in the recently discovered regulatory elements which are physically associated with one another and with expressed genes. Since an individual’s susceptibility to drug use and addiction is strongly influenced by the psychological and social context in which drug exposure occurs, and since these genomic regulatory elements can be modified by environmental exposures, it is not surprising that the search for key transcription factors, as well as efforts to identify the molecular events that can regulate transcription through modifications of DNA and chromatin structure in specific brain areas have become a major focus for research (Robison & Nestler, 2011; Wood & Lipovich, 2012). For example, Robison and Nestler (2011) recently reviewed some of the multiple mechanisms by which drugs of abuse can alter the expression of key transcription factors, i.e., proteins that,

153

in response to cell signaling pathways, bind to specific sequences of DNA to increase or repress the expression of downstream genes. Their review focuses on four specific transcription factors: ΔFOSB, cyclic AMP-responsive element binding protein (CREB), nuclear factor-κB (NF-κB), and multiple myocyte-specific enhancer factor 2 (MEF2) and their roles in addiction. They also usefully reviewed the current state of knowledge concerning various epigenetic mechanisms including histone tail modifications which alter the accessibility of genes within their native chromatin structure to the molecules needed for transcription, DNA methylation which can have enduring effects on gene expression as well as non-coding microRNAs. MicroRNAs are found in all mammalian cells and serve as post-translational regulators that bind to complementary sequences on target mRNAs to repress translation, thus silencing gene expression. Like histone modifications and DNA methylation, expression of microRNAs can alter gene transcription in the absence of any change to the DNA coding sequence. It is reasonable to anticipate that as future insights on transcriptional and epigenetic mechanisms of addiction accumulate, a more complete understanding of how addictive substances modify specific neural circuits will emerge. Hopefully this in turn will contribute to the identification of new targets for the treatment of addictive disorders. In sum, addiction and substance use disorders provide a compelling example of how our evolutionary history has left us vulnerable to certain forms of psychopathology (Nesse & Berridge, 1997). Specifically, the conserved neural circuitry involved in reward processing and salience determination in vertebrate species can be co-opted or hijacked by exogenous substances, leading to addiction and associated behavioral dysfunction and impairment. An Evolutionary Arms Race: Infections and Autoimmunity At first glance, the mention of infectious processes in a listing of evolutionary mechanisms might seem odd. However, our long-standing evolutionary relationship with bacterial and viral pathogens and the endless arms race between our species (theirs and ours) has probably done much to shape aspects of our genome and so our vulnerability to disease (Chapman & Hill, 2012; Dawkins, 1987; Dawkins & Krebs, 1979). In psychiatry these effects include post-infectious autoimmune disorders that have neuropsychiatric sequelae such as Sydenham’s chorea (Husby, van de Rijn, Zabriskie, Abdin, & Williams, 1976; Kiessling, Marcotte, & Culpepper, 1993; Swedo et al., 1989).

154

Understanding Developmental Psychopathology

PANDAS

Other Evolutionary-Based Explanations

Swedo and coworkers (1998) described a group of children with PANDAS characterized by the sudden overnight onset of OCD as well as at least two other neuropsychiatric symptoms (i.e., anxiety, emotional lability, irritability, aggression or oppositional behavior, behavioral regression, deterioration in school performance, sensory or motor abnormalities including motor or vocal tics, and somatic signs and symptoms like sleep difficulties, enuresis, and urinary frequency) in children who had been previously developing normally. Currently, the predominating theory to explain the pathophysiology behind PANDAS is molecular mimicry whereby antibodies formed to target Group A beta-hemolytic streptococci (GABHS) also targets proteins in the cells of the basal ganglia. However, since the introduction of PANDAS as a clinical entity, it has been a source of a fair amount of controversy, partly stemming from how the original PANDAS criteria have been interpreted by subsequent teams of investigators (Singer, Gilbert, Wolf, Mink, & Kurlan, 2012; Swedo, Leckman, Rose, 2012). Despite this controversy cogent models of disease pathogenesis have been put forward in which the neurotransmitter dopamine is influencing the functional status of the immune system (Besser, Ganor, & Levite, 2005; Kipnis et al. 2004; Sarkar, Basu, Chakroborty, Dasgupta, & Basu, 2010) and that antibodies directed against GABHS can increase the amount of dopamine released and affect the functional status of dopamine receptors (Kirvan, Swedo, Heuser, & Cunningham, 2003; Kirvan, Swedo, Kurahara, & Cunningham, 2006; Murphy, Kurlan, & Leckman, 2010). Additional clinical, translational, and basic science investigations are needed. Strikingly, animal models are again providing a proof of concept for PANDAS as a clinical entity (Birmberg et al., 2012; Yaddanapudi et al., 2010). For example, GABHS infections can trigger the production of cross-reactive antibodies that not only lead to some of the same neurologic and behavioral symptoms in the originally infected mice, but also evoke symptoms when they are passively transferred from donor to wild-type mice. Expanding these investigations to evaluate the effects of a variety of infectious agents and other immune stimulants would be useful, particularly if the animal models can be used to evaluate potential therapeutic interventions. From an evolutionary perspective since humans are the only known natural host for GABHS, the persistence of Sydenham’s chorea and PANDAS in our species is attributable to our continued vulnerability, as a species, to this pathogen that has made its permanent home in our bodies.

Several other evolutionary-based accounts of psychiatric disorders have been proposed. They include the limited power of natural selection to eliminate mutations, particularly recessive mutations, if their adverse effects on reproductive success are modest. This may be particularly germane for mild or late-onset disorders. For example, in Huntington’s disease the initial symptoms do not appear until after many of the affected individuals have had offspring. In contrast, inbreeding and consanguineous marriages which are commonplace in many cultures do have the potential to disrupt neurodevelopmental pathways and cause early onset neuropsychiatric disorders. A recent example, cited above, concerns inactivating mutations in the gene BCKDK in consanguineous families with ASD, epilepsy, and intellectual disability (Novarino et al., 2012). The heterozygote carriers of this mutation develop normally and their reproductive fitness is not otherwise affected. However, individuals can develop this syndrome when the mutation is transmitted within extended families so some descendants are homozygous for the same mutated recessive gene.

CONCLUSIONS AND CRITIQUE The study of evolutionary principles suggests that many of the mental and neurodevelopmental disorders encountered by professionals working with infants, children, and adolescents (e.g., clinical psychologists, pediatricians, primary care clinicians, child and adolescent psychiatrists, social workers, teachers) are likely to persist despite their being associated with a reduction in reproductive fitness. In many instances their ultimate causes are built into the genetic and neurobiological mechanisms that underlie highly conserved features of our basic behavioral and cognitive repertoires. While we usually confine our discussion of these conserved features to humanoid, primate, or mammalian lineages, the fossil record is clear that some of the most defining events in our neurobiological history date back more than 500 million years with the emergence of our first chordate ancestors. Examination of the cranial endocasts of our earliest vertebrate ancestors indicate that by 480 million years ago they already had the major divisions of the central nervous system (CNS): telencephalon, diencephalon, mesencephalon, metencephalon, myelencephalon, and spinal cord as well as olfactory capsules, orbital cavities, a hypophysis, labyrinths, and the exit

Conclusions and Critique

foramina for the 10 cranial nerves (Miklos, 1993; Karen, 1970). Remarkably, our CNS is simply a variant on the same theme. A second point focuses on the intrinsic population-based nature of evolution. Variation within populations is an essential ingredient for natural selection (Pujadas & Feinberg, 2012). This is true at each level of understanding from linear arrays of genes in genomes, to populations of molecules within cells, to collections of neurons, to organized behavioral repertoires, and even to collections of people in social communities. Several corollaries to this point are also apparent. First, at each level it appears that any single function can usually be carried out by more than one configuration of units or individuals. For example, many genes are pleiotropic with many functions. These functions can overlap with one another so that the loss of one gene can be made up by another (Finch & Rose, 1995). Similar arguments are possible at each of the other levels as well. Second, there is the potential for pleiotropy at more proximate levels. For example, central oxytocin pathways appear to play an important role in the initiation of a wide range of complex behaviors from pair bonding and sexual activity to uterine contractions during delivery and milk release during nursing. In the body, oxytocin is synthesized in the heart, thymus, gastrointestinal tract, as well as reproductive organs. The distribution of the oxytocin receptors in both the brain and periphery is even more far-reaching and its expression is subject to changes over the course of development. Oxytocin receptor expression is also sensitive to changes in the external environment and the internal somatic world. The net result is a vast, complex, and developmentally sensitive bio-behavioral system (Gordon, Martin, Feldman, & Leckman, 2011). Other elements include sensory inputs, the salience, reward, and threat detection pathways, the hypothalamic-pituitary-gonadal axis, and the hypothalamic-pituitary-adrenal stress response axis. A third inescapable lesson from the study of evolutionary mechanisms is the importance of the environment (ancestral, intrauterine, and the infant’s early nurturing environment) in shaping our genome and its capacity to be programmed to maximize adaptation and diversity. For each proposed mechanism some aspect of the environment, past or present, is playing or has played a crucial role. Perceived threats to survival, emotional attachments to others, the availability of drugs of abuse, the existence of reservoirs of beta-hemolytic streptococci, are just a few examples. A fourth point concerns the self-organizing character of these bio-behavioral systems—namely, that they have

155

the capacity to modify and reorganize in response to new environmental perturbations and to settle into one or a few modes of organization (Thelen & Smith, 1995). Behaviorally this is seen throughout the animal kingdom as we tend to settle into familiar habits. Indeed, one way of conceptualizing some forms of psychopathology is to consider them to be a reflection of individuals getting stuck in a restricted set of otherwise normal behaviors. Examples include the washing and cleaning rituals of some forms of OCD or the motor stereotypes seen in individuals with ASD. The basis for this fractal property of evolution (the ability of populations to self-organize regardless of the temporal or spatial scaling) remains speculative, but it would seem to be a requirement for evolution to have occurred in the first place. A fifth conclusion concerns the incomplete nature of evolutionary explanations. They may provide a plausible explanation of why certain vulnerabilities persist within human populations, but they do not account for why a particular individual is affected. Here the study of individual differences and of particular developmental histories is required. Special attention should be directed at events occurring early as they can constrain the later development of the CNS (Edelman & Tonini, 1995). Models of pathogenesis of early-onset disorders must embrace both developmental and evolutionary perspectives. Finally, we must acknowledge that although some of these evolutionary accounts have a certain degree of face validity (co-optation of conserved neurobiological systems, evolutionary arms races) many of the evolutionary accounts discussed above are speculative and largely untestable. Given the historical nature of the proposed accounts, the fragmentary fossil record, and the limits of paleobiology to document brain evolution, we must admit that the actual course of these events remains inexplicable. While computational models may be useful in establishing the plausibility of some of these mechanisms, many of these accounts may be justly characterized as just so stories that provide explanations that cannot be disproved. We must be content then with (1) their heuristic value to raise questions concerning data already available to us, (2) their ability to provide a framework for the integration of new knowledge from a broad range of scientific disciplines, and (3) their value to clinicians. Future Prospects The conceptual framework underlying emerging models of brain evolution and development provides a powerful framework for understanding aspects of disease

156

Understanding Developmental Psychopathology

pathogenesis in developmental psychopathology. This framework is consistent with the pioneering efforts of investigators such as Baldwin (1902), Bowlby (1969, 1973), Darwin (1872), Ekman (1982), and Waddington (1977) among others who have applied evolutionary and developmental principles to the study of normal variation or psychopathology. Although progress has been slow, due in part to the scope and complexity of the emerging scientific knowledge base, incremental progress can be anticipated that should enrich and refine our emerging models of disease pathogenesis. Future progress is likely to follow in the wake of powerful and promising technologies in genomics, epigenomics, neurodevelopment, neuroimmunology, longitudinal neuroimaging, computational modeling, and information processing, as well as taking greater advantage of population-based registries. Despite the relative failure of most GWAS it is clear that many rare genetic variants acting alone or together can be lethal for the normal emergence of species-typical behavioral patterns (as may occur in some forms of ASD). Despite many disappointments, it is reasonable to expect the eventual identification of multiple alleles of major and minor effect that act in concert (co-action) that contribute to a number of heritable disorders including some forms of intellectual disability, dyslexia, ASD, schizophrenia, Tourette’s syndrome, and antisocial behavior (Bentley et al., 2013). Future translational work with animal models should focus on the role that genes and/or environmental factors play in the development of circuits that regulate specific physiological and behavioral outcomes. This emphasis on circuit formation, as a fundamental unit for understanding behavior, is distinct from current approaches of modeling psychiatric illnesses in animals (Kaffman & Krystal, 2012). The work over the past several decades of Capecchi and his team at the University of Utah on the role of homeobox genes on development provides a clear illustration of this point. Homeobox genes are a class of regulatory genes that act primarily as transcription factors in the promoter regions of the genome. They are also known as master control genes because of the key role they play in the control of embryonic development along the longitudinal axis of the developing embryo from head to toe in all multicellular animals from fruit flies to humans. Capecchi and his team have for some time been pursuing the systematic analysis of the mouse homeobox genes by monitoring the animals’ somatic and behavioral phenotypes after a specific homeobox gene was knocked out. In 2002, Capecchi and his team knocked out the Hoxb8 gene. Remarkably, although this gene is anatomically expressed no higher than the lumbar region

of the developing embryonic spinal cord, all of the Hoxb8 (–/–) mice displayed excessive grooming behaviors leading to hair removal and skin lesions (Greer & Capecchi, 2002). Why did this behavioral phenotype emerge? Eight years later, they provided the answer. The Hoxb8 cell lineage is associated with the bone marrow-derived microglia that migrates into the brain during embryonic development from the yolk sac (Chen et al., 2010). The conclusion is clear: a behavioral disorder, with some resemblance to trichotillomania, is associated with mutant microglia in the brain. The role of microglia in normal brain development and circuit formation has also become increasingly clear given their active surveillance of synaptic activity and their capacity to prune presynaptic inputs based on their level of neural activity (Schafer et al., 2012). This brings us to another conclusion, namely, that we still have a great deal to learn about the role of the immune system and its role interconnecting the brain and body as well as its interface with other bio-behavioral systems. Based upon our limited knowledge, one can safely predict that additional stories will emerge that will strengthen the link between psychoneuroimmunology and stress response, maternal bonding, and sexual behaviors, as well as salience and reward pathways. Recently, data also suggest that the immune system will play a role in modulating feeding behavior and energy homeostasis, and perhaps even in the multi-directional microbiome-gut-brain network (Poutahidis et al., 2013). All of which point to the importance of the central role the immune system plays at the dynamic interface of physical and mental health (see Leckman, 2014). Although animal model systems may permit the testing of the effectiveness of novel pharmaceutical agents as well as a deeper understanding of how the expression of these particular alleles constrains the developing nervous system (Andersen & Navalta, 2011), species differences and ethical constraints place a natural limit on the testability of evolutionary and developmental explanations of human psychopathology. The recent finding for the various ENCODE studies concerning the many differences in noncoding regulatory regions of the human genome compared with other species reinforce the view that animal models will always be incomplete. Our ability to genotype accurately individual patients with regard to known vulnerability alleles as well as regulatory regions of the human genome may lead to more accurate clinical predictions of course, outcome, and treatment response. Similarly, the results of neuropsychological, neuroimaging, and other biological studies may become more interpretable by classifying patients according to their respective genotypes (Bansal et al., 2012).

References

For developmental neurobiologists the ongoing studies comparing genomes across species as well as comparing spatiotemporal differences of gene expression should provide a clearer understanding of the transcriptional foundations of human brain development (Johnson et al., 2009; Ward & Kellis, 2012). Future studies of human postmortem brain material that include recently discovered regulatory elements in the human genome should also allow investigators to generate expression trajectories of the interactive genes present in large-scale networks. Future progress may also depend on the refinement of psychopathological phenotypes and endophenotypes. Psychopathological states vary in their complexity and are often best seen as multidimensional composites. The Research Domain Criteria project launched by NIMH in 2009 seeks to implement this strategy by defining basic dimensions of functioning (such as the fear circuits or working memory) to be studied across multiple units of analysis, from genes to neural circuits to behaviors, cutting across disorders as traditionally defined (Cuthbert & Insel, 2010). Finally, there is the need to continue to examine the factors and mechanisms that promote and weaken resilience particularly among individuals experiencing significant adversity (Cicchetti & Toth, 2009). Resilience is an evolutionarily conserved trait that supports reproductive fitness. Consequently gaining a deeper understanding of the factors that contribute to resilience is an important area of continuing research in the area of developmental psychopathology. Given the multi-finality in developmental processes and the diversity of ways in which individuals respond to and interact with vulnerability and protective factors at each ecological level, continuing longitudinal efforts are needed to understand the complex matrix of the individual’s biological and psychological organization and its response to adversity over the course of development (Cicchetti, 2013). Thus far, it is clear that an individual’s resilience resides in a broad range of genomic, epigenomic, neural, neuroendocrine, and behavioral mechanisms, but more work is needed (Grigorenko & Cicchetti, 2012; Karatoreos & McEwen, 2013; Chapter 8, this volume). From an evolutionary perspective, although certain genetic variants may confer an increased vulnerability in the face of adversity, such as maltreatment, it is equally probable that other variants at the same locus may serve a protective function against environmental insults (Belsky, Jonassaint, Pluess, Stanton, Brummett, & Williams, 2009; Cicchetti & Blender, 2006). There is also evidence, in the case of maltreated children, while some genetic variants had a negligible effect for the maltreated group in predicting resilience, the same genotype contributed to higher resilient

157

functioning in non-maltreated children (Cicchetti & Rogosch, 2012). Clinical Implications The present conceptualization of neurodevelopment and the factors that selectively influence its course has important implications for the prevention, treatment, and care of individuals with various forms of developmental psychopathology. The identification of risk factors, both environmental and genetic, is likely to lead to screening procedures and preventive interventions such as public health programs aimed at reducing lead exposure in children. This conceptualization emphasizes the environment as a crucial factor in designing treatment interventions. Identifying features in the home and educational environments that will allow children to build on their strengths and improve their adaptive capacities should continue to be a major priority for research (Leckman, 2013). Advances in our understanding of those components of the environment that are most crucial for resilience and sustaining successful adaptations will doubtless refine interventions designed to prevent and treat various forms of psychopathology (Panter-Brick & Leckman, 2013; Rutter, 2013). Finally, aspects of this approach may permit a deeper empathetic understanding of individuals with early onset disorders. For example, if some forms of OCD bear some relationship to the mental states associated with highly conserved behavioral repertoires typically encountered in expectant parents (intrusive worries about some misfortune befalling the fetus or infant), it should be easier for clinicians to have a deeper emotional empathy for the anguish a patient is experiencing if they can relate the patient’s symptoms to emotional experiences in their own lives as parents and members of extended families (Kim et al., 2013; Leckman & Mayes, 1998; Leckman et al., 1999). Although these models can appear to be reductionistic approaches that neglect the inner worlds of children, some of the emerging models of mental development (Edelman & Tononi, 1995; Fricchione, 2011) are fully compatible with the rich, dynamic complexity of intrapsychic states that we encounter in ourselves, in the consulting room, in our families, and throughout our social ecological systems. REFERENCES Abikoff, H., Gittelman-Klein, R., & Klein, D. F. (1977). Validation of a classroom observation code for hyperactive children. Journal of Consulting and Clinical Psychology, 45(5), 772–783.

158

Understanding Developmental Psychopathology

Abrahams, B. S., & Geschwind, D. H. (2008). Advances in autism genetics: On the threshold of a new neurobiology. Nature Reviews Genetics, 9, 341–345. doi: 10.1038/nrg2346 Abrahams, B. S., & Geschwind, D. H. (2010). Connecting genes to brain in the autism spectrum disorders. Archives of Neurology, 67(4), 395–399. doi: 10.1001/archneurol.2010.47 Abramowitz, J. S., Meltzer-Brody, S., Leserman, J., Killenberg, S., Rinaldi, K., Mahaffey, B. L., & Pedersen, C. (2010). Obsessional thoughts and compulsive behaviors in a sample of women with postpartum mood symptoms. Archives of Women’s Mental Health, 13, 523–530. doi: 10.1007/s00737–010–0172–4 Achenbach, T. M., Rescorla, L. A., & Ivanova, M. Y. (2012). International epidemiology of child and adolescent psychopathology I: Diagnoses, dimensions, and conceptual issues. Journal American Academy of Child and Adolescent Psychiatry, 51(12), 1261–1272. doi: 10.1016/j.jaac.2012.09.010 Adolphs, R. (1999). Social cognition and the human brain. Trends in Cognitive Science, 13, 469–479. Akiskal, H. S., & McKinney, W. T. (1975). Overview of recent research in depression: Integration of ten conceptual models into a comprehensive frame. Archives of General Psychiatry, 32, 285–305. Alaghband-Rad, J., McKenna, K., Gordon, C. T., Albus, K. E., Hamburger, S. D., Rumsey, J. M., . . . Rapoport, J. L. (1995). Childhood-onset schizophrenia: the severity of premorbid course. Journal of the American Academy of Child and Adolescent Psychiatry, 34(10), 1273–1283. American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Washington, DC: Author. Andersen, S. L., & Navalta, C. P. (2011). New frontiers in developmental neuropharmacology: Can long-term therapeutic effects of drugs be optimized through carefully timed early intervention? Journal of Child Psychology and Psychiatry, 52(4), 476–503. doi: 10.1111 /j.1469–7610.2011.02376.x Andrews, P. W., & Thompson, J. A. (2009). The bright side of being blue: Depression as an adaptation for analyzing complex problems. Psychological Review, 116(3), 620–654. doi: 10.1037/a0016242 Anthony, J. C., Warner, L. A., & Kessler, R. C. (1994). Comparative epidemiology of dependence on tobacco, alcohol, controlled substances, and inhalants: Basic findings from the National Comorbidity Survey. Experimental Clinical Psychopharmacology, 2, 244–268. Arnsten, A. F. T., Berridge, C. W., & McCracken, J. T. (2009). The neurobiological basis of attention-deficit/hyperactivity disorder. Primary Psychiatry, 16(7), 47–54. Asarnow, J. R., Tompson, M. C., & McGrath, E. P. (2004). Childhoodonset schizophrenia: clinical and treatment issues. Journal of Child Psychology and Psychiatry, 45(2), 180–194. Asperger, H. (1944). Die autistischen Psychopathen im Kindesalter. Archiv fur Psychiatrie und Nervenkrankheiten, 117, 76–136. Ayuso-Mateos, J. (2006). Global burden of obsessive-compulsive disorder in the year 2000. Geneva, Switzerland: World Health Organization. Baldwin, J. M. (1902). Development and evolution. New York, NY: MacMillan. Bansal, R., Staib, L. H., Laine, A. F., Hao, X., Xu, D., Liu, J., Weissman, M., & Peterson, B. S. (2012). Anatomical brain images alone can accurately diagnose chronic neuropsychiatric illnesses. PLoS One, 7(12), e50698. doi: 10.1371/journal.pone.0050698 Barbosa-Morais, N. L., Irimia, M., Pan, Q., Xiong, H. Y., Gueroussov, S., Lee, L. J., Slobodeniuc, V., Kutter, C., . . . Blencowe, B. J. (2012). The evolutionary landscape of alternative splicing in vertebrate species. Science, 338(6114), 1587–1593. doi: 10.1126/science.1230612 Barkley, R. A. (1997). ADHD and the nature of self-control. New York, NY: Guilford Press.

Barkow, J. H., Cosmides, L., & Tooby, J. (Eds.). (1995). The adapted mind: Evolutionary psychology and the generation of culture. New York, NY: Oxford University Press. Baron-Cohen, S. (Ed.) (1997). The maladapted mind: Classical readings in evolutionary psychopathology. East Sussex, UK: Psychology Press. Baron-Cohen, S. (2002). The extreme male brain theory of autism. Trends in Cognitive Neuroscience, 6, 248–254. Baron-Cohen, S., Leslie, A. M., & Frith, U. (1985). Does the autistic child have a “theory of mind”? Cognition, 21, 37–46. Baron-Cohen, S., Lombardo, M. V., Auyeung, B., Ashwin, E., Chakrabarti, B., & Knickmeyer, R. (2011). Why are autism spectrum conditions more prevalent in males? PLoS Biology, 9(6), e1001081. doi: 10.1371/journal.pbio.1001081 Barry, E., & Gill, M. (2007) Environmental risk factors and gene– environment interaction in attention deficit hyperactivity disorder. In M. Fitzgerald, M. Bellgrove, & M. Gill (Eds.), Handbook of attention deficit hyperactivity disorder (pp. 149–182). West Sussex, UK: Wiley. Belsky, J. (1997). Variation in susceptibility to rearing influences: An evolutionary argument. Psychological Inquiry, 8, 182–186. Belsky, J., & Pluess, M. (2013). Beyond risk, resilience, and dysregulation: Phenotypic plasticity and human development. Development and Psychopathology, 25, 1243–1261. doi: 10.1017/S095457941300059X Belsky, J., Jonassaint, C., Pluess, M., Stanton, M., Brummett, B., & Williams, R. (2009). Vulnerability genes or plasticity genes? Molecular Psychiatry, 14, 746–754. doi: 10.1038/mp. 2009.44 Belsky, J., Steinberg, L., & Draper, P. (1991). Childhood experience, interpersonal development, and reproductive strategy: An evolutionary theory of socialization. Child Development, 62, 647–670. Bentley, M. J., Lin, H., Fernandez, T. V., Lee, M., Yrigollen, C. M., Pakstis, A. J., . . . Leckman, J. F. (2013). Gene variants associated with antisocial behaviour: A latent variable approach. Journal of Child Psychology and Psychiatry, 54(10), 1074–1085. doi: 10.1111/jcpp. 12109 Besser, M.J., Ganor, Y., & Levite, M. (2005). Dopamine by itself activates either D2, D3 or D1/D5 dopaminergic receptors in normal human T-cells and triggers the selective secretion of either IL-10, TNFalpha or both. Journal of Neuroimmunology, 169, 161–171. Betancur, C. (2011). Etiological heterogeneity in autism spectrum disorders: More than 100 genetic and genomic disorders and still counting. Brain Research, 1380, 42–77. doi: 10.1016/j.brainres.2010.11.078 Bideau, A., Desjardins B., & Brignoli, H.P. (Eds.). 1997. Infant and child mortality in the past. Oxford, UK: Clarendon Press. Blackford, J. U., & Pine, D. S. (2012). Neural substrates of childhood anxiety disorders: A review of neuroimaging findings. Child and Adolescent Psychiatric Clinics of North America, 21(3), 501–525. doi: 10.1016/j.chc.2012.05.002 Blake, R., & Shiffrar, M. (2007). Perception of human motion. Annual Review Psychology, 58, 47–73. Doi: 10.1146/annurev.psych.57 .102904.190152 Bowlby, J. (1969). Attachment and loss, vol. I: Attachment. New York, NY: Basic Books. Bowlby, J. (1973). Attachment and loss, vol. II: Separation, anxiety and anger. New York, NY: Basic Books. Bowlby, J. (1980). Attachment and loss, vol. III: Loss, sadness and depression. New York, NY: Basic Books Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary–developmental theory of the origins and functions of stress reactivity. Development and Psychopathology, 17, 271–301. Brimberg, L., Benhar, I., Mascaro-Blanco, A., Alvarez, K., Lotan, D., Winter, C., . . . Joel, D. (2012). Behavioral, pharmacological, and immunological abnormalities after streptococcal exposure: A novel

References rat model of Sydenham chorea and related neuropsychiatric disorders. Neuropsychopharmacology, 37(9), 2076–2087. doi: 10.1038 /npp. 2012.56. Brown, A. S. (2011). The environment and susceptibility to schizophrenia. Progress in Neurobiology, 93(1), 23–58. doi: 10.1016/j.pneurobio .2010.09.003 Brüne, M. (2005). “Theory of mind” in schizophrenia: A review of the literature. Schizophrenia Bulletin, 31, 21–42. Brüne, M. (2008). Textbook of evolutionary psychiatry: The origins of psychopathology. New York, NY: Oxford University Press. Burns, J. (2006). The social brain hypothesis of schizophrenia. World Psychiatry, 5, 77–81. Buss, D. (2011). Evolutionary psychology: The new science of the mind (4th ed.). New York, NY: Pearson. Caldji, C., Hellstrom, I. C., Zhang, T. Y., Diorio, J., & Meaney, M. J. (2011). Environmental regulation of the neural epigenome. FEBS Letters, 585(13), 2049–2058. doi: 10.1016/j.febslet.2011.03.032 Campbell, F., Conti, G., Heckman, J. J., Moon, S. H., Pinto, R., Pungello, E., & Pan, Y. (2014). Early childhood investments substantially boost adult health. Science, 343(6178), 1478–1485. doi: 10.1126/science.1248429 Canteras, N. S., Mota-Ortiz, S. R., & Motta, S. C. (2012). What ethologically based models have taught us about the neural systems underlying fear and anxiety. Brazilian Journal of Medical and Biological Research, 45(4), 321–327. doi: 10.1590/S0100–879X2012007500042 Caspi, A., Hariri, A. R., Holmes, A., Uher, R., & Moffitt, T. E. (2010). Genetic sensitivity to the environment: the case of the serotonin transporter gene and its implications for studying complex diseases and traits. American Journal of Psychiatry, 167(5), 509–527. doi: 10.1176/appi.ajp. 2010.09101452 Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., McClay, J., . . . Poulton, R. (2003). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301(5631), 386–389. Cassidy, S. B., Schwartz, S., Miller, J. L., & Driscoll, D. J. (2012). Prader-Willi syndrome. Genetics in Medicine, 14(1), 10–26. doi: 10.1038/gim.0b013e31822bead0 Chapman, S. J., & Hill, A. V. (2012). Human genetic susceptibility to infectious disease. Nature Reviews Genetics, 13(3), 175–188. doi: 10.1038/nrg3114 Chen, C. Y., Chen, K. H., Liu, C. Y., Huang, S. L., & Lin, K. M. (2009). Increased risks of congenital, neurologic, and endocrine disorders associated with autism in preschool children: cognitive ability differences. Journal of Pediatrics, 154(3), 345–350. doi: 10.1016/j.jpeds.2008.09.043 Chen, S. K., Tvrdik, P., Peden, E., Cho, S., Wu, S., Spangrude, G., & Capecchi, M. R. (2010). Hematopoietic origin of pathological grooming in Hoxb8 mutant mice. Cell, 141(5), 775–785. doi: 10.1016/j.cell.2010.03.055 Cherkasova, M., Sulla, E. Dalena, K., Pondé, M. P., & Hechtman, L. (2013). Developmental course of Attention Deficit Hyperactivity Disorder and its predictors. Canadian Journal of Child and Adolescent Psychiatry, 22(1), 47–54. Cicchetti, D. (2006). Development and psychopathology. In D. Cicchetti & D. Cohen (Eds.), Developmental psychopathology: Theory and method (Vol. 1) 1–23. New York, NY: Wiley. Cicchetti, D. (2013). Resilient functioning in maltreated children—Past, present, and future perspectives. Journal of Child Psychology and Psychiatry, 54(4), 402–422. doi: 10.1111/j.1469–7610.2012.02608.x Cicchetti, D., & Blender, J. A. (2006). A multiple-levels-of analysis perspective on resilience: Implications for the developing brain, neural

159

plasticity, and preventive interventions. Annals of the New York Academy of Sciences, 1094, 248–258. Cicchetti, D., & Rogosch, F. A. (2012). Gene by environment interaction and resilience: Effects of child maltreatment and serotonin, corticotropin releasing hormone, dopamine, and oxytocin genes. Development and Psychopathology, 24, 411–427. doi: 10.1017/S0954579412000077 Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50, 16–25. doi: 10.1111/j.1469–7610.2008.01979.x Courchesne, E., Mouton, P. R., Calhoun, M. E., Semendeferi, K., Ahrens-Barbeau, C., Hallet, M. J., Barnes, C. C., & Pierce, K. (2010). Neuron number and size in prefrontal cortex of children with autism. Journal of the American Medical Association, 306(18), 2001–2010. doi: 10.1001/jama.2011.1638 Crespi, B. J., & Thiselton, D. L. (2011). Comparative immunogenetics of autism and schizophrenia. Genes Brain and Behavior, 10(7), 689–701. doi: 10.1111/j.1601–183X.2011.00710.x Cuthbert, B. N., & Insel, T. R. (2010). Toward new approaches to psychotic disorders: The NIMH Research Domain Criteria project. Schizophrenia Bulletin, 36(6), 1061–1062. doi: 10.1093/schbul/sbq108 Daly, M., & Wilson, M. I. (1981). Child maltreatment from a sociobiological perspective. New Directions for Child Development, 198, 93–112. Danese, A., Moffitt, T. E., Harrington, H., Milne, B. J., Polanczyk, G., Pariante, C. M., Poulton, R., & Caspi, A. (2009). Adverse childhood experiences and adult risk factors for age-related disease: Depression, inflammation, and clustering of metabolic risk markers. Archives of Pediatrics and Adolescent Medicine, 163(12), 1135–1143. doi: 10.1001/archpediatrics.2009.214 Darnell, J. C., Van Driesche, S. J., Zhang, C., Hung, K. Y., Mele, A., Fraser, C. E., . . . Darnell, R. B. (2011). FMRP stalls ribosomal translocation on mRNAs linked to synaptic function and autism. Cell, 146, 247–261. doi: 10.1010/jcell.2011.06.013 Darwin, C. R. (1859). On the origin of species by means of natural selection or the preservation of favoured races in the struggle for life. London, UK: Murray. Darwin, C. R. (1872). The expression of emotions in man and animals. London, UK: Murray. Dawkins, R., & Krebs, J. R. (1979). Arms races between and within species. Proceedings of the Royal Academy of London, Series B—Biological Sciences, 205(1161), 489–511. Dawkins, R. (1987). The blind watchmaker: Why the evidence of evolution reveals a universe without design. New York, NY: Norton. Devlin, A. M., Brain, U., Austin, J., & Oberlander, T. F. (2010). Prenatal exposure to maternal depressed mood and the MTHFR C677T variant affect SLC6A4 methylation in infants at birth. PLoS One, 5(8), e12201. doi: 10.1371/journal. pone.0012201 Edelman, G. M., & Tononi, G. (1995). Neural Darwinism: The brain as a selectional system. In J. Cornwell (Ed.), Natures imagination: The frontiers of scientific vision (pp. 78–100). New York, NY: Oxford University Press. Ekman, P. (Ed.). (1982). Emotion in the human face (2nd ed.). New York, NY: Cambridge University Press. Ellis, B. J., & Bjorklund, D. F. (2012). Beyond mental health: An evolutionary analysis of development under risky and supportive environmental conditions: an introduction to the special section. Developmental Psychology, 48(3), 591–597. doi: 10.1037/a0027651 Ellis, B. J., & Boyce, W. T. (2011). Differential susceptibility to the environment: toward an understanding of sensitivity to developmental experiences and context. Developmental Psychopathology, 23(1), 1–5. doi: 10.1017/S095457941000060X

160

Understanding Developmental Psychopathology

ENCODE Project Consortium, Dunham, I., Kundaje, A., Aldred, S. F., Collins, P. J., Davis, C. A., . . . Birney, E. (2012). An integrated encyclopedia of DNA elements in the human genome. Nature, 489(7414), 57–74. doi: 10.1038/nature11247 Erhardt, D., & Hinshaw, S. P. (1994). Initial sociometric impressions of attention-deficit hyperactivity disorder and comparison boys: Predictions from social behaviors and from nonbehavioral variables. Journal of Consulting and Clinical Psychology, 62(4), 833–842. Erikson, E. H. (1968). Identity: Youth and crisis. London, UK: Faber & Faber. Evans, D. W., & Leckman, J. F. (2006). Origins of obsessive-compulsive disorder: Developmental and evolutionary perspectives. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology (2nd ed., pp. 404–435). New York, NY: John Wiley and Sons. Evans, D. W., Gray, F. L., & Leckman, J. F. (1999). Rituals, fears and phobias: Insights from development, psychopathology and neurobiology. Child Psychiatry and Human Development, 29, 261–276. Evans, D. W., Leckman, J. F., Carter, A., Reznick, J. S., Henshaw, D., King, R. A., & Pauls, D. (1997). Ritual, habit and perfectionism: The prevalence and development of compulsive-like behaviors in normal young children. Child Development, 68, 58–68. Eyler, L. T., Pierce, K., & Courchesne, E. (2012). A failure of left temporal cortex to specialize for language is an early emerging and fundamental property of autism. Brain, 135(Pt 3), 949–960. doi: 10.1093/brain/awr364 Faraone, S. V., & Mick, E. (2010). Molecular genetics of attention deficit hyperactivity disorder. Psychiatric Clinics of North America, 33, 159–180. doi: 10.1016/j.psc.2009.12.004.review Feygin, D. L., Swain, J. E., & Leckman, J. F. (2006). The normalcy of neurosis: evolutionary origins of obsessive-compulsive disorder and related behaviors. Progress in Neuropsychopharmacology and Biological Psychiatry, 30(5), 854–864. Finch, C. E., & Rose, M. R. (1995). Hormones and the physiological architecture of life history evolution. Quarterly Review of Biology, 70(1), 1–52. Flory, K., Molina, B. S., Pelham Jr., W. E., Gnagy, E., & Smith, B. (2006). Childhood ADHD predicts risky sexual behavior in young adulthood. Journal of Clinical Child Adolescent Psychology, 35, 571–577. Forbes, E. E., & Dahl, R. E. (2012). Altered reward function in adolescent depression: What, when and how? Journal of Child Psychology and Psychiatry, 53(1), 3–15. doi: 10.1111/j.1469–7610.2011.02477.x Fricchione, G. L. (2011). Compassion and healing in medicine and society: On the nature and use of attachment solutions and separation challenges. Baltimore, MD: Johns Hopkins University Press. Fu, Y. H., Kuhl, D. P., Pizzuti, A., Pieretti, M., Sutcliffe, J. S., Richards, S., . . . Caskey, C. T. (1991). Variation of the CGG repeat at the fragile X site results in genetic instability: Resolution of the Sherman paradox. Cell, 67(6), 1047–1058. Garmezy, N., & Streitman, S. (1974). Children at risk: The search for the antecedents of schizophrenia: Conceptual models and research methods. Schizophrenia Bulletin, 8, 14–90. Gesell, A. (1928). Infancy and human growth. New York, NY: Macmillan. Gillessen-Kaesbach, G., Robinson, W., Lohmann, D., Kaya-Westerloh, S., Passargem, E., & Horsthemke, B. (1995). Genotype-phenotype correlation in a series of 167 deletion and non-deletion patients with Prader-Willi syndrome. Human Genetics, 96, 638–643. Gogtay, N. (2008). Cortical brain development in schizophrenia: Insights from neuroimaging studies in childhood-onset schizophrenia. Schizophrenia Bulletin, 34, 30–36. doi: 10.1093/schbul/sbm103 Goldman, D., Oroszi, G., & Ducci, F. (2005). The genetics of addictions: Uncovering the genes. Nature Reviews Genetics, 6, 521–532.

Gordon, I., Martin, C., Feldman, R., & Leckman, J. F. (2011). Oxytocin and social motivation. Developmental and Cognitive Neuroscience, 1(4), 471–493. doi: 10.1016/j.dcn.2011.07.007 Greer, J. M., & Capecchi, M. R. (2002). Hoxb8 is required for normal grooming behavior in mice. Neuron, 33(1), 23–34. Greer, P. L., Hanayama, R., Bloodgood, B. L., Mardinly, A. R., Lipton, D. M., Flavell, S. W., . . . Greenberg, M. E. (2010). The Angelman syndrome protein Ube3A regulates synapse development by ubiquitinating arc. Cell, 140, 704–716. doi: 10.1016/j cell 2010.01.026 Grigorenko, E. L., & Cicchetti, D. (2012). Genomic sciences for developmentalists: The current state of affairs. Development and Psychopathology, 24(4), 1157–1164. doi: 10.1017/S0954579412000612 Gullone, E. (2000). The development of normal fear: a century of research. Clinical Psychology Review, 20(4), 429–451. Gurney, J., McPheeters, M., & Davis, M. (2006). Parental report of health conditions and health care use among children with and without autism: National survey of children’s health. Archives of Pediatric and Adolescent Medicine, 160(8), 825–830. Hagerman, R. J., & Hagerman, P. J. (2002). Fragile X syndrome, diagnosis treatment and research (3rd ed.). Baltimore, MD: Johns Hopkins University Press. Hamlyn, J., Duhig, M., McGrath, J., & Scott, J. (2013). Modifiable risk factors for schizophrenia and autism—Shared risk factors impacting on brain development. Neurobiology of Disease, 53, 3–9. doi: 10.1016/j.nbd.2012.10.023 Hartley, C. A., & Phelps, E. A. (2012). Anxiety and decision-making. Biological Psychiatry, 72(2), 113–118. doi: 10.1016/j.biopsych.2011 .12.027 Hatton, D. D., Sideris, J., Skinner, M., Mankowski, J., Bailey, D. B., Jr., Roberts, J., & Mirrett, P. (2006). Autistic behavior in children with fragile X syndrome: prevalence, stability, and the impact of FMRP. American Journal of Medical Genetics A, 140(A), 1804–1813. Heckman, J. J., Moon, S. H., Pinto, R., Savelyev, P. A., & Yavitz, A. (2010). The rate of return to the High/Scope Perry Preschool Program. Journal of Public Economics, 94(1–2), 114–128. doi:10.1016/j.jpubeco.2009 .11.001 Heller, T. (1908). Ueber dementia infantilis. Zeitschrift fuer die Erforschung und Behandlung des Jugendlichen Schwachsinns, 2, 17–28. Hettema, J. M., Neale, M. C., & Kendler, K. S. (2001). A review and meta-analysis of the genetic epidemiology of anxiety disorders. American Journal of Psychiatry, 158(10), 1568–78. Hinde, R. A. (1987). Can nonhuman primates help us understand human behavior? In B. B. Smuts, D. L. Cheney, R. M. Seyfarth, R. W. Wrangham, & T. T. Struhsaker (Eds.), Primate societies (pp. 413–420). Chicago, IL: University of Chicago Press. Hocutt, M. (1974). Aristotle’s four becauses. Philosophy, 49, 385–399. Hollis, C. (1995). Child and adolescent (juvenile onset) schizophrenia: A case control study of premorbid developmental impairments. British Journal of Psychiatry, 166(4), 489–595. Hosain, G. M., Berenson, A. B., Tennen, H., Bauer, L. O., & Wu, Z. H. (2012). Attention deficit hyperactivity symptoms and risky sexual behavior in young adult women. Journal of Women’s Health (Larchmont), 21(4), 463–468. doi: 10.1089/jwh.2011.2825 Hsiao, E. Y., & Patterson, P. H. (2012). Placental regulation of maternal–fetal interactions and brain development. Developmental Neurobiology, 72(10), 1317–1326. doi: 10.1002/dneu.22045 Husby, G., van de Rijn, I., Zabriskie, J. B., Abdin, Z. H., & Williams, R. C. (1976). Antibodies reacting with cytoplasm of subthalamic and caudate nuclei neurons in chorea and acute rheumatic fever. Journal of Experimental Medicine, 144, 1094–1110.

References Ingram, R. E., & Luxton, D. D. (2005). Vulnerability–stress models. In B. L. Hankin & J. R. Z. Abela (Eds.), Development of psychopathology: A vulnerability stress perspective (pp. 32–46). Thousand Oaks, CA: Sage Publications Inc. Iossifov, I., Ronemus, M., Levy, D., Wang, Z., Hakker, I., Rosenbaum, J., . . . Wigler, M.. (2012). De novo gene disruptions in children on the autistic spectrum. Neuron, 74(2), 285–299. doi: 10.1016/j neuron 2010 04.009 Jamain, S., Quach, H., Betancur, C., Råstam, M., Colineaux, C., Gillberg, I. C., . . . Bourgeron, T. (2003). Paris autism research international sibpair study. Mutations of the X-linked genes encoding neuroligins NLGN3 and NLGN4 are associated with autism. Nature Genetics, 34(1), 27–29. Janet, P. (1903). Les obsessions et la psychasthenie (2nd ed., Vol. 1 and 2). Paris, France: Alcan. Jensen, P. S. Mrazek, D., Knapp, P. K., Steinberg, L., Pfeffer, C., Schowalter, J., & Shapiro, T. (1997). Evolution and revolution in child psychiatry: ADHD as a disorder of adaptation. Journal of the American Academy of Child and Adolescent Psychiatry, 36, 1672–1679. Johnson, M. B., Kawasawa, Y. I., Masonm, C. E., Krsnik, Z., Coppola, G., Bogdanovi´c, D., . . . Sestan N. (2009). Functional and evolutionary insights into human brain development through global transcriptome analysis. Neuron, 62(4), 494–509. doi: 10.1016/j neuron.2009.03.027 Kaffman, A., & Krystal, J. H. (2012). New frontiers in animal research of psychiatric illness. Methods Molecular Biology, 829, 3–30. doi: 10.1007/978–1–61779–4582_1 Kaiser, M. D., Hudac, C. M., Shultz, S., Lee, S. M., Cheung, C., Berken, A. M., . . . Pelphreym, K. A. (2010). Neural signatures of autism. Proceedings of the National Academy of Sciences U.S.A., 107(49), 21223–21228. doi: 10.1073/pnas.1010412107 Kang, H., J., Kawasawa, Y. I., Cheng, F., Zhu, Y., Xu, X., Li, M., . . . Sestan, N. (2012). Spatio-temporal transcriptome of the human brain. Nature, 478(7370), 483–489. doi: 10.1038/nature10523 Kanner, L. (1943). Autistic disturbances of affective contact. Nervous Child, 2, 217–250. Karatoreos, I. N., & McEwen, B. S. (2013). The neurobiology and physiology of adaptation across the lifecourse. Journal of Child Psychology and Psychiatry, 54(4), 337–347. doi: 10.111/jcpp. 12054 Kendler, K. S. (2012). The dappled nature of causes of psychiatric illness: replacing the organic-functional/hardware-software dichotomy with empirically based pluralism. Molecular Psychiatry, 17(4), 377–388. doi: 10.1038/mp. 2011.182 Kendler, K. S., Aggen, S. H., Knudsen, G. P., Røysamb, E., Neale, M. C., & Reichborn-Kjennerud, T. (2011). The structure of genetic and environmental risk factors for syndromal and subsyndromal common DSM IV axis I and all axis II disorders. American Journal of Psychiatry, 168(1), 29–39. doi: 10.1176/appi.ajp.2010.10030340. Kendler, K. S., Eaves, L. J., Loken, E. K., Pedersen, N. L., Middeldorp, C. M., Reynolds, C., . . . Gardner, C. O. (2011). The impact of environmental experiences on symptoms of anxiety and depression across the life span. Psychological Science, 22(10), 1343–1352. doi: 10.1177/0956797611417255 Kendler, K. S., Gatz, M., Gardner, C. O., & Pedersen, N. L. (2006). A Swedish national twin study of lifetime major depression. American Journal of Psychiatry, 2163(1), 109–114. Kendler, K. S., Myers, J., & Prescott, C. A. (2007). Specificity of genetic and environmental risk factors for symptoms of cannabis, cocaine, alcohol, caffeine, and nicotine dependence. Archives of General Psychiatry, 64, 1313–1320. doi:10.1001/archpsyc.64.11.1313 Kendler, K. S., Prescott, C. A., Myers, J., & Neale, M. C. (2003). The structure of genetic and environmental risk factors for common psychiatric and substance use disorders in men and women. Archives of General Psychiatry, 60(9): 929–37.

161

Kennedy, D. P., & Adolphs, R. (2012). The social brain in psychiatric and neurological disorders. Trends in Cognitive Science, 16(11), 559–572. doi: 10.1016/j.tics.2012.09.006 Kessler, R. C., Avenevoli, S., Costello, E. J., Georgiades, K., Green, J. G., Gruber, M. J., . . . Merikangas, K. R. (2012). Prevalence, persistence, and sociodemographic correlates of DSM-IV disorders in the National Comorbidity Survey Replication Adolescent Supplement. Archives of General Psychiatry, 69(4), 372–380. doi: 10.1001/archgenpsychiatry.2011.160 Kessler, R. C., Avenevoli, S., McLaughlin, K. A., Green, J. G., Lakoma, M. D., Petukhova, M., . . . Merikangas, K. R. (2012). Lifetime co-morbidity of DSM-IV disorders in the US National Comorbidity Survey Replication Adolescent Supplement. Psychological Medicine, 42(9), 1997–2010. doi: 10.1017/S0033291712000025 Keverne, E. B. (2011). Epigenomics and brain evolution. Epigenomics, 3(2), 183–191. doi: 10.2217/epi.11.10 Keverne, E. B., Fundele, R., Narasimha, M., Baton, S. C., & Surani, M. A. (1996). Genomic imprinting and the differential roles of parental genomes in brain development. Brain Research and Developmental Brain Research, 92, 91–100. Kiessling, L., Marcotte, A., & Culpepper, L. (1993). Antineuronal antibodies: Tics and obsessive compulsive symptoms. Journal of Developmental and Behavioral Pediatrics, 14, 281–282. Killeen, P. R., Tannock, R., & Sagvolden, T. (2012). The four causes of ADHD: A framework. Current Topics in Behavioral Neurosciences, 9, 391–425. doi: 10.1007/7854_2011_160 Kim, P., Mayes, L., Feldman, R., Leckman, J. F., & Swain, J. E. (2013). Early postpartum parental preoccupation and positive parenting thoughts: Relationship with parent-infant interaction, Infant Mental Health Journal, 34(2), 104–116. doi: 10.1002/imhj.21359 King, M. C., & Wilson, A. C. (1975). Evolution at two levels in humans and chimpanzees. Science 188, 107–116. Kipnis, J., Cardon, M., Avidan, H., Lewitus, G. M., Mordechay, S., Rolls, A., Shani, Y., & Schwartz, M. (2004). Dopamine, through the extracellular signal-regulated kinase pathway, downregulates CD4-CD25+ regulatory T-cell activity: Implications for neurodegeneration. Journal of Neuroscience, 24, 6133–6143. Kirvan, C. A., Swedo, S. E., Heuser, J. S., & Cunningham, M. W. (2003). Mimicry and autoantibody-mediated neuronal cell signaling in Sydenham chorea. Nature Medicine, 9, 914–920. Kirvan, C. A., Swedo, S. E., Kurahara, D., & Cunningham, M. W. (2006). Streptococcal mimicry and antibody-mediated cell signaling in the pathogenesis of Sydenham’s chorea. Autoimmunity, 39, 21–29. Klei, L., Sanders, S. J., Murtha, M. T., Hus, V., Lowe, J. K., Willsey, A. J., . . . Devlin, B. (2012). Common genetic variants, acting additively, are a major source of risk for autism. Molecular Autism, 3(1), 9. doi:10.1186/2040–2392–3–9 Klein, D. F. (1993). False suffocation alarms, spontaneous panics, and related conditions. Archives of General Psychiatry, 50, 306–317. Klein, R. G., & Mannuzza, S. (1991). Long-term outcome of hyperactive children: A review. Journal of the American Academy of Child and Adolescent Psychiatry, 30(3), 383–387. Koob, G. F., & Volkow, N. D. (2010). Neurocircuitry of addiction. Neuropsychopharmacology, 35(1), 217–238. doi: 10.1038/npp. 2009.110 Kupfer, D. J., Frank, E., & Phillips, M. L. (2012). Major depressive disorder: New clinical, neurobiological, and treatment perspectives. Lancet, 379(9820), 1045–1055. doi: 10.1016/S0140–6736(11)60602–8 Kwan, K. Y., Lam, M. M., Johnson, M. B., Dube, U., Shim, S., Rašin, M. R., . . . Sestan, N. (2012). Species-dependent posttranscriptional regulation of NOS1 by FMRP in the developing cerebral cortex. Cell, 149(4), 899–911. doi: 10.1016/j cell_2012.02.060

162

Understanding Developmental Psychopathology

Lasker, G. W. (1969). Human biological adaptability: The ecological approach in physical anthropology. Science, 166(3912), 1480–1486.

Marks, I. M., & Nesse, R. M. (1994). Fear and fitness: An evolutionary analysis of anxiety disorders. Ethology Sociobiology, 15, 247–261.

Le Moal, M. (2010). Vulnerability factors in addiction disorders. In G. Koob, M. Le Moal, & R. Thompson (Eds.), Encyclopedia of behavioral neuroscience. 459–469 Online encyclopedia. Amsterdam, The Netherlands: Elsevier.

Masi, G., Mucci, M., & Pari, C. (2006). Children with schizophrenia: clinical picture and pharmacological treatment. CNS Drugs, 20(10), 841–866.

Leckman, J. F. (2013). What’s next for developmental psychiatry? World Psychiatry, 12(2), 125–126. doi: 10.1002/wps.20029

Mataix-Cols, D., Rosario-Campos, M. C., & Leckman, J. F. (2005). A multidimensional model of obsessive-compulsive disorder. American Journal of Psychiatry, 162(2), 228–238.

Leckman, J. F. (2014). What does immunology have to do with brain development and psychopathology? Journal of Child Psychology and Psychiatry, 55(6), 632–634. doi: 10.1111/jcpp. 12259

Mathews, C. A., & Reus, V. I. (2001). Assortative mating in the affective disorders: A systematic review and meta-analysis. Comprehensive Psychiatry, 42(4), 257–262.

Leckman, J. F., & Mayes, L. C. (1998). Understanding developmental psychopathology: How useful are evolutionary perspectives? Journal of the American Academy of Child and Adolescent Psychiatry, 37, 1011–1021.

Matson, J. L., & Nebel-Schwalm, M. S. (2007). Comorbid psychopathology with autism spectrum disorder in children: an overview. Research in Developmental Disabilities, 28(4), 341–352. doi: 10.1016/j.ridd.2005.12.004

Leckman, J. F., Denys, D., Mataix-Cols, D., Hollander, E., Rauch, S. L., Saxena, S., . . . Stein, D. J. (2010). Obsessive-compulsive disorder: A review of the diagnostic criteria and subtype and dimensional and specifiers for DSM-V. Depression and Anxiety, 27(6), 507–527. doi: 10.1002/da.20669

McEwen, B. S., & Stellar. E. (1993). Stress and the individual: Mechanisms leading to disease. Archives of Internal Medicine, 153(18), 2093–2101.

Leckman, J. F., Grice, D. E., Boardman, J., Zhang, H., Vitale, A., Bondi, C., & Pauls, D. L. (1997). Symptoms of obsessive-compulsive disorder. American Journal of Psychiatry, 154, 911–917. Leckman, J. F., Mayes, L. C., Feldman, R., Evans, D., King, R. A., & Cohen, D. J. (1999). Early parental preoccupations and behaviors and their possible relationship to the symptoms of obsessive-compulsive disorder. Acta Psychiatrica Scandinavica, 100, (Supplementum No. 396), 1–26. Leckman, J. F., Walker, D. E., Goodman, W. K., Pauls, D. L., & Cohen, D. J. (1994). “Just right” perceptions associated with compulsive behavior in Tourette’s syndrome. American Journal of Psychiatry, 151, 675–680.

McGuinness, M., Blissett, J., & Jones, C. (2011). OCD in the perinatal period: Is postpartum OCD a distinct subtype? A review of the literature. Behavioural and Cognitive Psychotherapy, 39, 285–310. doi: 10.1017/S1352465810000718. McGuire, M. T., & Troisi, A. (1998). Darwinian psychiatry. New York, NY: Oxford University Press. McGuire, M. T., Marks, I., Nesse, R. M., & Troisi, A. (1992). Evolutionary biology: A basic science for psychiatry? Acta Psychiatrica Scandinavica, 86, 89–96. McLennan, Y., Polussa, J., Tassone, F., & Hagerman, R. (2011). Fragile X syndrome. Current Genomics, 12(3), 216–224. doi: 10.2174 /138920211795677886

Leonard, H., Goldberger E. L., Rapoport, J. L., Cheslow, B. S., & Swedo, S. (1990). Childhood rituals: Normal development or obsessive-compulsive symptoms? Journal of the American Academy of Child and Adolescent Psychiatry, 29, 17–23.

Merikangas, K. R., He, J. P., Burstein, M., Swanson, S. A., Avenevoli, S., Cui, L., . . . Swendsen, J. (2010). Lifetime prevalence of mental disorders in U.S. adolescents: Results from the National Comorbidity Survey Replication—Adolescent Supplement. Journal of the American Academy of Child and Adolescent Psychiatry, 49(10), 980–989. doi: 10.1016/j.jaac.2010.05.017

Lin, M., Hrabovsky, A., Pedrosa, E., Wang, T., Zheng, D., Lachman, H. M. (2012). Allele-biased expression in differentiating human neurons: implications for neuropsychiatric disorders. PLoS One, 7(8), e44017. doi: 10.1371/journal.pone.0044017

Miklos, G. L. (1993). Molecules and cognition: Evolutionary overview of brain structure and function in some vertebrates and invertebrates. Journal of Neurobiology, 24(6), 842–890. doi: 10.1111/j.1469– 7610.2008.01909.x

Lombardo, M. V., Ashwin, E., Auyeung, B., Chakrabarti, B., Lai, M. C., Taylor, K., . . . Baron-Cohen, S. (2012). Fetal programming effects of testosterone on the reward system and behavioral approach tendencies in humans. Biological Psychiatry, 72(10), 839–847. doi: 10.1016/j.biopsych.2012.05.027

Mill, J., & Petronis, A. (2008). Pre-and peri-natal environmental risks for attention-deficit hyperactivity disorder: The potential role of epigenetic processes in mediating susceptibility. Journal of Child Psychology and Psychiatry 49, 1020–1030. doi: 10.1111/j.1469–7610.2008 .01909.x

Luthar, S. S., Cicchetti, D., & Becker, B. (2000). The construct of resilience: A critical evaluation and guidelines for future work. Child Development, 71(3), 543–562.

Miller, B., Messias, E., Miettunen, J., Alaräisänen, A., Järvelin, M. R., Koponen, H., . . . Kirkpatrick, B. (2011). Meta-analysis of paternal age and schizophrenia risk in male versus female offspring. Schizophrenia Bulletin, 37(5), 1039–1047. doi: 10.1093/schbul/sbq011

Mabb, A. M., Judson, M. C., Zylka, M. J., & Philpot, B. D. (2011). Angelman syndrome: Insights into genomic imprinting and neurodevelopmental phenotypes. Trends in Neuroscience, 34(6), 293–303. doi: 10.1016/j.tins.2011.04.001 Maia, T. V., Cooney, R. E., & Peterson, B. S. (2008). The neural bases of obsessive-compulsive disorder in children and adults. Developmental Psychopathology, 20(4), 1251–1283. doi: 10.1017/S0954579408000606 Makedonski, K., Abuhatzira, L., Kaufman, Y., Razin, A., & Shemer, R. (2005). MeCP2 deficiency in Rett syndrome causes epigenetic aberrations at the PWS/AS imprinting center that affects UBE3A expression. Human Molecular Genetics, 14(8), 1049–1058. Marks, I. M. (1987). Fears, phobias, and rituals. Oxford, UK: Oxford University Press.

Mineka, S., & Ohman, A. (2002). Phobias and preparedness: the selective, automatic, and encapsulated nature of fear. Biological Psychiatry, 52(10), 927–937. Monk, C., Spicer, J., & Champagne, F. A. (2012). Linking prenatal maternal adversity to developmental outcomes in infants: The role of epigenetic pathways. Developmental Psychopathology, 24(4), 1361–1376. doi: 10.1017/S0954579412000764 Morán, E. F. (2008). Human adaptability: An introduction to ecological anthropology. Boulder, CO: Westview Press. Mueller, B. R., & Bale, T. L. (2008). Sex-specific programming of offspring emotionality after stress early in pregnancy. Journal of Neuroscience, 28, 9055–9065. doi: 10.1523/JNEUROSCI.1424.08.2008

References Murphy, T. K., Kurlan, R., & Leckman, J. (2010). The immunobiology of Tourette’s disorder, pediatric autoimmune neuropsychiatric disorders associated with Streptococcus, and related disorders: A way forward. Journal of Child Adolescent Psychopharmacology, 20(4), 317–331. doi: 10.1089/cap. 2010.0043 Murray, C. J., & Lopez, A. D. (Eds.). (1996). The global burden of disease: A comprehensive assessment of mortality and disability from diseases, injuries, and risk factors in 1990 and projected to 2020. Cambridge, MA: Harvard University Press. Naura, W. J. H., & Karen, H. J. (1970). A general profile of the vertebrate brain with sidelights on the ancestry of the cerebral cortex. In F. O. Schmit (Ed.), Evolution of brain and behavior (pp. 7–26). New York, NY: Rockefeller University Press. Neale, B. M., Kou, Y., Liu, L., Ma’ayan, A., Samocha, K. E., Sabo, A., . . . Daly, M. J. (2012). Patterns and rates of exonic de novo mutations in autism spectrum disorders. Nature, 485(7397), 242–245. doi: 10.1038/nature11011 Neale, B. M., Medland, S. E., Ripke, S., Asherson, P., Franke, B., Lesch, K. P., Faraone, S. V., . . . Psychiatric GWAS Consortium: ADHD Subgroup. (2010). Meta-analysis of genome-wide association studies of attention-deficit/hyperactivity disorder. Journal of the American Academy of Child and Adolescent Psychiatry, 49(9), 884–897. doi: 10.1016/j.jaac.2010.06.008 Nesse, R. (1989). Panic disorders: An evolutionary view. Psychiatric Annals 18, 478–483. Nesse, R. M. (1990). Evolutionary explanations of emotions. Human Nature, 1, 261–289. Nesse, R. M. (1999). Proximate and evolutionary studies of anxiety, stress, and depression: Synergy at the interface. Neuroscience and Biobehavioral Reviews, 23, 895–903. Nesse, R. M. (2009). Evolutionary origins and functions of emotions. In K. Scherer & D. Sander (Eds.), The Oxford companion to emotion and the affective sciences (pp. 159–164). Oxford, UK: Oxford University Press. Nesse, R. M. (2000). Is depression an adaptation? Archives of General Psychiatry, 57(1), 14–20. Nesse, R. M., & Berridge, K. C. (1997). Psychoactive drug use in evolutionary perspective. Science, 278, 63–66. Nesse, R. M., & Stein, D. J. (2012). Towards a genuinely medical model for psychiatric nosology. BMC Medicine, 10:5. doi: 10.1186/1741–7015–10–5. Nesse, R. M., & Williams, G. C. (1994). Why we get sick: The new science of Darwinian medicine. New York, NY: Random House. Nettle, D. (2004). Evolutionary origins of depression: A review and reformulation. Journal of Affective Disorders, 81(2), 91–102. Nicolson, R., & Rapoport, J. L. (1999). Childhood-onset schizophrenia: Rare but worth studying. Biological Psychiatry, 46(10), 1418–1428. Novarino, G., El-Fishawy, P., Kayserili, H., Meguid, N. A., Scott, E. M., Schroth, J., . . . Gleeson, J. G. (2012). Mutations in BCKD-kinase lead to a potentially treatable form of autism with epilepsy. Science, 338(6105), 394–397. doi: 10.1126/science.1224631 Oberlander, T. F., Weinberg, J., Papsdorf, M., Grunau, R., Misri, S., & Devlin, A. M. (2008). Prenatal exposure to maternal depression, neonatal methylation of human glucocorticoid receptor gene (NR3C1) and infant cortisol stress responses. Epigenetics, 3(2), 97–106. doi: 10.4161/epi.3.2.6034 O’Connor, T. G., Moynihan, J. A., & Caserta, M. T. (2014). The neuroinflammation hypothesis for stress and psychopathology in children—Developmental psychoneuroimmunology. Journal of Child Psychology and Psychiatry, 55(6), 615–631. doi: 10.1111/jcpp. 12187

163

O’Roak, B. J., Deriziotis, P., Lee, C., Vives, L., Schwartz, J. J., Girirajan, S., . . . Eichler, E. E. (2011). Exome sequencing in sporadic autism spectrum disorders identifies severe de novo mutations. Nature Genetics, 43(6), 585–589. doi: 10.1038/ng.835 Panter-Brick, C., & Leckman, J. F. (2013). Resilience in child development—interconnected pathways to wellbeing. Journal of Child Psychology and Psychiatry, 54(4), 333–336. doi: 10.1111/jcpp. 12057 Pauls, D. L. (2010). The genetics of obsessive-compulsive disorder: A review. Dialogues in Clinical Neuroscience, 12(2), 149–163. Pawlby, S., Hay, D. F., Sharp, D., Waters, C. S., & O’Keane, V. (2009). Antenatal depression predicts depression in adolescent offspring: Prospective longitudinal community-based study. Journal of Affective Disorders, 113, 236–243. doi: 10.1016/j.jad.2008.05.018 Peça, J., & Feng, G. (2012). Cellular and synaptic network defects in autism. Current Opinions in Neurobiology, 22(5), 866–872. doi: 10.1016/j.conb.2012.02.015 Piaget, J. (1962). Play, dreams and imitation in childhood. New York, NY: Norton. Polanczyk, G., de Lima, M. S., Horta, B. L., Biederman, J., & Rohde, L. A. (2007). The worldwide prevalence of ADHD: A systematic review and metaregression analysis. American Journal of Psychiatry, 164, 942–948. doi:10.1176/appi.ajp. 164.6.942 Power, R. A., Kyaga, S., Uher, R., Maccabe, J. H., Långström, N., Landen, M., . . . Svensson, A. C. (2102). Fecundity of patients with schizophrenia, autism, bipolar disorder, depression, anorexia nervosa, or substance abuse vs. their unaffected siblings. Archives of General Psychiatry, 12, 1–8. doi: 10.1001/jamapsychiatry.2013.268 Poutahidis, T., Kearney, S. M., Levkovich, T., Qi, P., Varian, B. J., Lakritz, J. R., . . . Erdman, S. E. (2013). Microbial symbionts accelerate wound healing via the neuropeptide hormone oxytocin. PLoS ONE, 8, e78898. doi: 10.1371/journal.pone.0078898 Price, J., Sloman, L., Gardner Jr., R., Gilbert, P., & Rohde, P. (1994). The social competition hypothesis of depression. British Journal of Psychiatry, 164(3), 309–315. Provençal, N., Suderman, M. J., Guillemin, C., Massart, R., Ruggiero, A., Wang, D., . . . Szyf M. (2012). The signature of maternal rearing in the methylome in rhesus macaque rrefrontal cortex and T cells. Journal of Neuroscience, 32(44), 15626–15642. doi: 10.1523/jneurosci.1470–12.2012 Pujadas, E., & Feinberg, A. P. (2012). Regulated noise in the epigenetic landscape of development and disease. Cell, 148(6), 1123–1131. doi: 10.1016/j.cell.2012.02.045 Rapoport, J. L., & Gogtay, N. (2008). Brain neuroplasticity in healthy, hyperactive and psychotic children: Insights from neuroimaging. Neuropsychopharmacology, 33, 181–197. doi:10.1038/sj.npp. 1301553 Reser, J. E. (2011). Conceptualizing the autism spectrum in terms of natural selection and behavioral ecology: The solitary forager hypothesis. Evolutionary Psychology, 9(2), 207–238. Rett, A. (1966). [On an unusual brain atrophy syndrome in hyperammonemia in childhood] [German]. Wiener medizinische Wochenschrift, 116(37), 723–726. Robison, A. J., & Nestler, E. J. (2011). Transcriptional and epigenetic mechanisms of addiction. Nature Reviews Neuroscience, 12(11), 623–637. doi: 10.1038/nrn3111 Ruscio, A. M., Stein, D. J., Chiu, W. T., & Kessler, R. C. (2010). The epidemiology of obsessive-compulsive disorder in the National Comorbidity Survey Replication. Molecular Psychiatry, 15, 53–63. doi: 10.1038/mp. 2008.94 Rutter, M. (1979). Protective factors in children’s responses to stress and disadvantage. In M. W. Kent & J. E. Rolf (Eds.), Primary prevention of psychopathology (Vol. 3, pp. 49–74). Hanover, NH: University Press of New England.

164

Understanding Developmental Psychopathology

Rutter, M. (2008). Developing concepts in developmental psychopathology. In J. J. Hudziak (Ed.), Developmental psychopathology and wellness: Genetic and environmental influences (pp. 3–22). Washington, DC: American Psychiatric Publishing. Rutter, M. (2013). Resilience: clinical implications. Journal of Child Psychology and Psychiatry; 54(4), 474–487. doi: 10.1111/j.1469–7610 .2012.02615 Sanders, S. J., Murtha, M. T., Gupta, A. R., Murdoch, J. D., Raubeson, M. J., Willsey, A. J., . . . State, M. W. (2012). De novo mutations revealed by whole-exome sequencing are strongly associated with autism. Nature, 485(7397), 237–241. doi: 10.1038/nature10945 Sarkar, C., Basu, B., Chakroborty, D., Dasgupta, P. S., & Basu, S. (2010). The immunoregulatory role of dopamine: an update. Brain, Behavior and Immunology, 24(4), 525–528. doi: 10.1016/j.bbi.2009.10.015 Schafer, D. P., Lehrman, E. K., Kautzman, A. G., Koyama, R., Mardinly, A. R., Yamasaki, R., . . . & Stevens, B. (2012). Microglia sculpt postnatal neural circuits in an activity and complement-dependent manner. Neuron, 74, 691–705. doi: 10.1016/j.neuron.03.026 Sebat, J., Lakshmi, B., Malhotra, D., Troge, J., Lese-Martin, C., Walsh, T., . . . Wigler, M. (2007). Strong association of de novo copy number mutations with autism. Science, 2316(5823), 445–449. Doi: 10.1126/science.1138659 Shaw, P., Eckstrand, K., Sharp, W., Blumenthal, J., Lerch, J. P., Greenstein, D., . . . Rapoport, J. L. (2007). Attention-deficit /hyperactivity disorder is characterized by a delay in cortical maturation. Proceedings of the National Academy of Sciences U S A, 104, 19649–19654. doi: 10.1073/pnas.0707741104 Shyn, S. I., Shi, J., Kraft, J. B., Potash, J. B., Knowles, J. A., Weissman, M. M., Garriock, H. A., . . . Hamilton, S. P. (2011). Novel loci for major depression identified by genome-wide association study of sequenced treatment alternatives to relieve depression and meta-analysis of three studies. Molecular Psychiatry, 16(2), 202–215. doi: 10.1038/mp. 2009.125 Silk, J. S., Davis, S., McMakin, D. L., Dahl, R. E., & Forbes, E. E. (2012). Why do anxious children become depressed teenagers? The role of social evaluative threat and reward processing. Psychological Medicine, 42(10), 2095–2107. doi: 10.1017/S0033291712000207 Singer, H. S., Gilbert, D. L., Wolf, D. S., Mink, J. W., & Kurlan, R. (2012). Moving from PANDAS to CANS. Journal of Pediatrics, 160(5), 725–731. doi: 10.1016/j.jpeds.2011.11.040 Smeets, E. E., Pelc, K., & Dan, B. (2012). Rett syndrome. Molecular Syndromology, 2(3–5), 113–127. doi: 10.1159/000337637 Sporn, A. L., Addington, A. M., Gogtay, N., Ordoñez, A. E., Gornick, M., Clasen, L., . . . Rapoport, J. L. (2004). Pervasive developmental disorder and childhood-onset schizophrenia: comorbid disorder or a phenotypic variant of a very early onset illness? Biological Psychiatry, 55(10), 989–994. State, M. W., & Šestan, N. (2012). Neuroscience: The emerging biology of autism spectrum disorders. Science, 337(6100), 1301–1033. doi: 10.1126/science Steffenburg, S., Gillberg, C. L., Steffenburg, U., & Kyllerman, M. (1996). Autism in Angelman syndrome: A population-based study. Pediatric Neurology, 14(2), 131–136. Stein, D. J., & Bouwer, C. (1997). A neuro-evolutionary approach to the anxiety disorders. Journal of Anxiety Disorders, 11(4), 409–429. Stein, D. J., & Nesse, R. M. (2011). Threat detection, precautionary responses, and anxiety disorders. doi: 10.1016/j.neubiorev.2010.11.012 Stewart, S. E., Yu, D., Scharf, J. M., Neale, B. M., Fagerness, J. A., Mathews, C. A., . . . Pauls, D. L. (2013). Genome-wide association study of obsessive-compulsive disorder. Molecular Psychiatry, 18(7), 788–798. doi: 10.1038/mp. 2012.85

Sullivan, P. F., Neale, M. C., & Kendler, K. S. (2000). Genetic epidemiology of major depression: review and meta-analysis. American Journal of Psychiatry, 157(10), 1552–1562. Swedo, S. E., Leonard, H. L., Garvey, M., Mittleman, B., Allen, A. J., Perlmutter, S., . . . Dubbert, B. K. (1998). Pediatric autoimmune neuropsychiatric disorders associated with streptococcal infections: clinical description of the first 50 cases. American Journal of Psychiatry, 155(2), 264–271. Swedo, S. E., Rapoport, J. L., Cheslow, D. L., Leonard, H. L., Ayoub, E. M., Hosier, D. M., & Wald, E. R. (1989). High prevalence of obsessive-compulsive symptoms in patients with Sydenham’s chorea. American Journal of Psychiatry, 146(2), 246–249. Swedo. S. E., Leckman, J. F., & Rose, N. R. (2012). From research subgroup to clinical syndrome: Modifying the PANDAS criteria to describe PANS (pediatric acute-onset neuropsychiatric syndrome). Pediatrics and Therapeutics, 2, 113. doi:10.4172/2161–0665.1000113 Talkowski, M. E., Rosenfeld, J. A., Blumenthal, I., Pillalamarri, V., Chiang, C., Heilbut, A., . . . Gusella, J. F. (2012). Sequencing chromosomal abnormalities reveals neurodevelopmental loci that confer risk across diagnostic boundaries. Cell, 149(3), 525–537. doi: 10.1016/j.cell.2012.03.028. Thelen, E., & Smith, L. B. (1995). A dynamical systems approach to the development of cognition and action. Cambridge, MA: MIT Press. Thompson, R. A., & Limber S. P. (1992). Social anxiety in infancy: Stranger and separation reactions. In H. Leitenberg (Ed.), Handbook of social and evaluation anxiety (pp. 85–137). New York, NY: Plenum Press. Thornhill, R., & Thornhill, N. W. (1989). The evolution of psychological pain. In R. W. Bell & N. J. Bell (Eds.), Sociobiology and the social sciences (pp. 73–103). Lubbock, TX: Texas Tech University Press. Tinbergen, N. (2010). On aims and methods of ethology. Zeitschrift für Tierpsychologie, 20(4), 410–433. doi: 10.1111/j.1439–0310.1963 .tb01161.x Tooby, J., & Cosmides, L. (1990). The past explains the present: Emotional adaptations and the structure of ancestral environments. Ethology and Sociobiology, 11(4–5), 375–424. Urcelay, G. P., & Dalley, J. W. (2012). Linking ADHD, impulsivity and drug abuse: A neuropsychological perspective. Current Topics Behavior Neuroscience, 9, 173–197. doi:10.1007/7854_2011_119 Vallender, E. J., Mekel-Bobrov, N., & Lahn, B. T. (2008). Genetic basis of human brain evolution. Trends in Neuroscience, 31, 637–644. doi: 10.1016/j.tins.2008.08.010 Varki, A., Geschwind, D. H., & Eichler, E. E. (2008). Explaining human uniqueness: Genome interactions with environment, behaviour and culture. Nature Reviews Genetics, 9, 749–763. doi: 10.1038/nrg2428 Waddington, C. H. (1977). Tools for thought. New York, NY: Basic Books. Walker, F. O. (2007). Huntington’s disease. Lancet, 369, 218–228. doi: 10.1016/S0140–6736(07)60111–1 Wallace, M. L., Burette, A. C., Weinberg, R. J., & Philpot, B. D. (2012). Maternal loss of Ube3a produces an excitatory/inhibitory imbalance through neuron type-specific synaptic defects. Neuron, 74(5), 793–800. doi: 10.1016/j.neuron.2012.03.036 Ward, L. D., & Kellis, M. (2012). Evidence of abundant purifying selection in humans for recently acquired regulatory functions. Science, 337(6102), 1675–1678. doi: 10.1126/science.1225057 Waters, E., Matas, L., & Sroufe, L. A. (1975). Infants’ reactions to an approaching stranger: description, validation, and functional significance of wariness. Child Development, 46(2), 348–356. Werner, E. E., & Smith, R. S. (2001). Journeys from childhood to midlife: Risk, resiliency, and recovery. Ithaca, NY: Cornell University Press.

References Werner, H. (1957). The concept of development from a comparative and organismic point of view. In D. B. Harris (Ed.), The concept of development (pp. 125–146). Minneapolis, MN: University of Minnesota Press. Williams, C. A., Beaudet, A. L., Clayton-Smith, J., Knoll, J. H., Kyllerman, M., Laan, L. A., . . . Wagstaff, J. (2006). Angelman syndrome 2005: Updated consensus for diagnostic criteria. American Journal of Medical Genetics A, 140, 413–418. Williams, J., & Taylor, E. (2006). The evolution of hyperactivity, impulsivity and cognitive diversity. Journal of the Royal Society Interface, 3(8), 399–413. Winnicott, D. W. (1958a). Transitional objects and transitional phenomena. In Winnicott, D. W. Collected papers: Through paediatrics to psycho-analysis (pp. 229–242). New York, NY: Basic Books. (Original work published 1951.) Winnicott, D. W. (1958b). Primary maternal preoccupation. In Winnicott, D. W. Collected papers: Through paediatrics to psycho-analysis (pp. 300–305). New York, NY: Basic Books. (Original work published 1956.) Wood, E. J., & Lipovich, L. (2012). MicroRNAs in opioid addiction: elucidating evolution. Frontiers of Genetics 3:241. doi: 10.3389 /fgene.2012.00241

165

Woody, S., & Rachman, S. (1994). Generalized anxiety disorder (GAD) as an unsuccessful search for safety. Clinical Psychology Review, 8, 743–753. World Health Organization. (2009a). Global health risks: Mortality and burden of diseases attributable to selected major risks. Geneva, Switzerland: WHO Press. World Health Organization. (2009b). WHO report on the global tobacco epidemic. Geneva, Switzerland: WHO Press. Yaddanapudi, K., Hornig, M., Serge, R., De Miranda, J., Baghban, A., Villar, G., & Lipkin, W. I. (2010). Passive transfer of streptococcusinduced antibodies reproduces behavioral disturbances in a mouse model of pediatric autoimmune neuropsychiatric disorders associated with streptococcal infection. Molecular Psychiatry, 7, 712–726. doi: 10.1038/mp. 2009.77 Zachariah, R. M., & Rastegar, M. (2012). Linking epigenetics to human disease and Rett syndrome: The emerging novel and challenging concepts in MeCP2 research. Neural Plasticity, 2012, 415825 doi: 10.1155/2012/415825 Zigler, E., & Glick, M. (1986). A developmental approach to adult psychopathology. New York, NY: Wiley. Zohar, A. H., & Felz, L. (2001). Ritualistic behavior in young children. Journal of Abnormal Child Psychology, 29(2), 121–128.

CHAPTER 5

Animal Models of Developmental Psychopathology BRITTANY R. HOWELL, GRETCHEN N. NEIGH, and MAR M. SÁNCHEZ

INTRODUCTION 166 HPA Axis: Development and Regulation in Humans and Animal Models 167 RODENT MODELS OF DEVELOPMENTAL PSYCHOPATHOLOGY 169 Prenatal and Postnatal Development: Critical Periods 169 Common Models of Prenatal Stress 170 Rodent Models of Postnatal Stress 172 Plasticity and Susceptibility of Adverse Early Life Experiences 173

Puberty: The Perfect Storm 173 Common Rodent Models of Stress During Puberty 175 NONHUMAN PRIMATE MODELS OF DEVELOPMENTAL PSYCHOPATHOLOGY 179 Primate Brain Development: Sensitive Periods 180 CONCLUSIONS 186 Translational Implications 186 Future Directions 187 REFERENCES 188

INTRODUCTION

includes understanding the underlying cellular and molecular mechanisms involved and critical sensitive periods. A better understanding of the neurobiological mechanisms underlying increased risk for developmental psychopathology has led not only to more targeted research and clinical questions in human studies, but also to the inception of effective interventions and preventative therapies. If we can determine what is changing, when, and how, and the specific subsequent alterations in behavior, we would be able to design treatments aimed at specific processes and developmental time points. An important, but often overlooked, step in this process is to have a strong understanding of the normative patterns of neurobehavioral development not only in children, but also in the animal models being used (Machado & Bachevalier, 2003), which requires models and experiences with ecological validity. Understanding each species’ typical experiences and developmental trajectories provides the critical context in which to place outcomes associated with disruptions of neurobehavioral systems early in life, such as those that occur in response to early life adversity. Thus, the goal of this chapter is not only to describe animal models of developmental psychopathology, but to also present what is known about species-typical development of biological systems thought to play an especially strong role in developmental psychopathology and how these developmental processes compare to those in humans. Clinical studies have identified a series of irregularities in structure and function of specific neural circuits

Early life adversity is a well-documented risk factor for developmental psychopathology. While there is no substitute for directly studying human populations with early experience-related high risk for developmental psychopathology (e.g., maltreated children, children adopted from orphanages; Ehlert, 2013; Neigh, Gillespie, & Nemeroff, 2009; Teicher, Tomoda, & Andersen, 2006), animal studies offer a unique opportunity to understand the basic neurobiological and developmental underpinnings of these disorders at a level not possible in humans (Kaufman, Plotsky, Nemeroff, & Charney, 2000). This We thank the members of the NIMH-funded Early Experience, Stress and Neurobehavioral Development Center (P50 MH078105) for the stimulating discussions that influenced some of the views presented here. This work was supported by grants P50 MH078105, K18 MH105098 and F31 MH086203 from the National Institute of Mental Health (NIMH), grants HD055255 and HD077623 from the National Institute of Child Health & Human Development (NICHD) and grant NR014886 from the National Institute of Nursing Research (NINR). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIMH, NICHD, NINR or the National Institutes of Health. The project was also supported by the Office of Research Infrastructure Programs/OD P51OD11132 (YNPRC Base grant). The YNPRC is fully accredited by the Association for the Assessment and Accreditation of Laboratory Care, International. 166

Introduction

related to developmental psychopathology (De Bellis, 2005; De Bellis, Baum, et al., 1999; De Bellis, Keshavan, et al., 1999; Drevets, Gadde, & Krishnan 2011; Teicher, Andersen, Polcari, Anderson, Navalta, & Kim, 2003; Tottenham & Sheridan, 2009), and thus the focus of much animal research has targeted how these regions mature. They include the prefrontal cortex (PFC), particularly the ventromedial PFC (vmPFC: medial and orbital areas), and subcortical limbic structures that communicate with it, including the amygdala and hippocampus (Bale et al., 2010; Pryce et al., 2005; Sanchez, Ladd, & Plotsky, 2001; Veenema, 2009). These corticolimbic circuits play very important roles in emotion regulation and control of the stress response, including the activity of the neuroendocrine hypothalamic-pituitary-adrenal (HPA) axis (Herman, Ostrander, Mueller, & Figueiredo, 2005; Herman et al., 2003; Ulrich-Lai & Herman, 2009). Exposure to elevated levels of circulating stress hormones due to repeated activations of the HPA axis is thought to be one of the primary biological mediators of the increased risk for developmental psychopathology associated with early life adversity (Doom & Gunnar, 2013). Our knowledge of the developing HPA axis is still growing, but thanks in great part to animal models we are beginning to understand how perturbations of this system contribute to increased risk for developmental psychopathology. Other mechanisms involved, including inflammation, changes in gut microbiota and epigenetics will also be reviewed in this chapter, but we will focus on the HPA axis in the next section. HPA Axis: Development and Regulation in Humans and Animal Models It is important to have an understanding of normative developmental patterns in both humans and animal models to be able to accurately translate research between species. Thus, we begin this section with a brief description of normative development of the HPA axis in humans and some commonly used animal models. Typically developing human newborns are able to activate the HPA axis in response to stressors, but do not exhibit a mature circadian cortisol rhythm. However, they do show two peaks of cortisol about 12 hours apart that are independent of the time of day (Gunnar, 1992). Some studies have reported an adult-like circadian cortisol rhythm emerging as early as 2 weeks of age while others were unable detect a mature rhythm until 3 months of age (Price, Close, & Fielding, 1983; Santiago, Jorge, & Moreira, 1996). At about 3 months of age, there is also a decrease in the cortisol response to physical examinations, an effect that continues throughout the first year of life (Gunnar, 1992) and is paralleled by decreases in basal cortisol (Tollenaar, Jansen, Beijers, Riksen-Walraven, & de Weerth, 2010). This period

167

of apparent stress insensitivity is thought to mirror the stress hyporesponsive period (SHRP) described in rodents and may be dependent on the quality of the caregiving provided (Gunnar & Quevedo, 2007). In rodents, the SHRP is characterized by a quiescent HPA axis, evidenced by both an inability to mount a cortisol response to stressors that are effective in activating the axis at other ages, and by failure to suppress circulating cortisol levels in response to challenge with the synthetic glucocorticoid (GC) dexamethasone, suggesting insensitivity to the inhibitory effects of high levels of GCs observed in adults (Sapolsky & Meaney, 1986; Vazquez, 1998). Given the broad effects of GCs on homeostasis and gene expression, the purpose of this hyporesponsive period is thought to be to maintain the low level of GCs necessary for normative brain growth and development (Sapolsky & Meaney, 1986). There is strong evidence that maternal presence plays a major role maintaining these low GC levels during the SHRP and that the mother regulates the development of the pup HPA axis and fear systems, as well as their transition out of the SHRP. During the first couple weeks of life rat pups will not mount a corticosterone response to a stressor in the presence of the dam, but when removed from her they are able to show stress-induced elevations in corticosterone (Moriceau & Sullivan, 2006; Stanton, Gutierrez, & Levine, 1988; Stanton, Wallstrom, & Levine, 1987). More recent research into this phenomenon has shown that a sensitive period for attachment learning in rodents coincides with the low corticosterone levels associated with the SHRP, at the same time that fear learning is turned off (Sullivan & Holman, 2010). Interestingly, the transition from this early sensitive period when attachment (approach) learning is turned on but fear (avoidance) learning is turned off to the next developmental period when the pup starts exploring and fear is turned on seems mediated by the presence of higher levels of corticosterone acting on the amygdala, as shown by elegant manipulations of the corticosterone system during this time by Sullivan and colleagues (Moriceau, Roth, Okotoghaide, & Sullivan, 2004; Moriceau & Sullivan, 2004; Sullivan & Holman, 2010). Thus, in the early period, when maternal presence is paired with an aversive stimulus (i.e., foot shock) attachment and proximity seeking behavior (i.e., odor preference) is induced in pups due to the suppression of corticosterone release and subsequent suppression of amygdala activation by the dam, while in her absence elevated corticosterone activates the amygdala resulting in fear and avoidance behavior (Moriceau & Sullivan, 2006). Similarly, in nonhuman primates and humans there is also evidence that the period of relative stress hyporesponsivity described earlier is dependent on the quality of parental care, and may extend through childhood (Gunnar & Fisher, 2006;

168

Animal Models of Developmental Psychopathology

McCormack, Newman, Higley, Maestripieri, & Sanchez, 2009), which is consistent with the hypothesis that GC levels need to be tightly regulated during brain development to support typical maturational processes. This is also supported by alterations in the stress response and brains of humans exposed to early life stress (Gunnar & Quevedo, 2008; Loman & Gunnar, 2010; Tarullo & Gunnar, 2006). The role of salient social interactions, such as caregiving in infancy and peer relationships during adolescence, in regulating a developing human’s response to stressors has been demonstrated in many studies. In infants, sensitive and responsive parenting is able to decrease levels of cortisol elicited in response to a mild stressor (Albers, Riksen-Walraven, Sweep, & de Weerth, 2008). The developmental trajectory of this ability of caregiving to buffer the stress response is exemplified by a child’s response to inoculations over the first year and a half of life. In the presence of a parent 2, 4, and 6 month old infants show increases in cortisol; however, as early as 12 months old and to at least 18 months, this increase is no longer detectable when the parent is present (Jansen, Beijers, Riksen-Walraven, & de Weerth, 2010). The neurobiological mechanisms that underlie this ability have yet to be completely elucidated, but emerging evidence suggests some likely candidates (Hostinar, Sullivan, & Gunnar, 2014). The hormone oxytocin, induced by caregivers and with some anxiolytic and antistress properties (Cardoso, Ellenbogen, Orlando, Bacon, & Joober, 2013), is one possible route. Another may be through recruitment of brain regions linked to safety signal learning, including the vmPFC, as suggested by a recent report where viewing of one’s mother’s face resulted in changes in functional connectivity between the amygdala and the vmPFC that was accompanied by an increased ability to regulate affect in children, but that did not work in adolescence (Gee et al., 2014). Exactly how these safety learning mechanisms change over time is unknown, although recent evidence suggests that the caregiver’s ability to modulate offspring’s stress response may decrease in adolescence (Hostinar, Johnson, & Gunnar, 2014). Other changes in stress physiology taking place during adolescence include an increase in basal HPA activity, with increased stress reactivity also being reported in some studies (Gunnar, Wewerka, Frenn, Long, & Griggs, 2009). Sex differences in HPA axis function are not apparent in childhood but begin to be reliably measured around adolescence, with girls having higher cortisol levels than boys (Gunnar et al., 2009; Reynolds et al., 2013). The development of the HPA axis in nonhuman primates follows a similar pattern as that described for humans, making them ideal model organisms when attempting to model complex sequelae related to stress

physiology. The presence of a basal HPA diurnal rhythm has been reported in infant rhesus monkeys, but it still appears to be immature at 5 months of age (which is the equivalent to 18–24 months in humans), with high cortisol levels early in the morning but only a slight decline from morning to afternoon, followed by a steep decline from afternoon to bedtime—around sunset for animals housed outdoors (Raper, Bachevalier, Wallen, & Sanchez, 2013). By the juvenile stage (starting around 12 months, equivalent to 4 years in humans) cortisol sharply declines between morning and afternoon, and between afternoon and night, demonstrating an adult-like diurnal pattern of cortisol secretion (Barrett, Noble, Hanson, Pine, Winslow, & Nelson, 2009; Raper et al., 2014; Sanchez et al., 2005). Although there is no solid evidence of a true SHRP in rhesus monkeys, socially familiar cues, specifically the presence of a nurturing mother, do buffer the HPA axis stress responses in infant macaques (McCormack, Newman, Higley, Maestripieri, & Sanchez, 2009; Sanchez, 2006). Thus, as described for humans, nonhuman primate mothers also function as external regulators of infant HPA axis activity, serving as strong social buffers that prevent stress-induced activations to potential threats. In contrast, highly rejecting, abusive, and low protecting rhesus mothers fail to buffer their infants’ HPA axis stress responses (McCormack, Newman, Higley, Maestripieri, & Sanchez, 2009). This highlights the importance of early social experiences, in particular the role of nurturing caregiving, on the development of the HPA axis across different mammalian species, as well as the utility of animal models to study the role of adverse caregiving on the etiology of early life stress-induced developmental psychopathology. The mammalian HPA axis is regulated by different neural pathways that transmit specific information about exteroceptive (psychological, environmental) and interoceptive threats (hemorrhage, infection, pain) to the hypothalamic paraventricular nucleus (PVN) (Herman et al., 2003; Herman et al., 2005; Ulrich-Lai & Herman, 2009). Psychological/emotional stressors are processed by brain regions that include corticolimbic structures such as the vmPFC, amygdala, and hippocampus, which convey information about threatening stimuli to the PVN via indirect projections, and undergo massive prenatal and postnatal developmental changes in mammals, making them particularly vulnerable to adverse early social experiences (Andersen, 2003). These brain regions, as well as their connectivity, have a protracted development. Recent research has focused on the development of vmPFC-amygdala circuits and its relevance for the development of the HPA axis and emotional regulation. One successful approach to disentangle the complex interplay of interconnected structures that develop together within

Rodent Models of Developmental Psychopathology

a brain circuit is to lesion one of the regions and observe what happens to the other parts of the circuit during development using animal models. For example, although in adult nonhuman primates the amygdala plays a stimulatory role on the HPA axis, so that lesions of the amygdala result in blunted HPA axis reactivity (Kalin, Shelton, & Davidson, 2004; Machado & Bachevalier, 2008), without effects on basal HPA axis activity (Kalin et al., 2004; Machado & Bachevalier, 2008; Norman & Spies, 1981; Sapolsky, Zola-Morgan, & Squire, 1991), recent studies by our group showed that when the amygdala is lesioned during the neonatal period (around 2 weeks of age) alterations in both HPA axis reactivity and basal function result, with lesions being associated with heightened—instead of blunted-basal and stress-induced HPA axis reactivity during infancy and the juvenile period (Raper, Bachevalier, et al., 2013; Raper, Wallen, Sanchez, et al., 2013; Raper, Wilson, Sanchez, Machado, & Bachevalier, 2013; Raper et al., 2014). These findings suggest that the amygdala may have a paradoxical inhibitory influence on HPA axis activity in primates during development, switching later on to have a stimulatory role on cortisol release. Both adult and neonatal lesions of these limbic circuits result in similar alterations in emotional reactivity though (Kalin et al., 2004; Machado & Bachevalier, 2008; Machado, Emery, Capitanio, Mason, Mendoza, & Amaral, 2008; Machado, Kazama, & Bachevalier, 2009; Raper, Wallen, et al., 2013; Raper, Wilson, et al., 2013), which combined with the differential effects of lesion timing on the HPA axis highlights the need for further carefully thought out experiments to pinpoint the underlying neural mechanisms. The previous findings in nonhuman primates support emerging evidence of developmental switches in the role the amygdala plays in the regulation of stress and fear responses in humans, where amygdala functional coupling with PFC seems to switch from positive (i.e., when the amygdala is active so is the PFC) during early childhood to a more adult-like negative coupling during the adolescent transition (Gee, Humphreys, et al., 2013). This developmental switch from positive to negative coupling between PFC and amygdala could have a potentially important functional role, as it has been associated with stronger emotional regulation (Gee, Gabard-Durnam, et al., 2013). The findings from our studies using neonatal amygdala lesions in infant rhesus monkeys are consistent with this evidence, supporting an intriguing role of developmental switches in the amygdala’s role in emotional and stress regulation. Altogether, this evidence highlights the need to challenge the assumption that developing systems will be affected in the same ways as their mature counterparts. Thus, while the study of developmental psychopathology in humans is somewhat limited to observing the

169

correlational relationships between environmental influences, alterations in brain structure and function during development, and the emergence of psychopathology, animal models, both rodent and nonhuman primate, allow for very precise and ingenious manipulations of the developing organism’s environment, as well as the organism itself (e.g., neuroendocrine stress systems, brain circuits), providing invaluable information regarding possible biological, neurodevelopmental, molecular and system level mechanisms. RODENT MODELS OF DEVELOPMENTAL PSYCHOPATHOLOGY Rodent models include mostly mice and rats, and are some of the most commonly used model organisms in biomedical research. They allow for precise manipulations of neural systems and a level of experimental control not possible in humans and even in nonhuman primate studies. A full description of the natural history and ecology of rodents used in biomedical research is beyond the scope of this chapter, but has been reviewed for mice elsewhere (see Fox, Barthold, Davisson, Newcomer, Quimby, & Smith, 2006), although not for other species (e.g., rats). One notable difference between the early environment in rodents commonly used for research and that of a human infant is that rodents are altricial, born at a much earlier developmental stage than primates (equivalent to the end of the second trimester) and in litters, so that pups are not only exposed to their mothers, but also to their siblings. This allows for valuable within- and between-litter comparisons not possible in humans (except perhaps in the case of twin studies), but may also represent conceptual challenges when trying to understand the underlying mechanisms of developmental changes. Although there are obvious differences between rodents and humans, there are also certainly commonalities in neurodevelopmental patterns and physiology that make them strong animal models. Prenatal and Postnatal Development: Critical Periods Extensive research has been conducted on the plasticity of the prenatal and postnatal periods in the rat and mouse. Nearly 6,000 peer reviewed research papers have been indexed in the U.S. National Library of Medicine and it is beyond the limits of this chapter to recount all of the findings here. Many of the models and early findings of manipulations during the prenatal period and postnatal period are thoroughly detailed elsewhere (Marco, Macri, & Laviola, 2011; Molet, Maras, Avishai-Eliner, & Baram, 2014; Sanchez et al., 2001). Here we will briefly define the major models of prenatal and postnatal stress used in

170

Animal Models of Developmental Psychopathology

the field to study mechanisms leading to developmental psychopathology and highlight the systems that have been reported to be affected. Finally, we will summarize some of the recent work examining the potential reversibility of early stress effects and explore the evidence that suggests that these early life stressors create an inherent susceptibility to insults later in life. Common Models of Prenatal Stress Although a wide range of stressors are employed with rodents including both psychological and physical stressors, those applied prenatally need to be moderate enough so that to they do not cause reabsorption of the pregnancy. Because of this concern, most stressors applied prenatally in the literature are of a psychological nature and typically do not include massive physical challenges such as temperature manipulations or forced physical activity like swimming. Several of the widely used paradigms are summarized herein, but there is a great deal of variability among research groups and the descriptions in this chapter are not exhaustive, as our focus is mainly in postnatal experiences. Repeated Restraint Repeated restraint has been commonly used as a model of prenatal stress. Restraint is generally facilitated through the use of plastic restraining devices sold for use in collection of unanaesthetized blood samples. These devices have sufficient numbers of holes to allow for heat exchange and guard against hyperthermia and are designed to confine the rodent such that it cannot make head to tail turns but the chest wall is in no way compressed so that breathing is unencumbered. Restraint devices are available in a variety of sizes such that larger sized restrainers can be used as pregnancies progress and the rat’s abdominal circumference increases. The most common paradigm for prenatal restraint stress involves exposing the pregnant dam to restraint three times a day for 30–60 minutes under bright lights during the final week of gestation (days 14–21) (Bowman, MacLusky, Sarmiento, Frankfurt, Gordon, & Luine, 2004; Herrenkohl & Whitney, 1976; Kinsley & Bridges, 1988; Szuran, Pliska, Pokorny, & Welzl, 2000; Weller, Glaubman, Yehuda, Caspy, & Ben-Uria, 1988). Also reported in the literature are variations on this model in which the dam is exposed for a longer duration of restraint (2 hours) once a day for the final week of gestation (Wiggins & Gottesfeld, 1986). Although much less frequently reported, some research groups have also used this paradigm during early gestational

periods (days 7–10 postconception; Glavin, 1984) or even throughout the duration of the pregnancy (Rojo, Marin, & Menendez-Patterson, 1985). An extensive number of publications exist which demonstrate that the application of repeated restraint stress alters both the offspring behaviors with relevance to psychopathology and related neurotransmitter systems. But, the main issue is the ecological validity and translational value for human prenatal conditions of this stress model. Despite that caveat, repeated restraint stress of the dam results in intriguing alterations of multiple neural systems and behaviors relevant to psychopathology and neuropsychiatric disorders. As with many perturbations during development, the results are not necessarily uniformly negative or positive, reminding us that most of these developmental changes are a result of adaptations to the early environment. In one report of repeated prenatal restraint, although the offspring developed the previously described phenotypes of increased anxiety-like and depressive-like behaviors, they also demonstrated a resistance to chemically induced seizures (Marrocco et al., 2012). These findings highlight the adaptability of the developing brain, which in this instance were attributed to alterations in glutamatergic signaling (Marrocco et al., 2012). Prenatal restraint stress also alters attentional processing in offspring, renders the pups more susceptible to additional stress challenges after parturition (Burton, Lovic, & Fleming, 2006), and leads to modifications in responsiveness to psychostimulant drugs such as cocaine (Kippin, Szumlinski, Kapasova, Rezner, & See, 2008). As one might expect, prenatal restraint stress also causes robust alterations in the HPA axis both in terms of output of corticosterone and signaling within the hypothalamus and pituitary. More recently abundant evidence has emerged that supports the claim that the effects of prenatal restraint stress are not uniform and vary depending on the sex of the offspring (Garcia-Caceres, Diz-Chaves, et al., 2010; Garcia-Caceres, Lagunas, et al., 2010; Louvart, Maccari, & Darnaudery, 2005; Louvart, Maccari, Lesage, Leonhardt, Dickes-Coopman, & Darnaudery, 2006; Louvart, Maccari, Vaiva, & Darnaudery, 2009) and genetic factors such as gene polymorphisms (C. Kinsley & Svare, 1987; Lovic et al., 2013; Schroeder, Sultany, & Weller, 2013; van den Hove et al., 2011). Noise Noise stress is an easily delivered and potent inescapable stressor, sometimes delivered in combination with other stressors, as will be discussed in the next section. One common model of noise stress is exposure to 85- to 90-decibel

Rodent Models of Developmental Psychopathology

noise for 1 hour daily during the last week of gestation (Kim et al., 2006; Sobrian, Vaughn, Ashe, Markovic, Djuric, & Jankovic, 1997). In some cases, noise (880 Hz; 77 decibels) is applied to the pregnant dam for 15 minutes per day during midgestation (days 10–18) and the rat is also exposed to forced swim for 15 minutes per day (Nishio, Kasuga, Ushijima, & Harada, 2001). Given that forced swim is a very physically demanding stressor, this is not frequently used in gestational stress studies, but some groups have had success, meaning that the pregnancy is not compromised. A very common combination with noise stress is the addition of intense light, based on the ecology of most rodent species used in research, which are nocturnal animals, and therefore photophobic. This model has been published extensively (Fride & Weinstock, 1989; Kay, Tarcic, Poltyrev, & Weinstock, 1998; Weinstock, Matlina, Maor, Rosen, & McEwen, 1992). Noise stress has been reported to impact both somatic and cognitive outcomes in the offspring (Kay et al., 1998; Nishio, Tokumo, & Hirai, 2006; Sobrian et al., 1997), but overall less research has been conducted with these noise models than with either prenatal restraint stress or the combination of stressors used in the chronic mild stress paradigm described in the next section. Chronic Mild Stress Chronic mild stress (CMS), also referred to as chronic variable stress (CVS), is the pseudorandom exposure to multiple mild stressors over a series of days. Frequently the battery of stressors includes: restraint, noise, bright light, light cycle disruptions, dirty bedding, wet bedding, cage tilt, and crowding. There is a high degree of variability among research groups and there is no standard set of individual stressors in the CMS battery that is universally applied. The general principle of CMS is that these stressors represent daily hassles or unpredictable mild stressors that will trigger a stress response but are unlikely to result in habituation or severe physical detriment. CMS was first reported by Willner (Willner, Towell, Sampson, Sophokleous, & Muscat, 1987) and variations of CMS during gestation have been reported in the literature over the past decade with the first report appearing in 1998 (Secoli & Teixeira, 1998). The precise effects of CMS/CVS on offspring outcome are highly dependent on the timing of the gestational stress exposure (Mueller & Bale, 2006) and some effects of prenatal CMS/CVS may be mediated by influences on maternal behavior (Bourke, Capello, et al., 2013). CMS during gestation also increases startle responses, a metric associated with anxiety, in male (Hougaard, Mandrup, Kjaer, Bogh, Rosenberg, & Wegener, 2011) and female

171

offspring (Kjaer, Wegener, Rosenberg, & Hougaard, 2010). In some cases the cognitive deficits in offspring exposed to prenatal stress can be improved with antidepressant exposure (Abdul Aziz, Kendall, & Pardon, 2012). Although prenatal CMS/CVS has been demonstrated to also alter neuronal dendritic spine density in cortical regions of the adult offspring, there is less evidence that this manipulation modifies amphetamine responses (Muhammad & Kolb, 2011). The high variability in behavioral and neurological effects of CMS/CVS during pregnancy are likely due to the wide variety of stressors used to compose the CMS/CVS paradigm among laboratories and the range of gestational days to which the paradigm is applied. When considering CMS/CVS, or any prenatal stressor, it is important to consider the neural processes involved in the neuropsychiatric disorder of interest and design the paradigm to best target the development of these systems. Diet and Immune Challenge as Models of Prenatal Stress In addition to the classic prenatal challenges of psychological and physical stress, more recently it has been established that both dietary challenges and immune challenges during gestation can result in neural and behavioral manifestations that are reminiscent of psychopathology. The additional strength of these models is that they have more ecological validity and translational relevance for human pregnancy. High fat diet during gestation has been shown to further impair the changes in both dopamine signaling and the HPA axis response that normally result from prenatal stress exposure (Naef, Gratton, & Walker, 2013; Vucetic, Kimmel, Totoki, Hollenbeck, & Reyes, 2010). Further, high fat diet leads to delayed physical and neural development, increased aggression, and depressive-like behaviors in the offspring (Giriko et al., 2013). A different model using protein restriction during pregnancy also results in neural and behavioral deficits. Offspring of protein-restricted dams during pregnancy demonstrate deficits in memory, elevations in depressive-like behavior, and deficits in neurogenesis (Godoy et al., 2013) when assessed in adulthood. Given the profound energetic demands of development during gestation and the survival needs that drive adaptation to the environment, diet is a robust signal for the developing organism that garners lasting changes in neurobiology and behavior and some of these may contribute to the development of psychopathology depending on the adult environment. In terms of immune challenges during pregnancy, several rodent models have been used to examine the consequences of pathogens on later offspring development. For example, prenatal exposure to lipopolysaccharide

172

Animal Models of Developmental Psychopathology

(LPS) results in depressive-like behaviors and reduced neurogenesis in the offspring (Lin & Wang, 2014). In an alternate model of maternal immune activation using the immunogenic agent Poly I:C, the pregnant dams’ physiologic response to the challenge predicted behavioral changes in the offspring. Thus, dams that gained weight after the immune challenge gave birth to offspring with normal behavioral responses to amphetamine challenge, locomotor activity, and prepulse inhibition, while those that lost weight postimmune challenge had offspring with abnormal behavioral responses in these metrics (Missault et al., 2014). These findings suggest that the developmental alterations resulting by these (and other) early insults are complex and highly dependent on many maternal physiological factors, including her metabolic state. Not only that, but there is strong evidence that an important mechanism mediating the impact of psychological prenatal stress on offspring development involves immune activation. As shown by Bronson and Bale (Bronson & Bale, 2014), prenatal stress results in placenta inflammation, which mediates behavioral alterations in the adult offspring, in particular locomotor hyperactivity—which reflects brain dopaminergic dysregulation—and that were prevented by maternal treatment with nonsteroidal anti-inflammatory drugs during gestation. This literature linking maternal stress with proinflammatory states and the variety of prenatal immune challenge protocols being used in rodent models have been recently compared and contrasted (Meyer, 2014). Given the growing recognition that immune activation during pregnancy contributes to the etiology of psychopathology in the offspring, including the recent evidence that stress-induced activation of proinflammatory pathways in the placenta and stress-induced changes in maternal vaginal microbiome (Bronson & Bale, 2014; Howerton & Bale, 2012) mediate prenatal stress programming effects on offspring behavioral development, rodent models of gestational immune stress are proving critical to understand this early programming and to test treatment options for human patient populations. Rodent Models of Postnatal Stress Multiple postnatal stress paradigms have been used in the literature and most capitalize either directly or indirectly on manipulation of the powerful dam-pup relationship. One of the reasons behind the broad use of these models is based on the remarkable developmental changes in HPA axis regulation during the early postnatal period, which includes the thoroughly characterized transition to the HPA axis hyporesponsive period, as well as the

critical role of the dam as an external regulator of pup behavior and physiology. This maternal regulatory role is not only evident for the pup neuroendocrine stress responses, but for the development and regulation of fear responses (Rincon-Cortes & Sullivan, 2014; Sanchez et al., 2001). The most well-known mechanisms mediating the effects of maternal care on offspring neurobehavioral development have been reviewed elsewhere (Howell & Sanchez, 2011; Sanchez et al., 2001) and are not the focus of this chapter. They include (1) feeding regulation of the pup’s heart rate through milk actions on gatrointestinal receptors, (2) tactile stimulation via licking, grooming, nursing and retrieving, which stimulates sensory pathways that modulate the activity and development of limbic emotional circuits involving amygdala function, and also (3) epigenetic modifications such as DNA methylation changes in many genes, including GR, estrogen receptor, neurotrophic factors (e.g., brain derived neurotrophic factor: BDNF) and other brain proteins that regulate synaptic development (e.g., protocadherins), which altogether cause long-term genomic programming effects that seem part of extensive and coordinated biological responses to adapt to the environment, and that can be transmitted to future generations through epigenetic modifications in the germ line. While multiple postnatal stress paradigms exist in the literature, the two most commonly used with strong implications for behavior and neurobiology relevant to developmental psychopathology are reviewed next. Maternal Separation Maternal separation is a long-standing postnatal stress manipulation that has appeared in a variety of forms. The first apparent report of maternal separation was in 1969 by La Barba and colleagues (La Barba, Martini, & White, 1969) and the paradigm consisted of separation from the dam for 21 hours/day for 18 consecutive days beginning 24 hours after birth. Each separation period lasted for 6 hours with a 1 hour feeding period in which the dam was returned to the pups. This rather extreme paradigm has been replaced by a relatively standard model of 180 minutes of daily separation (maternal separation; MS) on postnatal days 2–14 that was originally described by Plotsky and Meaney (Plotsky & Meaney, 1993). This model has a much higher ecological validity than the previous longer separations as it is based on the fact that rat mothers in the wild leave the litter in the nest for short periods of time to forage, but when food sources are scarce, that time can increase to 2–3 hours (Fleming, 1986). As previously reviewed (Molet et al., 2014; Sanchez et al., 2001), this separation paradigm has robust and long lasting effects

Rodent Models of Developmental Psychopathology

on both behavioral and neurobiological endpoints with implications for developmental psychopathology, including increased anxiety/fear, stress reactivity and anhedonia. The effects of maternal separation have been shown to be mediated not really by separation-induced stress activation in the pups, but mostly by alterations in the dam’s maternal behavior following reunion after these prolonged separations, which becomes disorganized and harsh (Huot, Gonzalez, Ladd, Thrivikraman, & Plotsky, 2004). Maternal Abuse of Pups Induced by Impoverished Environment A recent rodent model of adverse caregiving induces abusive and neglectful maternal behaviors by reducing the bedding of the cage where the rat dam lives with her litter. This simple manipulation triggers drastic alterations in the mother’s behavior, which becomes very disorganized, and includes trampling the pups and carrying them around in a very rough manner. This impoverished environment model developed by Baram and colleagues (Avishai-Eliner, Gilles, Eghbal-Ahmadi, Bar-El, & Baram, 2001; Brunson et al., 2005; Gilles, Schultz, & Baram, 1996; Ivy et al., 2008), directly affects maternal behavior by manipulating the nesting environment and without separating the pups from her, leading to dysfunctional nurturing as well as physically abusive behaviors in the dam (Ivy, Brunson, Sandman, & Baram, 2008 Rice, Sandman, Lenjavi, & Baram, 2008; Roth & Sullivan, 2005). This paradigm induces cognitive and emotional alterations in the offspring (Baram et al., 2012), including the emergence of a behavioral phenotype later in life that includes depressive-like behavior as well as disrupted HPA-axis and amygdala function (Rincon-Cortes & Sullivan, 2014). Some of these behavioral and emotional alterations seem, at least in part, mediated by reduced BDNF expression in the PFC via epigenetic modifications (DNA methylation changes) in the BDNF gene (Roth, Lubin, Funk & Sweatt, 2009). Altogether this body of literature underscores the critical role of competent and nurturing parental behavior for proper neurobehavioral development of the offspring across mammalian species and points to some of the underlying biological and molecular mechanism involved. Plasticity and Susceptibility of Adverse Early Life Experiences The high degree of plasticity early in life in terms of both brain and behavior can work either for or against the ultimate adaptive development of an organism. Thus, if the immature animal grows in a generally positive

173

environment, with a nurturing mother and environmental enrichment, following early life adversity some effects of the stressor exposure can be mitigated. However, if an additional challenge is experienced, commonly referred to as a double-hit, then the outcome for the organism could be highly maladaptive and lead to overt somatic or behavioral deficits and potentially psychopathology. Models of either reversal of early insults by enrichment or further insult have begun to be applied in the laboratory setting with rodent models. Thus, alterations in synaptic proteins that are created by maternal separation have been reported to be reversible through exercise exposure, a type of enrichment, in early adulthood (Dimatelis et al., 2013). In addition, housing of maternally separated rat offspring in enriched environments postweaning was sufficient to prevent the manifestation of anxiety-like behaviors that were evident in maternally separated offspring that were not afforded the benefit of postweaning enrichment (Vivinetto, Suarez, & Rivarola, 2013). Similarly, both the HPA axis over-activation and deficits in play behavior that are induced by prenatal stress can be reversed by housing in an enriched environment during adolescence (Morley-Fletcher, Rea, Maccari, & Laviola, 2003). Conversely, exposure to LPS followed by repeated restraint in early development creates a more profound anxiety-like phenotype than either manipulation in isolation (Walker et al., 2009). Prenatal immune challenge also creates a susceptibility to restraint stress applied in the juvenile period such that mice exposed to both challenges develop deficits in prepulse inhibition and dopaminergic signaling. This susceptibility can be prevented by treatment with alphalipoic acid during the juvenile period (Deslauriers, Racine, Sarret, & Grignon, 2014), which is an antioxidant essential for aerobic metabolism. This suggests that factors in the diet can palliate the deleterious impact of stress. Thus, dietary modifications during stressor exposure during pregnancy (such as fish and coconut oil) can improve the behavioral phenotype of the offspring (Borsonelo, Suchecki, Calil, & Galduroz, 2011). However, there are also negative diet by stress interactions such that rats who have been exposed to stress prenatally are more susceptible to the negative effects of a high fat and high sugar diet in adulthood (Paternain, de la Garza, Batlle, Milagro, Martinez, & Campion, 2013). Puberty: The Perfect Storm Major depressive disorder (MDD) can have devastating consequences; it can be particularly detrimental if it develops during adolescence because it interrupts the

174

Animal Models of Developmental Psychopathology

development of social skills and causes potentially life-long alterations in physiology and behavior. The incidence of depression in the child and adolescent population is rapidly increasing in prevalence with an estimated rate of 8% near the beginning of adolescence and 25% by the end of adolescence (Kessler, Avenevoli, & Ries Merikangas, 2001). Furthermore, an adolescent episode of an anxiety or depressive disorder is associated with a two- to threefold increased risk for adult affective disorders (Pine, Cohen, Gurley, Brook, & Ma, 1998). Subthreshold depression during adolescence produces an increased risk of adult depression and suicidal behaviors similar to that of individuals who meet the criteria for major depression during adolescence (Fergusson, Horwood, Ridder, & Beautrais, 2005; Leuner, Mendolia-Loffredo, Kozorovitskiy, Samburg, Gould, & Shors, 2004). Finally, treatment plans for adolescents suffering from MDD are based largely on adult studies, and are not as efficacious in adolescents as they are in adults. The mechanisms of these differences are not understood but may be linked to the rapid physiological changes that occur during puberty and the concomitant demands on energy availability. Although ketones are a significant fuel source prenatally, glucose is the quintessential cerebral fuel source for the majority of the life span (Clarke & Sokoloff, 1994). In the resting state, the adult brain uses approximately 20% of whole-body glucose for brain metabolism (Cremer, 1982). While the cerebral demand for glucose is clearly substantial in the adult, the requirement is heightened during development. In fact, the demand for brain glucose is 3–4 times higher in children and adolescents, with as much as 80% of whole-body glucose utilization being demanded by the brain (Cremer, 1982). This substantial allocation of energetic resources is particularly remarkable in light of the substantial peripheral demands during childhood and adolescence for overall growth and maturation of the reproductive organs. The introduction of additional energetic demands during this life phase, in the form of stress, can be energetically costly and cause developmental delays (Laugero & Moberg, 2000). In addition to the demands of growth, there are dramatic and dynamic changes in sex steroids during puberty, and sex differences in physiology and behavior become more robust at this life phase. The precise timing of growth demands, sex steroid changes, and stress appear to result in a situation where intersectionality is important. Studies suggest that the timing of puberty influences the impact of stressful life events in humans (Ge, Conger, & Elder, 2001) and similar findings have been reported for rodents. Exposure during development appears to be fundamentally

different than exposure in adulthood. Exposure to a single stressor episode postweaning in rats, close to the time of the onset of puberty, increases anxiety-like behavior following a single stressor in adulthood, to a greater degree than exposure to two stressors in adulthood (Avital & Richter-Levin, 2005). Exposure to chronic variable stress during puberty enhances the acoustic startle response both at the end of puberty and in adulthood (Maslova, Bulygina, & Popova, 2002), and short-term exposure (3 days) to a psychogenic stressor (predator odor and elevated platform) during puberty produces sustained changes in fear-related behavior (Toledo-Rodriguez & Sandi, 2007). Finally, chronic stress exposure during adolescence results in sustained depressive-like and anxiety-like behaviors that are still evident in adult female rats but not male rats further suggesting that the hormonal milieu interacts with other demands of puberty to cause variable outcomes (Bourke & Neigh, 2011). Defining Adolescence in the Rodent As in humans, adolescence in rodents is difficult to precisely define and definitions can vary depending on the endpoint in question. Adolescence is loosely defined as PND 22–59 with less certainty placed on the ends of the time zone as the pubertal process is subject to environmental influence. Factors to consider when defining adolescence are social changes, sex steroid changes, and neurogenesis changes. In regard to social behavior, it is commonly accepted that infancy and early childhood in the rodent ends at weaning (PND 21) and adulthood begins at PND 60 (Beckman & Feuston, 2003; Marty, Chapin, Parks, & Thorsrud, 2003). This milestone is marked by development of social and motor behaviors which allow the rat/mouse to be independent of the mother including food seeking, reaching, play fighting, and maturation of sensory systems (Whishaw & Kolb, 2005). When using sex steroids to define adolescence, the peripubertal period is generally divided into three phases governed by the maturation of the sex organs: (1) early puberty (PND 22–35), which ends prior to the beginning of sexual organ maturation (i.e., vaginal opening in the female and balanopreputial separation in the male); (2) midpuberty (PND 35–49), which ends with the completion of sex organ maturation and the origination of sexual behavior; and (3) late puberty/early adult (PND 49–60), characterized by additional neurodevelopmental changes and cognitive behavioral changes (McCormick & Mathews, 2007; Spear, 2000). In regard to the third criteria, neurogenesis, while it is well described that neurogenesis occurs in humans

Rodent Models of Developmental Psychopathology

(Eriksson, Perfilieva, Bjork-Eriksson, Alborn, Nordborg, Peterson, & Gage, 1998), developmental changes in neurogenesis have been best characterized in rodent species. Androgens and estrogens are known to influence neurogenesis (Galea, 2008; Galea, Spritzer, Barker, & Pawluski, 2006; Li & Shen, 2005; Tanapat, Galea, & Gould, 1998) and more than twice the number of newly dividing cells in the brain are described during puberty as opposed to in adulthood (Cameron & McKay, 2001). Although many of the newly divided cells die in both the adolescent and adult brain, new cells can be retained if the animal is exposed to learning around the time of the birth of the new cells (Leuner et al., 2004). This process was first described in adults (Leuner et al., 2004), but has also been identified in adolescents (Curlik, Difeo, & Shors, 2014). Given the greater number of cells that are born in the adolescent brain, this may predispose the adolescent brain to be more likely to retain new information than the adult brain. In the context of psychopathology, this is a potentially interesting mechanism to explain why adolescent individuals exposed to the exact same traumatic event as adults are more likely to develop posttraumatic stress disorder (PTSD). An illustration of this phenomenon is the Ehime Maru sea accident in which 78% of adolescent survivors developed PTSD and only 12% of adults developed PTSD (Maeda, Kato, & Maruoka, 2009).

175

abnormal behavior associated with alterations of the normal physiology and genetic regulation of the stress response (Gutman & Nemeroff, 2003; Ladd, Huot, Thrivikraman, Nemeroff, Meaney, & Plotsky, 2000). Although prenatal and postnatal stress have been studied for several decades, studies of stress during puberty are more recent. Many of the stressors employed during adolescence function off of principles of isolation and aggression. Social isolation in mice is profoundly detrimental and causes prolonged changes in aggressive behavior and neurogenesis (Ibi et al., 2008). Social stress exposure during puberty alters the formation of agonistic behaviors in both male and female golden hamsters (Taravosh-Lahn & Delville, 2004; Wommack, Taravosh-Lahn, David, & Delville, 2003). Finally, social stress during puberty can alter subsequent stress responses (McCormick & Mathews, 2007), behavior (Barnum, Pace, Hu, Neigh, & Tansey, 2012; Bourke et al., 2013; Bourke & Neigh, 2011; McCormick, Smith, & Mathews, 2008), neuroanatomy (Weathington & Cooke, 2012; Weathington, Hamki, & Cooke, 2014), and increase substance abuse behaviors in animals (Ferris & Brewer, 1996; Ferris, Messenger, & Sullivan, 2005; McCormick, Robarts, Gleason, & Kelsey, 2004). Here we will describe several of the most commonly published models of adolescent stress currently used in rodents, as well as the most relevant findings in each model. Social Isolation

Common Rodent Models of Stress During Puberty Because of the limitations associated with conducting experiments in human subjects, animal models are invaluable in understanding the pathologic consequences of stress. Although rats cannot be queried to determine their affective state and are unlikely to experience depression in the sense that humans do, it is possible to study similar behavioral patterns and neurochemical phenomena in rodents. The neurobiology of rats is very well characterized and careful behavioral assessment can identify patterns of behavior that parallel depressive/anxiety-like behaviors in humans (Whishaw & Kolb, 2005). A multitude of preclinical studies have demonstrated that early life stressful experiences, such as maternal deprivation, exert both acute and long-term effects on neuroendocrine, cognitive, and behavioral systems (Gutman & Nemeroff, 2002; Weinstock, 2007). Consistent with the clinical literature, early life stress-induced changes in affective-like behavior and the HPA axis are more prevalent in female rats than male rats (Weinstock, 2007). Laboratory animals exposed to stressful conditions during development manifest adverse short and long-term cognitive dysfunction and

Isolation can be a powerful stressor in social animals regardless of the species. Although individual housing has profound health and behavioral effects on adult rodents, the positive influences of social housing have been mostly reported in the context of recovery from brain injury and peripheral wound healing (DeVries, Craft, Glasper, Neigh, & Alexander, 2007). However, in the context of development, social isolation has strong effects even in the absence of an additional physical challenge such as an injury. For example, a history of isolation housing starting in early puberty (postnatal day (PND 22) significantly increases the vulnerability to cocaine self-administration in adulthood. These data suggest that social isolation during puberty alters the development of reward brain circuits, predisposing the organism to greater drug abuse potential later in life. The specific effects of timing of social isolation during puberty is another current research question. An interesting approach to these social isolation stress paradigms involves starting the isolation towards the end of early puberty (PND 30) and to regroup the rats following the isolation. One such approach confines the isolation to

176

Animal Models of Developmental Psychopathology

just 5 days after PND30, returning the rats to group housing at PND 35. Although a very succinct period of isolation, this manipulation results in striking behavioral alterations in the forced swim test and elevated plus maze suggesting elevated anxiety- and depressive-like behaviors, as well as reductions in synaptic neuroplasticity (Leussis & Andersen, 2008). Furthermore, the reductions in synaptic plasticity are still evident when assessed weeks later, in adulthood (Leussis, Lawson, Stone, & Andersen, 2008). Extending the time of isolation to PND 30–50 results in further alterations in the HPA axis both at the level of output (corticosterone) and in terms of central neuronal changes (Weintraub, Singaravelu, & Bhatnagar, 2010). In addition, this prolonged separation paradigm results in sex-specific alterations in behavior and brain endpoints such that both males and females exhibit alterations following the separation, but the patterns are distinct (Weintraub et al., 2010). If the window of isolation is extended even further, PND 30–60, alterations in social learning are also observed. Thus, when tested for social recognition with a conspecific, mice that were group-housed their entire life performed better than mice housed alone from PND 30–60, independently of whether or not the isolated mice were returned to social groups for several weeks before they were tested (PND 80) (Kercmar, Budefeld, Grgurevic, Tobet, & Majdic, 2011). Altogether, this body of recent research highlights the strong long-term impact that even brief periods of social isolation during adolescence have on brain and behavioral development, emphasizing the important role of social relationships during this developmental period for humans. Restraint Stress Although the ecological and ethological validity of this model is questionable, restraint stress has been one of the most commonly used stressors used in rodent research. Plastic restraints commonly used for collection of blood samples are commercially available, and placement in a restrainer will generate a robust HPA axis response (Neigh, Bowers, Pyter, Gatien, & Nelson, 2004). Restraint is sometimes applied as a single acute stressor to challenge the HPA axis and examine the physiology of the activated state and in other cases is used as a repeated chronic stressor with multiple daily exposures. The range of days used in chronic restraint stress paradigms vary widely and can occur anywhere from days to multiple months. Repeated restraint was one of the first stressors applied in the context of adolescence and there is a wealth of data regarding the response of the adolescent rat to both acute and chronic restraint. Acute restraint increases anxiety-like behavior in

the elevated plus maze as compared to adult rats exposed to the same restraint paradigm (Slawecki, 2005). Five days of repeated restraint (90 minutes/day) cause increased social anxiety in adolescent rats and the HPA axis response to restraint did not habituate as efficiently in adolescent as it did in adult rats (Doremus-Fitzwater, Varlinskaya, & Spear, 2009). Similar divergences in effects of repeated restraint on HPA axis function between adolescents and adults have also been reported by other research groups (Lui et al., 2012). Ten days of restraint (1 hour/day) from PND 25–35 caused increased basal corticosterone and increased cocaine-induced locomotion (Lepsch, Gonzalo, Magro, Delucia, Scavone, & Planeta, 2005). Restraint during adolescence has also been demonstrated to alter ethanol-induced behaviors differentially from similar exposure in adulthood (Acevedo, Pautassi, Spear, & Spear, 2013; Varlinskaya, Truxell, & Spear, 2013; Willey & Spear, 2013). Neural function and morphology are also impacted by chronic restraint stress during adolescence with reports of reduced neurogenesis in females, but not males, (Barha, Brummelte, Lieblich, & Galea, 2011), enhanced hypothalamic glutamatergic transmission in males (Kusek, Tokarski, & Hess, 2013), evidence of lower synaptic transmission in the prelimbic cortex (Negron-Oyarzo, Perez, Terreros, Munoz, & Dagnino-Subiabre, 2014), and alterations in conditioned fear behaviors (Zhang & Rosenkranz, 2013). While restraint is a frequently used stress paradigm during adolescence, this approach has some limitations. For restraint to serve as an effective stressor, the rodents generally have to be individually housed. While this does not discount the value of restraint as a stressor, it is important that investigators acknowledge that they are using a combination of restraint and the more chronic stressor of isolation, which has strong developmental effects on brain, behavior, and physiology. This fact has strong implications for the interpretation of findings, particularly for translation to humans. Another concern is that rodents have the ability to habituate to restraint stress (Gray, Bingham, & Viau, 2010), which may create a situation in which the investigator is actually documenting the effects of habituation to a repeated stressor as opposed to the effects of chronic stress. Again, this does not necessarily invalidate the use of restraint as a chronic stressor, but it is essential to understand and differentiate which phenomenon is being studied to adequately and accurately interpret the experimental findings. Finally, repeated restraint has been criticized due to its lack of ethological relevance (Figueiredo, Bodie, Tauchi, Dolgas, & Herman, 2003) as this is not a stressor that a rodent would likely encounter

Rodent Models of Developmental Psychopathology

repeatedly in its natural environment and therefore the resulting physiological and behavioral processes may be contrived. Given these concerns and limitations of restraint as a model of chronic adolescent stress, additional stress paradigms have been recently generated that use more ethologically relevant stimuli or combine restraint with additional stressors. Social Instability Stress An adolescent stress approach that combines social isolation with instability is the social instability stress model. In this model, the rodent is isolated for one hour and then returned to a social group containing unfamiliar conspecifics. Social instability stress applied for 16 days during puberty (PND 30–45) has many behavioral, cognitive, emotional, and neural consequences. For example, it increases the risk for addiction, as shown by alterations in the locomotor effects of nicotine, with more pronounced neural correlates in male than female rats (McCormick & Ibrahim, 2007). In contrast, females appear to be more sensitive to deficits in spatial memory, which are paralleled by reduced hippocampal cell proliferation following social instability stress. Assessment of memory in an object spatial location task demonstrated that although female rats exposed to social instability had similar spatial memory to control rats shortly after the social stress paradigm in adolescence, their memory was impaired in adulthood (McCormick, Nixon, Thomas, Lowie, & Dyck, 2010). Some similar behavioral effects were also found in males exposed to social instability stress during adolescence, with memory appearing to be intact at the end of adolescence but a deficit in spatial memory evident in adulthood (McCormick, Thomas, Sheridan, Nixon, Flynn, & Mathews, 2012). Although social instability stress resulted in fewer newly generated cells in the hippocampus of females, the study design did not allow for the assessment of causation or determination of whether fewer cells were born or fewer cells survived (McCormick et al., 2010). The effects of social instability stress also extend to fear conditioning, with adolescent males (Morrissey, Mathews, & McCormick, 2011) and females (McCormick, Merrick, Secen, & Helmreich, 2007) exposed to this social stress during adolescence showing deficits in fear conditioning that could not be replicated when adult animals of either sex were exposed to the same chronic stressor. Collectively, these data highlight the plasticity of the brain and behavior in adolescent animals in response to exposure to complex social stressors and demonstrate the potential for these effects to be both delayed in onset, adolescence-specific, and enduring.

177

Social Defeat Stress Social defeat is a robust and ethologically relevant stressor that has been applied to adult male rodents for over three decades (Van de Poll, De Jonge, Van Oyen, & Van Pelt, 1982). More recently, this stress paradigm has begun to be applied to adolescent rodents. Although there is a broad range of precise experimental designs, the protocols typically involve using large adult male Long Evans rats as an aggressor against an adolescent rat (either male or female). The adult aggressor is frequently housed with a female rat to increase territorial behavior. The female is removed from the cage prior to the defeat episode. A partially obstructive barrier that divides the cage in half but allows exchange of olfactory and visual cues is placed in the center of the home cage of the aggressor and then the adolescent experimental rat is placed on the opposite side of the barrier. After a short period of acclimation (generally 5–10 minutes, depending on the specific laboratory protocol), the barrier is removed and the adult and adolescent are allowed to freely interact. There is a large degree of laboratory specific variation for how long the interaction lasts, but after a predetermined period of time or a specific number of pins, the rats are separated by the barrier and remain in visual and olfactory contact for an additional exposure period (30 minutes to 24 hours depending on the lab). Social defeat stress is a powerful stressor with high ecological and ethological validity with demonstrated wide range of effects in both male and female adolescents following exposure. Males exposed to social defeat for the entirety of puberty develop depressive-like behavior in the forced swim test and anhedonic behavior, and these effects are only partially reversible by either fluoxetine or a CRF antagonist (Bourke, Glasper, & Neigh, 2014). The demonstration of depressive-like and anxiety-like behavior following chronic adolescent social defeat has been replicated by multiple research groups with varying defeat parameters and in both rats and mice (Bourke et al., 2014; Huang et al., 2013; Iniguez et al., 2014; Vidal, Buwalda, & Koolhaas, 2011; Weathington, Arnold, & Cooke, 2012). The impact of chronic social defeat stress also impacts neural function and structure (Coppens et al., 2011; Watt et al., 2014; Weathington et al., 2012). Further, these changes in neural endpoints appear to have implications for drug abuse (Burke, Forster, Novick, Roberts, & Watt, 2013; Burke, Watt, & Forster, 2011; Novick, Forster, Tejani-Butt, & Watt, 2011), nicotine use (Zou, Funk, Shram, & Le, 2014), and alcohol consumption (Rodriguez-Arias et al., 2014). Finally, consistent with the effects of the previously discussed adolescent stress exposures, exposure to adolescent social defeat

178

Animal Models of Developmental Psychopathology

causes impairments in spatial memory (Novick, Miiller, Forster, & Watt, 2013). Altogether, these recent findings support the validity of this model as an emerging model of social stress during adolescence, which some research groups argue could potentially serve to model bullying and physical abuse during human adolescence. Chronic Mixed Modality Stress Some stress paradigms in the literature use a combination of multiple types of stressor exposures to avoid habituation to a specific stress stimulus or situation. A classic example of this strategy is the use of the chronic unpredictable stress model which consists of pseudo-random exposures to multiple types of stressors including restraint, isolation, extended light cycle exposure, damp bedding, cage tilt, cage crowding, and other rodent-specific stress manipulations. A less variable model that has been more recently used with adolescent rats is the chronic mixed modality stress model. This model combines isolation housing, social defeat, and restraint exposures. The isolation component of the stressor begins at PND 35 and then social defeat and restraint exposures occur over 12 days in midpuberty (PND 37–48) with a single exposure a day. This paradigm causes robust and long-lasting depressive-like and anxiety-like behaviors in females and increased body mass (Bourke & Neigh, 2011), and exaggerated inflammatory responses in males (Pyter, Kelly, Harrell, & Neigh, 2013). The combination of these individual stressors provides a context-rich stress paradigm that may have more pervasive effects than any one stressor given repeatedly. However, this is a newer stress paradigm and it has not yet been determined whether it also has implications for drug addiction-related behaviors, physiology or learning and memory. Predatory Stress In the wild, rats and mice have multiple predators and exposure to a predator or cues signaling the presence of a predator can be used to elicit a robust stress response. For both rats and mice, cat or cat-related signals and fox-related signals are frequently employed as predatory stressors, and they are now being applied during adolescence. For instance, in a study comparing adolescent and adult rat exposure to cat odor, adolescent rats were exposed to cat odor on PND 38–42 and adult rats were exposed on PND 60–64. Three weeks later, all adolescent and adult rats were assessed and the findings indicated that the adolescent rats were more impacted in terms of HPA axis response, anxiety-like behavior, and social behavior, than adult rats exposed to the predatory stress paradigm (Wright, Muir, & Perrot, 2012). In a separate

study, adolescent rats were subjected to 12 alternate days of cat fur exposure from PND 33 to 57. During exposure, adolescent rats demonstrated defensive huddling (all rats in the cage congregating away from cat fur) and reduced weight gain over the days of exposure. However, when the long-term effects of this experience were examined in adulthood, the rats exposed to cat fur actually showed reduced passive coping in the forced swim test (commonly interpreted as reduced depressive-like behaviors, although the actual functional meaning of these behaviors are a matter of debate in the field) and increased social behaviors, perhaps suggesting resilience (Kendig, Bowen, Kemp, & McGregor, 2011). It is important to note that these rats were not housed in isolation and the social housing may have provided social buffering from the repeated stress, again highlighting the powerful effect of positive social relationships during adolescence. An additional predatory stress model that is specific to mice, is the use of a rat as the predator. In this paradigm, a mouse is placed in a four inch plastic activity ball and then the ball is placed in the home cage of a pair of adult male Long Evans rats. The rats typically engage in physical interactions with the activity ball with the mouse housed inside. In this paradigm the mouse is exposed to the smell and sight of the rat as well as the unpredictable physical movement of the rats manipulating the activity ball. Although no physical injury is experienced, this predatory stress paradigm initiated during adolescence has been shown to increase anxiety-like and depressive-like behaviors as well as measures of inflammation (Barnum et al., 2012). The pro-inflammatory phenotype induced by this intense stressor exposure is even more potent than a chronic mild stress paradigm (Barnum et al., 2012; Burgado et al., 2014). However, the robust effects of these stressful exposure may not be specific of adolescence because effects of the same stressor have also been demonstrated when the predatory stress is applied in adulthood (Burgado et al., 2014). Brain Injury during Adolescence as a Model of Psychopathology Sports-related traumatic brain injury (TBI) is a significant public health concern in the US and the incidence has risen with estimates as high as 3.8 million cases in the US each year (Langlois, Rutland-Brown, & Wald, 2006). The number of child and adolescent emergency room visits for sports-related TBI have more than doubled in the past decade (Mitka, 2010) and the majority of this increase is attributable to adolescent sports-related TBI (Graves, Whitehill, Stream, Vavilala, & Rivara, 2013). TBI is associated with a myriad of somatic

Nonhuman Primate Models of Developmental Psychopathology

and psychological pathologies (Losiniecki & Shutter, 2010) and the long-term consequences of adolescent TBI are often overlooked (Graves et al., 2013; Ives, Alderman, & Stred, 2007). Further complicating the effects of adolescent TBI is the immature nature of the central nervous system. One example is the HPA axis, which undergoes substantial changes during adolescence, making this developmental stage a time of particular HPA axis susceptibility such that modifications to the HPA axis have profound and prolonged effects that are distinct from similar insults during adulthood or even early childhood (Avital & Richter-Levin, 2005; Bourke & Neigh, 2011; Greco, Hovda, & Prins, 2013, 2014; Maslova et al., 2002; McCormick & Mathews, 2007; McCormick et al., 2008; Pyter, Kelly, Harrell, & Neigh, 2013; Toledo-Rodriguez & Sandi, 2007). Adolescents are also susceptible to second-impact syndrome. Second-impact syndrome is rare in adults, but in adolescents, a second impact before complete resolution of the original injury, can lead to serious neurologic injury and death due to prolonged neuroinflammation and edema (Bruce, Alavi, Bilaniuk, Dolinskas, Obrist, & Uzzell, 1981; Cantu, 1995; McCrory & Berkovic, 1998). Even in the absence of a secondary injury, the prolonged neuroinflammation following TBI in an adolescent can cause profound changes, including possible dysregulation of the HPA axis. This is significant as the consequences of a dysregulated HPA axis response have been well documented in paving the way for a myriad of diseases including neurodegeneration and may underlie the progression of TBI-associated pathologies (Barnum et al., 2012; Bourke, Harrell, & Neigh, 2012; Bourke, Raees, Malviya, Bradburn, Binder, & Neigh, 2013; DeVries et al., 2007; McEwen, 1998, 2000, 2001, 2008; McEwen & Seeman, 1999; Neigh et al., 2009; Neigh & Nemeroff, 2006). Because of the clinical relevance of studies of TBI and evidence of developmental TBI involvement in the development of psychopathology, researchers have begun to focus on laboratory animal models of adolescent TBI, and potential treatments. This is a relatively new area of research with most studies focused on somatic outcomes. To date, research in adolescent rats has demonstrated implications of either single or repeated concussive injury on pituitary function (Greco et al., 2013, 2014) and sexual dysfunction (Greco et al., 2014). Much more work has been completed in adults with demonstrations of TBI-induced anxiety-like and depressive-like behaviors (Jones, Cardamone, Williams, Salzberg, Myers, & O’Brien, 2008; Pandey, Yadav, Mahesh, & Rajkumar, 2009; Rodgers et al., 2014) and deficits in learning and memory (Gorman, Shook, & Becker, 1993;

179

Shear, Galani, Hoffman, & Stein, 2002; Sweet, Eakin, Munyon, & Miller, 2014). Given the general susceptibility of the adolescent brain to injury and the high incidence of TBI in adolescents, it will be important to extend these studies of TBI to its impact on adolescent neurobehavioral development.

NONHUMAN PRIMATE MODELS OF DEVELOPMENTAL PSYCHOPATHOLOGY Nonhuman primates are more closely related to humans phylogenetically than rodents while still allowing for experimental manipulations not possible in studies investigating humans directly, making them crucial when studying the neural mechanisms of developmental psychopathology (Gibbs et al., 2007). Rhesus monkeys are a nonhuman primate species commonly used to study the neurobehavioral effects of early experience due to several advantages. In the wild they live in large troops made up of several multi-generational female-headed families and a few immigrant adult males, resulting in a complex social environment. They also have well-developed brains that closely resemble humans’ in respect to organization and neurochemistry (Barbas, 2000; Croxson et al., 2005; Reep, 1984; Thiebaut de Schotten, Dell’Acqua, Valabregue, & Catani, 2012), particularly regarding the PFC, which is very different in primates versus rodents (Öngür & Price, 2000; Preuss, 1995; Preuss & Goldman-Rakic, 1991a, 1991b). Stress physiology, including the regulation and development of the HPA axis, as well as the distributions of glucocorticoid receptors (GR) and corticotropin releasing factor (CRF) receptors are comparable in rhesus monkey and human brains, particularly regarding their high expression in PFC (Pryce, 2008; Sanchez, Young, Plotsky, & Insel, 1999, 2000; Seckl, Dickson, Yates, & Fink, 1991), suggesting that some effects of early life adversity on the developing brain induced by elevations in stress hormones and neuropeptides (e.g., GCs and CRH) could be similar in both species. Neuromaturational processes also occur with similar temporal and anatomical patterns in humans and monkeys, although over a condensed period of time in monkeys (approximately 1:4 years, monkey to human age) (Diamond, 1991; Gibson, 1991; Huttenlocher & Dabholkar, 1997). To provide a translational framework on which to place the findings from nonhuman primate studies presented in the following sections we will first present a brief overview of normative nonhuman primate brain development as it relates to human brain development.

180

Animal Models of Developmental Psychopathology

Primate Brain Development: Sensitive Periods Early adversity is thought to be so detrimental because it occurs at a time of rapid developmental changes in the brain, creating windows of vulnerability in which adverse experience can be encoded (Andersen, 2003; Knudsen, 2004; Rice & Barone, 2000). In both developing monkey and human brains cellular processes (e.g., synapse formation, pruning, and dendritic changes) are revealed by gross region specific decreases in gray matter (GM) and increases in white matter (WM) due to pruning of synaptic connections and myelination of axons (Lenroot & Giedd, 2006; Malkova, Heuer, & Saunders, 2006). Although early reports in humans of GM volume changes with age suggested linear decreases (Reiss, Abrams, Singer, Ross, & Denckla, 1996), more recent studies of GM volumes describe a region-specific inverted U shape during development (that is, an initial developmental increase that, after reaching a postnatal peak, follows a developmental decrease), with frontal and parietal GM peaking at 12 years of age, temporal GM peaking at 16 years, and occipital GM continuing to increase through 20 years (Giedd, 2004; Giedd, et al., 1999). These developmental patterns suggest region-specific sensitive periods, with some lasting for several years as a result of protracted developmental trajectories (Andersen, 2003). More detailed analyses of longitudinal data suggest that cortical maturation (i.e., GM loss) occurs in low-order regions (e.g., regions that process visual or somatosensory stimuli) prior to the association cortices that integrate this sensory input such as portions of the temporal and frontal cortices (Gogtay et al., 2004). Cortical development also seems to follow the maxim “ontogeny recapitulates phylogeny,” with phylogenetically older regions such as the piriform and entorhinal cortex maturing before evolutionarily newer regions such as the inferior temporal cortex and PFC (Giedd, 2004; Gogtay et al., 2004; Shaw et al., 2008). These data suggest the specific brain regions undergo increased developmental changes early in life, and those with protracted developmental trajectories, including regions such as the PFC, may be particularly vulnerable to environmental insult, and thus may contribute to developmental psychopathology. In monkeys the elimination of axons occurs both prenatally and postnatally in a region specific manner that coincides with the emergence of related behavioral and cognitive functions, with some regions such as the corpus callosum (CC) not reaching adult axon numbers until postnatal day 60 (LaMantia & Rakic, 1990). In both monkeys and humans myelination begins in some regions prenatally

(e.g., motor cortex) but it takes place postnatally in most other cortical regions and continues in association areas, namely temporal, prefrontal and parietal regions, into early adulthood (Gibson, 1991). Thus, cerebellar, temporal and prefrontal tract myelination occurs massively during childhood, with the cerebellum achieving adult levels of myelin between 3 and 6 months of age in monkeys and 1 to 2 years in humans, while WM of commissural and association tracts only reach adult levels of myelination at 3.5 years in monkeys (equivalent of 14 years in humans, peripuberty) and around 20 years in humans (Gibson, 1991). This suggests that commissural and association tracts, which are important for integrating information between brain regions, such as the uncinate fasciculus (UF), a WM tract connecting PFC with temporal regions and the last WM tract to myelinate in humans (Lebel, Walker, Leemans, Phillips, & Beaulieu, 2008), are particularly sensitive to critical aspects of maternal care as well as early adverse experiences because of their protracted development. This, combined with the important role the UF plays in emotion and stress regulation (Kim et al., 2011) via prefrontal–amygdala connectivity (i.e., prefrontal top-down control on amygdala reactivity to threat) suggest that adverse early experience-related alterations in the development of these networks are a likely neurobiological mechanism contributing to developmental psychopathology. Much of what we know about the rapid structural development that occurs early in life in the amygdala comes from studies of rhesus monkeys. The amygdala is composed of several heterogeneous nuclei that differ in cellular composition, connectivity, and function, which is why it is often referred to as the amygdaloid complex. Development of the amygdala in rhesus monkeys follows nuclei-specific patterns. The medial nucleus (which is involved in sensory processing) is near adult size at birth, while the lateral and basal nuclei, which are involved in emotional processing and learning, increase in volume from birth through 3 months of age (Chareyron, Lavenex, Amaral, & Lavenex, 2012). The volume of the central nucleus (which has direct connections to the autonomic nervous system, and is involved in modulating the stress response through these connections) is about half of adult size at birth and continues to increase beyond one year of age, during the juvenile period (Chareyron et al., 2012). These nuclei specific patterns are reflected in a postnatal increase in total amygdala volume, which occurs at the highest rates during the first 4 months of life (Payne, Machado, Bliwise, & Bachevalier, 2010). Thus different amygdala nuclei, and the specific functions and behaviors that they support,

Nonhuman Primate Models of Developmental Psychopathology

may be vulnerable at specific times during development. Numbers of oligodendrocytes in the amygdala increase in parallel with the increase in amygdala volume after 3 months of age, while neuronal size, astrocyte number, and neuron number do not change (Chareyron et al., 2012). This suggests that the increase in amygdala volume is not only due to neuronal changes, but also in the glia that produce the myelin in WM, and that these cells may also be vulnerable to early life adversity. This also suggests that WM in amygdala circuits supporting specific behaviors may be affected by early life stress. For example, in monkeys, amygdala afferents from portions of temporal cortex that relay visual information mature at week 3 when an animal first begins to respond appropriately to social cues, while efferents from these same temporal regions to orbital PFC do not become mature until 2 months when curiosity and frustration become apparent (Machado & Bachevalier, 2003). Thus, early adverse experiences in primates, particularly alterations in proper maternal care, occurring during the first few weeks of life may interrupt the ability to process social information, while adversity occurring during the second month of life might affect detection of stimuli salience and, therefore, exploratory behaviors. The complexity of the interplay of timing of these adverse experiences and neurodevelopment is one challenge the field of developmental psychopathology faces, and one in which nonhuman primate models may be especially informative given the similarities in neurobehavioral development highlighted earlier. Although there is still much to be learned about normative neurobehavioral development, we are beginning to appreciate the importance of first understanding how these brain circuits and processes develop under species typical conditions when trying to disentangle the many factors that contribute to developmental psychopathology. During normative development drastic changes occur at the cellular, molecular, structural, functional, and systems levels in brain networks that have been consistently linked with developmental psychopathological symptomatology. It is this rapid maturation early in life that is thought to make these regions sensitive to environmental influence (Andersen, 2003). Therefore, several nonhuman primate models of early adversity and social stress have been used to study these effects, and to determine the underlying mechanisms of these experiences on the brain and behavior (Kalin & Shelton, 2003; Machado & Bachevalier, 2003a; Nelson & Winslow, 2008; Parker & Maestripieri, 2011; Sanchez, 2006; Sanchez et al., 2001; Stevens, Leckman, Coplan, & Suomi, 2009), which will be reviewed in the next sections.

181

Social Deprivation and Nursery Rearing Arguably the most salient experience for a newborn primate is the relationship with its primary caregiver. Thus disruption of species typical early interaction with the primary caregiver, either through experimental manipulations or as sometimes occurs in nature, can serve as potent paradigms for studying the neurobiological underpinnings of developmental psychopathology. Studies of the effect of this type of early life stress (ELS) began in the 1960’s with Dr. Harry Harlow’s seminal work using experimental manipulations of the early rearing environment to model ELS including rearing in partial social isolation. Removing the vast majority of social interactions typically experienced by young primates resulted in severe behavioral and social deficits including stereotypies, self-injurious behavior, inability to read social cues or develop normal social relationships (including maternal behavior), and hyperaggressiveness (Suomi, Harlow, & Kimball, 1971). Rearing in complete social isolation led to similar outcomes, but with increased severity of developmental outcomes, and also included increased fear and anxiety (Seay & Gottfried, 1975). Alterations in the HPA axis have also been reported in this model (for review, see Sanchez et al., 2001), and recently several studies have reported evidence supporting possible genetic mechanisms mediating individual vulnerability to the severity of outcomes (i.e., gene-by-environment interactions that highlight strong effects of the serotonin transporter and monoamine oxidase A gene polymorphisms) as well as ELS-induced epigenetic alterations (Barr et al., 2003, 2004; Newman et al., 2005; Provençal et al., 2012). One recent study looked at whether levels of cortisol accumulated in hair (a measure of chronic glucocorticoid exposure) were related to anxious behavior and early experience in the form of peer-rearing. This study found that animals reared in the nursery were more anxious and had higher levels of accumulated cortisol (Dettmer, Novak, Suomi, & Meyer, 2012). These manipulations of the early social environment have been invaluable in understanding the importance of social stimuli on socioemotional and physiological development, but these are all somewhat artificial paradigms with limited ecological and ethological validity (except for extreme cases of social deprivation in children, such as orphanage rearing), which are complemented by other early adversity models discussed in the following sections. Experimental Alterations in Maternal Care Maternal separation has also been used to model developmental psychopathology in both New and Old World

182

Animal Models of Developmental Psychopathology

nonhuman primates, including rhesus monkeys, squirrel monkeys (Samiri sciureus), and common marmosets (Callithrix jacchus). In rhesus monkeys, separation from the mother is a potent stressor for infant monkeys that leads to increased activity, vocalizations, and HPA axis activations (Bayart, Hayashi, Faull, Barchas, & Levine, 1990; Harlow, Harlow, & Suomi, 1971). Maternal separation has been related to decreased left dorsolateral PFC activation, as well as increased right dorsolateral and right ventral temporal/occipital lobe activation, effects related to concurrent elevations in plasma cortisol concentrations (Rilling, Winslow, O’Brien, Gutman, Hoffman, & Kilts, 2001). Repeated maternal separation in infancy has been related to alterations in stress responsivity including increased cortisol reactivity in females and flattened diurnal cortisol rhythms, as well as increased acoustic startle responses (i.e., reflecting increased anxiety) later in life (Sanchez et al., 2005). These effects are predicted by both exposure to chronically elevated levels of stress hormones and impaired maternal care following mother–infant reunion after the separation (Sanchez et al., 2001). Cameron and colleagues have gone one step further and looked at the long-term effects of complete maternal deprivation on infant behavior as a function of when the mother and infant are separated. The primary findings of these studies are consistent with those reported in children. That is, rhesus infants separated from their mothers at earlier ages (beginning at 1 week of life) showed the most detrimental effects (O’Connor & Cameron, 2006). These effects included social disturbances and alterations in fear behaviors that appeared to depend on the timing of separation. For example, animals that were separated at 1 week of age did not seek comfort when anxious, while those separated later at 1 month of age did seek comfort but were hypervigilant to social cues (O’Connor & Cameron, 2006). These adverse early experiences were also related to alterations in both PFC and amygdala (O’Connor & Cameron, 2006). Marmosets are useful model organisms when studying the effects of alterations in early caregiving for a few reasons: (1) they are small, making them logistically easier to work with; (2) they are biparental; and (3) they typically give birth to twins, allowing for twin experimental designs to tease apart genetic versus experiential effects. Studies using marmosets have demonstrated the importance of considering parental factors when studying the effects of adverse early experience (Dettling, Schnell, Maier, Feldon, & Pryce, 2007). For example, in a model using parental-infant separation consisting of daily separations between days 2 and 28 (Dettling, Feldon, & Pryce,

2002a, 2002b), the adolescent animals that experienced parental separation showed decreases in GR mRNA, hippocampal growth-associated protein-43 (GAP-43) mRNA, serotonin 1A receptor (5-HT1A R) mRNA and binding ([3 H]WAY100635), and increased vesicular GABA transporter mRNA in the hippocampus (Arabadzisz et al., 2010; Law et al., 2008). They also showed decreased levels of the dendrite-enriched protein spinophilin in the subgenual anterior cingulate cortex (Law, Pei, Feldon, Pryce, & Harrison, 2009). Increases in both urinary norepinephrine and dopamine, along with increased systolic blood pressure have also been reported (Pryce, Dettling, Spengler, Spaete, & Feldon, 2004; Pryce, Dettling, Spengler, Schnell, & Feldon, 2004), and these physiological alterations may be related to alterations in both the physiological and behavioral responses to stress reported (Dettling et al., 2002a, 2002b). Cognitive effects of this adverse early experience include increased sensitivity to perceived loss of control, reduction in reward motivation, impairments in behavioral inhibition, and impairments in reversal learning (Pryce, Dettling, Spengler, Spaete, et al., 2004; Pryce, Dettling, Spengler, Schnell, et al., 2004). All of the aforementioned studies corroborate the importance of parental care in neurobehavioral development, and support the involvement of several systems in the etiology of development psychopathology including the HPA axis, the PFC, and subcortical brain regions such as the amygdala and hippocampus. The aforementioned nonhuman primate models have certainly demonstrated the critical role of the early social environment (including the mother and the father, in biparental species like humans) on primate neurobehavioral development, but they are based on artificial experimental manipulations (maternal deprivation since birth, repeated mother infant separations), which these species may not otherwise experience in the wild (e.g., in contrast to rats, rhesus monkey mothers don’t leave their infants behind to forage), making it difficult to interpret their developmental consequences, as well as determining the generalizability and ethological relevance of the biological mechanisms involved for the human experience. Thus, we now discuss the complementary information provided by more naturalistic nonhuman primate models with stronger ecological and ethological validity. Variable Foraging Demand The variable foraging demand model contrasts the previously described models in that it was specifically developed as an ethologically valid model of early life adversity that exploits species typical behavior (i.e., foraging for food).

Nonhuman Primate Models of Developmental Psychopathology

This model was first developed by Leonard Rosenblum and colleagues and consists of variable, and most importantly, unpredictable, availability of food (i.e., periods of intense foraging alternated with periods of free access to food), and has been used to study disruption of the early social environment in bonnet macaques (Macaca radiata) (Rosenblum & Paully, 1984). It is important to point out that even under conditions where the mother had to spend more time foraging enough food was available to maintain her and her infant’s health. Thus, this manipulation does not involve malnutrition, but alters maternal behavior by reducing the amount of time spent responding to infant solicitations for care (Rosenblum & Paully, 1984). These alterations in mother–infant interactions resulted in several behavioral alterations in the infants including increased clinging to the mother, reductions in exploratory and play behavior, as well as behavioral disturbances such as despair-like behaviors involving self-clasping (Andrews & Rosenblum, 1994; Rosenblum & Paully, 1984). At 2.5 years of age animals raised under high foraging conditions showed decreases in affiliative behaviors and reductions in reactivity in response to a mildly threatening stimulus (Andrews & Rosenblum, 1994; Rosenblum, Forger, Noland, Trost, & Coplan, 2001). Long-term behavioral alterations in the infant also include social incompetence, increased fearfulness, increased social subordination, and both behavioral and physiological hyperresponsiveness to stressful stimuli (Andrews & Rosenblum, 1994; Coplan et al., 1996). This environmentally induced change in mother–infant interactions does not only result in the behavioral alterations described already, but also alters neurobiological development. For example, behavioral and hormonal responses to neurochemical manipulation using the alpha2-adrenergic agonist clonidine, the alpha2-antagonist yohimbine, and the mixed serotonin 2C/1B receptor agonist m-chlorophenylpiperazine were reported in infants raised under the high foraging demand condition as compared to those raised under low foraging demand (Rosenblum, Coplan, Friedman, Bassoff, Gorman, & Andrews, 1994; Smith, Coplan, Trost, Scharf, & Rosenblum, 1997). The high foraging demand animals also had elevated levels of CRF and somatostatin, as well as monoamine metabolites including the serotonin metabolite 5-hydroxyindoleacetic acid (5HIAA), the dopamine metabolite homovanillic acid (HVA), but reduced levels of cortisol compared with those raised under low foraging demand (Coplan et al., 1996, 1998, 2001). The increase in CSF CRF in high foraging demand offspring was due to higher levels found in those carrying the short allele of the

183

serotonin transporter polymorphism, paralleling findings in maternally separated animals (Coplan, et al., 2011). In vivo imaging studies of these animals in adulthood have reported reductions in corpus callosum area, hippocampal and temporal gyrus volumes, the neuronal viability marker N-acetylaspartate in the anterior cingulate and medial temporal lobe (reflecting signs of neuropathology in these brain regions), as well as reductions in tract microstructural integrity of the anterior limb of the internal capsule (Coplan et al., 2010; Jackowski et al., 2011; Mathew et al., 2003). Just as the previously described models that involved more invasive manipulations of maternal care, this more ecologically and ethologically valid model points to the involvement of similar brain regions in the alterations associated with changes in mother–infant interactions. They also provide evidence that specific neurochemical systems and brain regions not traditionally thought to play an active role in developmental psychopathology, such as white matter tracts, may also contribute to the etiology of developmental psychopathology. Social Subordination Nonhuman primates live in highly structured societies that provide unique opportunities to model human socioeconomic status (SES), particularly during development (Hackman, Farah, & Meaney, 2010). Rhesus monkeys, for example, live in societies organized by a strict dominance hierarchy that functions to maintain group stability (Bernstein, 1976; Bernstein, Gordon, & Rose, 1974). This is true of other nonhuman primate species as well, such as cynomolgus macaques (Macaca fascicularis), another species of Old World monkey often used when studying the effects of social subordination (Shively, Laber-Laird, & Anton, 1997; Shively, Register, Friedman, Morgan, Thompson, & Lanier, 2005). An animal’s place within the hierarchy determines its access to resources, such as food and mates (Sade, 1967). These matrilineal hierarchies are maintained through contact and noncontact aggression and threats of aggression from dominant to subordinate animals (Bernstein, 1976; Bernstein & Gordon, 1974; Bernstein et al., 1974). The defining feature of social subordination is the presentation of submissive behaviors from subordinate animals in response to agonistic interactions with dominant animals (Bernstein, 1976; Bernstein & Gordon, 1974; Bernstein et al., 1974). These agonistic encounters are unpredictable, and often unprovoked, increasing uncertainty and stress in subordinates (Silk, 2002). This presumably reduces a subordinate animal’s sense of control over the social and physical environment. Infant rhesus monkeys acquire the relative rank of their

184

Animal Models of Developmental Psychopathology

mothers within their social group (Sade, 1967), a process thought to be attributable to latent observational learning as well as harassment from higher ranking females (Holekamp & Smale, 1991). Thus, infants of low ranking mothers are subject to the pressures of their rank beginning very early in development. While we know social subordination affects pubertal timing (Wilson, Bounar, Godfrey, Michopoulos, Higgins, & Sanchez, 2013; Wilson, Gordon, & Collins, 1986; Zehr, Van Meter, & Wallen, 2005), an effect that is associated with increased emotional reactivity in these animals (Wilson et al., 2013), the subordinate phenotype is less well understood in developing monkeys. It is important to recognize that this form of social stress is experienced throughout the life of the animal, and not just during early life, which is a main difference between this and other animal models used to study the underlying neurobiological mechanisms of stress-related developmental psychopathology, apart from the different adverse experience in comparison to animal models of disrupted maternal care. Consequences of this unpredictable, continual harassment in adult females include stress-related phenotypes, such as HPA dysregulation, evidenced by hypercortisolemia and impaired GC negative feedback (Jarrell, Hoffman, Kaplan, Berga, Kinkead, & Wilson, 2008; Kaplan et al., 2010; Michopoulos, Higgins, Toufexis, & Wilson, 2012; Shively et al., 1997). Subordinate animals receive higher rates of aggression and lower rates of affiliation (Michopoulos et al., 2012). It is interesting then to note that these two behaviors were the best predictors of increased cortisol in subordinates found in a meta-analysis of several primate species (Abbott et al., 2003). Thus, social subordination in rhesus monkeys is a well-established nonhuman primate model of chronic psychosocial stressor exposure, with high ethological validity. This model has been used to study the adverse effects of chronic social stress on a broad range of adult behavioral, neuroendocrine, neuromodulatory, immune and health outcomes (Gust et al., 1991; Kaplan, Adams, Clarkson, Manuck, Shively, & Williams, 1996; Kaplan et al., 2010; Michopoulos, Berga, Kaplan, & Wilson, 2009; Morgan et al., 2002; Paiardini et al., 2009; Tung et al., 2012); however many of these studies were performed on experimentally derived groups, meaning that animals from a single rank were taken and used to form new groups under well-controlled conditions, or animals were raised in nurseries beginning in infancy. While this approach can answer many valuable questions about the effects of chronic psychosocial stress on adult animals, it is not an ideal design when trying to answer questions relating to the effects of ecologically valid and species-specific adverse early experiences. Given

the success of this model in studying effects of chronic social stress in adults, though, recent investigations are just now beginning to focus on the effects of this chronic stressor during development. For example, a recent study by our group examined the effects of social status on WM microstructure development during the juvenile period and found increased microstructural integrity in prepubertal subordinate animals in frontal (motor and sensorimotor regions) and prefrontal WM (Howell. Godfrey, et al., 2014). In this study animals with higher fractional anisotropy (increased cortical WM structural integrity) in cortical regions involved in emotional and sensory processing were observed to be also more submissive in their home social groups and more emotionally reactive (fearful) during a standardized emotional reactivity task (the Human Intruder Paradigm) (Kalin & Shelton, 1989), suggesting that increased integrity in these cortical WM regions is related to increases in socially adaptive behavior (i.e., exhibiting submissive behavior and more emotional reactivity if you are a subordinate) (Howell, Godfrey, et al., 2014). Future studies will need to address genetic versus experiential contributions to these outcomes. Infant Maltreatment Child maltreatment is not a uniquely human phenomenon, but has also been reported in both wild and captive populations of nonhuman primate species, including macaques, baboons, and marmosets (Brent, Koban, & Ramirez, 2002; Carroll & Maestripieri, 1998; Johnson, Kamilaris, Calogero, Gold, & Chrousos, 1996; Maestripieri, 1998; Maestripieri & Carroll, 1998; Maestripieri, Wallen, & Carroll, 1997; Troisi & D’Amato, 1984; Troisi, D’Amato, Fuccillo, & Scucchi, 1982). In macaque species infant maltreatment is an adverse early experience that includes two comorbid types of behavior in the mother that both result in overt signs of infant distress (vocalizations, tantrums, etc.): (1) physical abuse in which the mother exhibits aberrant violent behaviors towards the infant that cause pain and distress (drags, crushes, sits on, or roughly grooms the infant) and typically occur during the first 2–3 months of life, and (2) high rates of infant rejection early in life, which consists of pushing the infant away when it solicits contact with the mother (Maestripieri, 1998; Maestripieri & Carroll, 2000; McCormack, Sanchez, Bardi, & Maestripieri, 2006). Infant physical abuse has been more widely studied and characterized in the literature. It occurs within specific families, and by cross-fostering females born to control mothers to abusive mothers and observing their behavior with their own infants, Maestripieri and colleagues were able to

Nonhuman Primate Models of Developmental Psychopathology

show that this type of infant maltreatment is experientially passed from generation to generation (Maestripieri, 2005; Maestripieri & Carroll, 1998; Maestripieri, Megna, & Jovanovic, 2000). Abusive mothers will often abuse successive infants, and primiparous mothers show higher rates of both abuse and neglect (Maestripieri & Carroll, 1998). Using sophisticated statistical methods Maestripieri and colleagues showed that physical abuse was not triggered by infant signals, such as crying, although abusive mothers were less likely to respond positively to their infants’ cries (Maestripieri, Jovanovic, & Gouzoules, 2000), possibly leading to alterations in vocal emotional communication later in life (Jovanovic & Maestripieri, 2010). Studies using this animal model have identified several alterations in the HPA axis associated with this adverse early experience, including increases in both basal and stress levels of cortisol, as well as blunted ACTH responses to a CRF pharmacological challenge that reflects down regulation of CRF receptors in the pituitary (Koch, McCormack, Sanchez, & Maestripieri, 2014; McCormack et al., 2003; Sanchez, 2006; Sanchez, McCormack, Grand, Fulks, Graff, & Maestripieri, 2010). Altogether these findings suggest that maltreatment is a stressful experience that results in elevated activity of the HPA axis, including a likely hypothalamic CRF overproduction that causes pituitary downregulation of its receptors. Socioemotional alterations include increased emotional reactivity, comparable to those seen in maltreated children (Howell, Grand, et al., 2014; McCormack et al., 2006; Sanchez & Pollak, 2009). Infant rejection was associated with decreases in brain serotonin and dopamine metabolite levels, and rejection and low serotonin function were, in turn, associated with increased activity of proinflammatory pathways (Maestripieri, Higley, Lindell, Newman, McCormack, & Sanchez, 2006; Maestripieri, McCormack, Lindell, Higley, & Sanchez, 2006; Sanchez et al., 2007), an interesting finding given that other studies have found modulation of some maltreatment related outcomes by serotonin transporter gene polymorphisms. For example, abused infants carrying a short (risk) allele showed increased vulnerability to exhibit increased anxiety and alterations in the HPA axis than abused infants with the long allele (McCormack et al., 2009). Delays in social development including delayed independence from the mother and less exploration and play have also been reported in this model (Maestripieri & Carroll, 1998; Maestripieri, Jovanovic, et al., 2000). The increased emotional reactivity caused by infant maltreatment persists throughout the juvenile and adolescent periods (Howell, Grand, et al., 2014). Recent neuroimaging studies have reported a positive correlation

185

between amygdala volume and abuse rates received by the animals during infancy, as well as negative correlations between infant basal cortisol (collected at 1 month of age, when abuse rates were highest) and juvenile WM integrity in tracts important for behavioral and emotional regulation (Howell, Grand, et al., 2014; Howell et al., 2013). These findings parallel long-term consequences found in maltreated humans, supporting the construct validity of this model to gain understanding of the mechanisms underlying the poor developmental outcomes associated with this adverse early experience in humans. Stress Inoculation Model: the Important Issue of Timing of Early Adverse Experiences for Resilience Versus Vulnerability to Psychopathology In contrast to the evidence presented in this chapter that early adverse experiences (prenatal and postnatal stress, immune/inflammatory challenges, and disruptions in the mother–infant bond) are strong risk factors for developmental psychopathology, there is an interesting body of literature showing evidence that moderate stress at specific times of development actually has the opposite effect. That is they increase resilience when the individual faces challenges later in life. Most of the evidence for this phenomenon in nonhuman primates comes from studies in squirrel monkeys showing that manageable mild stress exposure at specific developmental times (weaning) may actually be protective, allowing the animal to better adapt to environmental challenges (Lyons & Parker, 2007; Lyons, Parker, & Schatzberg, 2010; Parker, Buckmaster, Schatzberg, & Lyons, 2004). This phenomenon has been termed stress inoculation and has been most extensively studied by exposing young squirrel monkeys during weaning (around 16 weeks of age, when they are no longer as dependent on maternal care) to mild intermittent stress consisting of separation from the mother and social group for 1 hour, once a week for ten weeks. Although rodent studies suggest that the quality of maternal care upon reunion following a prolonged (3 hours) mother-pup separation is altered and constitutes the critical factor in the increased risk for developmental psychopathology in the offspring, the stress resilience observed in the stress inoculation model does not seem to be caused by changes in maternal care (Parker, Buckmaster, Sundlass, Schatzberg, & Lyons, 2006). Although seemingly contradictory to everything presented so far, both the resilience outcomes and the lack of maternal mediation on the effects could be explained by the fact that these infants are in a later developmental stage (weaning), and are less dependent on maternal care. These studies have shown several

186

Animal Models of Developmental Psychopathology

positive behavioral outcomes related to stress inoculation including increased exploratory behavior in adolescence and enhancements in inhibitory control of behavior lasting into early adulthood (Parker, Buckmaster, Justus, Schatzberg, & Lyons, 2005; Parker, Buckmaster, Lindley, Schatzberg, & Lyons, 2012). The HPA stress response also seems to be better calibrated to handle challenges, as evidenced by reductions in basal levels of stress hormones and the ability to mount a comparable neuroendocrine response to a moderate stressor as compared to animals that had not been inoculated (Parker et al., 2005; Parker et al., 2012). Neurobiological effects of this early experience include alterations in PFC morphology and GC receptor distributions in regions related to alterations in reward related-learning, but without accompanying effects on the hippocampus (Lyons, Afarian, Schatzberg, Sawyer-Glover, & Moseley, 2002; Lyons & Schatzberg, 2003; Patel, Katz, Karssen, & Lyons, 2008). These studies highlight how important the timing of the experience is in determining risk and resilience for developmental psychopathology (Parker & Maestripieri, 2011).

CONCLUSIONS Animal models, both rodent and nonhuman primate, are indispensable when trying to determine the underlying biological mechanisms of developmental psychopathology. These models have led to significant progress in our understanding of not only maladaptation and dysfunction as it relates to developmental psychopathology, but by utilizing more ethologically valid models they have also begun to provide the normative developmental and ecological context necessary to draw meaningful conclusions. Animal models also allow us to compare and contrast findings to gain an even deeper understanding of the specific role of the early environment and caregiving in neurobehavioral development. Despite the many evolutionary, developmental, neurobiological, behavioral, and ecological differences between rodents, nonhuman primates, and humans, some striking cross-species parallelisms have emerged in studies of developmental psychopathology. These include developmental alterations of PFC and limbic regions like the amygdala and hippocampus in the etiology of developmental psychopathology, as well as the complex interplay between these regions as they change across development resulting in specific sensitive periods and windows of vulnerability. Shifting research paradigms to use more naturalistic and ethologically valid animal models has highlighted the unique role of competent and sensitive

early caregiving in mammalian neurodevelopment. These paradigm shifts have also provided compelling evidence that there is a strong evolutionary, adaptive purpose for the developmental changes triggered by early life adversity, which are advantageous as they prepare the offspring to live in a harsh and challenging environment, but also result in significant and costly poor developmental outcomes. Animal models will remain invaluable as the field of developmental psychopathology moves towards understanding even more complex mechanisms involving epigenetic modifications and transgenerational transmission of experience-sensitive behaviors. Translational Implications The goal of translational research using animal models of developmental psychopathology is to understand the biological mechanisms, predictors, developmental processes, and critical periods by which early adversity gets under the skin, resulting in life long alterations. More recently the field has moved towards understanding resilience factors and developing interventions to positively impact human populations with developmental psychopathology. But, as in all translational animal research, it is difficult to definitively quantify its impact. We are just beginning to understand the neurobiological underpinnings of dysfunctions that cut across species and suggest highly conserved phenomena, which is the first step in designing effective treatments and preventative therapies. The work presented here has identified key biological mechanisms (not only the classical stress hormones, but inflammatory processes, epigenetic modifications and alterations in gut and vaginal microbiomes), brain regions and early experiences that are now being targeted to not only treat children at risk, but to ideally prevent symptoms from developing altogether (see Gunnar & Fisher 2006 for review). As highlighted in the previous examples from animal research, and from the expansive body of human literature that has been built over the past century, the primary caregiver is especially important in shaping the development of the neurobehavioral repertoire of a child. Thus, one striking translational consequence of animal research since the seminal studies by Harlow (1958) and other work in animals that led to attachment theory (Ainsworth, 1969, 1979; Bowlby, 1958) is to encourage and emphasize warm and sensitive caregiving in our society, as especially critical early in life. This has led to the development of interventions and therapies that significantly improve child outcomes, such as educating the primary caregiver on how to properly and attentively care for their child.

Conclusions

One such intervention developed by Dozier and colleagues stemmed directly from some of the animal work finding alterations in the HPA axis in response to disruption of the caregiver–offspring relationship described already. Based on the evidence in animal models that HPA axis dysfunction is part of the etiology of developmental psychopathology, the explicit goal of this intervention was to normalize HPA axis function among children in foster care. In this intervention caregivers were shown how to provide an environment for their child that encourages development of the child’s ability to regulate their emotions and behaviors, as well as teaching the caregiver how to interpret the child’s difficult behaviors (Dozier, Higley, Albus, & Nutter, 2002). This intervention resulted in the normalization of cortisol levels, as well as improved behavioral outcomes (Dozier et al., 2006; Dozier, Peloso, Lewis, Laurenceau, & Levine, 2008), constituting a beautiful example of how animal research can directly inform strategies to improve the lives of children facing early life adversity. Another example of the direct impact of research using animal models of developmental psychopathology to improve outcomes in children early in life is the emphasis on skin-to-skin contact and tactile stimulation, or kangaroo care, often prescribed to parents of low birth weight babies and babies being treated in the neonatal intensive care unit (NICU; Jefferies 2012). This strategy was supported by rodent work that demonstrated tactile stimulation of pups, both via naturally high rates of maternal licking and grooming as well as experimental stroking of pups, was related to neurodevelopmental changes in the brain, including arousal–amygdala pathways, that were related to improved performance on several behavioral tasks (Kaffman & Meaney 2007; Richards, Mychasiuk, Kolb, & Gibb 2012). Kangaroo care has been shown to decrease infant mortality, illness, and infection, and also to improve cardiorespiratory and temperature stability, sleep organization, and neurodevelopmental outcomes (Head 2014; Jefferies 2012). This practice is a prime example of how animal research can be used to not only understand the mechanisms underlying developmental psychopathology but also to develop interventions, as described in the previous example, and even to prevent these poor outcomes. Future Directions Translation of findings from animal models into direct treatments and applications that can be used to positively impact affected human populations is a challenge for

187

researchers in all health related disciplines, but it is particularly problematic for mental and behavioral health, which has been the focus of recent discussions (see Insel, 2007). In addition to these challenges, the field of developmental psychopathology has a unique challenge because the human population involved is undergoing massive normative changes in brain, behavior, physiology and social relationships on which psychopathology is superimposed. Without a clear understanding of normative development and pathology in humans and in each animal species being utilized as a model organism it is difficult to know exactly how to model human conditions and experiences. An important way to address how to model processes that translate to humans is to consider the natural history of each species, both the commonalities that suggest direct translation to humans (as exemplified by some of the animal work on maternal care highlighted earlier), and the differences that highlight possible specializations (as shown by the wonderful work explaining maternal control over the stress response in rodents during specific developmental times). This idea of ecological validity has been a recent discussion target for research using animal models (or model animals, as Dr. Thomas Insel emphasizes in a recent article (Insel, 2007), and we think it is critical for translation. Another point we want to emphasize as we reach the closing remarks of this review is that the best way to accomplish advances in the future is through interdisciplinary collaborations that allow us to understand human phenotypes and experiences to design studies with model animals that tap into similar constructs or developmental processes (see the Research Domain Criteria published by the National Institute of Mental Health; Cuthbert & Insel, 2010), also termed reverse translation, to then apply what we learn with animal models to guide human studies. This would allow for direct comparisons of results in human and animal studies, and through these direct comparisons a much clearer idea of underlying mechanisms and translational value of the animal model. The question of how early experience is translated to epigenetic changes is currently being investigated in parallel in both humans and animal studies. Due to an explosion in biotechnological capabilities to study the mindbending complexity of genetic regulation, we are beginning to appreciate how deep early experience can get under the skin, not only of the individual, but how these changes can be transmitted to future generations (Anacker, O’Donnell, & Meaney, 2014; Kundakovic & Champagne, 2014). Work showing epigenetic alterations (heritable traits not due to changes in the DNA sequence itself) related to

188

Animal Models of Developmental Psychopathology

early experience, including variations in maternal care in rats (Champagne & Curley, 2009; Weaver et al., 2004) and nursery rearing in nonhuman primates (Dettmer & Suomi, 2014), are leading the way for in depth neurobiological studies that go beyond the immediate impact of early adversity on the individual’s neural systems to expand our understanding of how that experience also causes changes at the molecular level, not only in somatic cells, but in the germ line, supporting the transgenerational transmission of phenotypes and memories. Consistent with other approaches and findings presented already, epigenetic modifications of genes and brain regions involved in the stress response have been identified in both animal models and humans (Anacker et al., 2014; McGowan et al., 2009). Using these techniques other cellular systems, including those involved in synaptogenesis, have been implicated in the brain’s adaptations to early adversity, providing new targets for future research. This is still a nascent field that is opening new avenues of research into the neurodevelopmental mechanisms of early life adversity underlying developmental psychopathology. New and fast advances are being made in other fields in addition to epigenetics which are impacting the way we understand how early experiences shape neurobehavioral development. The new emphasis on the gut-brain axis, for example, highlights the powerful role that something once thought to be static and unrelated to central nervous system functioning, the gut microbiome, plays in the control of neural processes and how early experiences can change this interaction, leading to developmental alterations in neurotransmitter function, such as the serotonin system (Cryan & Dinan, 2015). The role of the vaginal microbiota during birth, which is extremely sensitive to prenatal stress (Jašarevi´c, Rodgers, & Bale, 2015), is to inoculate the neonate’s gut with commensal microbiota shown to be important in neurobehavioral development (Borre, O’Keeffe, Clarke, Stanton, Dinan, & Cryan 2014). This is not only a growing field in the study of developmental psychopathology but is of particular relevance for the current review because much of the evidence for the role of the gut microbiota in normative neurobehavioral development comes from studies of animal models (i.e., germ-free animals raised without any microbiota; Clarke, O’Mahony, Dinan, & Cryan, 2014). These animals show alterations in the behavioral and physiological stress responses, as well as alterations in socioemotional behavior (Clarke, et al. 2014). Based on this evidence, physicians are now inoculating babies delivered via cesarean section with their mother’s vaginal microbiota (De Jesus-Laboy et al., 2014). By incorporating measures of the gut and vaginal

microbiome and other related physiological systems, and molecular techniques including epigenetic methods, we are now expanding our understanding of the complex etiology of developmental psychopathology to the systems biology level. These are very exciting times of change, both technological and conceptual, that are opening opportunities to advance our understanding and potential for intervention and prevention of developmental psychopathology. We are able now more than ever to study the etiology of developmental psychopathology using multiple levels of analysis, all the way from the molecular to the behavioral level. Animal models have been instrumental in identifying some of the systems and underlying neurodevelopmental mechanisms involved in the brain’s responses to early life adversity, but there is still much more for us to learn from them. The future of developmental psychopathology animal research is taking shape, and brings much hope for the development of efficacious treatments and preventative therapies. By combining knowledge gained in animal models and human studies with the ever-expanding frontiers of science and engineering, such as nanotechnology, the brain–machine interface, and biofeedback research (revolutionary technological and conceptual advances in the way we can manipulate the brain to treat behavioral and mental disorders), there is great potential for alleviating the burden of early life adversity and developmental psychopathology.

REFERENCES Abbott, D. H., Keverne, E. B., Bercovitch, F. B., Shively, C. A., Mendoza, S. P. l., Saltzman, W., . . . Garland, T. (2003). Are subordinates always stressed? A comparative analysis of rank differences in cortisol levels among primates. Hormones and Behavior, 43(1), 67–82. doi: 10.1016/S0018-506X(02)00037-5 Abdul Aziz, N. H., Kendall, D. A., & Pardon, M. C. (2012). Prenatal exposure to chronic mild stress increases corticosterone levels in the amniotic fluid and induces cognitive deficits in female offspring, improved by treatment with the antidepressant drug amitriptyline. Behavioural Brain Research, 231(1), 29–39. doi: 10.1016/j.bbr.2012.02.040 Acevedo, M. B., Pautassi, R. M., Spear, N. E., & Spear, L. P. (2013). Age-dependent effects of stress on ethanol-induced motor activity in rats. Psychopharmacology, 230(3), 389–398. doi: 10.1007/s00213– 013–3163–0 Ainsworth, M. D. S. (1969). Object relations, dependency, and attachment: A theoretical review of the infant-mother relationship. Child Development, 40(4), 969–1025. doi: 10.2307/1127008 Ainsworth, M. S. (1979). Infant–mother attachment. American Psychologist, 34(10), 932. Albers, E. M., Riksen-Walraven, J. M., Sweep, F. C. G. J., & de Weerth, C. (2008). Maternal behavior predicts infant cortisol recovery from a mild everyday stressor. Journal of Child Psychology and Psychiatry, 49(1), 97–103. doi: 0.1111/j.1469-7610.2007.01818.x

References Anacker, C., O’Donnell, K. J., & Meaney, M. J. (2014). Early life adversity and the epigenetic programming of hypothalamic-pituitary-adrenal function. Dialogues in Clinical Neuroscience, 16(3), 321–333. Andersen, S. L. (2003). Trajectories of brain development: Point of vulnerability or window of opportunity? Neuroscience & Biobehavioral Reviews, 27(1–2), 3–18. doi:10.1016/S0149-7634(03)00005-8 DOI:10.1016/S0149-7634(03)00005-8#doilink Andrews, M. W., & Rosenblum, L. A. (1994). The development of affiliative and agonistic social patterns in differentially reared monkeys. Child Development, 65(5), 1398–1404. doi: 10.1111/j.1467-8624 .1994.tb00824.x Arabadzisz, D., Diaz-Heijtz, R., Knuesel, I., Weber, E., Pilloud, S., Dettling, A. C., . . . Pryce, C. R. (2010). Primate early life stress leads to long-term mild hippocampal decreases in corticosteroid receptor expression. Biological Psychiatry, 67(11), 1106–1109. doi: 10.1016/j.biopsych.2009.12.016 Avishai-Eliner, S., Gilles, E. E., Eghbal-Ahmadi, M., Bar-El, Y., & Baram, T. Z. (2001). Altered regulation of gene and protein expression of hypothalamic-pituitary-adrenal axis components in an immature rat model of chronic stress. Journal of Neuroendocrinology, 13(9), 799–807. Doi: 10.1046/j.1365-2826.2001.00698.x Avital, A., & Richter-Levin, G. (2005). Exposure to juvenile stress exacerbates the behavioural consequences of exposure to stress in the adult rat. International Journal of Neuropsychopharmacology, 8(2), 163–173. doi: 10.1017/S1461145704004808 Bale, T. L., Baram, T. Z., Brown, A. S., Goldstein, J. M., Insel, T. R., McCarthy, M. M., . . . Susser, E. S. (2010). Early life programming and neurodevelopmental disorders. Biological Psychiatry, 68(4), 314–319. doi: 10.1016/j.biopsych.2010.05.028 Baram, T. Z., Davis, E. P., Obenaus, A., Sandman, C. A., Small, S. L., Solodkin, A., & Stern, H. (2012). Fragmentation and unpredictability of early-life experience in mental disorders. American Journal of Psychiatry, 169(9), 907–915. doi: 10.1176/appi.ajp. 2012.11091347 Barbas, H. (2000). Connections underlying the synthesis of cognition, memory, and emotion in primate prefrontal cortices. Brain Research Bulletin, 52(5), 319–330. doi: 10.1016/S0361-9230(99)00245-2 Barha, C. K., Brummelte, S., Lieblich, S. E., & Galea, L. A. (2011). Chronic restraint stress in adolescence differentially influences hypothalamic-pituitary-adrenal axis function and adult hippocampal neurogenesis in male and female rats. Hippocampus, 21(11), 1216–1227. doi: 10.1002/hipo.20829 Barnum, C. J., Pace, T. W., Hu, F., Neigh, G. N., & Tansey, M. G. (2012). Psychological stress in adolescent and adult mice increases neuroinflammation and attenuates the response to LPS challenge. Journal of Neuroinflammation, 9(9). doi: 10.1186/1742–2094–9–9 Barr, C. S., Newman, T. K., Becker, M. L., Parker, C. C., Champoux, M., Lesch, K. P., . . . Higley, J. D. (2003). The utility of the non-human primate model for studying gene by environment interactions in behavioral research. Genes, Brain and Behavior, 2(6), 336–340. doi: 10.1046/j.1601-183X.2003.00051.x Barr, C. S., Newman, T. K., Shannon, C., Parker, C., Dvoskin, R. L., Becker, M. L., . . . Higley, J. D. (2004). Rearing condition and rh5-HTTLPR interact to influence limbic-hypothalamicpituitary-adrenal axis response to stress in infant macaques. Biological Psychiatry, 55(7), 733–738. doi: 10.1016/j.biopsych.2003.12.008 Barrett, C. E., Noble, P., Hanson, E., Pine, D. S., Winslow, J. T., & Nelson, E. E. (2009). Early adverse rearing experiences alter sleep–wake patterns and plasma cortisol levels in juvenile rhesus monkeys. Psychoneuroendocrinology, 34(7), 1029–1040. doi:10.1016/j.psyneuen .2009.02.002 doi:10.1016/j.psyneuen.2009.02.002 Bayart, F., Hayashi, K. T., Faull, K. F., Barchas, J. D., & Levine, S. (1990). Influence of maternal proximity on behavioral and

189

physiological responses to separation in infant rhesus monkeys (Macaca mulatta). Behavioral Neuroscience, 104(1), 98–107. doi: 10.1037/0735-7044.104.1.98 Beckman, D. A., & Feuston, M. (2003). Landmarks in the development of the female reproductive system. Birth Defects Research. Part B, Developmental and Reproductive Toxicology, 68(2), 137–143. doi: 10.1002/bdrb.10016 Bernstein, I. S. (1976). Dominance, aggression and reproduction in primate societies. Journal of Theoretical Biology, 60(2), 459–472. doi: 10.1177/053901883022001005 Bernstein, I. S., & Gordon, T. P. (1974). The function of aggression in primate societies. American Scientist, 62(3), 304–311. http://www .jstor.org/stable/27844884 Bernstein, I. S., Gordon, T. P., & Rose, R. M. (1974). Aggression and social controls in rhesus monkey (Macaca mulatta) groups revealed in group formation studies. Folia Primatologica; International Journal of Primatology (Basel), 21(2), 81–107. doi: 10.1159/000155607 Borre, Y. E., O’Keeffe, G. W., Clarke, G., Stanton, C., Dinan, T. G., & Cryan, J. F. (2014). Microbiota and neurodevelopmental windows: implications for brain disorders. Trends in Molecular Medicine, 20(9), 509–518. doi: 10.1016/j.molmed.2014.05.002 Borsonelo, E. C., Suchecki, D., Calil, H. M., & Galduroz, J. C. (2011). Supplementation with fish oil and coconut fat prevents prenatal stress-induced changes in early postnatal development. International Journal of Developmental Neuroscience, 29(5), 521–527. doi: 10.1016/j.ijdevneu.2011.04.003 Bourke, C. H., Capello, C. F., Rogers, S. M., Yu, M. L., Boss-Williams, K. A., Weiss, J. M., Stowe, Z. N., & Owens, M. J. (2013). Prenatal exposure to escitalopram and/or stress in rats: a prenatal stress model of maternal depression and its treatment. Psychopharmacology (Berl), 228(2), 231–241. doi: 10.1007/s00213–013–3030-z Bourke, C. H., Glasper, E. R., & Neigh, G. N. (2014). SSRI or CRF antagonism partially ameliorate depressive-like behavior after adolescent social defeat. Behavioural Brain Research, 270, 295–299. doi: 10.1016/j.bbr.2014.05.035 Bourke, C. H., Harrell, C. S., & Neigh, G. N. (2012). Stress-induced sex differences: Adaptations mediated by the glucocorticoid receptor. Hormones and Behavior, 62(3), 210–218. doi: 10.1016/j.yhbeh.2012 .02.024 Bourke, C. H., & Neigh, G. N. (2011). Behavioral effects of chronic adolescent stress are sustained and sexually dimorphic. Hormones and Behavior, 60(1), 112–120. doi: 10.1016/j.yhbeh.2011.03.011 Bourke, C. H., Raees, M. Q., Malviya, S., Bradburn, C. A., Binder, E. B., & Neigh, G. N. (2013). Glucocorticoid sensitizers Bag1 and Ppid are regulated by adolescent stress in a sex-dependent manner. Psychoneuroendocrinology, 38(1), 84–93. doi: 10.1016/j.psyneuen.2012 .05.001 Bowlby, J. (1958). The nature of the child’s tie to his mother. International Journal of Psychoanalysis, 39, 359–373. Bowman, R. E., MacLusky, N. J., Sarmiento, Y., Frankfurt, M., Gordon, M., & Luine, V. N. (2004). Sexually dimorphic effects of prenatal stress on cognition, hormonal responses, and central neurotransmitters. Endocrinology, 145(8), 3778–3787. doi: 10.1210/en.2003–1759 Brent, L., Koban, T., & Ramirez, S. (2002). Abnormal, abusive, and stress-related behaviors in baboon mothers. Biological Psychiatry, 52(11), 1047–1056. doi: 10.1016/S0006-3223(02)01540-8 Bronson, S. L., & Bale, T. L. (2014). Prenatal stress-induced increases in placental inflammation and offspring hyperactivity are male-specific and ameliorated by maternal antiinflammatory treatment. Endocrinology, 155(7), 2635–2646. doi: 10.1210/en.2014–1040 Bruce, D. A., Alavi, A., Bilaniuk, L., Dolinskas, C., Obrist, W., & Uzzell, B. (1981). Diffuse cerebral swelling following head injuries in children:

190

Animal Models of Developmental Psychopathology

The syndrome of “malignant brain edema.” Journal of Neurosurgery, 54(2), 170–178. doi: 10.3171/jns.1981.54.2.0170 Brunson, K. L., Kramar, E., Lin, B., Chen, Y., Colgin, L. L., Yanagihara, T. K., Lynch, G., & Baram, T. Z. (2005). Mechanisms of late-onset cognitive decline after early-life stress. Journal of Neuroscience, 25(41), 9328–9338. doi: 10.1523/JNEUROSCI.2281– 05.2005 Burgado, J., Harrell, C. S., Eacret, D., Reddy, R., Barnum, C. J., Tansey, M. G., . . . Neigh, G. N. (2014). Two weeks of predatory stress induces anxiety-like behavior with co-morbid depressive-like behavior in adult male mice. Behavioural Brain Research, 275C, 120–125. doi: 10.1016/j.bbr.2014.08.060 Burke, A. R., Forster, G. L., Novick, A. M., Roberts, C. L., & Watt, M. J. (2013). Effects of adolescent social defeat on adult amphetamine-induced locomotion and corticoaccumbal dopamine release in male rats. Neuropharmacology, 67, 359–369. doi: 10.1016/j.neuropharm.2012.11.013 Burke, A. R., Watt, M. J., & Forster, G. L. (2011). Adolescent social defeat increases adult amphetamine conditioned place preference and alters D2 dopamine receptor expression. Journal of Neuroscience, 197, 269–279. doi: 10.1016/j.neuroscience.2011.09.008 Burton, C., Lovic, V., & Fleming, A. S. (2006). Early adversity alters attention and locomotion in adult Sprague-Dawley rats. Behavioral Neuroscience, 120(3), 665–675. doi: 10.1037/0735–7044.120.3.665 Cameron, H. A., & McKay, R. D. (2001). Adult neurogenesis produces a large pool of new granule cells in the dentate gyrus. Journal of Comparative Neurology, 435(4), 406–417. doi: 10.1002/cne.1040 Cantu, R. C. (1995). Head and spine injuries in youth sports. Clinics in Sports Medicine, 14(3), 517–532. Cardoso, C., Ellenbogen, M. A., Orlando, M. A., Bacon, S. L., & Joober, R. (2013). Intranasal oxytocin attenuates the cortisol response to physical stress: A dose–response study. Psychoneuroendocrinology, 38(3), 399–407. doi: 10.1016/j.psyneuen.2012.07.013 Carroll, K. A., & Maestripieri, D. (1998). Infant abuse and neglect in monkeys—A discussion of definitions, epidemiology, etiology, and implications for child maltreatment: Reply to Cicchetti (1998) and Mason (1998). Psychological Bulletin, 123(3), 234–237. doi: 10.1037/ 0033-2909.123.3.234 Champagne, F. A., & Curley, J. P. (2009). Epigenetic mechanisms mediating the long-term effects of maternal care on development. Neuroscience & Biobehavioral Reviews, 33(4), 593–600. doi: 10.1016/ j.neubiorev.2007.10.009 Chareyron, L. J., Lavenex, P. B., Amaral, D. G., & Lavenex, P. (2012). Postnatal development of the amygdala: A stereological study in macaque monkeys. Journal of Comparative Neurology, 520(9), 1965–1984. doi: 10.1002/cne.23023 Clarke, G., O’Mahony, S. M., Dinan, T. G., & Cryan, J. F. (2014). Priming for health: gut microbiota acquired in early life regulates physiology, brain and behaviour. Acta Paediatrica. 103(8), 812–819. doi: 10.1111/apa.12674 Clarke, D. D., & Sokoloff, L. (1994). Circulation and energy metabolism of the brain. Basic neurochemistry: molecular, cellular and medical aspects. New York, NY: Raven Press. Coplan, J. D., Abdallah, C. G., Kaufman, J., Gelernter, J., Smith, E. L., Perera, T. D., . . . Nemeroff, C. B. (2011). Early-life stress, corticotropin-releasing factor, and serotonin transporter gene: A pilot study. Psychoneuroendocrinology, 36(2), 289–293. doi: 10.1016/ j.psyneuen.2010.07.011 Coplan, J. D., Abdallah, C. G., Tang, C. Y., Mathew, S. J., Martinez, J., Hof, P. R., . . . Gorman, J. M. (2010). The role of early life stress in development of the anterior limb of the internal capsule in nonhuman primates. Neuroscience Letters, 480(2), 93–96. doi: S0304–3940(10)00745–7

Coplan, J. D., Andrews, M. W., Rosenblum, L. A., Owens, M. J., Friedman, S., Gorman, J. M., & Nemeroff, C. B. (1996). Persistent elevations of cerebrospinal fluid concentrations of corticotropinreleasing factor in adult nonhuman primates exposed to early-life stressors: implications for the pathophysiology of mood and anxiety disorders. Proceedings of the National Academy of Sciences, 93(4), 1619–1623. Coplan, J. D., Smith, E. L., Altemus, M., Scharf, B. A., Owens, M. J., Nemeroff, C. B., Gorman, J. M., & Rosenblum, L. A. (2001). Variable foraging demand rearing: sustained elevations in cisternal cerebrospinal fluid corticotropin-releasing factor concentrations in adult primates. Biological Psychiatry, 50(3), 200–204. doi: 10.1016/S0006-3223(01)01175-1 Coplan, J. D., Trost, R. C., Owens, M. J., Cooper, T. B., Gorman, J. M., Nemeroff, C. B., & Rosenblum, L. A. (1998). Cerebrospinal fluid concentrations of somatostatin and biogenic amines in grown primates reared by mothers exposed to manipulated foraging conditions. Archives of General Psychiatry, 55(5), 473–477. doi: 10.1001/archpsyc.55.5.473 Coppens, C. M., Siripornmongcolchai, T., Wibrand, K., Alme, M. N., Buwalda, B., de Boer, S. F., Koolhaas, J. M., & Bramham, C. R. (2011). Social defeat during adolescence and adulthood differentially induce BDNF-regulated immediate early genes. Frontiers in Behavioral Neuroscience, 5, 72. doi: 10.3389/fnbeh.2011.00072 Cremer, J. E. (1982). Substrate utilization and brain development. Journal of Cerebral Blood Flow and Metabolism, 2(4), 394–407. doi: 10.1038/jcbfm.1982.45 Croxson, P. L., Johansen-Berg, H., Behrens, T. E. J., Robson, M. D., Pinsk, M. A., Gross, C. G., . . . Rushworth, M. F. S. (2005). Quantitative investigation of connections of the prefrontal cortex in the human and macaque using probabilistic diffusion tractography. Journal of Neuroscience, 25(39), 8854–8866. doi: 10.1523/JNEUROSCI.1311-05.2005 Cryan, J. F., & Dinan, T. G. (2015). More than a gut feeling: The microbiota regulates neurodevelopment and behavior. Neuropsychopharmacology, 40(1), 241–242. doi: 10.1038/npp. 2014.224 Curlik, D. M., Difeo, G., & Shors, T. J. (2014). Preparing for adulthood: thousands upon thousands of new cells are born in the hippocampus during puberty, and most survive with effortful learning. Frontiers in Neuroscience, 8, 70. doi: 10.3389/fnins.2014.00070 Cuthbert, B. N., & Insel, T. R. (2010). Toward new approaches to psychotic disorders: The NIMH research domain criteria project. Schizophrenia Bulletin, 36(6), 1061–1062. doi: 10.1093/schbul/sbq108 De Bellis, M. D. (2005). The psychobiology of neglect. Child Maltreatment, 10(2), 150–172. doi: 10.1177/1077559505275116 De Bellis, M. D., Baum, A. S., Birmaher, B., Keshavan, M. S., Eccard, C. H., Boring, A. M., Jenkins, F. J., & Ryan, N. D. (1999). A.E. Bennett Research Award. Developmental traumatology. Part I: Biological stress systems. Biological Psychiatry, 45(10), 1259–1270. doi: 10.1016/S0006-3223(99)00044-X De Bellis, M. D., Keshavan, M. S., Clark, D. B., Casey, B. J., Giedd, J. N., Boring, A. M., Frustaci, K., & Ryan, N. D. (1999). A.E. Bennett Research Award. Developmental traumatology. Part II: Brain development. Biological Psychiatry, 45(10), 1271–1284. doi: 10.1016/S0006-3223(99)00045-1 De Jesus-Laboy, K. M., Cox, L. M., Rodriguez-Rivera, S. M., Rivera-Vinas, J., Mendez, K., Clemente, J. C., Knight, R., & Dominguez-Bello, M. G. (2014, May). Restoring the normal microbiota of Cesarean-section born infants. Poster presented at the Meeting of the American Society of Microbiology, Boston, MA. Deslauriers, J., Racine, W., Sarret, P., & Grignon, S. (2014). Preventive effect of alpha-lipoic acid on prepulse inhibition deficits in a juvenile two-hit model of schizophrenia. Neuroscience, 272, 261–270. doi: 10.1016/j.neuroscience.2014.04.061

References Dettling, A. C., Feldon, J., & Pryce, C. R. (2002a). Early deprivation and behavioral and physiological responses to social separation/novelty in the marmoset. Pharmacology Biochemistry and Behavior, 73(1), 259–269. doi: 10.1016/S0091-3057(02)00785-2 Dettling, A. C., Feldon, J., & Pryce, C. R. (2002b). Repeated parental deprivation in the infant common marmoset (Callithrix jacchus, primates) and analysis of its effects on early development. Biological Psychiatry, 52(11), 1037–1046. doi: 10.1016/S0006-3223(02)01460-9 Dettling, A. C., Schnell, C. R., Maier, C., Feldon, J., & Pryce, C. R. (2007). Behavioral and physiological effects of an infant-neglect manipulation in a bi-parental, twinning primate: Impact is dependent on familial factors. Psychoneuroendocrinology, 32(4), 331–349. doi: 10.1016/j.psyneuen.2007.01.005 Dettmer, A. M., Novak, M. A., Suomi, S. J., & Meyer, J. S. (2012). Physiological and behavioral adaptation to relocation stress in differentially reared rhesus monkeys: Hair cortisol as a biomarker for anxiety-related responses. Psychoneuroendocrinology, 37(2), 191–199. doi: 10.1016/j.psyneuen.2011.06.003 Dettmer, A. M., & Suomi, S. J. (2014). Nonhuman primate models of neuropsychiatric disorders: Influences of early rearing, genetics, and epigenetics. ILAR Journal, 55(2), 361–370. doi: 10.1093/ilar/ilu025 DeVries, A. C., Craft, T. K., Glasper, E. R., Neigh, G. N., & Alexander, J. K. (2007). 2006 Curt P. Richter award winner: Social influences on stress responses and health. Psychoneuroendocrinology, 32(6), 587–603. doi: 10.1016/j.psyneuen.2007.04.007 Diamond, A. (1991). Frontal lobe involvement in cognitive changes during the first year of life. Brain Maturation and Cognitive Development: Comparative and Cross-Cultural Perspectives (pp. 127–180). Hawthorne, NY: Aldine de Gruyter. Dimatelis, J. J., Hendricks, S., Hsieh, J., Vlok, N. M., Bugarith, K., Daniels, W. M., & Russell, V. A. (2013). Exercise partly reverses the effect of maternal separation on hippocampal proteins in 6-hydroxydopamine-lesioned rat brain. Experimental Physiology, 98(1), 233–244. doi: 10.1113/expphysiol.2012.066720 Doom, J. R., & Gunnar, M. R. (2013). Stress physiology and developmental psychopathology: past, present, and future. Developmental Psychopathology, 25(4 Pt 2), 1359–1373. doi: 10.1017/ S0954579413000667 Doremus-Fitzwater, T. L., Varlinskaya, E. I., & Spear, L. P. (2009). Social and non-social anxiety in adolescent and adult rats after repeated restraint. Physiology & Behavior, 97(3–4), 484–494. doi: 10.1016/j.physbeh.2009.03.025 Dozier, M., Higley, E., Albus, K. E., & Nutter, A. (2002). Intervening with foster infants’ caregivers: Targeting three critical needs. Infant Mental Health Journal, 23(5), 541–554. doi: 10.1002/imhj.10032 Dozier, M., Peloso, E., Lewis, E., Laurenceau, J. P., & Levine, S. (2008). Effects of an attachment-based intervention on the cortisol production of infants and toddlers in foster care. Development and Psychopathology, 20(3), 845–859. doi: 10.1017/S0954579408000400 Dozier, M., Peloso, E., Lindhiem, O., Gordon, M. K., Manni, M., Sepulveda, S., . . . Levine, S. (2006). Developing evidence-based interventions for foster children: An example of a randomized clinical trial with infants and toddlers. Journal of Social Issues, 62(4), 767–785. Drevets, W. C., Gadde, K. M., & Krishnan, K. R. R. (2011). Neuroimaging studies of mood. In D. S. Charney & E. J. Nestler (Eds.), Neurobiology of mental illness (pp. 461–490). New York, NY: Oxford University Press. Ehlert, U. (2013). Enduring psychobiological effects of childhood adversity. Psychoneuroendocrinology, 38(9), 1850–1857. doi: 10.1016/ j.psyneuen.2013.06.007 Eriksson, P. S., Perfilieva, E., Bjork-Eriksson, T., Alborn, A. M., Nordborg, C., Peterson, D. A., & Gage, F. H. (1998). Neurogenesis in the

191

adult human hippocampus. Nature Medicine, 4(11), 1313–1317. doi: 10.1038/3305 Fergusson, D. M., Horwood, L. J., Ridder, E. M., & Beautrais, A. L. (2005). Subthreshold depression in adolescence and mental health outcomes in adulthood. Archives of General Psychiatry, 62(1), 66–72. doi: 10.1001/archpsyc.62.1.66 Ferris, C. F., & Brewer, J. (1996). Adolescent stress alters ethanol ingestion and agonistic behavior in the male golden hamster. Annals of the New York Academy of Sciences, 794, 348–351. doi: 10.1111/j.1749-6632 .1996.tb32539.x Ferris, C. F., Messenger, T., & Sullivan, R. (2005). Behavioral and neuroendocrine consequences of social subjugation across adolescence and adulthood. Frontiers in Zoology, 2(1), 7–18. doi: 10.1186/ 1742-9994-2-7 Figueiredo, H. F., Bodie, B. L., Tauchi, M., Dolgas, C. M., & Herman, J. P. (2003). Stress integration after acute and chronic predator stress: differential activation of central stress circuitry and sensitization of the hypothalamo-pituitary-adrenocortical axis. Endocrinology, 144(12), 5249–5258. doi: 10.1210/en.2003–0713 Fleming, A. (1986). Psychobiology of rat maternal behavior: how and where hormones act to promote maternal behavior at parturition. Annals of the New York Academy of Sciences, 474, 234–251. doi: 10.1111/j.1749-6632.1986.tb28015.x Fox, J. G., Barthold, S., Davisson, M., Newcomer, C. E., Quimby, F. W., & Smith, A. (2006). The mouse in biomedical research: Normative biology, husbandry, and models (Vol. 3). New York, NY: Academic Press. Fride, E., & Weinstock, M. (1989). Alterations in behavioral and striatal dopamine asymmetries induced by prenatal stress. Pharmacology Biochemistry and Behavior, 32(2), 425–430. doi: 10.1016/0091 -3057(89)90174-3 Galea, L. A. (2008). Gonadal hormone modulation of neurogenesis in the dentate gyrus of adult male and female rodents. Brain Research Reviews, 57(2), 332–341. doi: 10.1016/j.brainresrev.2007.05.008 Galea, L. A., Spritzer, M. D., Barker, J. M., & Pawluski, J. L. (2006). Gonadal hormone modulation of hippocampal neurogenesis in the adult. Hippocampus, 16(3), 225–232. doi: 10.1002/hipo.20154 Garcia-Caceres, C., Diz-Chaves, Y., Lagunas, N., Calmarza-Font, I., Azcoitia, I., Garcia-Segura, L. M., . . . Chowen, J. A. (2010). The weight gain response to stress during adulthood is conditioned by both sex and prenatal stress exposure. Psychoneuroendocrinology, 35(3), 403–413. doi: 10.1016/j.psyneuen.2009.08.001 Garcia-Caceres, C., Lagunas, N., Calmarza-Font, I., Azcoitia, I., Diz-Chaves, Y., Garcia-Segura, L. M., . . . Chowen, J. A. (2010). Gender differences in the long-term effects of chronic prenatal stress on the HPA axis and hypothalamic structure in rats. Psychoneuroendocrinology, 35(10), 1525–1535. doi: 10.1016/j.psyneuen.2010.05.006 Ge, X., Conger, R. D., & Elder, G. H., Jr. (2001). Pubertal transition, stressful life events, and the emergence of gender differences in adolescent depressive symptoms. Developmental Psychology, 37(3), 404–417. doi: 10.1037/0012-1649.37.3.404 Gee, D. G., Gabard-Durnam, L., Telzer, E. H., Humphreys, K. L., Goff, B., Shapiro, M., . . . Tottenham, N. (2014). Maternal buffering of human amygdala-prefrontal circuitry during childhood but not during adolescence. Psychological Science, 25(11), 2067–2078. doi: 10.1177/0956797614550878 Gee, D. G., Gabard-Durnam, L. J., Flannery, J., Goff, B., Humphreys, K. L., Telzer, E. H., . . . Tottenham, N. (2013). Early developmental emergence of human amygdala-prefrontal connectivity after maternal deprivation. Proceedings of the National Academy of Sciences USA, 110(39), 15638–15643. doi: 10.1073/pnas.1307893110 Gee, D. G., Humphreys, K. L., Flannery, J., Goff, B., Telzer, E. H., Shapiro, M., . . . Tottenham, N. (2013). A developmental shift from

192

Animal Models of Developmental Psychopathology

positive to negative connectivity in human amygdala–prefrontal circuitry. Journal of Neuroscience, 33(10), 4584–4593. doi: 10.1523/ JNEUROSCI.3446-12.2013 Gibbs, R. A., Rogers, J., Katze, M. G., Bumgarner, R., Weinstock, G. M., Mardis, E. R., . . . Wilson, R. K. (2007). Evolutionary and biomedical insights from the rhesus macaque genome. Science, 316(5822), 222–234. doi: 10.1126/science.1139247 Gibson, K. R. (1991). Myelination and behavioral development: a comparative perspective on questions of neotony, altriciality, and intelligence. In K. R. Gibson & A. C. Petersen (Eds.), Brain maturation and cognitive development: Comparative and cross-cultural perspectives (pp. 29–63). Hawthorne, NY: Aldine de Gruyter. Giedd, J. N. (2004). Structural magnetic resonance imaging of the adolescent brain. Annals of the New York Academy of Sciences, 1021(1), 77–85. doi: 10.1196/annals.1308.009 Giedd, J. N., Blumenthal, J., Jeffries, N. O., Castellanos, F. X., Liu, H., Zijdenbos, A., . . . Rapoport, J. L. (1999). Brain development during childhood and adolescence: A longitudinal MRI study. Nature Neuroscience, 2(10), 861–863. doi: 10.1038/13158 Gilles, E. E., Schultz, L., & Baram, T. Z. (1996). Abnormal corticosterone regulation in an immature rat model of continuous chronic stress. Pediatric Neurology, 15(2), 114–119. doi: 10.1016/0887-8994(96)00153-1 Giriko, C. A., Andreoli, C. A., Mennitti, L. V., Hosoume, L. F., Souto Tdos, S., Silva, A. V., & Mendes-da-Silva, C. (2013). Delayed physical and neurobehavioral development and increased aggressive and depression-like behaviors in the rat offspring of dams fed a high-fat diet. International Journal of Developmental Neuroscience, 31(8), 731–739. doi: 10.1016/j.ijdevneu.2013.09.001 Glavin, G. B. (1984). Prenatal maternal stress: Differential effects upon male and female offspring responses to restraint stress as adults. Pavlovian Journal of Biological Science, 19(3), 157–159. doi: 10.1007/ BF03003589 Godoy, M. A., Souza, A. S., Lobo, M. A., Sampaio, O. V., Moraes, L., Baldanza, M. R., . . . Santiago, M. F. (2013). Effects of protein restriction during gestation and lactation on cell proliferation in the hippocampus and subventricular zone: Functional implications. Protein restriction alters hippocampal/SVZ cell proliferation. Brain Research, 1496, 10–27. doi: 10.1016/j.brainres.2012.10.047 Gogtay, N., Giedd, J. N., Lusk, L., Hayashi, K. M., Greenstein, D., Vaituzis, A. C., . . . Toga, A. W. (2004). Dynamic mapping of human cortical development during childhood through early adulthood. Proceedings of the National Academy of Sciences USA, 101(21), 8174–8179. doi: 10.1073/pnas.0402680101 Gorman, L. K., Shook, B. L., & Becker, D. P. (1993). Traumatic brain injury produces impairments in long-term and recent memory. Brain Research, 614(1–2), 29–36. doi: 10.1016/0006-8993(93)91014-J Graves, J. M., Whitehill, J. M., Stream, J. O., Vavilala, M. S., & Rivara, F. P. (2013). Emergency department-reported head injuries from skiing and snowboarding among children and adolescents, 1996–2010. Injury Prevention, 19(6), 399–404. doi: 10.1136/injuryprev-2012– 040727 Gray, M., Bingham, B., & Viau, V. (2010). A comparison of two repeated restraint stress paradigms on hypothalamic-pituitary-adrenal axis habituation, gonadal status and central neuropeptide expression in adult male rats. Journal of Neuroendocrinology, 22(2), 92–101. doi: 10.1111/j.1365–2826.2009.01941.x Greco, T., Hovda, D., & Prins, M. (2013). The effects of repeat traumatic brain injury on the pituitary in adolescent rats. Journal of Neurotrauma, 30(23), 1983–1990. doi: 10.1089/neu.2013.2990 Greco, T., Hovda, D. A., & Prins, M. L. (2014). Adolescent TBI-induced hypopituitarism causes sexual dysfunction in adult male rats. Developmental Neurobiology. 75(2), 193–202. doi: 10.1002/dneu.22218

Gunnar, M., & Quevedo, K. (2007). The neurobiology of stress and development. Annual Review of Psychology, 58, 145–173. doi: 10.1146/annurev.psych.58.110405.085605 Gunnar, M. R. (1992). Reactivity of the hypothalamic-pituitaryadrenocortical system to stressors in normal infants and children. Pediatrics, 90(3), 491–497. Gunnar, M. R., & Fisher, P. A. (2006). Bringing basic research on early experience and stress neurobiology to bear on preventive interventions for neglected and maltreated children. Developmental Psychopathology, 18(3), 651–677. doi: 10.1017/S0954579406060330 Gunnar, M. R., & Quevedo, K. M. (2008). Early care experiences and HPA axis regulation in children: A mechanism for later trauma vulnerability. Progress in Brain Research, 167, 137–149. doi: 10.1016/S0079-6123(07)67010-1 Gunnar, M. R., Wewerka, S., Frenn, K., Long, J. D., & Griggs, C. (2009). Developmental changes in hypothalamus–pituitary–adrenal activity over the transition to adolescence: Normative changes and associations with puberty. Developmental Psychopathology, 21(1), 69–85. doi: 10.1017/S0954579409000054 Gust, D. A., Gordon, T. P., Wilson, M. E., Ahmed-Ansari, A., Brodie, A. R., & McClure, H. M. (1991). Formation of a new social group of unfamiliar female rhesus monkeys affects the immune and pituitary adrenocortical systems. Brain, Behavior, and Immunity, 5(3), 296–307. doi: 10.1016/0889-1591(91)90024-5 Gutman, D. A., & Nemeroff, C. B. (2002). Neurobiology of early life stress: rodent studies. Seminars in Clinical Neuropsychiatry, 7(2), 89–95. doi: 10.1053/scnp. 2002.31781 Gutman, D. A., & Nemeroff, C. B. (2003). Persistent central nervous system effects of an adverse early environment: Clinical and preclinical studies. Physiology & Behavior, 79(3), 471–478. doi: 10.1016/S0031-9384(03)00166-5 Hackman, D. A., Farah, M. J., & Meaney, M. J. (2010). Socioeconomic status and the brain: mechanistic insights from human and animal research. Nature Reviews Neuroscience, 11(9), 651–659. doi: 10.1038/nrn2897 Harlow, H. F. (1958). The nature of love. American Psychologist, 13(12), 673. doi: 10.1037/h0047884 Harlow, H. F., Harlow, M. K., & Suomi, S. J. (1971). From thought to therapy: Lessons from a primate laboratory. American Scientist, 59, 538–549. Head, L. M. (2014). The effect of kangaroo care on neurodevelopmental outcomes in preterm infants. Journal of Perinatal & Neonatal Nursing, 28(4), 290–299. doi: 10.1097/JPN.0000000000000062 Herman, J. P., Figueiredo, H., Mueller, N. K., Ulrich-Lai, Y., Ostrander, M. M., Choi, D. C., & Cullinan, W. E. (2003). Central mechanisms of stress integration: hierarchical circuitry controlling hypothalamopituitary-adrenocortical responsiveness. Frontiers in Neuroendocrinology, 24(3), 151–180. doi: 10.1016/j.yfrne.2003.07.001 Herman, J. P., Ostrander, M. M., Mueller, N. K., & Figueiredo, H. (2005). Limbic system mechanisms of stress regulation: hypothalamopituitary-adrenocortical axis. Progress in Neuropsychopharmacology & Biological Psychiatry, 29(8), 1201–1213. doi: 10.1016/j.pnpbp. 2005 .08.006 Herrenkohl, L. R., & Whitney, J. B. (1976). Effects of prepartal stress on postpartal nursing behavior, litter development and adult sexual behavior. Physiology & Behavior, 17(6), 1019–1021. doi: 10.1016/0031-9384(76)90025-1 Holekamp, K. E., & Smale, L. (1991). Dominance acquisition during mammalian social development: The “inheritance” of maternal rank. American Zoologist, 31(2), 306–317. doi: 10.1093/icb/31.2.306 Hostinar, C. E., Johnson, A. E., & Gunnar, M. R. (2014). Parent support is less effective in buffering cortisol stress reactivity for adolescents

References compared to children. Developmental Science.18(2), 281–297. doi: 10.1111/desc.12195 Hostinar, C. E., Sullivan, R. M., & Gunnar, M. R. (2014). Psychobiological mechanisms underlying the social buffering of the hypothalamic–pituitary–adrenocortical axis: A review of animal models and human studies across development. Psychological Bulletin, 140(1), 256–282. doi: 10.1037/a0032671 Hougaard, K. S., Mandrup, K. R., Kjaer, S. L., Bogh, I. B., Rosenberg, R., & Wegener, G. (2011). Gestational chronic mild stress: Effects on acoustic startle in male offspring of rats. International Journal of Developmental Neuroscience, 29(4), 495–500. doi: 10.1016/ j.ijdevneu.2011.01.002 Howell, B. R., Godfrey, J., Gutman, D. A., Michopoulos, V., Zhang, X., Nair, G., . . . Sanchez, M. M. (2014). Social subordination stress and serotonin transporter polymorphisms: Associations with brain white matter tract integrity and behavior in juvenile female macaques. Cerebral Cortex, 24(12), 3334–3349. doi: 10.1093/cercor/bht187 Howell, B. R., Grand, A. P., McCormack, K. M., Shi, Y., LaPrarie, J. L., Maestripieri, D., Styner, M. A., & Sanchez, M. M. (2014). Early adverse experience increases emotional reactivity in juvenile rhesus macaques: Relation to amygdala volume. Developmental Psychobiology, 56(8), 1735–1746. doi: 10.1002/dev.21237 Howell, B. R., McCormack, K. M., Grand, A. P., Zhang, X., Maestripieri, D., Hu, D., & Sanchez, M. M. (2013). Brain white matter microstructure alterations in adolescent rhesus monkeys exposed to early life stress: Associations with high cortisol during infancy. Biology of Mood & Anxiety Disorders, 3(21). doi: 10.1186/2045– 5380–3–21 Howerton, C. L., & Bale, T. L. (2012). Prenatal programing: at the intersection of maternal stress and immune activation. Hormones and Behavior, 62(3), 237–242. doi: 10.1016/j.yhbeh.2012.03.007 Huang, G. B., Zhao, T., Muna, S. S., Bagalkot, T. R., Jin, H. M., Chae, H. J., & Chung, Y. C. (2013). Effects of chronic social defeat stress on behaviour, endoplasmic reticulum proteins and choline acetyltransferase in adolescent mice. International Journal of Neuropsychopharmacology, 16(7), 1635–1647. doi: 10.1017/S1461145713000060 Huot, R. L., Gonzalez, M. E., Ladd, C. O., Thrivikraman, K. V., & Plotsky, P. M. (2004). Foster litters prevent hypothalamic-pituitaryadrenal axis sensitization mediated by neonatal maternal separation. Psychoneuroendocrinology, 29(2), 279–289. doi: 10.1016/S0306 -4530(03)00028-3 Huttenlocher, P. R., & Dabholkar, A. S. (1997). Regional differences in synaptogenesis in human cerebral cortex. Journal of Comparative Neurology, 387(2), 167–178. doi: 10.1002/(SICI)1096 -9861(19971020)387:23.0.CO;2-Z Ibi, D., Takuma, K., Koike, H., Mizoguchi, H., Tsuritani, K., Kuwahara, Y., . . . Yamada, K. (2008). Social isolation rearing-induced impairment of the hippocampal neurogenesis is associated with deficits in spatial memory and emotion-related behaviors in juvenile mice. Journal of Neurochemistry, 105(3), 921–932. doi: 10.1111/j.1471– 4159.2007.05207.x Iniguez, S. D., Riggs, L. M., Nieto, S. J., Dayrit, G., Zamora, N. N., Shawhan, K. L., Cruz, B., & Warren, B. L. (2014). Social defeat stress induces a depression-like phenotype in adolescent male c57BL/6 mice. Stress, 17(3), 247–255. doi: 10.3109/10253890.2014.910650 Insel, T. R. (2007). From animal models to model animals. Biological Psychiatry, 62(12), 1337–1339. doi: 10.1016/j.biopsych.2007.10.001 Ives, J. C., Alderman, M., & Stred, S. E. (2007). Hypopituitarism after multiple concussions: A retrospective case study in an adolescent male. Journal of Athletic Training, 42(3), 431–439. Ivy, A. S., Brunson, K. L., Sandman, C., & Baram, T. Z. (2008). Dysfunctional nurturing behavior in rat dams with limited access to nesting

193

material: A clinically relevant model for early-life stress. Neuroscience, 154(3), 1132–1142. doi: 10.1016/j.neuroscience.2008.04.019 Jackowski, A., Perera, T. D., Abdallah, C. G., Garrido, G., Tang, C. Y., Martinez, J., . . . Kaufman, J. (2011). Early-life stress, corpus callosum development, hippocampal volumetrics, and anxious behavior in male nonhuman primates. Psychiatry Research, 192(1), 37–44. doi: 10.1016/j.pscychresns.2010.11.006 Jansen, J., Beijers, R., Riksen-Walraven, M., & de Weerth, C. (2010). Cortisol reactivity in young infants. Psychoneuroendocrinology, 35(3), 329–338. doi: 10.1016/j.psyneuen.2009.07.008 Jarrell, H., Hoffman, J. B., Kaplan, J. R., Berga, S., Kinkead, B., & Wilson, M. E. (2008). Polymorphisms in the serotonin reuptake transporter gene modify the consequences of social status on metabolic health in female rhesus monkeys. Physiology & Behavior, 93(4–5), 807–819. doi: 10.1016/j.physbeh.2007.11.042 Jašarevi´c, E., Rodgers, A. B., & Bale, T. L. (2015). A novel role for maternal stress and microbial transmission in early life programming and neurodevelopment. Neurobiology of Stress, 1, 81–88. doi: 10.1016/j.ynstr.2014.10.005 Jefferies, A. L., & Fetus and Newborn Committee. (2012). Kangaroo care for the preterm infant and family. Paediatrics & Child Health, 17(3), 141. Johnson, E. O., Kamilaris, T. C., Calogero, A. E., Gold, P. W., & Chrousos, G. P. (1996). Effects of early parenting on growth and development in a small primate. Pediatric Research, 39(6), 999–1005. doi: 10.1203/00006450-199606000-00012 Jones, N. C., Cardamone, L., Williams, J. P., Salzberg, M. R., Myers, D., & O’Brien, T. J. (2008). Experimental traumatic brain injury induces a pervasive hyperanxious phenotype in rats. Journal of Neurotrauma, 25(11), 1367–1374. doi: 10.1089/neu.2008.0641 Jovanovic, T., & Maestripieri, D. (2010). Effects of early traumatic experience on vocal expression of emotion in young female rhesus macaques. Developmental Psychobiology, 52(8), 794–801. doi: 10.1002/dev.20475 Kaffman, A., & Meaney, M. J. (2007). Neurodevelopmental sequelae of postnatal maternal care in rodents: clinical and research implications of molecular insights. Journal of Child Psychology and Psychiatry, 48(3–4), 224–244. doi: 10.1111/j.1469-7610.2007.01730.x Kalin, N. H., & Shelton, S. E. (1989). Defensive behaviors in infant rhesus monkeys: environmental cues and neurochemical regulation. Science, 243(4899), 1718–1721. doi: 10.1126/science.2564702 Kalin, N. H., & Shelton, S. E. (2003). Nonhuman primate models to study anxiety, emotion regulation, and psychopathology. Annals of the New York Academy of Sciences, 1008, 189–200. doi: 10.1196/annals.1301.021 Kalin, N. H., Shelton, S. E., & Davidson, R. J. (2004). The role of the central nucleus of the amygdala in mediating fear and anxiety in the primate. Journal of Neuroscience, 24(24), 5506–5515. doi: 10.1523/JNEUROSCI.0292-04.2004 Kaplan, J. R., Adams, M. R., Clarkson, T. B., Manuck, S. B., Shively, C. A., & Williams, J. K. (1996). Psychosocial factors, sex differences, and atherosclerosis: lessons from animal models. Psychosomatic Medicine, 58(6), 598–611. Kaplan, J. R., Chen, H., Appt, S. E., Lees, C. J., Franke, A. A., Berga, S. L., . . . Clarkson, T. B. (2010). Impairment of ovarian function and associated health-related abnormalities are attributable to low social status in premenopausal monkeys and not mitigated by a high-isoflavone soy diet. Human Reproduction, 25(12), 3083–3094. doi: 10.1093/humrep/deq288 Kaufman, J., Plotsky, P. M., Nemeroff, C. B., & Charney, D. S. (2000). Effects of early adverse experiences on brain structure and function: Clinical implications. Biological Psychiatry, 48(8), 778–790. doi: 10.1016/S0006–3223(00)00998–7

194

Animal Models of Developmental Psychopathology

Kay, G., Tarcic, N., Poltyrev, T., & Weinstock, M. (1998). Prenatal stress depresses immune function in rats. Physiology & Behavior, 63(3), 397–402. doi: 10.1016/S0031-9384(97)00456-3 Kendig, M. D., Bowen, M. T., Kemp, A. H., & McGregor, I. S. (2011). Predatory threat induces huddling in adolescent rats and residual changes in early adulthood suggestive of increased resilience. Behavioural Brain Research, 225(2), 405–414. doi: 10.1016/j.bbr .2011.07.058 Kercmar, J., Budefeld, T., Grgurevic, N., Tobet, S. A., & Majdic, G. (2011). Adolescent social isolation changes social recognition in adult mice. Behaviourial Brain Research, 216(2), 647–651. doi: 10.1016/j.bbr .2010.09.007 Kessler, R. C., Avenevoli, S., & Ries Merikangas, K. (2001). Mood disorders in children and adolescents: an epidemiologic perspective. Biological Psychiatry, 49(12), 1002–1014. doi: 10.1016/S0006 -3223(01)01129-5 Kim, H., Lee, M. H., Chang, H. K., Lee, T. H., Lee, H. H., Shin, M. C., . . . Kim, C. J. (2006). Influence of prenatal noise and music on the spatial memory and neurogenesis in the hippocampus of developing rats. Brain & Development, 28(2), 109–114. doi: 10.1016/ j.braindev.2005.05.008 Kim, M. J., Loucks, R. A., Palmer, A. L., Brown, A. C., Solomon, K. M., Marchante, A. N., & Whalen, P. J. (2011). The structural and functional connectivity of the amygdala: from normal emotion to pathological anxiety. Behavioural Brain Research, 223(2), 403–410. doi: 10.1016/j.bbr.2011.04.025 DOI:10.1016/j.bbr.2011.04.025#doilink Kinsley, C., & Svare, B. (1987). Genotype modulates prenatal stress effects on aggression in male and female mice. Behavioral Neural Biology, 47(2), 138–150. doi: 10.1016/S0163-1047(87)90243-3 DOI:10.1016/ S0163-1047(87)90243-3#doilink Kinsley, C. H., & Bridges, R. S. (1988). Prenatal stress and maternal behavior in intact virgin rats: response latencies are decreased in males and increased in females. Hormones and Behavior, 22(1), 76–89. doi: 10.1016/0018-506X(88)90032-3 DOI:10.1016/ 0018-506X(88)90032-3#doilink Kippin, T. E., Szumlinski, K. K., Kapasova, Z., Rezner, B., & See, R. E. (2008). Prenatal stress enhances responsiveness to cocaine. Neuropsychopharmacology, 33(4), 769–782. doi: 10.1038/sj.npp. 1301447 Kjaer, S. L., Wegener, G., Rosenberg, R., & Hougaard, K. S. (2010). Reduced mobility but unaffected startle response in female rats exposed to prenatal dexamethasone: Different sides to a phenotype. Developmental Neuroscience, 32(3), 208–216. doi: 10.1159/000313914 Knudsen, E. I. (2004). Sensitive periods in the development of the brain and behavior. Journal of Cognitive Neuroscience, 16(8), 1412–1425. doi: 10.1162/0898929042304796 DOI:10.1162/ 0898929042304796#blank Koch, H., McCormack, K., Sanchez, M. M., & Maestripieri, D. (2014). The development of the hypothalamic-pituitary-adrenal axis in rhesus monkeys: Effects of age, sex, and early experience. Developmental Psychobiology, 56(1), 86–95. doi: 10.1002/dev.21093 Kundakovic, M., & Champagne, F. A. (2014). Early-life experience, epigenetics, and the developing brain. Neuropsychopharmacology.40, 141–153. doi: 10.1038/npp. 2014.140 Kusek, M., Tokarski, K., & Hess, G. (2013). Repeated restraint stress enhances glutamatergic transmission in the paraventricular nucleus of the rat hypothalamus. Journal of Physiology and Pharmacology, 64(5), 565–570. La Barba, R. C., Martini, J., & White, J. (1969). The effect of maternal-separation on the growth of Ehrlich carcinoma in the Balb/c mouse. Psychosomatic Medicine, 31(2), 129–133. Ladd, C. O., Huot, R. L., Thrivikraman, K. V., Nemeroff, C. B., Meaney, M. J., & Plotsky, P. M. (2000). Long-term behavioral and

neuroendocrine adaptations to adverse early experience. Progress in Brain Research, 122, 81–103. doi: 10.1016/S0079-6123(08)62132-9 DOI:10.1016/S0079-6123(08)62132-9#doilink LaMantia, A. S., & Rakic, P. (1990). Axon overproduction and elimination in the corpus callosum of the developing rhesus monkey. Journal of Neuroscience, 10(7), 2156–2175. Langlois, J. A., Rutland-Brown, W., & Wald, M. M. (2006). The epidemiology and impact of traumatic brain injury: A brief overview. Journal of Head Trauma Rehabilitation, 21(5), 375–378. Laugero, K. D., & Moberg, G. P. (2000). Effects of acute behavioral stress and LPS-induced cytokine release on growth and energetics in mice. Physiology & Behavior, 68(3), 415–422. doi: 10.1016/ S0031-9384(99)00206-1 DOI:10.1016/S0031-9384(99)00206-1#doilink Law, A. J., Pei, Q., Feldon, J., Pryce, C. R., & Harrison, P. J. (2009). Gene expression in the anterior cingulate cortex and amygdala of adolescent marmoset monkeys following parental separations in infancy. International Journal of Neuropsychopharmacology, 12(6), 761–772. doi: 10.1017/S1461145708009723 Law, A. J., Pei, Q., Walker, M., Gordon-Andrews, H., Weickert, C. S., Feldon, J., Pryce, C. R., & Harrison, P. J. (2008). Early parental deprivation in the marmoset monkey produces long-term changes in hippocampal expression of genes involved in synaptic plasticity and implicated in mood disorder. Neuropsychopharmacology, 34(6), 1381–1394. doi: 10.1038/npp. 2008.106 Lebel, C., Walker, L., Leemans, A., Phillips, L., & Beaulieu, C. (2008). Microstructural maturation of the human brain from childhood to adulthood. Neuroimage, 40(3), 1044–1055. doi: 10.1016/j.neuroimage .2007.12.053 DOI:10.1016/j.neuroimage.2007.12.053#doilink Lenroot, R. K., & Giedd, J. N. (2006). Brain development in children and adolescents: insights from anatomical magnetic resonance imaging. Neuroscience and Biobehavioral Reviews, 30(6), 718–729. doi: 10.1016/j .neubiorev.2006.06.001 DOI:10.1016/j.neubiorev.2006.06.001#doilink Lepsch, L. B., Gonzalo, L. A., Magro, F. J., Delucia, R., Scavone, C., & Planeta, C. S. (2005). Exposure to chronic stress increases the locomotor response to cocaine and the basal levels of corticosterone in adolescent rats. Addiction Biology, 10(3), 251–256. doi: 10.1080/13556210500269366 Leuner, B., Mendolia-Loffredo, S., Kozorovitskiy, Y., Samburg, D., Gould, E., & Shors, T. J. (2004). Learning enhances the survival of new neurons beyond the time when the hippocampus is required for memory. The Journal of Neuroscience, 24(34), 7477–7481. doi: 10.1523/JNEUROSCI.0204–04.2004 Leussis, M. P., & Andersen, S. L. (2008). Is adolescence a sensitive period for depression? Behavioral and neuroanatomical findings from a social stress model. Synapse, 62(1), 22–30. doi: 10.1002/syn.20462 Leussis, M. P., Lawson, K., Stone, K., & Andersen, S. L. (2008). The enduring effects of an adolescent social stressor on synaptic density, part II: Poststress reversal of synaptic loss in the cortex by adinazolam and MK-801. Synapse, 62(3), 185–192. doi: 10.1002/syn.20483 Li, R., & Shen, Y. (2005). Estrogen and brain: synthesis, function and diseases. Frontiers in Bioscience, 10, 257–267. doi: 10.2741/1525 Lin, Y. L., & Wang, S. (2014). Prenatal lipopolysaccharide exposure increases depression-like behaviors and reduces hippocampal neurogenesis in adult rats. Behaviourial Brain Research, 259, 24–34. doi: 10.1016/j.bbr.2013.10.034 Loman, M. M., & Gunnar, M. R. (2010). Early experience and the development of stress reactivity and regulation in children. Neuroscience & Biobehavioral Reviews, 34(6), 867–876. doi: 10.1016/j.neubiorev .2009.05.007 DOI:10.1016/j.neubiorev.2009.05.007#doilink Losiniecki, A., & Shutter, L. (2010). Management of traumatic brain injury. Current Treatment Options Neurology, 12(2), 142–154. doi: 10.1007/s11940–010–0063-z

References Louvart, H., Maccari, S., & Darnaudery, M. (2005). Prenatal stress affects behavioral reactivity to an intense stress in adult female rats. Brain Research, 1031(1), 67–73. doi: 10.1016/j.brainres.2004.10.025 Louvart, H., Maccari, S., Lesage, J., Leonhardt, M., Dickes-Coopman, A., & Darnaudery, M. (2006). Effects of a single footshock followed by situational reminders on HPA axis and behaviour in the aversive context in male and female rats. Psychoneuroendocrinology, 31(1), 92–99. doi: 10.1016/j.psyneuen.2005.05.014 Louvart, H., Maccari, S., Vaiva, G., & Darnaudery, M. (2009). Prenatal stress exacerbates the impact of an aversive procedure on the corticosterone response to stress in female rats. Psychoneuroendocrinology, 34(5), 786–790. doi: 10.1016/j.psyneuen.2008.12.002 Lovic, V., Belay, H., Walker, C. D., Burton, C. L., Meaney, M. J., Sokolowski, M., & Fleming, A. S. (2013). Early postnatal experience and DRD2 genotype affect dopamine receptor expression in the rat ventral striatum. Behaviourial Brain Research, 237, 278–282. doi: 10.1016/j.bbr.2012.09.046 Lui, P., Padow, V. A., Franco, D., Hall, B. S., Park, B., Klein, Z. A., & Romeo, R. D. (2012). Divergent stress-induced neuroendocrine and behavioral responses prior to puberty. Physiology & Behavior, 107(1), 104–111. doi: 10.1016/j.physbeh.2012.06.011 Lyons, D. M., Afarian, H., Schatzberg, A. F., Sawyer-Glover, A., & Moseley, M. E. (2002). Experience-dependent asymmetric variation in primate prefrontal morphology. Behavioural Brain Research, 136(1), 51–59. roi: 10.1016/S0166-4328(02)00100-6 DOI:10.1016/ S0166-4328(02)00100-6#doilink Lyons, D. M., & Parker, K. J. (2007). Stress inoculation-induced indications of resilience in monkeys. Journal of Traumatic Stress, 20(4), 423–433. doi: 10.1002/jts.20265

195

Maestripieri, D. (2005). Early experience affects the intergenerational transmission of infant abuse in rhesus monkeys. Proceedings of the National Academy of Sciences USA, 102(27), 9726–9729. doi: 10.1073/pnas.0504122102 Maestripieri, D., & Carroll, K. A. (1998). Risk factors for infant abuse and neglect in group-living rhesus monkeys. Psychological Sciences, 9, 143–145. http://www.jstor.org/stable/40063266 Maestripieri, D., & Carroll, K. A. (2000). Causes and consequences of infant abuse and neglect in monkeys. Aggression and Violent Behavior, 5(3), 245–254. doi: 10.1016/S1359-1789(98)00019-6 DOI:10.1016/S1359-1789(98)00019-6#doilink Maestripieri, D., Higley, J. D., Lindell, S. G., Newman, T. K., McCormack, K. M., & Sanchez, M. M. (2006). Early maternal rejection affects the development of monoaminergic systems and adult abusive parenting in rhesus macaques (Macaca mulatta). Behavioral Neuroscience, 120(5), 1017–1024. doi: 10.1037/0735-7044.120.5.1017 Maestripieri, D., Jovanovic, T., & Gouzoules, H. (2000). Crying and infant abuse in rhesus monkeys. Child Development, 71(2), 301–309. doi: 10.1111/1467-8624.00145 Maestripieri, D., McCormack, K., Lindell, S. G., Higley, J. D., & Sanchez, M. M. (2006). Influence of parenting style on the offspring’s behaviour and CSF monoamine metabolite levels in crossfostered and noncrossfostered female rhesus macaques. Behaviourial Brain Research, 175(1), 90–95. doi: 10.1016/j.bbr.2006.08.002 DOI:10.1016/j.bbr.2006.08.002#doilink Maestripieri, D., Megna, N. L., & Jovanovic, T. (2000). Adoption and maltreatment of foster infants by rhesus macaque abusive mothers. Developmental Science, 3(3), 287–293. doi: 10.1111/1467–7687.00122

Lyons, D. M., Parker, K. J., & Schatzberg, A. F. (2010). Animal models of early life stress: implications for understanding resilience. Developmental Psychobiology, 52(7), 616–624. doi: 10.1002/dev.20500

Maestripieri, D., Wallen, K., & Carroll, K. A. (1997). Genealogical and demographic influences on infant abuse and neglect in group-living sooty mangabeys (Cercocebus atys). Developmental Psychobiology, 31(3), 175–180.

Lyons, D. M., & Schatzberg, A. F. (2003). Early maternal availability and prefrontal correlates of reward-related memory. Neurobiology of Learning and Memory, 80(2), 97–104. doi: 10.1016/S1074–7427 (03)00044–3

Malkova, L., Heuer, E., & Saunders, R. C. (2006). Longitudinal magnetic resonance imaging study of rhesus monkey brain development. European Journal of Neuroscience, 24(11), 3204–3212. doi: 10.1111/j.1460-9568.2006.05175.x

Machado, C. J., & Bachevalier, J. (2003). Non-human primate models of childhood psychopathology: The promise and the limitations. Journal of Child Psychology and Psychiatry, 44(1), 64–87. doi: 10.1111/1469 -7610.00103

Marco, E. M., Macri, S., & Laviola, G. (2011). Critical age windows for neurodevelopmental psychiatric disorders: Evidence from animal models. Neurotoxicity Research, 19(2), 286–307. doi: 10.1007/s12640–010–9205-z

Machado, C. J., & Bachevalier, J. (2008). Behavioral and hormonal reactivity to threat: effects of selective amygdala, hippocampal or orbital frontal lesions in monkeys. Psychoneuroendocrinology, 33(7), 926. doi: 10.1016/j.psyneuen.2008.04.012 DOI:10.1016/j.psyneuen .2008.04.012#doilink

Marrocco, J., Mairesse, J., Ngomba, R. T., Silletti, V., Van Camp, G., Bouwalerh, H., . . . Morley-Fletcher, S. (2012). Anxiety-like behavior of prenatally stressed rats is associated with a selective reduction of glutamate release in the ventral hippocampus. Journal of Neuroscience, 32(48), 17143–17154. doi: 10.1523/JNEUROSCI.1040–12.2012

Machado, C. J., Emery, N. J., Capitanio, J. P., Mason, W. A., Mendoza, S. P., & Amaral, D. G. (2008). Bilateral neurotoxic amygdala lesions in rhesus monkeys (Macaca mulatta): Consistent pattern of behavior across different social contexts. Behavioral Neuroscience, 122(2), 251. doi: 10.1037/0735-7044.122.2.251

Marty, M. S., Chapin, R. E., Parks, L. G., & Thorsrud, B. A. (2003). Development and maturation of the male reproductive system. Birth Defects Research Part B Developmental and Reproductive Toxicology, 68(2), 125–136. doi: 10.1002/bdrb.10015

Machado, C. J., Kazama, A. M., & Bachevalier, J. (2009). Impact of amygdala, orbital frontal, or hippocampal lesions on threat avoidance and emotional reactivity in nonhuman primates. Emotion, 9(2), 147–163. doi: 10.1037/a0014539 Maeda, M., Kato, H., & Maruoka, T. (2009). Adolescent vulnerability to PTSD and effects of community-based intervention: Longitudinal study among adolescent survivors of the Ehime Maru sea accident. Psychiatry and Clinical Neurosciences, 63(6), 747–753. doi: 10.1111/j.1440–1819.2009.02031.x Maestripieri, D. (1998). Parenting styles of abusive mothers in groupliving rhesus macaques. Animal Behaviour, 55(1), 1–11. doi: 10.1006/ anbe.1997.0578 DOI:10.1006/anbe.1997.0578#doilink

Maslova, L. N., Bulygina, V. V., & Popova, N. K. (2002). Immediate and long-lasting effects of chronic stress in the prepubertal age on the startle reflex. Physiology & Behavior, 75(1–2), 217–225. doi: 10.1016/S0031-9384(01)00648-5 DOI:10.1016/S0031-9384(01) 00648-5#doilink Mathew, S. J., Shungu, D. C., Mao, X., Smith, E. L., Perera, G. M., Kegeles, L. S., . . . Coplan, J. D. (2003). A magnetic resonance spectroscopic imaging study of adult nonhuman primates exposed to early-life stressors. Biological Psychiatry, 54(7), 727–735. doi: 10.1016/S0006-3223(03)00004-0 DOI:10.1016/S0006-3223(03)00004 -0#doilink McCormack, K., Newman, T. K., Higley, J. D., Maestripieri, D., & Sanchez, M. M. (2009). Serotonin transporter gene variation, infant

196

Animal Models of Developmental Psychopathology

abuse, and responsiveness to stress in rhesus macaque mothers and infants. Hormones and Behavior, 55(4), 538–547. doi: 10.1016/ j.yhbeh.2009.01.009 DOI:10.1016/j.yhbeh.2009.01.009#doilink McCormack, K., Sanchez, M. M., Bardi, M., & Maestripieri, D. (2006). Maternal care patterns and behavioral development of rhesus macaque abused infants in the first 6 months of life. Developmental Psychobiology, 48(7), 537–550. doi: 10.1002/dev.20157 McCormack, K. M., Grand, A., LaPrairie, J., Fulks, R., Graff, A., Maestripieri, D., Plotsky, P. M., & Sanchez, M. M. (2003). Behavioral and neuroendocrine outcomes of infant maltreatment in rhesus monkeys: The first four months. Paper presented at the Society for Neuroscience, Washington, DC. McCormick, C. M., & Ibrahim, F. N. (2007). Locomotor activity to nicotine and Fos immunoreactivity in the paraventricular nucleus of the hypothalamus in adolescent socially-stressed rats. Pharmacology Biochemistry and Behavior, 86(1), 92–102. doi: 10.1016/j.pbb.2006 .12.012 McCormick, C. M., & Mathews, I. Z. (2007). HPA function in adolescence: Role of sex hormones in its regulation and the enduring consequences of exposure to stressors. Pharmacology Biochemistry and Behavior, 86(2), 220–233. doi: 10.1016/j.pbb.2006.07.012 DOI:10.1016/j.pbb.2006.07.012#doilink McCormick, C. M., Merrick, A., Secen, J., & Helmreich, D. L. (2007). Social instability in adolescence alters the central and peripheral hypothalamic-pituitary-adrenal responses to a repeated homotypic stressor in male and female rats. Journal of Neuroendocrinology, 19(2), 116–126. doi: 10.1111/j.1365-2826.2006.01515.x McCormick, C. M., Nixon, F., Thomas, C., Lowie, B., & Dyck, J. (2010). Hippocampal cell proliferation and spatial memory performance after social instability stress in adolescence in female rats. Behavioural Brain Research, 208(1), 23–29. doi: 10.1016/j.bbr.2009.11.003 McCormick, C. M., Robarts, D., Gleason, E., & Kelsey, J. E. (2004). Stress during adolescence enhances locomotor sensitization to nicotine in adulthood in female, but not male, rats. Hormones and Behavior, 46(4), 458–466. doi: 10.1016/j.yhbeh.2004.05.004 DOI:10.1016/j.yhbeh.2004.05.004#doilink McCormick, C. M., Smith, C., & Mathews, I. Z. (2008). Effects of chronic social stress in adolescence on anxiety and neuroendocrine response to mild stress in male and female rats. Behavioural Brain Research, 187(2), 228–238. doi: 10.1016/j.bbr.2007.09.005 DOI:10.1016/j.bbr.2007.09.005#doilink McCormick, C. M., Thomas, C. M., Sheridan, C. S., Nixon, F., Flynn, J. A., & Mathews, I. Z. (2012). Social instability stress in adolescent male rats alters hippocampal neurogenesis and produces deficits in spatial location memory in adulthood. Hippocampus, 22(6), 1300–1312. doi: 10.1002/hipo.20966 McCrory, P. R., & Berkovic, S. F. (1998). Second impact syndrome. Neurology, 50(3), 677–683. doi: 10.1212/WNL.50.3.677 McEwen, B. S. (1998). Stress, adaptation, and disease. Allostasis and allostatic load. Annals of the New York Academy of Sciences, 840, 33–44. doi: 10.1111/j.1749-6632.1998.tb09546.x McEwen, B. S. (2000). The neurobiology of stress: from serendipity to clinical relevance. Brain Research, 886(1–2), 172–189. doi: 10.1016/S0006 -8993(00)02950-4 DOI:10.1016/S0006-8993(00)02950-4#doilink McEwen, B. S. (2001). Stress, sex, hippocampal plasticity: relevance to psychiatric disorders. Clinical Neuroscience Research, 1(2001), 19–34. doi: 10.1016/S1566-2772(00)00004-9 DOI:10.1016/S1566-2772 (00)00004-9#doilink McEwen, B. S. (2008). Central effects of stress hormones in health and disease: Understanding the protective and damaging effects of stress and stress mediators. European Journal of Pharmacology, 583(2–3), 174–185. doi: 10.1016/j.ejphar.2007.11.071

McEwen, B. S., & Seeman, T. (1999). Protective and damaging effects of mediators of stress. Elaborating and testing the concepts of allostasis and allostatic load. Annals of the New York Academy of Sciences, 896, 30–47. doi: 10.1111/j.1749-6632.1999.tb08103.x McGowan, P. O., Sasaki, A., D’Alessio, A. C., Dymov, S., Labonté, B., Szyf, M., Turecki, G., & Meaney, M. J. (2009). Epigenetic regulation of the glucocorticoid receptor in human brain associates with childhood abuse. Nature Neuroscience, 12(3), 342–348. doi: 10.1038/nn.2270 Meyer, U. (2014). Prenatal poly(i:C) exposure and other developmental immune activation models in rodent systems. Biological Psychiatry, 75(4), 307–315. doi: 10.1016/j.biopsych.2013.07.011 Michopoulos, V., Berga, S. L., Kaplan, J. R., & Wilson, M. E. (2009). Social subordination and polymorphisms in the gene encoding the serotonin transporter enhance estradiol inhibition of luteinizing hormone secretion in female rhesus monkeys. Biology of Reproduction, 81(6), 1154–1163. doi: 10.1095/biolreprod.109.079038 Michopoulos, V., Higgins, M., Toufexis, D., & Wilson, M. E. (2012). Social subordination produces distinct stress-related phenotypes in female rhesus monkeys. Psychoneuroendocrinology, 37(7), 1071–1085. doi: 10.1016/j.psyneuen.2011.12.004 Missault, S., Van den Eynde, K., Vanden Berghe, W., Fransen, E., Weeren, A., Timmermans, J. P., Kumar-Singh, S., & Dedeurwaerdere, S. (2014). The risk for behavioural deficits is determined by the maternal immune response to prenatal immune challenge in a neurodevelopmental model. Brain, Behavior, and Immunity. 42, 138–146. doi: 10.1016/j.bbi.2014.06.013 Mitka, M. (2010). Reports of concussions from youth sports rise along with awareness of the problem. Journal of the American Medical Association, 304(16), 1775–1776. doi: 10.1001/jama.2010.1487 Molet, J., Maras, P. M., Avishai-Eliner, S., & Baram, T. Z. (2014). Naturalistic rodent models of chronic early-life stress. Developmental Psychobiology, 56(8), 1675–1688. doi: 10.1002/dev.21230 Morgan, D., Grant, K. A., Gage, H. D., Mach, R. H., Kaplan, J. R., Prioleau, O., . . . Nader, M. A. (2002). Social dominance in monkeys: Dopamine D2 receptors and cocaine self-administration. Nature Neuroscience, 5(2), 169–174. doi: 10.1038/nn798 Moriceau, S., Roth, T. L., Okotoghaide, T., & Sullivan, R. M. (2004). Corticosterone controls the developmental emergence of fear and amygdala function to predator odors in infant rat pups. International Journal of Developmental Neuroscience, 22(5–6), 415–422. doi: 10.1016/j.ijdevneu.2004.05.011 Moriceau, S., & Sullivan, R. M. (2004). Corticosterone influences on Mammalian neonatal sensitive-period learning. Behavioral Neuroscience, 118(2), 274. doi: 10.1037/0735-7044.118.2.274 Moriceau, S., & Sullivan, R. M. (2006). Maternal presence serves as a switch between learning fear and attraction in infancy. Nature Neuroscience, 9(8), 1004–1006. doi: 10.1038/nn1733 Morley-Fletcher, S., Rea, M., Maccari, S., & Laviola, G. (2003). Environmental enrichment during adolescence reverses the effects of prenatal stress on play behaviour and HPA axis reactivity in rats. European Journal of Neuroscience, 18(12), 3367–3374. doi: 10.1111/j.1460-9568.2003.03070.x Morrissey, M. D., Mathews, I. Z., & McCormick, C. M. (2011). Enduring deficits in contextual and auditory fear conditioning after adolescent, not adult, social instability stress in male rats. Neurobiology of Learning and Memory, 95(1), 46–56. doi: 10.1016/j.nlm.2010.10.007 Mueller, B. R., & Bale, T. L. (2006). Impact of prenatal stress on long term body weight is dependent on timing and maternal sensitivity. Physiology & Behavior, 88(4–5), 605–614. doi: 10.1016/j.physbeh.2006 .05.019 Muhammad, A., & Kolb, B. (2011). Mild prenatal stress-modulated behavior and neuronal spine density without affecting amphetamine

References sensitization. Developmental Neuroscience, 33(2), 85–98. doi: 10.1159/000324744 Naef, L., Gratton, A., & Walker, C. D. (2013). Exposure to high fat during early development impairs adaptations in dopamine and neuroendocrine responses to repeated stress. Stress, 16(5), 540–548. doi: 10.3109/10253890.2013.805321 Negron-Oyarzo, I., Perez, M. A., Terreros, G., Munoz, P., & Dagnino-Subiabre, A. (2014). Effects of chronic stress in adolescence on learned fear, anxiety, and synaptic transmission in the rat prelimbic cortex. Behavioural Brain Research, 259, 342–353. doi: 10.1016/j.bbr.2013.11.001 Neigh, G. N., Bowers, S. L., Pyter, L. M., Gatien, M. L., & Nelson, R. J. (2004). Pyruvate prevents restraint-induced immunosuppression via alterations in glucocorticoid responses. Endocrinology, 145(9), 4309–4319. doi: 10.1210/en.2003–1748 Neigh, G. N., Gillespie, C. F., & Nemeroff, C. B. (2009). The neurobiological toll of child abuse and neglect. Trauma Violence Abuse, 10(4), 389–410. doi: 10.1177/1524838009339758 Neigh, G. N., & Nemeroff, C. B. (2006). Reduced glucocorticoid receptors: Consequence or cause of depression? Trends in Endocrinology & Metabolism, 17(4), 124–125. doi: 10.1016/j.tem.2006.03.002 DOI:10.1016/j.tem.2006.03.002#doilink Nelson, E. E., & Winslow, J. T. (2008). Non-human primates: model animals for developmental psychopathology. Neuropsychopharmacology, 34(1), 90–105. doi: 10.1038/npp. 2008.150 Newman, T. K., Syagailo, Y. V., Barr, C. S., Wendland, J. R., Champoux, M., Graessle, M., . . . Lesch, K. (2005). Monoamine oxidase A gene promoter variation and rearing experience influences aggressive behavior in rhesus monkeys. Biological Psychiatry, 57(2), 167–172. doi: 10.1016/j.biopsych.2004.10.012 DOI:10.1016/j.biopsych.2004.10.012 #doilink Nishio, H., Kasuga, S., Ushijima, M., & Harada, Y. (2001). Prenatal stress and postnatal development of neonatal rats—Sex-dependent effects on emotional behavior and learning ability of neonatal rats. International Journal of Developmental Neuroscience, 19(1), 37–45. doi: 10.1016/S0736-5748(00)00070-8 DOI:10.1016/S0736 -5748(00)00070-8#doilink Nishio, H., Tokumo, K., & Hirai, T. (2006). Effects of perinatal stress on the anxiety-related behavior of the adolescence mouse. International Journal of Developmental Neuroscience, 24(4), 263–268. doi: 10.1016/j.ijdevneu.2006.02.003 Norman, R. L., & Spies, H. G. (1981). Brain lesions in infant female rhesus monkeys: Effects on menarche and first ovulation and on diurnal rhythms of prolactin and cortisol. Endocrinology, 108(5), 1723–1729. doi: 10.1210/endo-108-5-1723 Novick, A. M., Forster, G. L., Tejani-Butt, S. M., & Watt, M. J. (2011). Adolescent social defeat alters markers of adult dopaminergic function. Brain Research Bulletin, 86(1–2), 123–128. doi: 10.1016/j.brainresbull.2011.06.009 Novick, A. M., Miiller, L. C., Forster, G. L., & Watt, M. J. (2013). Adolescent social defeat decreases spatial working memory performance in adulthood. Behavioral and Brain Functions, 9, 39. doi: 10.1186/ 1744–9081–9–39 O’Connor, T. G., & Cameron, J. L. (2006). Translating research findings on early experience to prevention: animal and human evidence on early attachment relationships. American Journal of Preventive Medicine, 31(6), 175–181. doi: 10.1016/j.amepre.2006.07.005 DOI:10.1016/ j.amepre.2006.07.005#doilink Öngür, D., & Price, J. L. (2000). The organization of networks within the orbital and medial prefrontal cortex of rats, monkeys and humans. Cerebral Cortex, 10(3), 206–219. doi: 10.1093/cercor/10.3.206 Paiardini, M., Hoffman, J., Cervasi, B., Ortiz, A. M., Stroud, F., Silvestri, G., & Wilson, M. E. (2009). T-cell phenotypic and

197

functional changes associated with social subordination and gene polymorphisms in the serotonin reuptake transporter in female rhesus monkeys. Brain, Behavior, and Immunity, 23(2), 286–293. doi: 10.1016/j.bbi.2008.10.006 DOI:10.1016/j.bbi.2008.10.006#doilink Pandey, D. K., Yadav, S. K., Mahesh, R., & Rajkumar, R. (2009). Depression-like and anxiety-like behavioural aftermaths of impact accelerated traumatic brain injury in rats: A model of comorbid depression and anxiety? Behavioural Brain Research, 205(2), 436–442. doi: 10.1016/j.bbr.2009.07.027 Parker, K. J., Buckmaster, C. L., Justus, K. R., Schatzberg, A. F., & Lyons, D. M. (2005). Mild early life stress enhances prefrontal-dependent response inhibition in monkeys. Biological Psychiatry, 57(8), 848–855. doi: 10.1016/j.biopsych.2004.12.024 Parker, K. J., Buckmaster, C. L., Lindley, S. E., Schatzberg, A. F., & Lyons, D. M. (2012). Hypothalamic-pituitary-adrenal axis physiology and cognitive control of behavior in stress inoculated monkeys. International Journal of Behavioral Development, 36(1), 45–52. doi: 10.1177/0165025411406864 Parker, K. J., Buckmaster, C. L., Schatzberg, A. F., & Lyons, D. M. (2004). Prospective investigation of stress inoculation in young monkeys. Archives of General Psychiatry, 61(9), 933–941. doi: 10.1001/archpsyc.61.9.933 Parker, K. J., Buckmaster, C. L., Sundlass, K., Schatzberg, A. F., & Lyons, D. M. (2006). Maternal mediation, stress inoculation, and the development of neuroendocrine stress resistance in primates. Proceedings of the National Academy of Sciences USA, 103(8), 3000–3005. doi: 10.1073/pnas.0506571103 Parker, K. J., & Maestripieri, D. (2011). Identifying key features of early stressful experiences that produce stress vulnerability and resilience in primates. Neuroscience & Biobehavioral Reviews, 35(7), 1466–1483. doi: 10.1016/j.neubiorev.2010.09.003 DOI:10.1016/ j.neubiorev.2010.09.003#doilink Patel, P. D., Katz, M., Karssen, A. M., & Lyons, D. M. (2008). Stressinduced changes in corticosteroid receptor expression in primate hippocampus and prefrontal cortex. Psychoneuroendocrinology, 33(3), 360–367. doi: 10.1016/j.psyneuen.2007.12.003 DOI:10.1016/ j.psyneuen.2007.12.003#doilink Paternain, L., de la Garza, A. L., Batlle, M. A., Milagro, F. I., Martinez, J. A., & Campion, J. (2013). Prenatal stress increases the obesogenic effects of a high-fat-sucrose diet in adult rats in a sex-specific manner. Stress, 16(2), 220–232. doi: 10.3109/10253890.2012.707708 Payne, C., Machado, C. J., Bliwise, N. G., & Bachevalier, J. (2010). Maturation of the hippocampal formation and amygdala in Macaca mulatta: a volumetric magnetic resonance imaging study. Hippocampus, 20(8), 922–935. doi: 10.1002/hipo.20688 Pine, D. S., Cohen, P., Gurley, D., Brook, J., & Ma, Y. (1998). The risk for early-adulthood anxiety and depressive disorders in adolescents with anxiety and depressive disorders. Archives of General Psychiatry, 55(1), 56–64. doi: 10.1001/archpsyc.55.1.56 Plotsky, P. M., & Meaney, M. J. (1993). Early, postnatal experience alters hypothalamic corticotropin-releasing factor (CRF) mRNA, median eminence CRF content and stress-induced release in adult rats. Molecular Brain Research, 18(3), 195–200. doi: 10.1016/ 0169-328X(93)90189-V DOI:10.1016/0169-328X(93)90189-V#doilink Preuss, T. M. (1995). Do rats have prefrontal cortex? The Rose-WoolseyAkert program reconsidered. Journal of Cognitive Neuroscience, 7(1), 1–24. doi: 10.1162/jocn.1995.7.1.1 DOI:10.1162/jocn.1995.7 .1.1#blank Preuss, T. M., & Goldman-Rakic, P. S. (1991a). Ipsilateral cortical connections of granular frontal cortex in the strepsirhine primate Galago, with comparative comments on anthropoid primates. Journal of Comparative Neurology, 310(4), 507–549. doi: 10.1002/cne .903100404

198

Animal Models of Developmental Psychopathology

Preuss, T. M., & Goldman-Rakic, P. S. (1991b). Myelo- and cytoarchitecture of the granular frontal cortex and surrounding regions in the strepsirhine primate Galago and the anthropoid primate Macaca. Journal of Comparative Neurology, 310(4), 429–474. doi: 10.1002/cne.903100402 Price, D. A., Close, G. C., & Fielding, B. A. (1983). Age of appearance of circadian rhythm in salivary cortisol values in infancy. Archives of Disease in Childhood, 58(6), 454–456. doi: 10.1136/adc.58.6.454 Provençal, N., Suderman, M. J., Guillemin, C., Massart, R., Ruggiero, A., Wang, D., . . . Côté, S. M. (2012). The signature of maternal rearing in the methylome in rhesus macaque prefrontal cortex and T cells. Journal of Neuroscience, 32(44), 15626–15642. doi: 10.1523/JNEUROSCI.1470-12.2012 Pryce, C. R. (2008). Postnatal ontogeny of expression of the corticosteroid receptor genes in mammalian brains: Inter-species and intraspecies differences. Brain Research Reviews, 57(2), 596–605. doi: 10.1016/j.brainresrev.2007.08.005 DOI:10.1016/j.brainresrev.2007.08 .005#doilink Pryce, C. R., Dettling, A., Spengler, M., Spaete, C., & Feldon, J. (2004). Evidence for altered monoamine activity and emotional and cognitive disturbance in marmoset monkeys exposed to early life stress. Annals of the New York Academy of Sciences, 1032(1), 245–249. doi: 10.1196/annals.1314.030 Pryce, C. R., Dettling, A. C., Spengler, M., Schnell, C. R., & Feldon, J. (2004). Deprivation of parenting disrupts development of homeostatic and reward systems in marmoset monkey offspring. Biological Psychiatry, 56(2), 72–79. doi: 10.1016/j.biopsych.2004.05.002 DOI:10.1016/j.biopsych.2004.05.002#doilink Pryce, C. R., Rüedi-Bettschen, D., Dettling, A. C., Weston, A., Russig, H., Ferger, B., & Feldon, J. (2005). Long-term effects of early-life environmental manipulations in rodents and primates: Potential animal models in depression research. Neuroscience & Biobehavioral Reviews, 29(4), 649–674. doi: 10.1016/j.neubiorev.2005.03.011 DOI:10.1016/j.neubiorev.2005.03.011#doilink Pyter, L. M., Kelly, S. D., Harrell, C. S., & Neigh, G. N. (2013). Sex differences in the effects of adolescent stress on adult brain inflammatory markers in rats. Brain, Behavior, and Immunity, 30, 88–94. doi: 10.1016/j.bbi.2013.01.075 Raper, J., Bachevalier, J., Wallen, K., & Sanchez, M. (2013). Neonatal amygdala lesions alter basal cortisol levels in infant rhesus monkeys. Psychoneuroendocrinology, 38(6), 818–829. doi: 10.1016/ j.psyneuen.2012.09.004 DOI:10.1016/j.psyneuen.2012.09.004#doilink Raper, J., Stephens, S. B. Z., Henry, A., Villarreal, T., Bachevalier, J., Wallen, K., & Sanchez, M. M. (2014). Neonatal amygdala lesions lead to increased activity of brain CRF systems and hypothalamicpituitary-adrenal axis of juvenile rhesus monkeys. Journal of Neuroscience, 34(34), 11452–11460. doi: 10.1523/JNEUROSCI.0269 -14.2014 Raper, J., Wallen, K., Sanchez, M. M., Stephens, S. B. Z., Henry, A., Villareal, T., & Bachevalier, J. (2013). Sex-dependent role of the amygdala in the development of emotional and neuroendocrine reactivity to threatening stimuli in infant and juvenile rhesus monkeys. Hormones and Behavior. 63(4), 646–658. doi: 10.1016/j.yhbeh.2013.01.010 Raper, J., Wilson, M., Sanchez, M., Machado, C. J., & Bachevalier, J. (2013). Pervasive alterations of emotional and neuroendocrine responses to an acute stressor after neonatal amygdala lesions in rhesus monkeys. Psychoneuroendocrinology, 38(7), 1021–1035. doi: 10.1016/j.psyneuen.2012.10.008 Reep, R. (1984). Relationship between prefrontal and limbic cortex: a comparative anatomical review. Brain Behavior and Evolution, 25(1), 65–80. doi: 10.1159/000118852 Reiss, A. L., Abrams, M. T., Singer, H. S., Ross, J. L., & Denckla, M. B. (1996). Brain development, gender and IQ in children: A

volumetric imaging study. Brain, 119(5), 1763–1774. doi: 10.1093/ brain/119.5.1763 Reynolds, R. M., Hii, H. L., Pennell, C. E., McKeague, I. W., Kloet, E., Lye, S., . . . Foster, J. K. (2013). Analysis of baseline hypothalamic-pituitary-adrenal activity in late adolescence reveals gender specific sensitivity of the stress axis. Psychoneuroendocrinology, 38(8), 1271–1280. doi: 10.1016/j.psyneuen.2012.11.010 DOI:10.1016 /j.psyneuen.2012.11.010#doilink Rice, C. J., Sandman, C. A., Lenjavi, M. R., & Baram, T. Z. (2008). A novel mouse model for acute and long-lasting consequences of early life stress. Endocrinology, 149(10), 4892–4900. doi: 10.1210/en.2008–0633 Rice, D., & Barone Jr, S. (2000). Critical periods of vulnerability for the developing nervous system: evidence from humans and animal models. Environmental Health Perspectives, 108(Suppl 3), 511–533. Richards, S., Mychasiuk, R., Kolb, B., & Gibb, R. (2012). Tactile stimulation during development alters behaviour and neuroanatomical organization of normal rats. Behavioural Brain Research, 231(1), 86–91. doi: 10.1016/j.bbr.2012.02.043 doi:10.1016/j.bbr .2012.02.043 Rilling, J. K., Winslow, J. T., O’Brien, D., Gutman, D. A., Hoffman, J. M., & Kilts, C. D. (2001). Neural correlates of maternal separation in rhesus monkeys. Biological Psychiatry, 49(2), 146–157. doi: 10.1016/S0006-3223(00)00977-X doi:10.1016/S0006-3223 (00)00977-X Rincon-Cortes, M., & Sullivan, R. M. (2014). Early life trauma and attachment: Immediate and enduring effects on neurobehavioral and stress axis development. Frontiers in Endocrinology, 5, 33. doi: 10.3389/fendo.2014.00033 Rodgers, K. M., Deming, Y. K., Bercum, F. M., Chumachenko, S. Y., Wieseler, J. L., Johnson, K. W., Watkins, L. R., & Barth, D. S. (2014). Reversal of established traumatic brain injury-induced, anxiety-like behavior in rats after delayed, post-injury neuroimmune suppression. Journal of Neurotrauma, 31(5), 487–497. doi: 10.1089/neu.2013.3090 Rodriguez-Arias, M., Navarrete, F., Blanco-Gandia, M. C., Arenas, M. C., Bartoll-Andres, A., Aguilar, M. A., . . . Manzanares, J. (2014). Social defeat in adolescent mice increases vulnerability to alcohol consumption. Addiction Biology. doi: 10.1111/adb.12184 Rojo, M., Marin, B., & Menendez-Patterson, A. (1985). Effects of low stress during pregnancy on certain parameters of the offspring. Physiology & Behavior, 34(6), 895–899. doi: 10.1016/0031-9384(85)90010-1 doi:10.1016/0031-9384(85)90010-1 Rosenblum, L. A., Coplan, J. D., Friedman, S., Bassoff, T., Gorman, J. M., & Andrews, M. W. (1994). Adverse early experiences affect noradrenergic and serotonergic functioning in adult primates. Biological Psychiatry, 35(4), 221–227. doi: 10.1016/0006-3223(94)91252-1 doi:10.1016/0006-3223(94)91252-1 Rosenblum, L. A., Forger, C., Noland, S., Trost, R. C., & Coplan, J. D. (2001). Response of adolescent bonnet macaques to an acute fear stimulus as a function of early rearing conditions. Developmental Psychobiology, 39(1), 40–45. doi: 10.1002/dev.1026 Rosenblum, L. A., & Paully, G. S. (1984). The effects of varying environmental demands on maternal and infant behavior. Child Development, 55(1), 305–314. doi: 10.2307/1129854 Roth, T. L., Lubin, F. D., Funk, A. J., & Sweatt, J. D. (2009). Lasting epigenetic influence of early-life adversity on the BDNF gene. Biological Psychiatry, 57(8), 823–831. doi: 10.1016/j.biopsych.2008.11.028 doi:10.1016/j.biopsych.2008.11.028 Roth, T. L., & Sullivan, R. M. (2005) Memory of early maltreatment: neonatal behavioral and neural correlates of maternal maltreatment within the context of classical conditioning. Biol Psychiatry, 57, 823–831. doi: 10.1016/j.biopsych.2005.01.032 doi:10.1016/ j.biopsych.2005.01.032

References

199

Sade, D. (1967). Determinants of dominance in a group of free ranging rhesus monkeys. In S. A. Altmann (Ed.), Social communication among primates (pp. 99–114). Chicago, IL: University of Chicago Press.

RNA expression in human postmortem hippocampus. Developmental Brain Research, 561(2), 332–337. doi: 10.1016/0006-8993(91)91612-5 doi:10.1016/0006-8993(91)91612-5

Sanchez, M. M. (2006). The impact of early adverse care on HPA axis development: Nonhuman primate models. Hormones & Behavior, 50, 623–631. doi: 10.1016/j.yhbeh.2006.06.012 doi:10.1016/ j.yhbeh.2006.06.012

Secoli, S. R., & Teixeira, N. A. (1998). Chronic prenatal stress affects development and behavioral depression in rats. Stress, 2(4), 273–280. doi: 10.3109/10253899809167291

Sanchez, M. M., Alagbe, O., Felger, J. C., Zhang, J., Graff, A. E., Grand, A. P., Maestripieri, D., & Miller, A. H. (2007). Activated p38 MAPK is associated with decreased CSF 5-HIAA and increased maternal rejection during infancy in rhesus monkeys. Molecular Psychiatry, 12(10), 895–897. doi: 10.1038/sj.mp. 4002025 Sanchez, M. M., Ladd, C. O., & Plotsky, P. M. (2001). Early adverse experience as a developmental risk factor for later psychopathology: evidence from rodent and primate models. Developmental Psychopathology, 13(3), 419–449. Sanchez, M. M., McCormack, K., Grand, A. P., Fulks, R., Graff, A., & Maestripieri, D. (2010). Effects of sex and early maternal abuse on adrenocorticotropin hormone and cortisol responses to the corticotropin-releasing hormone challenge during the first 3 years of life in group-living rhesus monkeys. Developmental Psychopathology, 22(1), 45–53. doi: 10.1017/S0954579409990253 Sanchez, M. M., Noble, P. M., Lyon, C. K., Plotsky, P. M., Davis, M., Nemeroff, C. B., & Winslow, J. T. (2005). Alterations in diurnal cortisol rhythm and acoustic startle response in nonhuman primates with adverse rearing. Biological Psychiatry, 57(4), 373–381. doi: 10.1016/j.biopsych.2004.11.032 doi:10.1016/j.biopsych.2004 .11.032 Sanchez, M. M., & Pollak, S. D. (2009). Socio-emotional development following early abuse and neglect: Challenges and insights from translational research. In M. de Haan & M. R. Gunnar (Eds.), Handbook of developmental social neuroscience (pp. 497–520). New York, NY: Guilford Press. Sanchez, M. M., Young, L. J., Plotsky, P. M., & Insel, T. R. (1999). Autoradiographic and in situ hybridization localization of corticotropinreleasing factor 1 and 2 receptors in nonhuman primate brain. Journal of Comparative Neurology, 408(3), 365–377. doi: 10.1002/ (SICI)1096-9861(19990607)408:33.3.CO;2-E Sanchez, M. M., Young, L. J., Plotsky, P. M., & Insel, T. R. (2000). Distribution of corticosteroid receptors in the rhesus brain: Relative absence of glucocorticoid receptors in the hippocampal formation. Journal of Neuroscience, 20(12), 4657–4668. doi: 20/12/4657 Santiago, L. B., Jorge, S. M., & Moreira, A. C. (1996). Longitudinal evaluation of the development of salivary cortisol circadian rhythm in infancy. Clinical Endocrinology, 44(2), 157–161. doi: 10.1046/ j.1365-2265.1996.645466.x Sapolsky, R. M., & Meaney, M. J. (1986). Maturation of the adrenocortical stress response: neuroendocrine control mechanisms and the stress hyporesponsive period. Brain Research Reviews, 11(1), 65–76. doi: 10.1016/0165-0173(86)90010-X doi:10.1016/0165 -0173(86)90010-X Sapolsky, R. M., Zola-Morgan, S., & Squire, L. R. (1991). Inhibition of glucocorticoid secretion by the hippocampal formation in the primate. Journal of Neuroscience, 11(12), 3695–3704. Schroeder, M., Sultany, T., & Weller, A. (2013). Prenatal stress effects on emotion regulation differ by genotype and sex in prepubertal rats. Developmental Psychobiology, 55(2), 176–192. doi: 10.1002/dev.21010 Seay, B., & Gottfried, N. W. (1975). A phylogenetic perspective for social behavior in primates. Journal of General Psychology, 92(1), 5–17. doi: 10.1080/00221309.1975.9711323 Seckl, J. R., Dickson, K. L., Yates, C., & Fink, G. (1991). Distribution of glucocorticoid and mineralocorticoid receptor messenger

Shaw, P., Kabani, N. J., Lerch, J. P., Eckstrand, K., Lenroot, R., Gogtay, N., . . . Rapoport, J. L. (2008). Neurodevelopmental trajectories of the human cerebral cortex. Journal of Neuroscience, 28(14), 3586–3594. doi: 10.1523/JNEUROSCI.5309-07.2008 Shear, D. A., Galani, R., Hoffman, S. W., & Stein, D. G. (2002). Progesterone protects against necrotic damage and behavioral abnormalities caused by traumatic brain injury. Experimental Neurology, 178(1), 59–67. doi: 10.1006/exnr.2002.8020 doi:10.1006/exnr.2002 .8020 Shively, C. A., Laber-Laird, K., & Anton, R. F. (1997). Behavior and physiology of social stress and depression in female cynomolgus monkeys. Biological Psychiatry, 41(8), 871–882. doi: 10.1016/S0006 -3223(96)00185-0 doi:10.1016/S0006-3223(96)00185-0 Shively, C. A., Register, T. C., Friedman, D. P., Morgan, T. M., Thompson, J., & Lanier, T. (2005). Social stress-associated depression in adult female cynomolgus monkeys (Macaca fascicularis). Biological Psychology, 69(1), 67–84. doi: 10.1016/j.biopsycho.2004.11.006 doi:10.1016/j.biopsycho.2004.11.006 Silk, J. B. (2002). Practice random acts of aggression and senseless acts of intimidation: the logic of status contests in social groups. Evolutionary Anthropology: Issues, News, and Reviews, 11(6), 221–225. doi: 10.1002/evan.10038 Slawecki, C. J. (2005). Comparison of anxiety-like behavior in adolescent and adult Sprague-Dawley rats. Behavioral Neuroscience, 119(6), 1477–1483. doi: 10.1037/0735–7044.119.6.1477 Smith, E. L., Coplan, J. D., Trost, R. C., Scharf, B. A., & Rosenblum, L. A. (1997). Neurobiological alterations in adult nonhuman primates exposed to unpredictable early rearing. Relevance to posttraumatic stress disorder. Annals of the New York Academy of Sciences, 821, 545–548. doi: 10.1111/j.1749-6632.1997.tb48326.x Sobrian, S. K., Vaughn, V. T., Ashe, W. K., Markovic, B., Djuric, V., & Jankovic, B. D. (1997). Gestational exposure to loud noise alters the development and postnatal responsiveness of humoral and cellular components of the immune system in offspring. Environmental Research, 73(1–2), 227–241. doi: 10.1006/enrs.1997.3734 Spear, L. P. (2000). The adolescent brain and age-related behavioral manifestations. Neuroscience & Biobehavioral Reviews, 24(4), 417–463. doi: 10.1016/S0149-7634(00)00014-2 doi:10.1016/S0149 -7634(00)00014-2 Stanton, M. E., Gutierrez, Y. R., & Levine, S. (1988). Maternal deprivation potentiates pituitary-adrenal stress responses in infant rats. Behavioral Neuroscience, 102(5), 692. doi: 10.1037/0735-7044 .102.5.692 Stanton, M. E., Wallstrom, J., & Levine, S. (1987). Maternal contact inhibits pituitary-adrenal stress responses in preweanling rats. Developmental Psychobiology, 20(2), 131–145. doi: 10.1002/dev.420200204 Stevens, H. E., Leckman, J. F., Coplan, J. D., & Suomi, S. J. (2009). Risk and resilience: early manipulation of macaque social experience and persistent behavioral and neurophysiological outcomes. Journal of the American Academy of Child & Adolescent Psychiatry, 48(2), 114–127. doi: 10.1097/CHI.0b013e318193064c doi:10.1097/CHI .0b013e318193064c Sullivan, R. M., & Holman, P. J. (2010). Transitions in sensitive period attachment learning in infancy: The role of corticosterone. Neuroscience & Biobehavioral Reviews, 34(6), 835–844. doi: 10.1016/ j.neubiorev.2009.11.010

200

Animal Models of Developmental Psychopathology

Suomi, S. J., Harlow, H. F., & Kimball, S. D. (1971). Behavioral effects of prolonged partial social isolation in the rhesus monkey. Psychological Reports, 29(3), 1171–1177. doi: 10.2466/pr0.1971.29.3f.1171

Ulrich-Lai, Y. M., & Herman, J. P. (2009). Neural regulation of endocrine and autonomic stress responses. Nature Reviews Neuroscience, 10(6), 397–409. doi: 10.1038/nrn2647

Sweet, J. A., Eakin, K. C., Munyon, C. N., & Miller, J. P. (2014). Improved learning and memory with theta-burst stimulation of the fornix in rat model of traumatic brain injury. Hippocampus, 24(12), 1592–1600. doi: 10.1002/hipo.22338

Van de Poll, N. E., De Jonge, F., Van Oyen, H. G., & Van Pelt, J. (1982). Aggressive behaviour in rats: Effects of winning or losing on subsequent aggressive interactions. Behavioural Processes, 7(2), 143–155. doi: 10.1016/0376–6357(82)90023–7

Szuran, T. F., Pliska, V., Pokorny, J., & Welzl, H. (2000). Prenatal stress in rats: effects on plasma corticosterone, hippocampal glucocorticoid receptors, and maze performance. Physiology & Behavior, 71(3–4), 353–362. doi: 10.1016/S0031-9384(00)00351-6 doi:10.1016/S0031-9384(00)00351-6

van den Hove, D. L., Jakob, S. B., Schraut, K. G., Kenis, G., Schmitt, A. G., Kneitz, S., . . . Lesch, K. P. (2011). Differential effects of prenatal stress in 5-HTT deficient mice: Towards molecular mechanisms of gene × environment interactions. PLoS ONE, 6(8), e22715. doi: 10.1371/journal.pone.0022715

Tanapat, P., Galea, L. A., & Gould, E. (1998). Stress inhibits the proliferation of granule cell precursors in the developing dentate gyrus. International of Journal of Developmental Neuroscience, 16(3–4), 235–239. doi: 10.1016/S0736-5748(98)00029-X doi:10.1016/S0736 -5748(98)00029-X

Varlinskaya, E. I., Truxell, E. M., & Spear, L. P. (2013). Repeated restraint stress alters sensitivity to the social consequences of ethanol differentially in early and late adolescent rats. Pharmacol Biochem Behav, 113, 38–45. doi: 10.1016/j.pbb.2013.10.016

Taravosh-Lahn, K., & Delville, Y. (2004). Aggressive behavior in female golden hamsters: development and the effect of repeated social stress. Hormones and Behavior, 46(4), 428–435. doi: 10.1016/j .yhbeh.2004.03.007 doi:10.1016/j.yhbeh.2004.03.007 Tarullo, A. R., & Gunnar, M. R. (2006). Child maltreatment and the developing HPA axis. Hormones and Behavior, 50(4), 632–639. doi: 10.1016/j.yhbeh.2006.06.010 doi:10.1016/j.yhbeh.2006 .06.010 Teicher, M. H., Andersen, S. L., Polcari, A., Anderson, C. M., Navalta, C. P., & Kim, D. M. (2003). The neurobiological consequences of early stress and childhood maltreatment. Neuroscience & Biobehavioral Reviews, 27(1–2), 33–44. doi: 10.1016/S0149-7634(03)00007-1 doi:10.1016/S0149-7634(03)00007-1 Teicher, M. H., Tomoda, A., & Andersen, S. L. (2006). Neurobiological consequences of early stress and childhood maltreatment: Are results from human and animal studies comparable? Annals of the New York Academy of Sciences, 1071, 313–323. doi: 10.1196/annals.1364.024 Thiebaut de Schotten, M., Dell’Acqua, F., Valabregue, R., & Catani, M. (2012). Monkey to human comparative anatomy of the frontal lobe association tracts. Cortex, 48(1), 82–96. doi: 10.1016/j.cortex .2011.10.001 doi:10.1016/j.cortex.2011.10.001 Toledo-Rodriguez, M., & Sandi, C. (2007). Stress before puberty exerts a sex- and age-related impact on auditory and contextual fear conditioning in the rat. Neural Plasticity, 71203.. doi: 10.1155/2007/71203 Tollenaar, M. S., Jansen, J., Beijers, R., Riksen-Walraven, J. M., & de Weerth, C. (2010). Cortisol in the first year of life: Normative values and intra-individual variability. Early Human Development, 86(1), 13–16. doi: 10.1016/j.earlhumdev.2009.12.003 doi:10.1016/j .earlhumdev.2009.12.003 Tottenham, N., & Sheridan, M. A. (2009). A review of adversity, the amygdala and the hippocampus: A consideration of developmental timing. Frontiers in Human Neuroscience, 3, 68. doi: 10.3389/neuro.09.068 .2009 doi:10.3389%2Fneuro.09.068.2009#pmc_ext Troisi, A., & D’Amato, F. R. (1984). Ambivalence in monkey mothering: infant abuse combined with maternal possessiveness. Journal of Nervous and Mental Disease, 172(2), 105–108. Troisi, A., D’Amato, F. R., Fuccillo, R., & Scucchi, S. (1982). Infant abuse by a wild-born group-living Japanese macaque mother. Journal of Abnormal Psychology, 91(6), 451–456. doi: 10.1037/0021 -843X.91.6.451 Tung, J., Barreiro, L. B., Johnson, Z. P., Hansen, K. D., Michopoulos, V., Toufexis, D., . . . Gilad, Y. (2012). Social environment is associated with gene regulatory variation in the rhesus macaque immune system. Proceedings of the National Academy of Sciences, 109(17), 6490–6495. doi: 10.1073/pnas.1202734109

Vazquez, D. M. (1998). Stress and the developing limbic-hypothalamicpituitary-adrenal axis. Psychoneuroendocrinology, 23(7), 663–700. doi: 10.1016/S0306-4530(98)00029-8 doi: 10.1016/S0306-4530(98) 00029-8 Veenema, A. H. (2009). Early life stress, the development of aggression and neuroendocrine and neurobiological correlates: What can we learn from animal models? Frontiers in Neuroendocrinology, 30(4), 497–518. doi: 10.1016/j.yfrne.2009.03.003 doi:10.1016/j.yfrne .2009.03.003#doilink Vidal, J., Buwalda, B., & Koolhaas, J. M. (2011). Differential long-term effects of social stress during adolescence on anxiety in Wistar and wild-type rats. Behavioural Processes, 87(2), 176–182. doi: 10.1016/ j.beproc.2011.03.004 Vivinetto, A. L., Suarez, M. M., & Rivarola, M. A. (2013). Neurobiological effects of neonatal maternal separation and post-weaning environmental enrichment. Behavioural Brain Research, 240, 110–118. doi: 10.1016/j.bbr.2012.11.014 Vucetic, Z., Kimmel, J., Totoki, K., Hollenbeck, E., & Reyes, T. M. (2010). Maternal high-fat diet alters methylation and gene expression of dopamine and opioid-related genes. Endocrinology, 151(10), 4756–4764. doi: 10.1210/en.2010–0505 Walker, A. K., Nakamura, T., Byrne, R. J., Naicker, S., Tynan, R. J., Hunter, M., & Hodgson, D. M. (2009). Neonatal lipopolysaccharide and adult stress exposure predisposes rats to anxiety-like behaviour and blunted corticosterone responses: Implications for the double-hit hypothesis. Psychoneuroendocrinology, 34(10), 1515–1525. doi: 10.1016/j.psyneuen.2009.05.010 Watt, M. J., Roberts, C. L., Scholl, J. L., Meyer, D. L., Miiller, L. C., Barr, J. L., . . . Forster, G. L. (2014). Decreased prefrontal cortex dopamine activity following adolescent social defeat in male rats: Role of dopamine D2 receptors. Psychopharmacology, 231(8), 1627–1636. doi: 10.1007/s00213–013–3353–9 Weathington, J. M., Arnold, A. R., & Cooke, B. M. (2012). Juvenile social subjugation induces a sex-specific pattern of anxiety and depression-like behaviors in adult rats. Hormones and Behavior, 61(1), 91–99. doi: 10.1016/j.yhbeh.2011.10.008 Weathington, J. M., & Cooke, B. M. (2012). Corticotropin-releasing factor receptor binding in the amygdala changes across puberty in a sex-specific manner. Endocrinology, 153(12), 5701–5705. doi: 10.1210/en.2012–1815 Weathington, J. M., Hamki, A., & Cooke, B. M. (2014). Sex- and region-specific pubertal maturation of the corticotropin-releasing factor receptor system in the rat. Journal of Comparative Neurology, 522(6), 1284–1298. doi: 10.1002/cne.23475 Weaver, I. C., Cervoni, N., Champagne, F. A., D’Alessio, A. C., Sharma, S., Seckl, J. R., . . . Meaney, M. J. (2004). Epigenetic programming

References by maternal behavior. Nature Neuroscience, 7(8), 847–854. doi: 10.1038/nn1276 Weinstock, M. (2007). Gender differences in the effects of prenatal stress on brain development and behaviour. Neurochemical Research, 32(10), 1730–1740. doi: 10.1007/s11064–007–9339–4 Weinstock, M., Matlina, E., Maor, G. I., Rosen, H., & McEwen, B. S. (1992). Prenatal stress selectively alters the reactivity of the hypothalamic-pituitary adrenal system in the female rat. Brain Research, 595(2), 195–200. doi: 10.1016/0006-8993(92)91049-K doi: 10.1016/0006-8993(92)91049-K Weintraub, A., Singaravelu, J., & Bhatnagar, S. (2010). Enduring and sex-specific effects of adolescent social isolation in rats on adult stress reactivity. Brain Research, 1343, 83–92. doi: 10.1016/j.brainres .2010.04.068 Weller, A., Glaubman, H., Yehuda, S., Caspy, T., & Ben-Uria, Y. (1988). Acute and repeated gestational stress affect offspring learning and activity in rats. Physiology & Behavior, 43(2), 139–143. doi: 10.1016/0031-9384(88)90229-6 doi:10.1016/0031-9384(88)90229-6 Whishaw, I. Q., & Kolb, B. (Eds.). (2005). The behavior of the laboratory rat. New York, NY: Oxford University Press. Wiggins, R. C., & Gottesfeld, Z. (1986). Restraint stress during late pregnancy in rats elicits early hypermyelination in the offspring. Metabolic Brain Disease, 1(3), 197–203. doi: 10.1007/BF01001781 Willey, A. R., & Spear, L. P. (2013). Effects of acute ethanol administration and chronic stress exposure on social investigation and 50kHz ultrasonic vocalizations in adolescent and adult male Sprague-Dawley rats. Pharmacology Biochemistry and Behavior, 105, 17–25. doi: 10.1016/j.pbb.2013.01.012 Willner, P., Towell, A., Sampson, D., Sophokleous, S., & Muscat, R. (1987). Reduction of sucrose preference by chronic unpredictable mild stress, and its restoration by a tricyclic antidepressant. Psychopharmacology, 93(3), 358–364. doi: 10.1007/BF00187257

201

Wilson, M. E., Bounar, S., Godfrey, J., Michopoulos, V., Higgins, M., & Sanchez, M. M. (2013). Social and emotional predictors of the tempo of puberty in female rhesus monkeys. Psychoneuroendocrinology, 38(1), 67–83. doi: 10.1016/j.psyneuen.2012.04.021 doi:10.1016/j.psyneuen.2012.04.021 Wilson, M. E., Gordon, T. P., & Collins, D. C. (1986). Ontogeny of luteinizing hormone secretion and first ovulation in seasonal breeding rhesus monkeys. Endocrinology, 118(1), 293–301. doi: 10.1210/endo-118–1–293 Wommack, J. C., Taravosh-Lahn, K., David, J. T., & Delville, Y. (2003). Repeated exposure to social stress alters the development of agonistic behavior in male golden hamsters. Hormones and Behavior, 43(1), 229–236. doi: 10.1016/S0018-506X(02)00029-6 doi:10.1016/S0018-506X(02)00029-6 Wright, L. D., Muir, K. E., & Perrot, T. S. (2012). Enhanced stress responses in adolescent versus adult rats exposed to cues of predation threat, and peer interaction as a predictor of adult defensiveness. Developmental Psychobiology, 54(1), 47–69. doi: 10.1002/dev .20575 Zehr, J. L., Van Meter, P. E., & Wallen, K. (2005). Factors regulating the timing of puberty onset in female rhesus monkeys (Macaca mulatta): Role of prenatal androgens, social rank, and adolescent body weight. Biology of Reproduction, 72(5), 1087–1094. Zhang, W., & Rosenkranz, J. A. (2013). Repeated restraint stress enhances cue-elicited conditioned freezing and impairs acquisition of extinction in an age-dependent manner. Behavioural Brain Research, 248, 12–24. doi: 10.1016/j.bbr.2013.03.028 Zou, S., Funk, D., Shram, M. J., & Le, A. D. (2014). Effects of stressors on the reinforcing efficacy of nicotine in adolescent and adult rats. Psychopharmacology, 231(8), 1601–1614. doi: 10.1007/s00213–013 –3314–3

CHAPTER 6

The Role of Early Nutritional Deficiencies in the Development of Psychopathology ADAM T. SCHMIDT, ERIN A. OSTERHOLM, and MICHAEL K. GEORGIEFF

THE ROLE OF EARLY NUTRIENT DEFICIENCIES IN THE DEVELOPMENT OF PSYCHOPATHOLOGIES 202 Background 202 Studies on Psychopathology Related to Early Life Nutrition 206 HUMAN STUDIES OF NUTRITION AND COGNITIVE DEVELOPMENT 213 NEUROBIOLOGY OF NUTRITIONAL EFFECTS: EVIDENCE FROM BENCH SCIENCE 220 Early Life Macronutrient Undernutrition 220

Early Life Micronutrient Undernutrition 224 Summary of Micronutrient Deficiencies 226 FUTURE DIRECTIONS 226 TRANSLATIONAL IMPLICATIONS 229 CONCLUDING REMARKS 230 REFERENCES 231

THE ROLE OF EARLY NUTRIENT DEFICIENCIES IN THE DEVELOPMENT OF PSYCHOPATHOLOGIES

when there is an overall lack of nutrition (as occurs in famine) or specific, as in the absence of specific nutrients in an otherwise sufficient diet. More recently, the term malnutrition has taken on a broader meaning more consistent with its literal roots: bad nutrition. In this conceptualization, overnutrition is also a risk to somatic and brain development and, like undernutrition, could be due to a global excess of nutrients or the presence of toxic amounts of a specific nutrient. Indeed, risk from nutrients represent a U-shaped curve, where the lowest risk and greatest efficacy are from a range of nutrient sufficiency, while deficiency or excess confer risks to developing systems such as the brain. Specific conditions such as phenylketonuria (PKU), where there is an excess of the amino acid phenylalanine due to a genetic inability to metabolize the amino acid, illustrate the importance of nutritional factors in brain development. If not addressed, PKU can result in mental retardation (Huttenlocher, 2000; Sullivan & Chang, 1999). The build-up of this amino acid during brain development results in a toxic reaction leading to brain damage (Huttenlocher, 2000). Reports hypothesizing a nutritional origin of some forms of mental retardation first appeared in the early 1950s when researchers noted high amounts of phenylketones in the urine of some mentally retarded individuals (Huttenlocher, 2000). Subsequent investigations linked PKU to various neuropathologies including diffuse

Background The influence of early nutrition on the developing brain has been a scientific issue over the past century (Susser & Stein, 1994). However, popular culture has recognized the importance of early nutrition in children’s health dating back to Roman times. The Nobel Prize–winning author Pearl S. Buck wrote about the devastating effects of starvation on young children in her 1931 classic book The Good Earth. In this novel, the first daughter of the main character is born at the height of a severe famine. Buck implies the significant malnutrition the child endured early in life resulted in devastating brain damage, leaving her profoundly retarded. While such a profound effect likely overstates the case, the role of undernutrition and malnutrition in developmental psychopathology has strong scientific basis. It is the goal of this chapter to review and update the evidence for this science and to highlight the importance of a developmental psychopathology perspective for studying this complex issue. Our species comes from a long history of inadequate access to food and thus malnutrition is generally conceptualized as undernutrition. Undernutrition can be global 202

The Role of Early Nutrient Deficiencies in the Development of Psychopathologies

white matter changes and reduced dendritic branching (Huttenlocher, 2000). Specific genetic disorders resulting in deficiencies of micronutrients (such as copper in Menkes disease), regional geographic deficiencies of trace elements (such as iron, iodine and selenium), the presence of inhibitors of nutrients from the diet (all divalent metals including zinc, copper and iron) and low prenatal levels of maternal nutrients (such as folate and iron) have all been linked to severe neurological abnormalities or intellectual deficits (Hibbard & Smithells, 1965; Lozoff, 1989; Pharoah, Buttfield, & Hetzel 1971; Prohaska, 2000; Shim & Harris, 2003; Yi-Ming, 1996). The famines and food shortages throughout Europe during the Second World War created an interest in studying the changes in psychological and physiological functioning caused by general undernutrition (Keys, 1946; Pollitt, 1988). Although many of the outcomes were chronicled as alterations to adult behaviors, it is pertinent in our context because (1) children likely exhibit similar, if not greater, psychological consequences and (2) by affecting the psychological state of their caregivers, parental malnutrition may exert indirect effects on children regardless of their nutritional status (Brozek, 1990). In a more controlled scientific approach to the general malnutrition question, Keys and colleagues (at the University of Minnesota) exposed 36 healthy adult males to 6 months of semistarvation followed by 3 months of controlled rehabilitation (Franklin, Scheile, Brozek, & Keys, 1948; Keys, 1946). Results indicated that during the period of semistarvation the participants became lethargic, apathetic, less sociable, uninterested in sexual activity, less alert (except with regard to hearing), and depressed (Franklin et al., 1948; Keys, 1946). Although the participants complained about changes in mental alertness and inability to think cogently, researchers noted no decreases on tests of memory or intellectual ability (Franklin et al., 1948; Keys, 1946). The participants in the study exhibited unusual behaviors involving food such as compulsive gum chewing and coffee/tea drinking (up to 40 packs of gum a day and 15 cups of coffee a day), souping (i.e., the practice of drinking the liquid in soup and then refilling the bowl with hot salted water and repeating the process before eating the vegetables in the soup), and unusual eating habits including novel spice concoctions, hoarding of food, and doting over every last piece of food on the plate (Franklin et al., 1948). Many of the participants greatly increased their daily consumption of water and either began smoking or substantially increased the amount they smoked (Franklin et al., 1948). Finally, participants displayed numerous physiological reactions such as decreased

203

sensitivity to heat, increased sensitivity to cold, edema, skin discoloration, and reduced hair growth (Franklin et al., 1948). During the controlled rehabilitation phase, participants were initially irritable and easily frustrated (Franklin et al., 1948). These effects were short-lived; however, weight gain and sexual interest did not immediately return to prestarvation levels. In fact, participants did not reach their prestarvation weight and level of physical fitness until almost a year after the conclusion of the semistarvation phase (Franklin et al., 1948). These findings are important because they indicate in adults periods of substantial undernutrition result in demonstrable acute but reversable alterations in behavior, physiology, and mood (Franklin et al., 1948; Keys, 1946). The reversibility of these symptoms suggests, in adults, behavioral changes occur secondarily to acute changes in neurotransmitter systems, as opposed to neuronal structure, induced by the nutrient deficiency. Research in humans and animals show dietary manipulations result in rapid changes in behavior and concentration of specific neurotransmitters (Fromentin, Gietzen, & Nicolaidis, 1997; Gietzen & Magrum, 2001; Phillips, Oxtoby, Langley, Bradshaw, & Szabadi, 2000; Pierucci-Lagha et al., 2004; Riedel, 2004). The developing brain of the child is at greater risk for the effects of undernutrition and these may not be as amenable to rehabilitation, thus resulting in long-term of lifelong behavioral abnormalities. The child’s brain is at greater risk for residual abnormalities (after the nutritional burden is lifted) because in contrast to the adult brain, malnutrition likely impacts neuroanatomic organization in addition to neurotransmitter systems during periods of sensitive or critical periods of regional brain development (Winick & Rosso, 1969). While neurotransmitter recovery may seem more likely to occur, disordered neuroanatomical organization in childhood is more difficult to overcome because of sensitive periods of development (Wachs et al., 2014). To test this question of long-term risk of psychopathology following early life global malnutrition, Stein et al. (1972) used a natural experiment to investigate whether prenatal exposure to famine in the Dutch hunger winter of 1944–1945 adversely affected intellectual abilities in a sample of 19-year-old males. In September 1944 the Netherlands cooperated with allied forces in an attempt by British paratroopers to force a bridgehead over the Rhine River (Stein et al., 1972, 1975; Susser et al., 1998). The attempt failed, and in retaliation the Nazi army imposed a trade embargo on Holland. Unfortunately, when the army lifted the embargo, previous railroad strikes and a severe

204

Nutritional Deficiencies and Psychopathology

winter that had frozen the shipping channels resulted in serious food shortages in much of Western Holland. At the height of the famine, daily food rations could be as low as 450 kilocalories per person, a quarter of the standard amount. The most significant period of the famine began in November 1944 and lasted until allied armies crossed the Rhine and liberated Holland in May 1945. The severity of famine was indicated by a sharply increased death rate in the affected cities (many of these deaths were attributed to starvation), substantial (up to 25%) loss in body weight, and the occurrence of physical signs associated with severe malnourishment such as osteomalacia and hunger edema (Stein et al., 1972, 1975). This event was unique because it occurred at a specified time and place in a population which before and after famine had good nutrition, there was good reporting on food rations provided to the population, and excellent public health records allowed researchers to divide individuals into groups of varying exposures based on birth date and birth location. This natural experiment has generated numerous investigations on the effects of prenatal famine exposure on a multitude of psychological and medical variables. Stein et al. (1972) grouped individuals into six birth cohorts based upon their intrauterine exposure to famine: (1) individuals born before the beginning of the famine; (2) individuals prenatally exposed to famine in the third trimester only; (3) individuals exposed to famine in the middle 6 months of gestation; (4) individuals conceived during the famine and exposed in the first and second trimesters; (5) individuals conceived during the famine and exposed during the third trimester only; and (6) individuals conceived and born after the famine occurred. Stein and colleagues used Raven Progressive Matrices scores gathered at age 19 during military induction testing to compare these cohorts with control cohorts born during the same time frame in similarly sized cities not affected by the famine. The researchers failed to find significant declines in average IQ or increases in the rates of mild or severe mental retardation in the four groups prenatally exposed or in the one group exposed postnatally. Because many births occurred at home, the researchers could not evaluate the influence of prenatal famine exposure on the birth weight of the entire cohort; however, a subsample of births occurring in hospitals in the famine and control cities indicated no relationship of birth weight to IQ at age 19. These findings cast doubt on the assertion that maternal malnutrition alone results in significantly declined intellectual abilities. However, the authors emphasize that their

findings should not be generalized to chronically malnourished populations where the mother is likely to be malnourished before and after parturition, the child is at high risk for postnatal nutritional deficits, or deficiencies in specific trace nutrients are common. The authors suggest their findings reveal either a high degree of protection afforded to the developing child in utero or the considerable resilience of many children in the face of prenatal insult. Over the years, many people have speculated that nutrition acutely affects behavior (e.g., sugar increases activity levels). Although careful research has not substantiated this particular assertion, there are many clearly documented short-term effects of specific nutrients (or substances) on behavior (Beyth & Baratta, 1996; Fernstrom, 2000; Fishbein & Pease, 1994; Lieberman, 2003). These acute effects are not the topic of this paper. Rather, the discussion focuses on the effects of generalized malnutrition or specific micronutrient deficiencies on the developing brain during fetal life, infancy and childhood and how these insults may contribute to the later emergence of psychopathology. Before beginning the discussion in detail, it is important to define several terms. First, in this chapter early undernutrition refers to nutrient restriction occurring prenatally (sometimes even prior to conception) or in the first few years of postnatal development (prior to age 3). Second, undernutrition refers to a general reduction in macronutrient intake, also known as protein energy malnutrition (PEM). Nutritional deficits can be characterized by PEM, by deficits in specific micronutrients such as iron (Fe), Zinc (Zn), or iodine (I), or by both, as is often the case. Most of the extant research concerns PEM/undernutrition, and this will be the focus of the first section of this paper; however, each of the three micronutrients is discussed in the second half of the chapter because of their high prevalence rates around the world and their profound impact on early brain and behavior development. An unambiguous definition of what constitutes psychopathology has proven elusive (Lilienfeld & Marino, 1999; Meehl, 1979; Spitzer, Fleiss, & Endicott, 1978; Wakefield, 1993, 1999), but is not necessary here. We will use major mental disorders as listed in the American Psychiatric Association’s Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5; American Psychiatric Association, 2013) as constituting a large subset of psychopathology pertinent to this discussion. Most research relevant to this review examines schizophrenia and schizophrenia spectrum disorders, internalizing disorders (such as major depression and anxiety disorders), and externalizing problems (such as

The Role of Early Nutrient Deficiencies in the Development of Psychopathologies

attention-deficit/hyperactivity disorder and antisocial personality disorder). Basic Principles When examining the effects of nutritional deficiencies on the developing brain and the role these may play in the genesis of psychopathology, it is important to keep in mind several core principles. First, brain development occurs in overlapping but dissociable stages, and each of these stages may be uniquely vulnerable to injury (Honig, Herrman, & Shatz, 1996; Monk, Webb, & Nelson, 2001; Morgane et al., 1993; Teicher, Andersen, Polcari, Anderson, & Navalta, 2002; Wachs et al., 2014; Webb, Monk, & Nelson, 2001). These stages include neurulation, neurogenesis, proliferation, migration, differentiation, synaptogenesis, myelination, apoptosis, and synaptic pruning (Andersen, 2003; Monk et al., 2001; Webb et al., 2001). Disturbances in some of these processes are associated with neurobehavioral deficits—as in the nonverbal impairments and psychiatric symptoms seen in the leukodystrophies—or the hypothesis that aberrant synaptic connections play a role in the etiology of autism and schizophrenia (Innocenti, Ansermet, & Parnas, 2003; Keller & Persico, 2003; Percy & Rutledge, 2001; Rosebush, Garside, Levinson, & Mazurek, 1999; Van Geel, Assies, Wanders, & Barth, 1997). Some of the initial processes of brain development (e.g., neurulation and neurogenesis) occur during a single early period whereas other later stages (e.g., myelination, apoptosis, and synaptic pruning) occur independently at different times in various brain regions (Monk et al., 2001; Webb et al., 2001). Thus, a brain structure or region will have its own time course through which development proceeds. Moreover, certain highly metabolic brain regions such as the hippocampus are more vulnerable to injury than other regions (Walsh & Emerich, 1988). Ultimately, the result of any insult to the developing brain depends on the time of onset of the insult and what developmental processes are occurring/brain structures are actively developing at that time, how innately vulnerable these processes/structures are to injury—a factor driven largely by the relationship between metabolic demand and blood supply—and the magnitude of the insult (i.e., the severity and duration of the injury). Thus, a specific nutrient deficiency can have different effects on the developing nervous system, largely dependent upon these variables. (Adams et al., 2000; Georgieff & Rao, 2001; Kretchmer, Beard, & Carlson, 1996; Morgane et al., 1993; Pollitt, 1996; Wachs et al., 2014). In this context, it is worth highlighting that the timing of rapid development of certain “major” systems spans fetal and early postnatal life. The hippocampus,

205

which is so central to discriminative recognition memory, undergoes its most rapid differentiation between 32 weeks gestation and 18 months postnatal age. Myelination, a mediator of speed of processing throughout the nervous system, also begins in earnest around 32 weeks gestation and proliferates rapidly until approximately 3 years of age. The dopamine neurotransmitter system, which is integral to reward, motivation and attention, starts to develop in mid-gestation and is solidly regulated by 3 years of age. Each is particularly sensitive to common nutrient deficiencies of pregnancy and young children such as protein malnutrition, iron deficiency and fatty acid deficiencies. Not only is there concern about the primary functions mediated by these three systems, but also proper function of these primary areas during development is key for establishing connections to more distal areas (e.g., frontal cortex) and for the efficiency of integrated circuits in mediating complex behaviors (e.g., prepulse inhibition). A second important principle underlying the current discussion is the concept that abnormal behavior (as in psychopathology) is the efferent expression of brain activity, and changes in neuroanatomy, neurochemistry, and neurophysiology can result in alterations in behavior (Brower & Price, 2001; Moffitt, 1993; Moffitt & Caspi, 2001; Nestor, Kimble, Berman, & Haycock, 2002; Pennington & Ozonoff, 1996; Raine, 2002). Thus, early malnutrition may lead to the expression of cognitive deficits or abnormal behavior stemming from alterations in fine neuroanatomy (delicate structures such as dendritic arbors and white matter tracks), neurochemistry (disturbances in neurotransmitter systems), or neurophysiology (the functioning of ion channels and specific receptors) (Arnold, 1999; Benes & Berretta, 2001; Bowley, Drevets, Öngür, & Price, 2002; Fuglestad et al., 2008; Konradi & Heckers, 2003; Molnar, Potkin, Bunney, & Jones, 2003). Although specific findings (e.g., white matter abnormalities or increased receptor density in a certain region) are sometimes difficult to replicate in new samples, disturbances in these three broad domains are consistently associated with various forms of psychopathology (Dwork, 1997; Flecher, 1998; Jones, 1997). The last core principle running through this discussion is the concept that genes will likely prove to be the single most important long-term determinant of behavior; however, their expression can be modified (repressed or activated) by interaction with the environment (Bunney et al., 2003; Jacobs et al., 2002; Murphy et al., 2001; Rutter, Pickles, Murray, & Eaves, 2001; Rutter & Silberg, 2002; Slutske et al., 1997; Wachs et al., 2014). Environmental influences causing anomalous neurological development

206

Nutritional Deficiencies and Psychopathology

(such as chronic and severe neglect or toxicity) can have especially salient and long-term effects (Barone, Das, Lassiter, & White, 2000; Castle et al., 1999; Schettler, 2001; Wachs et al., 2014; Weiss, 2000). Similarly, nutrition is an important environmental factor that can affect the developing brain (Fuglestad et al., 2008; Georgieff & Rao, 2001; Morgane et al., 1993; Morgane, Mokler, & Galler, 2002; Wachs et al., 2014). This does not imply that environmental insults such as malnutrition have uniform effects on people of differing genetic characteristics. It is likely that individual differences in the capacity to maintain homeostasis in response to challenges (thereby allowing for reasonably normal development) and in—at present—poorly explicated concepts such as neurological integrity, plasticity, and resilience are genetically or experientially determined, and such factors likely modulate the effects of environmental challenges such as undernutrition (Bhutta & Anand, 2002; Engle, Castle, & Menon, 1996; Morgane et al., 1993; Nelson, 2000a; Sapolsky, 2001; Wachs et al., 2014; Walker et al., 2011). Furthermore, it is rarely the case that undernutrition exists independently of other risk factors such as maternal stress, genetic predisposition, poverty, and dangerous living conditions (e.g., poor sanitation or in the presence of toxins such as lead) (Guesry, 1998; Pollitt, 1969; ; Pollitt et al., 1996; Schettler, 2001; Snodgrass, 1994; Stein & Susser; 1985; Tanner & Finn-Stevenson, 2002; Trope, Lopez-Villegas, Cecil, & Lenkinski, 2001). Pollitt (1969) reported that in developing countries children with the greatest nutritional deficits had the poorest living conditions (i.e., lack of running water or adequate plumbing, cramped housing, and large families). Thus, malnutrition may generally exert its effects by acting in concert (additively or synergistically) with other risk factors (Guesry, 1998; Morgane et al., 1993, 2002; Stein & Susser, 1985; Zeskind & Ramey, 1981). For example, Caspi and coworkers (2002) found that abused children with a genotype leading to higher levels of a neurotransmitter metabolizing enzyme (monoamine oxidase) displayed lower rates of antisocial behavior. The authors interpreted these findings to be evidence of a gene/environment interaction in which the genotype influenced an individual’s response to a particular environmental stimulus. It is possible that analogous mechanisms or interactions could be at work in determining the outcome of early malnutrition. Overview The current chapter endeavors to examine the effect of early nutritional deficits on brain development and the role

this may play in the emergence of psychopathology. We will begin the discussion by examining the possible link between early nutritional insults and the development of psychopathology in the context of those studies that have specifically researched this question. After this, we will step back and delve into other investigations that assess the effects of early malnutrition on broader issues such as cognitive development in humans and developmental neurobiology. These factors are important because they characterize a developmental path (resulting from early malnutrition) in which psychopathology may exist as part of the milieu or which may increase an individual’s risk for the later development of mental illness. Finally, the chapter will conclude with a brief hypothetical discussion of potential causal mechanisms that may account for the relationship between early nutritional insults and the later development of psychopathology. Studies on Psychopathology Related to Early Life Nutrition Studies in this section are presented with regard to the spectrum of psychopathology they examine. Thus, studies investigating the effects of early malnutrition on schizophrenia or schizophrenia spectrum disorders, neurodevelopmental disorders (e.g., attention-deficit/ hyperactivity disorder [ADHD], autism), externalizing disorders such as conduct problems and antisocial behavior, and internalizing disorders (e.g., depression) are discussed. Krueger and colleagues (in addition to other researchers) have proposed the broad dimensions of externalizing and internalizing to capture more accurately the structure of common mental disorders (Krueger, 2002; Krueger, Caspi, Moffitt, & Silva, 1998; Krueger, McGue, & Iacono, 2001). These dimensions are used in the current discussion for clarity and succinctness. Despite numerous investigations into the cognitive effects of malnutrition on children, most of the research dealing with the effects of malnutrition on the development of psychopathology concerns individuals older than eighteen. This is a significant shortcoming of our knowledge in this area because many psychiatric disorders will likely prove to have developmental origins. Thus, it is probable that these conditions have perceptible manifestations during childhood and adolescence. Similarly, early malnutrition may influence the developmental course (e.g., age of onset, severity, or progression) of psychiatric disorders. Moreover, malnutrition may be expected to exert more powerful immediate effects upon the developing brain—thereby suggesting the effects of malnutrition may be more salient in individuals younger than 18.

The Role of Early Nutrient Deficiencies in the Development of Psychopathologies

Thus, children affected by early nutritional deficits may display additional or significantly different behavioral effects than mature adults (Keys, 1946). Furthermore, it is possible that some of the cognitive manifestations of undernutrition (discussed later) may have resulted from impairments secondary to various psychiatric symptoms such as inattention or anxiety. Much of the information regarding the relationship between early malnutrition and later psychopathology comes from data collected on cohorts prenatally exposed to famine during the Dutch hunger winter of 1944–1945 and the later Chinese Great Leap Forward famine of 1959–1961 (Brown & Susser 2008; Brown et al., 1996; Butler, Susser, Brown, Kaufmann, & Gorman 1994; Hoek, Brown, & Susser, 1998; Lumey, Stein, & Susser, 2011; McGrath, Brown, & St Clair, 2011; Song, Wang, & Hu, 2009; St Clair et al., 2005; Susser, Hoek, & Brown, 1998; Xu et al., 2009). As discussed previously, Stein et al. (1972) found no evidence suggesting decreased intellectual abilities or increased rates of severe or mild mental retardation following prenatal famine exposure. However, years later, Susser (the son of the original investigators) and his collaborators revisited the famine data to examine if individuals with a history of prenatal famine exposure exhibited higher rates of psychopathology in adulthood (Brown, Susser, Lin, Neugebauer, & Gorman, 1995; Brown et al., 2000; Hulshoff-Pol et al., 2000; Insel et al., 2008; Neugebauer, Hoek, & Susser, 1999; Susser et al., 1996; Susser & Lin, 1992). Several major admonitions regarding these natural experiment should be pointed out. First, the amount of food redistribution in families is not known (Stein et al., 1972). Thus, pregnant women who occasionally received greater rations might have been supplied more calories via sacrifice of other family members (Stein et al., 1972, 1975). Second, although probably not substantially significant, it is not known how much individuals supplemented their diets with other materials, which may have actually contained toxins (Stein et al., 1972, 1975). Third, except for records indicating profession of father within the Dutch sample, there is no information regarding familial background. Finally, because many of the births occurred at home, it is impossible to gather information on birth or gestational complications, which along with maternal infection and stress may play etiological roles in the emergence of behavioral deficits and psychopathological conditions (Anand & Scalzo, 2000; Brown, Cohen, Greenwald, & Susser, 2000; Brown & Susser, 2002; Brown et al., 2000; McNeil, Cantor-Graae, & Weinberger, 2000; Teicher et al., 2002). Regardless of these essentially unavoidable

207

imperfections, the studies discussed below provide valuable insights regarding the possible role of prenatal malnutrition in the emergence of later psychiatric disorders. Moreover, the consistency of the findings from the two episodes serves to support the contention that early exposure to severe prenatal nutritional deprivation increases the risk for later development of various psychopathologies. In an initial 1992 study of the Dutch sample, Susser and Lin found that women, but not men, exhibited an increased prevalence of schizophrenia following prenatal exposure to famine in the first trimester. However, they obtained these results through a broad inclusion criteria (that is, first trimester of gestation during periods of low food rations), and review of admission records to psychiatric hospital units from 1978 to 1989. Thus, they based exposure on a single criterion that may not have completely captured children who experienced the highest levels of exposure. Furthermore, individuals were at least 32 years old at the start of the review period. This age tends to be past the age that most males have their first psychotic episode; therefore, it is likely many men would have already made initial psychiatric contact prior to this age and were not re-hospitalized during the review period (Brown et al., 1996). In a more extensive follow-up investigation, Susser et al. (1996) used stringent inclusion criteria to define a maximal exposure cohort. The inclusion criteria for the other cohorts were the same as in the 1992 study. However, they defined the maximal exposure cohort to include only those individuals conceived at the height of the famine. This cohort consisted of individuals born when the general population exhibited increases in adverse health effects and new births demonstrated excess congenital malformations (which according to Stein, et al. (1975) are associated with severe prenatal starvation). These criteria ensured that the researchers dealt exclusively with the maximally exposed cohort (e.g., the cohort that underwent the most significant early insult due to the famine). These restrictions led to the identification of a cohort born between October 15, 1945 and December 31, 1945 (Hoek et al., 1996, 1998; Susser et al., 1998). The authors extended the period of record review back to 1970, thereby more accurately reflecting lifetime history of psychiatric admission. Based upon these data, the researchers found a significantly increased prevalence of schizophrenia in both males and females of the maximal exposure cohort. This cohort displayed a prevalence rate for schizophrenia nearly two times higher than that of the unexposed cohort. Further, the other exposed cohorts did not differ in prevalence rate from the unexposed cohort. Susser et al. demonstrated

208

Nutritional Deficiencies and Psychopathology

these findings using a narrow and a broad diagnostic definition of schizophrenia. Individuals within the maximal exposure cohort had mothers malnourished during the periconceptional phase, and their prenatal famine exposure occurred early in pregnancy, generally entirely within the first trimester. Using magnetic resonance imaging, Hulshoff-Pol and colleagues (2000) found increased brain abnormalities (specifically focal white matter hyperintensities) in a subgroup of individuals from the maximal exposure cohort (Hulshoff-Pol et al., 2000). They showed that individuals from this cohort diagnosed with schizophrenia evidenced reduced intracranial volume compared to non-exposed individuals with schizophrenia and non-diagnosed individuals with the same level of famine exposure. The authors speculated that the decreased intracranial volumes observed only in the famine exposed individuals with schizophrenia resulted from an interaction between genetic risk factors and early brain stunting caused by gestational malnutrition—eventually resulting in the development of schizophrenia. They concluded that exposure to famine early in gestation is associated with increased focal brain abnormalities—possibly secondary to specific micronutrient deficiencies. In some individuals, these aberrations may result in a higher risk for developing schizophrenia. The researchers urge discretion when examining their findings, stressing they obtained the results using a limited number of participants and other confounding factors could not be completely taken into account. Nevertheless, these data provide an initial glimpse into specific structural brain changes present following severe first trimester malnutrition. Another study by the same group expanded their inquiry to determine if males conceived at the height of the hunger winter evidenced increased prevalence of schizophrenia spectrum personality disorders (SSPD)—as defined by ICD-6 (Hoek et al., 1996). This study used methodology similar to the aforementioned study. Thus, the authors compared the maximal exposure cohort (see Susser et al., 1996, for a description of this cohort) to an unexposed cohort and a cohort with first trimester exposure but without maternal malnutrition at the time of conception. The authors obtained data regarding the diagnosis of SSPD from military induction records; thus, males only are represented. This maximal exposure group demonstrated a higher prevalence of SSPDs. This increase appeared to be unique to SSPDs because the three groups did not differ from one another in the prevalence of psychoneurosis (i.e., general anxiety). They noted that this relationship remains consistent when estimated SES is

taken into account. However, the authors hasten to point out that despite the increased risk in the maximal exposure group, relatively few individuals actually met criteria for SSPD in the entire study. Furthermore, they cautioned that despite the apparent care taken in making these diagnoses, no formalized diagnostic system existed at the time of these evaluations. To some extent, this caution should be applied to both of the studies previously discussed. Using information obtained from birth and medical records in a population-based cohort born between 1924 and 1932 in Helsinki, Finland, Wahlbeck, Forsén, Osmond, Barker, and Eriksson (2001) reported on the relationship between factors indicative of prenatal undernutrition and the later development of schizophrenia. This study found that children born to mothers with a low body mass index (BMI) prior to parturition, who were born with small placentas and were below-average length and weight at birth, had an increased risk of developing schizophrenia in adulthood. Furthermore, this study indicated that, independent of other predictors, children thin at 7 years of age had an increased risk for developing schizophrenia in later life. The data also indicated that children short at birth who remained lean during childhood had a fourfold risk of developing schizophrenia. The authors observed no differences in the risk for schizophrenia between groups of varying SES, replicating findings from other studies. These results are consistent with the hypothesis that fetal and childhood factors (specifically undernutrition) play a role in the pathogenesis of schizophrenia. However, given the relatively low and consistent prevalence of schizophrenia within the population and the relatively high rate of early malnutrition, it may be more plausible to conjecture that early nutritional deficits act as predisposing or triggering circumstances. This study suggests malnutrition statistically increases the risk for later development of schizophrenia. Wahlbeck and colleagues suggest, in their sample, nutritional status (based upon maternal BMI, birth length, and slenderness in childhood) is independent of SES; however, they provide no evidence to bolster this claim. Thus, it is impossible to discern (given the information available) if these measures reflect actual undernutrition, or are simply corollaries of low SES or a genetic diathesis that increases risk for schizophrenia. Similar results were obtained in later studies using a cohort of individuals affected by the Great Leap Forward famine occurring in China between 1959 and 1961. In the initial study with this cohort, St Clair and colleagues (2005) found that in the most affected province, the adjusted risk of developing schizophrenia increased from 0.84% for

The Role of Early Nutrient Deficiencies in the Development of Psychopathologies

those born in 1959 to 2.15% and 1.81% for those born in the height of the famine in 1960 and 1961 respectively. A later investigation by Xu and colleagues (2009) confirmed these increases using a cohort from a different region of China. This study found that this result appeared largely driven by individuals from a rural population suggesting that the increase was observed in only the most affected areas (Xu et al., 2009). Taken together, these studies suggest that the prevalence of schizophrenia and schizophrenia spectrum disorders increases in individuals who are exposed to severe generalized malnutrition early in gestation. The findings of Wahlbeck et al. (2001) also provide evidence that schizophrenia is (either directly or indirectly) related to suboptimal growth in utero through early childhood. Hoek et al. (1998) speculated about potential mechanisms to explain the findings of their group. First, they noted that it was unlikely that this finding was related solely to nutritional restriction because if such a correlation existed, one would expect to find significantly higher rates of schizophrenia and schizophrenia spectrum disorders in developing countries that commonly have serious food shortages. Rather, it is more likely that the increased prevalence was related to a restriction in a specific micronutrient, stating that these types of deficiencies are frequently observed in both developed and developing areas. For example, early iron deficiency (a very common micronutrient deficiency in both developed and developing countries) causes impairments in prepulse inhibition (PPI) in rodents (Pisansky et al., 2013). PPI is considered a reliable endophenotype of schizophrenia (Swerdlow, Braff, & Geyer, 2000), suggesting that early iron deficiency is a plausible mechanism for the development of psychosis in some circumstances (Pisansky et al., 2013). These initial findings underscore the importance of additional research using animal models of specific micronutrient deficiencies occurring at circumscribed times during brain development that reflect a proposed pathophysiology of a disorder—for example, migration defects in schizophrenia secondary to folate deficiency. This approach could be useful in elucidating the effects of micronutrient deficiencies on neuroanatomy, neurochemistry, and neurophysiology; however, it would fall short in illustrating possible behavioral correlates because there are currently no irrefutable animal models of mental disorders, especially schizophrenia. Finally, Hoek and colleagues (1998) stated, their findings could have resulted from the increased level of birth complications and morbidity associated with early birth found in the most affected famine cohort. They noted that areas less

209

affected by the famine experienced similar increases in obstetric complications without the concomitant rise in schizophrenia spectrum disorders, making this potential confound less problematic. Neugebauer and colleagues (1999) again used the Dutch famine data to investigate possible relationships between prenatal nutritional restriction and the subsequent development of antisocial personality disorder (ASPD). The researchers used diagnostic information regarding ASPD from Dutch military induction records. According to the authors, Dutch psychiatrists considered ASPD to include a constellation of behaviors such as aggression, history of criminal activity, disregard for the truth, impulsivity, volatile temperament, unreliability, and interpersonal difficulties. Furthermore, psychiatrists categorized individuals under the subclassifications of violent or nonviolent ASPD. Individuals exposed to severe malnutrition in utero during the first or second trimesters had an increased risk for the emergence of later ASPD (specifically violent ASPD) compared to either unexposed individuals, individuals with only moderate levels of caloric restriction, or individuals severely malnourished in the third trimester alone. Malnutrition remained predictive, even after they included other proven or potential risk factors (e.g., low social class, low IQ, or large sibship size) for ASPD in multiple regression equations. The authors comment that the most likely explanation for this pattern of results is the lack of a particular unspecified micronutrients during this critical period of brain development. They argued that similarities in infant morbidity between cases and noncases, specificity of time and place of the findings, and the lack of third trimester effects indicate severe malnutrition (during the first and second trimester) supersedes other environmental insults. However, they also proposed a possible mechanism in which malnutrition potentiates preexisting genetic or environmental vulnerabilities. Additional research with the Dutch cohort indicates that exposure to the famine during the first trimester of gestation is associated with an increased risk of development of an addictive disorder in adulthood (Franzek, Sprangers, Janssens, Van Duijn, & Van De Wetering, 2008). These investigators did not find any evidence of increased risk for those individuals with exposures limited to the second or third trimesters, although those individuals with first trimester exposure during the peak of the famine continued to be at increased risk despite reintroduction of adequate nutrition during the remainder of their gestation. Of note, this effect appeared to hold only in males and was not observed in females with exposures during similar times in development.

210

Nutritional Deficiencies and Psychopathology

Using a cohort of low birth weight (LBW) children, Breslau, Chilcoat, Johnson, Andreski, and Lucia (2000), examined the relationship between LBW, soft neurological signs, and subsequent intellectual/behavioral functioning at ages 6 and 11. Soft neurological signs refer to a number of various subtle neurological abnormalities in sensory, motor, and integrative skills that do not significantly impact an individual’s overall functioning but may still result in suboptimal abilities in these domains (Breslau et al., 2000). Importantly, soft signs are not thought to reflect focal brain damage, but rather a diffuse neurological pathology resulting from a variety of sources including genetic anomalies such as microdeletions or mosaics (Breslau et al., 2000). Furthermore, studies have linked soft neurological signs to early malnutrition and deficits in IQ and behavioral functioning (Breslau et al., 2000; Galler, Ramsey, Solimano, Kucharski, & Harrison, 1984). Breslau et al. (2000) observed a relationship between LBW and the occurrence of soft neurological signs in which LBW children are nearly twice as likely to exhibit soft signs as normal birth weight (NBW) children. The researchers found a linear relationship between LBW and the frequency of soft signs. Thus, the lower a child’s weight at birth, the more likely the child was to display soft signs. Moreover, the data indicated children evidencing soft signs consistently performed below average on tests of intellectual abilities regardless of birth weight; however, LBW children tended to score lower on the IQ test in general, thus indicating additive influences. They also found a relationship between soft signs in both LBW and NBW children and the occurrence of learning disorders (at age 11) and the occurrence of internalizing problems at six and eleven. In the LBW group, the investigators found a relationship between soft signs and the occurrence of externalizing behaviors including attention problems at age 6 but not at age 11. Even though this study did not strictly deal with nutritional deficits, it is interesting because LBW and soft signs may be associated with undernutrition. This is particularly germane to the Breslau et al. (2000) study because many of the participants, especially those within the LBW group, lived in urban settings and came from low SES backgrounds and thus had an elevated risk for prenatal or postnatal malnutrition. Regardless of the nexus between LBW and soft neurological signs, the findings suggest LBW children may be at increased risk for reductions in IQ, learning disabilities, internalizing problems, and, when younger, externalizing difficulties including attention deficits. The reason for the apparent evaporation of differences in the externalizing domain may be increased

levels of externalizing behaviors exhibited by the control groups (Moffitt, 1993), or the beneficial effects of behavioral/medical interventions. In a follow-up to their 1986 study (to be discussed in a subsequent section of the chapter), Galler and Ramsey (1989) assessed a group of previously malnourished Barbadian children for symptoms of hyperactivity. Using parent and teacher reports, they found increased levels of hyperactivity and distractibility. These relationships remained consistent after taking into account environmental factors such as parental involvement, parental career/knowledge of the world, stability of the home environment, number of siblings, and quality of living conditions (e.g., quantity of modern conveniences and durability of dwelling construction). Furthermore, these results concur with earlier findings (Galler, Ramsey, Solimano, & Lowell, 1983), which suggested increased attention difficulties in previously malnourished children between 5 and 11 years of age. The current study revealed an increased prevalence of speech difficulties in the formerly malnourished children as noted by teacher report. However, difficulties with socialization, emotional reactivity, and immaturity previously observed in these children (Galler, Ramsey, Solimano, & Lowell, 1983) were not seen in the current investigation. According to the authors, this may have resulted from slight changes in the assessment tools (necessary to reflect the older age of the children) or interventions targeted at ameliorating some of the effects of early malnutrition. If the latter is the case, it is meaningful because it suggests some aspects of early malnutrition (e.g., social skills) may be more amenable to educational interventions than others (e.g., attention problems) (Galler & Ramsey, 1989). However, it is possible that effects such as emotional instability are more characteristic of younger children who were previously malnourished and would disappear in older children regardless of interventions (Galler & Ramsey, 1989). A shortcoming of this study is the researchers’ failure to control for intellectual abilities. Thus, it is unknown if the increased attention problems observed in the experimental group are a unique aspect of undernutrition in the first year of life, or rather are a corollary of the lower intellectual skills previously reported in these children (Galler et al., 1986). In a long-term follow-up investigation, Galler and colleagues (Galler, Bryce, Waber, Zichlin, Fitzmaurice, & Eaglesfield, 2012) determined that those individuals previously malnourished continued to experience increased symptoms of attention problems in adulthood. This investigation used both self-report and computer-based assessments of attention difficulties and demonstrated that

The Role of Early Nutrient Deficiencies in the Development of Psychopathologies

the impact of early malnutrition was not attenuated by environmental factors in childhood and that self-report of attention concerns was not influenced by childhood intellectual abilities. Other recent studies by this same group (Galler, Bryce, Waber, Medford, Eaglesfield, & Fitzmaurice, 2011; Galler, Bryce, Zichlin, Fitzmaurice, Eaglesfield, & Waber, 2012) found a relationship between a history of proteinenergy-malnutrition in the first year of life and attenuated executive skills via parent report, peer directed aggression when the cohort was relatively young, and the development of conduct problems via self-report as the cohort aged. These investigators stated that findings indicated continuity between the reports of the youth and earlier obtained teacher reports (Galler, Bryce, Waber, et al., 2012). Results suggested a relative decrease in parent reported peer aggression as the children aged and also showed that the later emergence of conduct problems was partly mediated by other factors such as childhood IQ and home environment. Although not as robust as the findings in schizophrenia, the studies discussed already imply that prenatal as well as postnatal malnutrition predisposes individuals for exhibiting a variety of externalizing behaviors. It is intriguing to speculate whether the brain is susceptible to insults causing externalizing behaviors for a longer period during development, or if there are distinct etiologies of externalizing disorders, potentially linked to disruptions at various stages of brain development. Nevertheless, this handful of studies provides an impetus for future investigations on the role of malnutrition and other neurobiological risk factors in the etiology of externalizing problems. Brown and colleagues investigated the relationship between prenatal famine exposure and the later development of affective disorders (affective psychosis and neurotic depression) (Brown, Susser, Lin, Neugebauer, & Gorman, 1995). The researchers defined affective psychosis and neurotic depression in accordance with ICD-9 definitions. Thus, individuals with affective psychosis displayed disturbances in mood (either depression or mania) accompanied by mood congruent delusions, whereas individuals with a diagnosis of neurotic depression displayed disturbances in mood (including symptoms of depression or anxiety) but evidenced no psychotic symptoms. They formed the groups by matching birth date and location with psychiatric admission records from 1978–1991. Prenatal exposure to famine in the second trimester increased the risk of developing affective psychosis in later life; however, the authors did not find increased risk of affective psychosis resulting from first or third trimester

211

famine exposure. These results remained significant when sex was taken into account; however, there was a significant interaction between sex and second semester famine exposure. Thus, males but not females appeared to be at greater risk for developing affective psychosis after second semester famine exposure—whereas females had higher prevalence rates in general. The data indicated no increased risk of neurotic depression following prenatal famine exposure in any trimester or in either sex. A possible reason for the disparity between the diagnostic categories may be that (when compared to neurotic depression) affective psychosis represents a more severe mood disorder with a greater risk of hospitalization. Furthermore, neurotic depression may be more likely triggered by life stressors and thus differences between groups could be submerged by random variation in the population. A limitation of this study is that females usually begin exhibiting mood symptoms in adolescence. Therefore, effects of prenatal famine exposure may be observed in females if the researchers reviewed earlier hospital admission records. Additionally, Brown and his collaborators caution, their findings may not represent increased prevalence of affective psychosis, but rather may reflect increased need for hospital admission or delayed onset of symptoms in prenatally famine-exposed males. Nevertheless, these results suggest that males are at greater risk for developing serious mood disorders following significant mid-gestational malnutrition. In a follow-up to their 1995 study, Brown and colleagues (2000) greatly increased their sample size by obtaining new diagnostic records. The researchers focused on unipolar and bipolar depression requiring hospitalization. Exposure to famine during the second and third trimesters related to increased hospital admissions in both males and females with unipolar and bipolar depression. The authors indicated that this effect was particularly robust if exposure occurred in the third trimester only. These results extend the findings of the 1995 study, and suggest prenatal famine exposure predisposes adults to develop serious mood disorders requiring hospitalization and characterized by either depression or mania/hypomania. Another more recent study with the Dutch cohort indicated increases in depressive symptoms in adulthood among those individuals whose mothers were exposed to famine prior to gestation (i.e., in the preconceptual period; Stein, Pierik, Verrips, Susser, & Lumey 2009). The authors also found decreases in self-reported quality of life among this cohort. Interestingly, differences in depressive scores did not appear dependent on trimester of exposure during pregnancy (Stein et al., 2009).

212

Nutritional Deficiencies and Psychopathology

Thompson and colleagues examined the relationship between low birth weight and the incidence of depression in old age (Thompson, Syddall, Rodin, Osmond, & Barker, 2001). These investigators used data obtained from individuals born between 1920 and 1930 in the English county of Hertfordshire and who still lived in the area in the early 1990s. The authors assessed all of the participants for depressive symptoms using two measures (a self-report questionnaire and a semistructured interview), and simultaneously asked about confounding factors such as SES, recent loss, general illness, and coronary heart disease (CHD). This study found a relationship between birth weight and depression in late life so that the lower a child’s birth weight, the greater the risk for developing late life depression. The findings were more robust in males and remained significant when the researchers statistically controlled for other complicating factors such as illness, SES, and bereavement. When they specifically took into account the occurrence of chronic heart disease (another risk factor for late life depression) the relationship grew stronger. The researchers that indicated males with low birth weight but high weight at one year exhibited the highest risk for late life depression; conversely, males with above average birth weight and comparatively low weight at one year exhibited the lowest risk for late life depression. They conjectured that this pattern indicated poor prenatal nutrition followed by compensatory growth and good prenatal nutrition followed by a regression to the mean respectively. Because this study focused on current and not early life depression, the results should not be generalized to depression occurring outside of old age. The authors hypothesized their findings provided support for aberrant fetal programming (see the conclusion section of this paper) caused by undernutrition as the major factor contributing to the future development of depression. Conversely, Waber et al. (2011) using the Barbadian cohort demonstrated that the direct relationship between depressive symptoms and early malnutrition was significantly attenuated when intellectual abilities were used as a covariate. In summary, severe prenatal undernutrition appears to increase the risk for developing internalizing disorders (specifically mood disorders). However, it is unclear if the role of postnatal undernutrition significantly increases the risk of depressive disorders once other contributing variables (e.g., cognitive functioning) are taken into account. It is interesting that the risk period for exposure to malnutrition in the externalizing and internalizing domains may be later than for the schizophrenia spectrum disorders. The stages of brain development affected during these periods of vulnerability concur with research on the neurological

sequelae of these various psychopathologies, and support the timing of onset and duration of insult hypothesis. For example, the research presented already on schizophrenia strongly suggests malnutrition exerts its effects in early stages of CNS development, neural tube formation, or initial neuronal migration, all of which have been implicated in the etiology of schizophrenia (Arnold, 1999; LaMantia, 1999; Rioux, Nissanov, Lauber, Bilker, & Arnold, 2003; Susser, Brown, Klonowski, Allen, & Lindenbaum, 1998). The later period of vulnerability observed in externalizing and internalizing disorders suggests malnutrition disrupts later processes of brain development such as neurotransmitter formation, late stage migration, or gliogenesis. Abnormalities in these processes have been implicated in the etiology of externalizing and internalizing disorders (Bowley et al., 2002; Cotter, Pariante, & Everall, 2001; Jenihe et al., 1996). Further, some conditions (such as schizophrenia) may be associated with a relatively acute injury, whereas behavioral disturbances in the externalizing and internalizing domains may be associated with insults extending over longer periods. The investigations undertaken with the Dutch and Chinese famine cohorts and similar samples suggest early severe malnutrition, even if relatively brief in duration, increases the risk for the later development of mental disorders. Furthermore, this effect is not a general increase in vulnerability, but rather appears to depend on the timing and duration of the nutritional insult. Because the Dutch famine studies lack confounds such as social class, deprived environments, parental education, and extended postnatal malnutrition (which frequently accompany early malnutrition) they cannot be generalized to a chronically malnourished population. However, they provide strong support for the unique contribution of severe prenatal undernutrition in the emergence of later psychopathology. Few studies have addressed the role of postnatal or prenatal–postnatal malnutrition in the development of psychopathology, even though it is clear from the National Collaborative Perinatal Project data base from the United States that prenatal malnutrition followed by postnatal failure to thrive has profound effects on IQ at 7 years of age (Pylipow et al., 2009). The lack of data on an effect of prenatal–postnatal malnutrition on psychopathology is particularly disheartening because postnatal malnutrition is common throughout the world, and, as discussed later in the studies on cognition, it is likely postnatal malnutrition is associated with deficits that differ from those of prenatal malnutrition. The previous discussion suggests that early malnutrition may lead to the later development of psychopathology. Moreover, the expression of a particular psychopathology

Human Studies of Nutrition and Cognitive Development

likely depends on when the insult occurred (e.g., prenatal vs. postnatal or first verses second and third trimester). Another contributing factor (discussed in greater detail in the concluding remarks section) is that malnutrition may exert its maximal affect if it occurs concurrently with another insult such as stress or toxicity. Now that a plausible link has been established, it is necessary to take a step back in order to examine what may underlie the development of these conditions. We will first examine human studies that have been designed to determine if early malnutrition results in adverse effects on cognitive/intellectual development. It is necessary to note that human protein energy malnutrition or undernutrition rarely exists without concomitant micronutrient deficiencies (Gorman, 1995; Kretchmer et al., 1996). Whether long-term effects are due to global or specific key nutrient undernutrition (e.g., iron, zinc) remains an interesting research question. HUMAN STUDIES OF NUTRITION AND COGNITIVE DEVELOPMENT One of the shortcomings of research on humans is it is difficult to control all potentially relevant variables, sometimes even known confounds. This limitation is particularly vexing in epidemiological research regarding the long-term consequences of early malnutrition. For example, as previously discussed, early malnutrition rarely occurs independently of other risk factors such as poverty and environmental stress (Engle et al., 1996; Pollitt, 1969; Stein & Susser, 1985; Tanner & Finn-Stevenson, 2002; Wachs et al., 2014). Although there are statistical techniques that can help compensate for this co-morbidity, one must still be careful when making inferences regarding experimental interventions or the interplay of particular risk factors. Secondarily, many studies using a population-based design are unable to account for other known or conjectured risk factors such as genetic predisposition, parental education, maternal stress, chaotic home environment, or other complicating physical factors such as toxin exposure, infection, or birth complications (Grantham-McGregor, 2002; Pollitt, 1969; Wachs et al., 2014). These caveats are not indictments of these types of studies, because what such studies lack in elegant controls they often make up for in direct relevance. However, the limited control over potential third variables should be kept in mind when evaluating such studies. There are numerous studies detailing the effects of early malnutrition on human cognitive development (Fuglestad et al., 2008; Georgieff & Rao, 2001; Gorman, 1995; Park et al., 2011; Pollitt, 1988; Walker et al., 2011). Many of

213

these studies were conducted in developing countries where malnutrition (or perhaps more frequently, undernutrition) is likely chronic and may span all of development—from the periconceptional phase to adolescence and young adulthood (Grantham-McGregor, 2002; Pongcharoen et al., 2012; Walker et al., 2011). Given the growing minority and immigrant populations of developed countries such as the United States and the fact that a vast portion of children worldwide suffer from early malnutrition, a careful consideration of these studies is warranted—even for those primarily interested in industrialized populations (Pollitt, 1994). Furthermore, knowledge of successful nutritional interventions in developing countries may help to improve current U.S. government programs such as the Women, Infants and Children (WIC) program (Pollitt, 1994; Tanner & Finn-Stevenson, 2002). It is useful to review this literature in this chapter in order to provide insight into the broad behavioral effects of early malnutrition. Further, research has linked intellectual/cognitive deficits with various types of psychopathologies (Moffitt & Caspi, 2001). Although this link may not be causal, these conditions (i.e., cognitive abilities and psychopathology) may nonetheless be connected. Studies on neurocognitive outcomes of children who suffered early malnutrition are presented in three sections: studies concerning prenatal malnutrition and effects of low birth weight (where the fetus is protected from certain environmental factors such as poverty), studies on postnatal or combined prenatal and postnatal malnutrition, and studies examining the effects of nutritional interventions with malnourished populations. LBW children (1,200 loci whose genetic variants were shown to be associated with >165 common human diseases/disorders and complex traits; collectively, these variants implicated many previously unknown or unconsidered roles for numerous biological pathways (Hirschhorn, 2009; Lander, 2011; Manolio, Brooks, & Collins, 2008). It was expected that these very first GWASes (measuring the genome in this case with single nucleotide polymorphisms, SNPs) would result in the translation of previously obtained

The Operationalization

heritability estimates into effect sizes for measured G (i.e., specific genetic variants). The GWAS-based estimate of heritability is the ratio of the heritability due to the variants used in the GWAS (specified as the numerator), estimated directly from their observed effects, to the total heritability (specified as the denominator), obtained either from previous studies or from the same study (if the sample is comprised of related individuals), or inferred indirectly from population data (Zuk, Hechter, Sunyaev, & Lander, 2012). However, these first GWASes and all subsequent ones (until recently, when the data-analytic approach changed and genomewide complex-trait analysis, GCTA, was introduced, e.g., Plomin et al., 2013), reportedly accounted for only a small proportion of the previously published heritability estimates of diseases/disorders and complex traits. This discrepancy between what was expected based on results from quantitative-genetic studies and what was observed based on the results from molecular-genetic studies was labeled as the “missing heritability” problem. As GWASes have become bigger (i.e., larger samples) and better (i.e., denser coverage of the genome), results have improved, with the referenced amount of variance reaching 20–30% or even 50% in isolated cases. However, the puzzle remains, as only a smaller portion of the previously obtained heritability estimates have been mapped onto specific genetic variants measured by GWASes (Lander, 2011). As the problem of missing heritability became apparent, the dominant explanation was that it arose from incomplete coverage of the genome, i.e., the assumption that there were still some (or many) undiscovered variants not included in the numerator, leading to estimates biased on the lower side. Yet another possibility pertains to the overestimation of the denominator: the phenomenon referred to as phantom heritability (Zuk et al., 2012). Such an overestimation can arise, for example, if models do not take into account epistatic genetic interactions within loci (in that effects of each loci may not merely be additive but interact with each other to predict disease/disorder risk) (Zuk et al., 2012). To illustrate their point, Zuk and colleagues (2012) referenced Crohn’s disease—an inflammatory bowel disease that impacts various components of the digestive tract. For this disease, in a traditional GWAS paradigm, 71 risk-associated loci have been identified (Franke et al., 2010). If an additive model is assumed, then these loci account for only 21.5% of the heritability, which was previously estimated at 50% (Halme, Paavola-Sakki, Turunen, Färkkilä, & Kontula, 2006), yet, if an epistatic

295

model is assumed, then these loci explain 80% of the adjusted heritability (the phantom heritability was estimated at 62.8%). Unfortunately, it is estimated that very large sample sizes (∼500,000) are required to detect genetic interactions even for such a relatively heritable disease as Crohn’s. The authors (Zuk et al., 2012) concluded that current estimates (and, therefore, the whole discussion) of missing heritability might not be meaningful, as these estimates were obtained without taking into account genetic interactions. This observation is quite congruent with the worry that most previously reported heritability estimates are, indeed, artificially elevated. In fact, most of these estimates come from twin studies, which tend to provide an upper limit to the genetic component of the variance of a disease/disorder or a trait, and thereby may lead to invalid conclusions (Wallace, 2006). The second concern relevant to the measurement of G and E, as mentioned above, is with the selection of the candidate genes and candidate environments used in GXE studies. For most, if not all, common complex diseases and disorders, even main effects of G and E are still rather mysterious (see the discussion for psychiatric disorders in Flint & Munafò, 2008). In fact, almost two decades of candidate gene association studies have produced, arguably, little or no unequivocally accepted findings regarding genetic effects (Burmeister, McInnis, & Zöllner, 2008). This observation is especially worrisome given that most (if not all) hypotheses about specific candidate genes emerged from strong neuroscience research that uncovered the putative properties and functions of these genes (Hebebrand, Scherag, Schimmelmann, & Hinney, 2010); these candidate genes, most of which are protein-coding, were selected because they were found to have functional variants that somehow changed the properties of synthesized proteins, which were implicated in complex diseases and disorders. Given the lack of prior success in uncovering main effects of G for a given complex disease or disorder, there is understandable skepticism regarding the possibility that more complex associations, such as between a specific variant in a specific candidate gene (or any other type of genotype) and a specific facet or a general characteristic of a particular environment (Duncan & Keller, 2011; Ioannidis, 2005), will also be uncovered.5 The GWAS approach is fundamentally different and hypothesis-free. Correspondingly, it is not surprising that 5

A prior probability distribution (the prior) of an uncertain quantity p, in Bayesian statistics, is a distribution capturing expectations about p prior to the accumulation of the relevant evidence.

296

The Trilogy of GxE: Conceptualization, Operationalization, and Application

thus far, results of GWAS do not align with hypotheses generated by candidate–gene research. In fact, out of 531 SNPs labeled as the most robustly associated (SNPs) to various medical and psychiatric phenotypes in GWAS studies, 45% are intronic, 43% are intergenic, and only 11% are exonic (Hindorff et al., 2009). Moreover, when specific polymorphisms in specific candidate genes are investigated (i.e., those that have been featured in numerous candidate gene studies, such as the serotonin transporter gene promoter polymorphism) in GWAS designs, they reportedly do not demonstrate performance above the level of chance (Bosker et al., 2011; Lasky-Su et al., 2008; Need et al., 2009; Sullivan et al., 2008). To summarize, GWASes do not appear to be converging on the expected candidate genes, and the SNPs that GWASes are illuminating are mostly not exonic. This state of affairs has been referred to as a failure of both candidate gene studies (Little et al., 2009) and GWASes (McClellan & King, 2010), as their respective findings do not converge. Yet several considerations are important to mention here. Many fashionable polymorphisms that are used in GxE candidate gene studies are not present (and cannot be present due to their biological makeup) in GWASes; thus, they need to be imputed. For example, the polymorphism in the promoter region of the serotonin transporter gene may be imputed in Caucasian samples with high (∼93–95%) accuracy using multiple SNPs present in some GWAS arrays (Knodt, 2012; Lu et al., 2012). Conducting such imputations for groups other than Caucasians is much more difficult. Importantly, GWASes are designed primarily to investigate main effects of G (or its specific variants). However, the polymorphism in the promoter region of the serotonin transporter gene has not been registered to exert substantial main effects on various aspects of psychopathology and is known primarily through publications on GxE. As GWASes routinely do not test for interactions, perhaps it is not surprising that this polymorphism (or its imputed proxy) is not associated with phenotypic variation in these studies. Finally, often phenotypes used in GWASes and phenotypes used in GxE studies are quite different, even if they might be stated to tap into the same disease/disorder (e.g., depression). GWASes’ phenotypes tend to be substantially less detailed and elaborate than the phenotypes showcased in GxE studies. Correspondingly, criticisms that GWASes have not implicated specific candidate genes that have been featured in GxE studies have to be examined carefully with specific caveats (e.g., what polymorphisms and what phenotypes) in mind. The third concern is self-evident: the nonconvergence of findings from different methodologies is not encouraging

and seems to indicate lack of insight into the genetic mechanisms underlying complex human diseases/disorders, at least at the level of formulating specific and verifiable hypotheses pertaining to candidate genetic polymorphisms and candidate genes (Hebebrand et al., 2010). It has been stated (Colhoun, McKeigue, & Davey Smith, 2003) that the overwhelming majority (up to 95%!) of main effect findings obtained in genetic association studies appear to be false positives. This estimate, under the assumption of statistical power between 10% and 90%, in turn, is translated to a prior probability of a true association being 0.3–3.0%. Furthermore, this estimate might well be inflated, as testing main effects demands less statistical power than testing GxE interaction effects, meaning the prior for the latter might be even lower than 0.3–3.0%. In light of these considerations, Duncan and Keller (2011) concluded that under the relatively expectant assumptions of a prior of 5% and power of 55% (which seem rather unrealistic, as small sample sizes are the norm rather than the exception among GxE studies), approximately 63% of positive findings are likely to represent Type I error. These researchers argued that if these assumptions are altered to be more realistic—i.e., a prior of 1% and statistical power of 10%—the false discovery rate is likely to be at 98%. This problem of dealing with false positive findings is quite familiar to epidemiologists, who for a long time and with little satisfaction have been engaged in detecting subtle effects, whether for environments (Taubes, 1995) or genes (Crow, 2011). Commenting on this situation, Clayton and McKeigue (2001, pp. 1357–1358) noted the following: If we could specify in advance that the effect of the environmental factor on disease risk would be restricted to a subgroup of individuals with a particular genotype, there would, of course, be a gain in power from testing only this subgroup for the effect of the environmental factor. In practice, such an extreme situation is unlikely to be frequently encountered in the study of complex diseases, and entails a level of knowledge of underlying biology that would probably render epidemiological studies redundant. In less extreme situations, and where previous knowledge is more limited, a combined test would need to be done for the main effect of environmental exposure and its interaction with genotype. Since such tests have multiple degrees of freedom, the gain in power is much reduced; indeed, power might even be lost.

Fourth, there is both conceptual and statistical uncertainty as to whether the terms, as defined, are really G, E, or GxE, particularly given the findings from earlier quantitative genetic studies (Burton, Tobin, & Hopper,

The Operationalization

2005). For example, it has been shown that if the environment, E, is separated into shared (common, C) and nonshared (unique, E) components, then the term GxC will inflate the estimate of G, whereas the term GxE will inflate the estimate of E (Molenaar, Boomsma, & Dolan, 1997). These uncertainties were brought to the forefront in GxE discussions: “Importantly, however, we must acknowledge an almost complete ignorance of the relevant gene–environment interactions—as data accumulate, causes that now seem to be environmental could turn out to be gene–environment interactions” (Hemminki et al., 2006, p. 961). Such interactions could erroneously inflate heritability estimates (Hemminki et al., 2006). A final concern is that both the genotypic frequency (i.e., the observed frequency of particular genotypes, including those associated with risk) and frequency of exposure (i.e., the observed frequency of encountering the risk environmental factor) are crucial for discovery as well as replication (Caspi, Hariri, Holmes, Uher, & Moffitt, 2010), as the same GxE phenomenon might (1) not manifest when the prevalence of exposure is very low; (2) manifest via statistical interaction when the prevalence is moderate; and (3) manifest via main effect when the prevalence is very high. To maximize statistical power, it is recommended that the distribution of genotypes and exposure within a given sample follow a so-called balanced design, when rates of both derived (i.e., minor) allele frequencies and exposure are at ∼50%. However, such a balance is often unrealistic, especially when utilizing a case–control design or considering more than one polymorphism. Moreover, the number and nature of subgroups resulting from the joint distribution of genotypes and environment is often unknown. Facing such a situation, researchers may be tempted (Flint & Munafò, 2008) to search for the best outcome, exhausting all analytic possibilities in a drive to register nominal statistical significance, in response to publication bias toward positive results (Ioannidis & Trikalinos, 2007). As the number of subgroups and sub-subgroups is large, so is the number of comparisons (Ioannidis, 2006; Patsopoulos, Tatsioni, & Ioannidis, 2007). Such comparisons have been used, for example, in situations when the original effect of GxE failed to be replicated in the whole sample, but was registered to operate in particular subgroups (Eley et al., 2004). Yet, extensive subgrouping is highly susceptible to false positive findings (Brookes et al., 2001). For the brave who are willing and able to take on GxE studies in spite of the above caveats, specific recommendations have been offered (Moffitt, 2005). These recommendations include seven distinct steps.

297

The first step assumes a survey of the existing quantitative-genetic literature. It is always helpful if there are studies in which the specific GxE interaction in question has been evaluated already and deemed to be substantive and important, particularly if there is an explicit biological mechanism that has been coupled with this statistically significant interaction. If the relevant publications have not explicitly tested for GxE, particular attention should be given to the estimates of G (in particular, the additive genetic component, A) and E (in particular, the nonshared component of E). As discussed earlier (Molenaar et al., 1997), if GxE is not explicitly modeled, GxC can look like G (A) and GxE can look like E (E). Therefore, moderate-to-large estimates of A and E can signify the presence of potentially large GxE interaction effects (Purcell, 2002). The second step is to identify a candidate environmental factor, the exposure to which is known to have a (preferably strong) main effect on the phenotype (trait, behavior, disease, or disorder) in question. To illustrate such a factor, Moffitt (2005) referenced early maltreatment for antisocial behavior, arguing that the former is particularly relevant through its association with biological correlates of the latter (DeBellis, 2001), although the behavior itself has been associated with multiple environmental risks (Loeber & Farrington, 1998). As a substep, Moffitt urged researchers to ensure the environmental mediation of the selected risk factor (for an illustation, see Fujisawa, Yamagata, Ozaki, & Ando, 2012). Otherwise, the selection of the environmental risk factor may be misguided, as it might capture the interaction between different genotypes rather than between genotype and environment (which could, in fact, be an issue with the maltreatment—antisocial behavior connection, Schulz-Heik et al., 2007). The third step provides recommendations for the selection of specific measurements to be used to capture the exposure. Indeed, minimizing measurement error is associated with enhanced power (Luan, Wong, Day, & Wareham, 2001; Wong, Day, Luan, & Wareham, 2004). However, Moffitt’s reference to the possibility of related reduction of sample size in situations when near flawless assessments of exposure are used is misguided. At the fourth step, attention is switched from environments to genes. The main recommendation here is to stay in touch with the literature, as the knowledge of various gene–behavior associations advances rapidly, through the addition of new candidates and the elimination of old candidates (via nonreplication) from the list (Insel & Collins, 2003). Other recommendations are aligned with those that are present in many writings on GxE, and are discussed in

298

The Trilogy of GxE: Conceptualization, Operationalization, and Application

this chapter. Specifically, the genotypes of interest should be relatively prevalent in the general population. Both high and low prevalence of genotypes of interest, although they may be reflective of particular evolutionary dynamics (Hill, 1999; Searle & Blackwell, 1999), can generate various statistical biases (Lachance & Tishkoff, 2013; Vinkhuyzen, Wray, Yang, Goddard, & Visscher, 2013). Similar to the expectations for exposure, it is desirable that the candidate genetic factor be previously implicated as exerting a main effect on the phenotype of interest, although this expectation may not be realistic, given the general landscape of nonreplicability of such main effects. Finally, it is always recommended to look for a candidate that permits the formulation of a biologically plausible hypothesis as a counterpart to the statistical hypotheses. In illustrating these points, Moffitt references the polymorphism in the promoter region of the serotonin transporter (see Illustrations), arguing that the plausibility of the selection of this candidate polymorphism was substantiated by evidence from studies in psychopathology (Caspi et al., 2003), animal models (Bennett et al., 2002; Murphy et al., 2001), and human brain imaging research (Hariri et al., 2002). She also references both experimental research (e.g., Sayette, Griffin, & Sayers, 2010) and large-scale efforts documenting the range of responses of different genotypes to various environmental risks (Kaiser, 2003) as particularly important sources of nomination of candidates for GxE studies. The fifth step is the statistical test itself. Although a reference is made, in passing, to a variety of designs used in the field (Moffitt, Caspi, & Rutter, 2005; Ottman, 1990; van Os & Sham, 2003; Yang & Khoury, 1997), the representative population-based cohort is endorsed as the most informative. It is argued that this type of design allows not only an appraisal of the presence or absence of the interaction (by the statistical test of GxE), but also an evaluation of the magnitude of this interaction. Of note is that such population-based cohorts require large sample sizes in order to capture a full distribution of E and provide enough power for various statistical tests. The sixth step is to ensure the specificity and robustness of the registered effect by exploring the model via substitution of different candidates, both genetic and environmental. It is argued that, although it is vital to be hypothesisdriven in setting up the evaluation of GxE, having registered it, it is important to evaluate the original hypothesis among other plausible hypotheses (Licinio, 2003). The seventh and final step calls for replication, although it is not specified whether this request pertains to replication within the same research effort (i.e., with a different sample by the same investigator), by different investigators

on the same data, or some other form of replication. This recommendation acknowledges the tentative nature of an isolated GxE discovery but argues that its presentation (whether the interaction is ultimately true or not) should trigger both attempts at replication and collateral research. So far, we have presented the literature on GxE focusing on the emergence and history of the concept and its operationalization. In the next part of this chapter, we will summarize the literature through the lens of the current state of affairs in the field of GxE. THE ANALYSES Almost 15 years of intensive research into GxE have generated many published reports, which, in turn, have provided the foundation for meta-analyses, systematic reviews, targeted literature reviews, and opinion pieces (e.g., Calati, Gressier, Balestri, & Serretti, 2013; Decoster, van Os, Myin-Germeys, De Hert, & van Winkel, 2012; Duncan & Keller, 2011; Eaves, 2006; Flint & Munafò, 2008; Gressier et al., 2013; Karg, Burmeister, Shedden, & Sen, 2011; Keller, 2014; Modinos et al., 2013; Munafò, Durrant, Lewis, & Flint, 2009; Munafò & Flint, 2009; Risch et al., 2009; Uher & McGuffin, 2008). In general, the field is in a curious state. On one hand, the number of empirical reports on GxE, whether positive or negative, has been growing both overall and annually, as exemplified by studies of the specific polymorphism in the promoter region of the serotonin transporter gene. Figure 8.1 captures the studies discussed in the following illustrations and listed in the Appendix; the x axis shows the number of publications, and the y axis shows the year of publication. In other words, this research is still widely funded, which motivates researchers to engage with it and 2012 2011 2010 2009 2008 2007 2006 2005 2004 2003 0

5

10

15

20

25

30

35

40

45

50

Figure 8.1 Number of publications GxE, by publication year. See footnote 1.

The Analyses

publish extensively on it. On the other hand, meta-analyses and literature reviews on GxE have generated results and comments that are unabashedly skeptical. It seems that the larger the literature grows, the more incredulity it generates; yet, it still keeps growing! This incredulity stems from concerns pertaining both to the content (definitional) and the formal (analytical) aspects of this literature. As the former has been discussed already, this section of the chapter will focus on the latter. As for analytical concerns, observations have been made with regard to the following. First, en masse, the literature on GxE is characterized by a low replication rate (Kaufman, Gelernter, Kaffman, Caspi, & Moffitt, 2010; Munafò, Durrant, Lewis, & Flint, 2010). Second, although low even in the published literature, the replication rate is, probably, even lower among all attempts to replicate due to publication bias toward positive findings but not null findings. That is, it is impossible to know how many attempts at replication have failed and therefore have not been published (i.e., the file drawer problem), and the specific magnitude of this capitalization is difficult to appraise (Munafò & Flint, 2009). In fact, it has been argued that the false discovery rate in the GxE literature is substantially higher than the Type I error rate of .05 (Duncan & Keller, 2011; Flint & Munafò, 2008) commonly utilized in inferential statistics. Third, as a whole, studies conducted on small samples are common in this literature, thereby adding additional complications associated with insufficient statistical power to detect effects, especially statistical interactions. This can, somewhat counterintuitively, serve to inflate the rate of false-positive results reported in the literature, as null findings are less likely to be published in small-sample underpowered studies relative to the likelihood that positive findings are reported in such samples, whereas the likelihood of reporting both null and positive findings is more equal in sufficiently powered, large sample studies (Burmeister et al., 2008). In essence, there is an expectation that the fate of GxE findings will be similar to that of GWAS findings (Murcray, Lewinger, & Gauderman, 2009) when the GWAS field leapt from small- to large-sample studies and was not able to replicate the majority of its previously celebrated findings (Bosker et al., 2011; Collins, Kim, Sklar, O’Donovan, & Sullivan, 2012; Need et al., 2009; Sanders et al., 2008; Sullivan et al., 2008). Although such a situation is anticipated, we certainly hope that it will not transpire. Yet if it does, it is important to understand why it has transpired and, moreover, be as accurate and comprehensive as possible in appraising the findings that are in the literature on GxE. In the next part of the chapter, we consider a number

299

of characteristics of GxE studies that may explain why successful replications have proven difficult to realize. Measurement Error As discussed already, GxE studies are highly susceptible to measurement error in the assessment of genetic (Wong et al., 2004) and environmental exposure (Caspi et al., 2010; Luan et al., 2001) indicators. The former is viewed as less threatening, as quality control for genetic data is typically set at a threshold of 1% or less, and error per se can be quantified exactly by genotyping the same individuals on the same markers twice. The latter is thought to be much more concerning, as the magnitude of measurement error in exposure can be large, especially if captured by retrospective self-reports. Moreover, if measurement error around the exposure variable is high, even relatively small genotyping errors can result in a discernible impact on interaction estimates. In turn, poor measurement leads to a substantial loss of information. For example, it has been demonstrated that the mode of measurement of exposure (e.g., stressful life events) can have an overpowering effect on indicators of both frequency of occurrence and predictive power of this exposure (Monroe & Reid, 2008), which can bias the results of the interaction regression. Thus, care must be exercised in determining how indicators of exposure are ascertained. There is a direct connection between reducing measurement error and improving the statistical power of a study, which suggests that minimizing measurement error may be a more cost effective alternative in conducting GxE analyses than increasing the sample size. In quantifying environment or exposure, it is suggested that precise objective measures be used, such as environmental sampling or observational measures and experimental techniques capturing biological indicators of stress, rather than subjective reports. It has been argued that smaller GxE studies tend to use these higher precision prospective measures, whereas larger ones tend to use lower precision retrospective reports (Caspi et al., 2010; Lotrich & Lenze, 2009). While this may be true, sufficient sample size for a study depends on a number of indicators, measurement error being only one of them. Even with the use of these more precise measurement tools, small studies can still be underpowered, and therefore more susceptible to either inflated false positive rates or publication biases, than larger ones, and it is generally difficult to justify a certain sample size without the use of comprehensive statistical power calculations. It is important to note that the presence of measurement error does not signify only one type of bias. The problem

300

The Trilogy of GxE: Conceptualization, Operationalization, and Application

is that if either factor (G or E) has been measured with error, the relation between them will be altered and biased toward a multiplicative model (Clayton & McKeigue, 2001). Specifically, Clayton and McKeigue illustrated that if the presence of interaction is established through lack of fit to a multiplicative model, the test for interaction will be conservative if there is measurement error, such that if the null hypothesis is correct, the test will not yield significant results more often than expected by chance. If any other definition of interaction is used, the bias of the test (conservative or liberal) in the presence of measurement error is difficult to predict. There are ways to correct for measurement error in environmental variables (Thomas, Stram, & Dwyer, 1993), but these corrections typically require clear ideas of what the errors are, how they arise, and what their time-based properties are. Of note, in large-scale studies, exposures might not even be measured at the individual level, but rather, may be assigned based on some other information indicating ecologic-level exposures (e.g., exposure to natural disaster as a mere fact of having a registered address in the area where the disaster occurred, even though a given person might or might have not been in the area at the moment of the disaster), or obtained from another prediction model. Such uncertainties can lead to the manifestation of unpredictable biases, which may be especially detrimental to the model’s accuracy if these biases are differential with respect to the phenotype. As a result, spurious interactions can be introduced (Holmans et al., 2009). It is worth noting that even though there are methods of correction for measurement errors in indicators of G and E that are well established in studies testing for main effects, they have rarely been used in studying GxE (Lobach, Carroll, Spinka, Gail, & Chatterjee, 2008; Wong, Day, Luan, & Wareham, 2004). Nevertheless, interactions are less likely to be biased than main effects, except when the measurement errors are differentially associated with both exposure and genotype and, as a result, the measurement error is not equal for G and E factors (Thomas, 2010a). Confounders Yet another relevant issue that has been comprehensively addressed in the recent literature (Keller, 2014) pertains to controlling for potential confounders. The general point that an interaction term of interest should be adjusted for the effects of confounding variables has been put to force in other behavioral sciences (Hull, Tedlie, & Lehn, 1992; Yzerbyt, Muller, & Judd, 2004), but has not penetrated the field of GxE research. In its typical form, the analytical

facet of GxE studies includes three variables—the genetic polymorphism (typically dummy-coded), the environmental factor (typically either continuous or somehow categorized), and the interaction effect (typically captured by the product of the two main effect variables). These variables are then placed in a linear or logistic regression with a predicted variable of conceptual interest (e.g., delinquent behavior, academic achievement, mental health indicator). Keller (2014) and others (Hebebrand et al., 2010) rightfully noticed, however, that there are additional variables—race/ethnicity, gender, age, socioeconomic status, education, IQ, among many—known to be predictive of outcomes of interest, as main effects or members of other interaction terms, that are often treated by GxE researchers as noise to be controlled for. To that end, these potentially important variables are either residualized out prior to fitting the regression, or else they are included as covariates. Keller, however, argued that the proper account for the confounding effects of these additional variables can be achieved only if the full factorial model is tested, specifically, if all of the covariate × environment and covariate × gene interaction terms are tested for in the same model where the GxE interaction is featured. For example, to adjust the GxE term for gender and general cognitive ability, the regression should include the following terms in addition to the two main effects (of G and E), and their interaction (GxE): the main effects of the covariates—gender, general cognitive ability, and the interaction of the covariates with both G and E— gender × G, gender × E, general cognitive ability × G, and general cognitive ability × E. Keller (2014) attributed his view on how best to adjust for confounding variables to Yzerbyt and colleagues (2004) but provided an application of this general solution to GxE models specifically. He also elaborated on the nature of related biases, and illustrated the profound gap between his expectations for adjustments to avoid confounding and the selected GxE literature he reviewed—in fact, not a single study that was featured as novel by Duncan and Keller (2011) met Keller’s expectation. Although such a lack of proper statistical treatment, on its own, might not mean that the previously published GxE findings were not real, it does generate uncertainty about them (Keller, 2014). In all fairness, Keller (2014) acknowledged that the recommended treatment could be objected to, specifically, via references to model overfitting and multicollinearity. The pros and cons related to either approach are discussed in the broader epidemiological and other related literature (e.g., Chen et al., 2008; Kalil, Mattei, Florescu, Sun, & Kalil, 2010; Kim, Watkinson, & Anastassiou, 2011; Mizushima,

The Analyses

Tsuchida, & Yamori, 1999; Singh, Repsilber, Liebscher, Taher, & Fuellen, 2013), with no unequivocal outcome.

TABLE 8.1 Contingency table of genotype (geno) and environmental exposure (env) in simulated dataset.

Scaling There is a debate in the literature on how GxE interaction effects should be scaled. In psychology, the presence of interaction is typically detected via adding an interaction (product) term to a regression model and establishing the statistical significance of this term. In linear regression analyses, the product term’s coefficient captures the degree of deviation from the additivity of main effects on the outcome (Kendler & Gardner, 2010). In logistic regression analyses, the product term’s coefficient captures the degree of departure from multiplicativity in the risk of the outcome (Knol, van der Tweel, Grobbee, Numans, & Geerlings, 2007). Although it has been argued (Darroch, 1997) that the additive model of interaction is preferable, in fact, the literature is replete with examples of both types of regression analyses. Because the field of GxE interaction, particularly as delineated in this chapter, is defined first and foremost by the statistical conceptualization of the interaction, making inferences regarding the mechanistic or biological relationships between the predictors and outcomes requires adherence to a set of conditions and assumptions (VanderWeele, Hernández-Díaz, & Hernán, 2010). One of the issues relates to the scaling of the interaction effects (Kendler & Gardner, 2010). Thus, simple scale transformations, e.g., logarithmic transformations that are commonly used for normalization purposes, can yield a statistically significant although spurious interaction, whereas bona fide interactions can disappear (Eaves, 2006; Kendler & Gardner, 2010; Thompson, 1991). The issue of scaling in the GxE field has been (Rothman et al., 1980) and remains (Eaves, 2006; Rothman & Greenland, 2005) disputed without an adequate resolution. To gain a more concrete appreciation for how the concept of scaling removes interactions in a GxE study, we performed two simple simulation studies. Simulation 1 We started by generating genotypes and environmental exposures for 500 participants. Genotypes, using the variable name geno, were sampled randomly from the set {0,1}, with 75% probability of a 1 and 25% probability of a 0. This represented a very simple case of a single gene under a dominant heritability model, in which the gene is captured by two alleles, B and b, with B being dominant over b. The genotypes BB and Bb were assigned the value 1, and the genotype bb was assigned the value 0.

301

env geno 0 1

0 69 172

1 60 199

We then generated an environment, with the variable name env, as a simple yes-no exposure to an environmental factor, with a 0 (no exposure) 50% of the time, and a 1 (exposure) 50% of the time. We assumed environmental exposure to be independent of genotype as both variables were generated independently. Table 8.1 is the contingency table showing the randomly generated data. This indicates that of the sample of 500 participants, there were 371 participants with either BB or Bb genotypes and 129 participants with the bb genotype; there were 241 unexposed and 259 exposed individuals. Of the BB and Bb participants, 172 were unexposed and 199 were exposed to the environmental factor; of the bb carriers, 69 were unexposed and 172 were exposed to the factor. Next, we created a continuous (noncategorical) phenotype from the genotype and environmental exposure of each participant. We defined the genotype effect, g, and the environment effect, e, as the contribution of each feature to the phenotype, and assigned them values of 2 and 1, respectively. We then defined phenotype as an additive function of genotype and environment with the expected value of phenotype for each subject, given the genotype and environment, is g*geno + e*env. Table 8.2, the contingency table of means was thus. Here we created an additive model. To do so, we added the same fixed value (in this case 2) to every column, going from the first row to the second row; and, we added the same fixed value (1) to each row, going from the first column to the second column. The actual phenotype of each participant was the expected mean for that participant, plus some randomly generated (Gaussian) noise. As we started with a standard deviation for a noise level of 0.05, the following TABLE 8.2 Table of expected means of the phenotype for each genotype/environment combination. env geno 0 1

0 0 2

1 1 3

302

The Trilogy of GxE: Conceptualization, Operationalization, and Application Histogram of pheno

40 0

0

10

20

20

30

Frequency

60 40

Frequency

50

80

60

70

100

Histogram of pheno

0.0

0.5

1.0

1.5 pheno

2.0

2.5

3.0

0

2

distribution of phenotype values was produced, as indicated in Figure 8.2. In this case, with a small level of noise in relation to the magnitudes of the G and E effects (indicated here by g and e, respectively), the values of genotype and phenotype were quite obvious for each participant. In the second step of our simulation, we increased the standard deviation of the noise to 1 so that the statistical problem was not so trivial. This increase in noise changed the distribution of phenotypes, as indicated in Figure 8.3.

Figure 8.4 of 0.2.

Histogram of phenotypes using a standard deviation

This increase in noise relative to the effect sizes (G and E) caused the distribution of phenotype values to resemble a continuous spectrum. In contrast to the previous example, the genotype and phenotype values were no longer clear. As a third simulation (Figure 8.4), we established a noise level somewhere in the middle, e.g., 0.2. In this scenario, the distribution of phenotypes was different enough to estimate the specific genotype and environment effects, as shown below. However, the issues related to noise were not trivial, even in this circumstance. Consider the average value of the phenotype for each genotype–environment value (Table 8.3). These data are consistent with the parameters used to construct the model. This model is an additive model with no interactions between gene and environment, because it was constructed that way. However, in practice, one never knows what the true underlying model is: one might have a good idea of what it should be, based on prior evidence, but all a researcher really has is the phenotype, genotype, and environment values and a guess for the model. We then guessed that we had an additive model with a GxE interaction, and used

20

40

60

80

Histogram of pheno

0

TABLE 8.3 Table of observed means of the phenotype for each genotype/environment combination in simulated data.

−2

0

2 pheno

3

pheno

Figure 8.2 Histogram of simulated phenotypes using a standard deviation of 0.05.

Frequency

1

4

6

Figure 8.3 Histogram of simulated phenotypes using a standard deviation of 1.

env geno 0 1

0 −0.0197 2.0133

1 1.0338 3.0059

The Analyses TABLE 8.4 Table of observed means of the phenotype for each genotype/environment combination in simulated data, after scaling the phenotypes by an exponential transformation.

303

200

Histogram of pheno2

0

regression to estimate the effects (Table 8.4); that is, we regressed pheno on geno, env, and geno X env. R, gives the following output:

100

1 11.82 1140.07

Frequency

0 1.04 113.97

50

geno 0 1

150

env

Call : lm(pheno ~ geno + env + geno:env) Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -0.01970 0.02487 -0.792 0.429 geno 2.03303 0.02883 70.519 |z|) (Intercept) -1.1273 0.6101 -1.848 0.064651 . g -1.4386 0.3931 -3.659 0.000253 *** e -0.5886 0.2946 -1.997 0.045772 * g:e 0.7752 0.1704 4.549 5.38e-06 ***

So was there an interaction in these examples, or not? As we have seen, a case could conceivably be made for both answers in both examples. However, if an explanation involving interaction relies on a particular scaling of the variables, and another reasonable scaling exists in which the interaction disappears and a simple additive model fits well, one might well prefer the simple additive model. In the first simulation, a very simple, ubiquitous transformation—taking logarithms—was sufficient to transform a multiplicative process that appeared to have a strong interaction to an additive process with no interaction. In the second example, a simple binary transformation generated an appearance of GxE, even though the main effects of G and E on an underlying quantitative trait (the liability) were purely additive. So the liability provided a simple additive description of the disease process with no GxE interaction, but as the liability is a latent variable it may require more ingenuity or sophisticated methods to uncover this simple description. These simulations raise a more general concern pertaining to the definition of GxE interaction (Flint & Munafò, 2008). Indeed, given that statistical interactions are susceptible to scaling effects, the issue is how useful this term actually is, especially if the null hypothesis, however formulated, has no discernible biological meaning. Types of Interactions It is assumed that only 1–2% of cancer syndromes can be explained by inherited cancer syndromes of high penetrance (Ponder, 2001). In fact, the population-attributable fractions of known environmental factors are considered to be up to 90% for cancer syndromes (Doll & Peto, 1981; Higginson, 1968). Similarly, the population-attributable fractions of known environmental factors are considered to be up to 70% for coronary heart disease, stroke, and type 2 diabetes (Willett, 2002). However, recent understanding of the related causation reflects a number of complexities. Specifically, it appears that some environmental factors

The Analyses

(e.g., tobacco smoking and asbestos exposure for lung cancer) interact and increase risk multiplicatively, not additively (IARC, 1990, 2004). Moreover, there appears to be a tremendous amount of individual variation in how environmental exposures are converted into cellular mechanisms and the role of genetic factors (e.g., carcinogen metabolism, DNA repair, cell-cycle control and apoptosis) in this conversion (Vogelstein & Kinzler, 2002). Thus, what were previously considered straight environmental factors might, in fact, reflect the co-action of G and E, which leads to the etiological conclusion that most diseases/disorders are interactions of G and E or multiple Gs and multiple Es (Guttmacher, Collins, & Carmona, 2004). Multiple dimensions have been discussed in the literature that allows the classification of GxE interactions. Here we exemplify some of these dimensions. First, Ottman (1990, 1996) discussed a number of biologically plausible types of relationships between G and E with regard to their differential effects on disease risk. In the first model, exemplified by PKU, the effect of G is to generate or magnify the role of a risk factor, which can also be generated by E. In the second model, G exacerbates the effect of the risk factor but has no effect if the person is not exposed to this factor (E). This model is exemplified by the relationship between an autosomal recessive disorder, Xeroderma Pigmentosum, and ultraviolet (UV) radiation with regard to skin cancer. Although, in the general population, excessive exposure to UV radiation increases risk for skin cancer, this exposure is substantially riskier (and, thus, results in an elevated odds ratio) for individuals with this disorder, as they are deficient in an enzyme required for the repair of DNA damage induced by UV radiation. Per the third model, E exacerbates the effect of G, but not in individuals with the low-risk genotype. For example, individuals with the autosomal dominant disorder, Porphyria Variegata, develop skin problems of different severity (i.e., excessive blistering, scarring, changes in pigmentation under exposure to sunlight). Although an exposure to barbiturates is inoffensive in the general population, such individuals respond to the same exposure with acute attacks that might result in paralysis or even death. In the fourth model, both G and E risk factors are required to increase risk. To illustrate, most individuals with an X-linked recessive disorder glucose-6-phosphate dehydrogenase (G6PD) deficiency are asymptomatic; yet the consumption of fava beans (an ingredient widely used by the general population) by these individuals might result in the development of severe hemolytic anemia. Finally, in the fifth model, both G and E risk factors have

307

some unique impact on the disease/disorder risk so that their co-occurrence either elevates or decreases the risk compared to their occurrence in isolation. An illustration of this model comes from the development of chronic obstructive pulmonary disease (COPD) in the context of a-1-antitrypsin deficiency (i.e., an inherited disorder causing dysfunction of the lungs and liver) and smoking; in fact, risk of COPD is elevated both in nonsmokers with a-1-antitrypsin deficiency and in smokers without a-1-antitrypsin deficiency, but is particularly increased in smokers with a-1-antitrypsin deficiency. Second, as in epidemiology (Gail & Simon, 1985), two types of interactions are differentiated in GxE research. Crossover (qualitative) interactions are stated to transpire when a particular level of a factor (either G or E) is superior for some subset (or subsets) of the sample, whereas a different level of a factor is superior for other subsets. Noncrossover (quantitative) interactions are said to manifest when there is variation in the magnitude, but not the direction of the effect. The theoretical and practical values of these interactions are different, with the former being associated with higher significance and the latter with lower significance. The interactions also differ with regard to associated methodological vulnerabilities, such as sample size and power as well as high rate of false-positive effects (Bogdan, Agrawal, Gaffrey, Tillman, & Luby, 2014). Third, another important typology pertains to the differentiation of essential and removable interactions (Wu et al., 2009). This differentiation occurred in the accrual of data from GWAS. While working with SNP × SNP interactions of different orders, these two types of interactions were defined such that an interaction is essential when the direction (and, possibly, but not critically, the magnitude) of the effect for at least one of the SNPs is changed in the presence of the other SNP (or SNPs). The interaction is removable when only the magnitude (but not the direction) of the effect of at least one SNP is changed in the presence of the other SNP (SNPs). This differentiation and systematic screening of all possible interactions is a chance to detect more interesting and stronger effects (Chen, Liu, Zhang, & Zhang, 2007; Marchini, Donnelly, & Cardon, 2005). The corresponding number of interactions, however, even limiting the scope of consideration to two-way interactions, is staggering, with a count of a million and more. The importance of such constellations of interactions is obvious as most candidate gene studies are embedded in conceptual models featuring a specific biochemical pathway (or often multiple pathways), including more than one

308

The Trilogy of GxE: Conceptualization, Operationalization, and Application

gene and more than one polymorphism. To illustrate, the Southern California Children’s Health Study, investigating the impact of air pollution on children’s health in general and the manifestation of asthma in particular, is based on a theory engaging such etiological factors as inflammation, oxidative stress, and anti-oxidant intake (Gilliland, McConnell, Peters, & Gong, 1999). Similar reasoning is applicable to environmental or to a mixture of environmental and genetic factors; although the original study of GxE in cannabis usage (Caspi et al., 2005) tested for a direct interaction, subsequent studies tested for three-way (Henquet et al., 2009; Henquet et al., 2006) and four-way (Peerbooms et al., 2012) interactions. The translation of this theoretical model into specific hypotheses assumed the simultaneous investigation of multiple genes representing such pathways. Indeed, polymorphisms in these genes, genes themselves, and pathways can interact; these interactions should be properly modelled and analyzed. Furthermore, these genetic interactions can be layered with interactions with specific environmental factors. There are various methods that are suitable for multivariate analysis of high-dimensional data. These methods include standard multiple regression techniques, various machine learning, and pattern recognition methods (Cordell, 2009; Hoh, Wille, & Ott, 2001; McKinney, Reif, Ritchie, & Moore, 2006; Moore & Williams, 2009; Ritchie & Motsinger, 2005). GxE Study Designs To explore main and interaction effects of genes and environments, studies of GxE utilize conventional epidemiological designs such as cohort, case–control, or a hybrid of the two (e.g., case–cohort; Andrieu & Goldstein, 1998; Manolio et al., 2006; Yang & Khoury, 1997). Just as in epidemiological research, the selection among designs is driven by weighing their strengths and weaknesses vis-à-vis factors such as biases and confounding variables, temporal sequences of exposure and disease, data accessibility and quality, and capacity to investigate common and rare diseases, disorders, and risk factors (Thomas, 2010a). In addition, the literature contains a specific line of discourse pertaining to the utilization of traditional epidemiological designs specifically for the purposes of GxE studies (e.g., Collins, 2004; Manolio et al., 2006). As of today, the most present design in the GxE literature is that of case–control. In this design, a sample of carefully chosen people with (cases) and without (controls) the disease/disorder should be ascertained for a specific primary outcome so that similarities and differences between

the two groups can be investigated with regard to the distributions of genetic (e.g., frequencies of specific polymorphisms) and environmental (e.g., frequencies of specific exposures) factors (Gordis, 2000). In other words, these studies aim at investigating all individuals who are cases of disease/disorder, or are a representative sample of cases compared with a representative sample of all individuals who are free of disease/disorder. The literature acknowledges many advantages of case–control studies, specifically their relative ease of administration and low costs, their suitability for studies of rare diseases/disorders, and their capacity to sample multiple exposures retrospectively, maximizing the success of identifying true risk factors. The majority of case–control studies are retrospective; thus, although they ascertain cases after the onset of the disease/disorder, they collect information about genetic and environmental risk factors that predates the onset of the disorder and, in so doing, make a priori assumptions about causality (Doll, 2002). It follows that they are exposed to multiple sources of bias. Indeed, as the corresponding literature has accumulated, the shortcomings of case–control studies have become a limiting factor in the utilization of this design. Among many such shortcomings, the most relevant to this discussion are the following: 1. The tendency for individuals with positive family history to participate at higher odds (Bhatti et al., 2005; Wang, Fridinger, Sheedy, & Khoury, 2001), which biases sample structure in a particular way 2. The tendency for clinically diagnosed cases to represent the most severe tail of the distribution (Guo, 1998), which biases the prevalence-incidence estimates (Neyman, 1955) and overlooks specific cases such as short-episode or fatal cases (Taube, 1968) 3. The difficulty for undiagnosed controls to constitute a bias-free group (Schlesselman, 1982; Wacholder, Silverman, McLaughlin, & Mandel, 1992) and the degree of comparability (by geographic and ethnical ancestry and by predominant environmental exposures) of cases and controls (Helgason, Yngvadottir, Hrafnkelsson, Gulcher, & Stefansson, 2005; Rosenberg, Li, Ward, & Pritchard, 2003) 4. The difficulty of accurately documenting exposure, as most of these studies rely on recalled rather than evidenced exposures to environmental (Feinstein, 1985) or genetic (Silberberg, Wlodarczyk, Fryer, Ray, & Hensley) risk factors. In realizing the magnitude of these biases and the difficulties associated with the qualification and quantification of risk at the population level (Austin, Hill, Flanders, &

The Analyses

Greenberg, 1994; Hill, 1965) even when specific adjustments for these biases can be applied (e.g., Ben-Shlomo, Smith, Shipley, & Marmot, 1993), it has been argued that the estimates of risk at the population level is best obtained through prospective, population-based cohort studies (Gordis, 2000). A number of arguments are put forth when the advantages of cohort studies are discussed (e.g., Manolio et al., 2006). First, in contrast to case–control studies, prospective cohort studies utilize representative samples of the population before disease/disorder onset. The underlying idea here is to follow a representative sample from before, throughout, and after specified time points (Manolio, 2003). These time points can be defined in a number of ways, e.g., as bracketing the age at onset of a particular disease/disorder (e.g., type 1 diabetes) or as developmental stages (e.g., infancy, childhood, adolescence, adulthood). The main aim of this design is to ascertain, in the population as a whole rather than among already affected individuals, risk factors for the manifestation of the disease/disorder or biomarkers for the disease/disorder’s development. Thus, reduction of many related types of bias is the chief consideration for choosing prospective cohort design over case–control design. Second, cohort studies are particularly important for understanding the etiology and course of diseases/disorders with regard to investigating risk factors that are subject to recall biases (Langholz, Rothman, Wacholder, & Thomas, 1999). Third, prospective cohort design allows a comprehensive and standardized collection of various indicators of premorbid exposure in accord with the main objectives of the study. The problem of recall bias is not relevant, as exposure information is collected prior to the onset of the disease/disorder (Colman & Jones, 2004). Fourth, all members of the cohort are recruited and followed in a systematic way, so that the resulting sample is truly representative and all types of cases of disease/disorder are marked by equal probability of detection. Thereby, the case identification bias that is so problematic in case–control studies is minimized (Manolio et al., 2006). Fifth, unlike the case with case–control studies, multiple diseases/disorders can be studied simultaneously and the time window for disease/disorder onset can be established more precisely (Manolio et al., 2006). Prospective cohort studies impose a number of requirements (Manolio et al., 2006). First, it is assumed that individuals ascertained into the cohort are characterized by similar genetic (e.g., ancestry) and environmental (e.g., dietary preferences) factors that are distinct from those who are not included in the cohort. Second, it is assumed

309

that participants who are excluded due to attrition are similar to the remaining participants with regard to disease/disorder risks, both genetic as well as environmental. Third, inclusion/exclusion criteria, recruitment, and definition of outcomes should be unified for all members of the cohort. In other words, to avoid biases and ensure the similarity of data collection between cohort members who are and are not exposed to risk factors, whether genetic or environmental, it is assumed that the probability of disease/disorder diagnosis is independent of the exposure to the environmental risk factors, as well as of potentially confounding factors such as age, access to care, and other critical exposures. It is particularly important to document and track changes in exposure history; thus, exposure information should be collected repeatedly (Zeger, Liang, & Albert, 1998). Fourth, it is assumed that all members of the cohort are systematically evaluated for the occurrence of diseases/disorders. The critical feature of prospective cohort studies is that all cohort members have equal probability for the detection and diagnoses of diseases/disorders, regardless of their access to medical care. Therefore, cohort studies cannot rely on the identification of outcomes in the course of everyday clinical care and must embed regular evaluations of the cohort participants into the study procedures. Such evaluations, given their time-consuming and resource-heavy implementation, are the target of criticism of cohort studies. Fifth, and most importantly, it is assumed that prospective cohort studies adequately capture both incidence and accumulation of diseases/disorders and, thus, are characterized by large sample sizes. Sixth, it is assumed that these studies provide an opportunity to comprehensively sample risk factors of interest prior to the onset of cases. To summarize, it has been argued that prospective cohort studies (Manolio et al., 2006) are particularly suited to (1) studying the full range of disease/disorder manifestations (e.g., diseases/disorders with high mortality at onset like pancreatic cancer or with a long preclinical phase such as type 2 diabetes; Collins, 2004); (2) the identification of predictive biomarkers manifesting prior to the clinical presentation of the disease/disorder (Langholz & Goldstein, 1996); (3) the identification of risk factors that transform after the onset of disease/disorder due to treatment, change in lifestyle, or imperfect or biased recall (Colman & Jones, 2004); (4) the investigation of common complex diseases/disorders of a polygenetic nature (Foster & Sharp, 2005); (5) the simultaneous investigation of multiple outcomes (ARIC Investigators, 1989; Colditz, Manson, & Hankinson, 1997; Kolonel et al., 2000; Leibowitz et al., 1980; Lloyd-Jones, Larson, Beiser, & Levy, 1999; Newman

310

The Trilogy of GxE: Conceptualization, Operationalization, and Application

et al., 2006; Troyer, Mubiru, Leach, & Naylor, 2004; Tsai et al., 2002; Women’s Health Initiative Study Group, 1998); (6) confirmation and extension of findings obtained by other means—i.e., via other designs, such as case–control studies (Aleksic et al., 2002; Ellenberg & Nelson, 1980; Kannel, 1995) and dispelling misconceptions (Kannel, 1995; Stamler, 1991); and (7) providing, with adequate protection, wide access to data and samples for analysis and reanalysis (Marshall, 1997). There are multiple examples of large-scale cohort studies that contribute to the world’s general understanding of the epidemiology of health and disease/disorders, including studies in the United Kingdom (Foster & Sharp, 2005; Harvey, Matthews, Collins, Cooper, & U.K. Biobank Musculoskeletal Advisory Group, 2013; Pramanik et al., 2012; Swanson, 2012; Ul-Haq et al., 2014), Iceland (Winickoff, 2001), Germany (Aleksic, Jahn, Heckenkamp, Wielckens, & Brunkwall, 2005; Wichmann, 2005; Wichmann & Gieger, 2007), Sweden (Abbott, 1999), Canada (Gibson et al., 2008; Godard, Ozdemir, Fortin, & Egalite, 2010; Kosseim et al., 2013; Webster, 2008), France (Goldberg & Zins, 2014; Spira, 2014), and Japan (Yuasa & Kishi, 2009). These samples are of considerable size. For example, the UK BioBank, with a case population of 10,000, is considered to be adequately powered to detect risks of 4,000, Surtees et al., 2006), they represent exceptions to the rule, and the average size of GxE samples is much smaller. For example, for the 18 studies of the serotonin transporter, the average sample size was 600 (Uher & McGuffin, 2008). The general rule is that if the interaction is of the same magnitude as the main effect and the power is maintained at the same level, the sample size for a study of GxE must be increased fourfold (Brookes et al., 2001; Smith & Day, 1984). This ratio, however, changes dramatically (to 100 or greater) if subtler ( 1,000) was examined with the following inclusion-exclusion criteria. The publications were included if the study examined an interaction between the serotonin transporter gene promoter polymorphism and an environmental exposure indicator deemed relevant to developmental psychopathology as a predictor of a psychological/psychiatric outcome (some of these studies also examined additional variants in other genes). The articles were excluded if (1) exposure indicators were obstetric complications, environmental toxins, diet, physical exercise, medical conditions, or hormonal intervention; (2) no

The Analyses

exposure indicator was identified and the interaction term included G and such factors as indicators of temperament, personality factors, or other psychiatric disorders (e.g., depression, anxiety, addiction); (3) outcome variables were other (nonpsychiatric) medical diseases or disorders (e.g., Crohn’s disease, obesity); (4) the sample included related individuals of any kind (twins, other sibling pairs, or any other family members); and (5) more than one variant in the SLC6A4 gene was investigated (e.g., haplotypes were analyzed). The application of inclusion–exclusion criteria resulted in the selection of 192 articles (see Appendix). These articles were reviewed to extrapolate answers to a number of specific questions systematically. In general, the intent of the analyses presented here was to see whether this literature contains indications of the utilization of the best practices in the field (i.e., regarding observations and recommendations discussed in this chapter) in the selected publications. Thus, both the extrapolations and interpretations are based on the descriptive analyses of these studies as a group. The first set of questions dealt with the sample size and the design of the analyzed studies. The sample size ranged from n = 24 to n = 4,334. Importantly, indicators of central tendency were quite variable: mean = 594.5, median = 301, mode = 118 (sd = 769.8). Taken together, this set of studies was still quite far from what is recommended as the standard in the field in terms of sample size. The majority of the samples were characterized cross sectionally (n = 123, 64.1%), but a substantial number of samples (n = 69, 35.9%) contained longitudinal data. All studies utilized a version of the traditional case–control design, even if there were previously collected longitudinal data. Second, the studies were diverse in terms of utilizing samples including only women (n = 27, 14.1%), only men (n = 9, 4.7%), or both genders (n = 156, 81.3%). The studies were also diverse in terms of including different races and ethnicities. The third block of questions pertained to issues of measurement error, both in G and E variables, however defined. In general, details of measurement were not well explicated in this set of studies, although many contained conventional psychometric indicators of the assessments that were utilized either to measure the outcomes or the exposure (E). The presentation of the measurement approach toward G was predominantly characterized by missing information with regard to the error rate for genotyping: only 34 studies (17.7%) reported error rates. Furthermore, only a handful of studies reflected on whether these rates met the expectations established in the field and discussed possible related biases. Notably, very

317

few publications contained information pertaining to the presence of missing data and the process by which this missing data was accounted for (n = 28, 14.6%). Fourth, only a very small proportion of outcome variables (6.8%) were transformed. The fifth set of questions had to do with the range of outcomes, exposures, and polymorphisms (see exclusion and inclusion criteria above) utilized in the articles. All three were observed to present considerable ranges: 1–59 for outcomes (mean = 2.7, median = 2, mode = 1, sd = 4.7), 1–35 for exposures (mean = 2.2, median = 1, mode = 1, sd = 2.9), and 1–14 for polymorphisms (mean = 1.8, median = 1, mode = 1, sd = 2.0). Importantly, only 23 studies (12%) exercised any type of correction for multiple comparisons. The most frequently used was the Bonferroni, with other methods such as false discovery rate (FDR) applied only rarely. Sixth, among the statistical techniques utilized for testing interaction effects, preferences were given to ANOVA/ANCOVA or MANOVA/MANCOVA (∼20%) or different types of regression approaches (linear or logistic, ∼60%). Typically, the statistical software used was not specified. When specified, the software was not tailored for GxE analyses in the majority of cases, and, therefore, was not necessarily the best vehicle to analyze the data. Seventh, relatively few studies (n = 45, 23.4%) reported the obtained effect sizes and discussed their practical meaning. Curiously, there were journals that were particularly receptive to the GxE studies involving the serotonin transporter gene during the period of time between 2000 and 2012. Thus, Biological Psychiatry published 15 and Journal of Affective Disorders published 12 of the 192 articles commented on here. In this section, we summarized studies utilizing the statistical concept of GxE. Clearly, this interpretation of GxE attracts a lot of attention in the field and appears to generate studies, with the rate of published studies increasing over time. Interestingly, there appears to be a delayed reaction of the field to the criticisms that have been explicated in multiple reviews and meta-analyses. As this summary of the 192 articles demonstrates, the majority of the studies are still underpowered, have not paid enough attention to the issue of multiple comparisons, may not be sensitive enough to issues of measurement error, variable transformations and tracking confounders, and, most importantly, do not discuss the robustness, meaningfulness, or practical significance of the established interactions.

318

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Understanding the Reaction Range in the Acquisition of Academic Skills It is impossible, in modern society, to overestimate the impact of key academic skills such as reading, mathematics, and science reasoning on subsequent life success and life outcomes. This impact is omnipresent worldwide, and the effectiveness of country-specific primary and secondary educational systems is judged, in part, by how well students score on international competitions, such as the OECD Programme for International Student Assessment (PISA). In the twenty-first century, PISA has become “the world’s premier yardstick for evaluating the quality, equity and efficiency of school systems in providing young people with these skills” (http://www.oecd.org/pisa/keyfindings/pisa-2012-resultsoverview.pdf, p. 2). While the importance of these skills is not disputed, there are two additional considerations that are also rather axiomatic at this stage. First, it is accepted, with the exception of a few rare developmental trajectories (i.e., extreme giftedness or circumscribed skills manifested in certain types of atypical development like autism), that all academic skills are constructed by a child as he or she experiences schooling. In other words, a child acquiring these skills builds a number of new cognitive (i.e., brain-based) representations that will guide the processing of relevant information. The formation of these representations requires a functional reorganization of the brain. Second, it is also known that, within any educational system (whether it is high-scoring such as Chinese, Taiwanese, or Finnish; mid-range-scoring such as German, Spanish, Russian, or American; or low-scoring such as Mexican or Indian, as per a number of PISA cycles), a tremendous amount of individual difference scatters student performance across the continuum. The presence of individual differences has been of interest to researchers around the world. As schooling was put in place only in the first half of the last century, the research into the student-based variation in learning that occurs even when teaching is homogenized has a history of only about 50 years. Yet these 50 years of research have unequivocally established that the efficiency of mastering the key academic skills of reading, mathematics, and scientific reasoning is tightly connected to variation in both the brain (in terms of its processing of the relevant information) and the genome (both structurally and functionally). Thus, the understanding of both the acquisition of, and individual variation in, core academic skills requires the investigation of how structurally diverse genomes assume the task of reorganizing the neural connections of the brain through schooling to acquire the academic skills

of reading, mathematics, and scientific reasoning under the pressure of particular educational systems. In other words, this understanding requires a differentiation of the mechanism (or mechanisms) that delivers the environment (teaching or schooling) under the skin, so that a biological machinery (i.e., within a child) that ensures both the initial acquisition of skills and their further automatization and that explains individual differences in both the acquisition and utilization of these skills may unfold. It is argued here that one such mechanism is that of epigenetic regulation in general and DNA methylation in particular. Five bodies of literature substantiate this argument. First, the literature substantiates a high degree of genetic control in both the acquisition and maintenance of these three academic skills (reading, mathematics, and scientific reasoning), with some variation across skills. The literature also offers evidence that, while the degree of genetic control seems to be stronger in individuals who exhibit low performance on the academic tasks that utilize these skills, there is no reason to believe that the mechanisms of genetic control for poorer performers are different from the mechanisms of genetic control for stronger performers. In fact, it appears that whatever these mechanisms are, they operate across the range of academic performance (Elliott & Grigorenko, 2014). Second, similar to other complex phenotypes, academic skills such as reading and reading-related components are highly heritable. However the currently considered candidate genes and corresponding polymorphisms do not account for meaningful portions of the previously estimated heritability. This phenomenon, which applies to these specific phenotypes among many others, has been referred to as the missing heritability problem. Concordantly, although limited, there is growing literature on the importance of GxE for the acquisition of academic skills (Pennington et al., 2009; Rosenberg, Pennington, Willcutt, & Olson, 2012; Taylor et al., 2010). Third, there are large literatures (Fields, 2011), although uneven for the three skills, substantiating the presence of distinct distributed brain signatures that differentiate individuals who (1) are engaged in particular tasks requiring skills of reading, mathematics, and scientific reasoning as compared to individuals who are engaged in other types of tasks (or when the same individual is engaged in different types of tasks); (2) have acquired the skills versus those who have not (i.e., either when comparing young children who have not yet begun their formal education to older children who have completed

Conclusions

of primary education, or adults who have corresponding functional skills against those who do not have them, for example, literate and illiterate adults, or adults who can and cannot perform operations involving quantitative or scientific reasoning); and (3) perform within the normal range (i.e., within a particular quantification, whether 1, 1.5, or 2—whatever the criterion dictates—standard deviations around the population mean) compared with those who perform outside the normal range (i.e., outside of the established criterion, as defined already). Fourth, there is a growing body of research indicating the role of epigenetic mechanisms in general and DNA methylation mechanisms in particular in all types of learning (Levenson & Sweatt, 2005). This research, however, has been conducted predominantly on animal models. The human epigenetic literature, however, has assessed the role of this mechanism only in social learning, and there is not a single study that investigates the role of this mechanism in cognitive/academic-related learning to date. Fifth, there is now a strengthening line of reasoning which connects the literatures on the missing heritability problem, GxE studies, and epigenetics (Slatkin, 2009). Thus, we argue that the acquisition of academic skills, all of which are both heritable and sensitive to environmental exposure (i.e., pedagogy), exemplifies the biological concept of GxE. As was the case 100 years ago, there is substantially less research on this biological interpretation of GxE compared to its statistical interpretation, but the reasoning presented in this chapter can and certainly should be translated into testable hypotheses. FUTURE DIRECTIONS Three observations appear to be instrumental in summarizing the material presented in this chapter in light of future developments. First, there is a critical mass of literature, spread across multiple fields of inquiry (epidemiology, genetics, psychology, and other fields) that indicates that designing, implementing, and interpreting GxE studies require conscientious consideration of a number of methodological caveats. Future studies of GxE should be judged being both aware and proactive about these caveats. The field is too advanced now to forgive any naïveté in those who wish to practice in it. Second, methodological rigor should be expected not only of publications, but also of proposals for GxE studies. In other words, funders should be educated along with researchers on the pros and cons of GxE studies. Finally, the whole field should

319

carefully consider the practical implementations of GxE research. Even when carried out as rigorously as possible, what can the field learn from this research and at what cost? Answers to these questions should surely frame the future of GxE research.

CONCLUSIONS There is no question that GxE studies have made an impact on the fields of developmental psychopathology, neuropsychiatric genetics, and genetic epidemiology, among others. As is obvious from the discourse above, this impact is complex and has instigated numerous kinds of actions and reactions. First, there are hopes. The appeal of studying GxE is in the possible applications of the most reliable findings (yet to be secured!). Given the state of the field, it is difficult to predict what findings will be deemed reliable and what applications might be derived from these findings. Specifically, such findings can form a foundation—or at least lead to a set of guidelines—for targeting interventions for individuals at high risk (Khoury & Wagener, 1995). Three types of such guidelines have been discussed. The first type of recommendation (following the established the presence of the statistical GxE interaction) involves avoiding specific exposure. For example, researchers (Vandenbroucke et al., 1994) investigated whether the occurrence of venous thrombosis in young women who use oral contraceptives might be explained by the factor V Leiden mutation (this mutation results in resistance to activated protein C and, therefore, increases susceptibility to thrombosis). The reported differential increases in risk (Vandenbroucke et al., 1994) were four-fold among users of oral contraceptives, eight-fold among carriers of the mutation,, and 30-fold among carriers who used contraceptives. Clearly, carriers should consider alternative methods of contraception. The second type of recommendation involves forming relevant public-health policies. For example, in a study in Rwanda (Kolassa et al., 2010), individuals with extremely high levels of trauma exposure were stratified based on their genotypes at the serotonin transporter gene promoter site. It was reported that the individuals with the ss genotype were marked by a higher chance of manifesting lifetime PTSD regardless of the number of traumatic experiences, whereas the individuals with the l allele (i.e., ll or sl) had an elevated chance of developing lifetime PTSD only when the number of traumatic experiences was elevated. In the Florida Hurricane Study, individuals with the s allele were characterized by (1) the elevated

320

The Trilogy of GxE: Conceptualization, Operationalization, and Application

risk of PTSD only when their exposure to hurricane was higher and their social support was lower, compared to the rest of the sample (Kilpatrick et al., 2007); and (2) the decreased risk of PTSD in low-risk environments, but with an increased risk in high-risk environments (Koenen et al., 2009). Assuming that the multiple inconsistencies of GxE studies may be resolved and the robust findings replicated, relevant policies can be established to diversify services for individuals in disaster zones. The third type of recommendation involves seeking and receiving particular interventions in accordance with a particular genotype. In general, it has been stated (Baird, 2001; Rose, 1992) that in the realm of public health, the predicted health gains are superior in situations when the whole population, not just the high-risk group is targeted. This superiority is justified (Wallace, 2006), in particular, by the relative unimportance of genetic mechanisms in the risk of common complex disorders/diseases (e.g., lung cancer); the complexity of various types of interactions, GxG and GxE (e.g., schizophrenia); and the simplicity of the representation of environmental exposure, as typically captured by a single environmental factor. At the present time, there is no convincing support for the idea that delivering interventions based on individual genotypes improves the desired outcome. Yet, there are hopes that such examples will appear (Benner et al., 2014; Tucker-Drob & Harden, 2012), as, at least in developmental psychopathology, it has been suggested that “differential susceptibility may ideally lead to differential intervention and thus more effective treatment” (van IJzendoorn et al., 2011, p. 50). There is also hope that the biological interpretation of GxE will help the field connect currently unconnected dots (e.g., to understand in detail the acquisition of highly heritable skills, for example, academic skills, which require a specific type of environmental exposure for each individual). Second, there are precautions. In particular, two scientific journals, Behavior Genetics (Hewitt, 2012) and the Journal of Abnormal Child Psychology (Johnston, Lahey, & Matthys, 2013), have established requirements to be met before manuscripts presenting candidate–gene main or interaction effects can be considered for review. These requirements reflect many of the caveats of the field that we have discussed in this chapter. Thus, 15 years of GxE studies have taught the field to be particularly sensitive to the methodological aspects of research in general and the reproducibility of the results in particular. Third, there are commentaries. This chapter was conceived as a constellation of observations from the literature on GxE, which is large and continues to grow, regardless of precautions. Although hopes for GxE remain high, so far,

their realization has, in general, not risen to the level of our expectations. If anything, this discrepancy, as well as the numerous issues discussed in this chapter, “provides significant reason to pause for reflection” (Eaves, 2006, p. 1).

REFERENCES Abbott, A. (1999). Sweden sets ethical standards for use of genetic “biobanks.” Nature, 400, 3. Albert, P. S., Ratnasinghe, D., Tangrea, J., & Wacholder, S. (2001). Limitations of the case-only design for identifying gene–environment interactions. American Journal of Epidemiology, 154, 687–693. Aleksic, M., Jahn, P., Heckenkamp, J., Wielckens, K., & Brunkwall, J. (2005). Comparison of the prevalence of APC-resistance in vascular patients and in a normal population cohort in Western Germany. European Journal of Vascular & Endovascular Surgery, 30, 160–163. Aleksic, N., Ahn, C., Wang, Y.-W., Juneja, H., Folsom, A. R., Boerwinkle, E., & Wu, K. K. (2002). Factor XIIIA Val34Leu polymorphism does not predict risk of coronary heart disease: The Atherosclerosis Risk in Communities (ARIC) Study. Arteriosclerosis, Thrombosis, and Vascular Biology, 22, 348–352. doi: 10.1161/hq0202.102874 Altshuler, D., Hirschhorn, J. N., Klannemark, M., Lindgren, C. M., Vohl, M. C., Nemesh, J., . . . Lander, E. S. (2000). The common PPARgamma Pro12Ala polymorphism is associated with decreased risk of type 2 diabetes. Nature Genetics, 26, 76–80. Andrieu, N., & Demenais, F. (1997). Interactions between genetic and reproductive factors in breast cancer risk in a French family sample. American Journal of Human Genetics, 61, 678–690. Andrieu, N., & Goldstein, A. M. (1998). Epidemiologic and genetic approaches in the study of gene–environment interaction: An overview of available methods. Epidemiologic Reviews, 20, 137–147. Andrieu, N., Goldstein, A. M., Thomas, D. C., & Langholz, B. (2001). Counter-matching in studies of gene–environment interaction: Efficiency and feasibility. American Journal of Epidemiology, 153, 265–274. ARIC Investigators. (1989). The atherosclerosis risk in communities (ARIC) study: Design and objectives. American Journal of Epidemiology, 129, 687–702. Aron, E. N., & Aron, A. (1997). Sensory-processing sensitivity and its relation to introversion and emotionality. Journal of Personality and Social Psychology, 73, 345–368. Austin, H., Hill, H. A., Flanders, W. D., & Greenberg, R. S. (1994). Limitations in the application of case–control methodology. Epidemiologic Reviews, 16, 65–76. Baird, P. (2001). The human genome project, genetics and health. Community Genetics, 4, 77–80. Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2006). Gene–environment interaction of the dopamine D4 receptor (DRD4) and observed maternal insensitivity predicting externalizing behavior in preschoolers. Developmental Psychobiology, 48, 406–409. Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2011). Differential susceptibility to rearing environment depending on dopamine-related genes: New evidence and a meta-analysis. Development & Psychopathology, 23, 39–52. Barbour, V. (2003). UK Biobank: a project in search of a protocol? Lancet, 361, 1734–1738. Belsky, J. (1997). Variation in susceptibility to environmental influence: An evolutionary argument. Psychological Inquiry, 8, 182–185.

References Belsky, J. (2014). Commentary: Beyond stressful life events and depression? Reflections on Bogdan et al. Journal of Child Psychology and Psychiatry, 55, 458–459. doi: 10.1111/jcpp. 12238 Belsky, J., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2007). For better and for worse: Differential susceptibility to environmental influences. Current Directions in Psychological Science, 16, 300–304. doi: 10.1111/j.1467-8721.2007.00525.x Belsky, J., & Pluess, M. (2009). Beyond diathesis stress: Differential susceptibility to environmental influences. Psychological Bulletin, 135, 885–908. Belsky, J., Steinberg, L., Houts, R. M., Halpern-Felsher, B. L., & the NICHD Early Child Care Research Network. (2010). The development of reproductive strategy in females: Early maternal harshness—Earlier menarche—Increased sexual risk-taking. Developmental Psychology, 46, 120–128. doi: 10.1037/a0015549 Ben-Shlomo, Y., Smith, G. D., Shipley, M., & Marmot, M. G. (1993). Magnitude and causes of mortality differences between married and unmarried men. Journal of Epidemiology and Community Health, 47, 200–205. Benner, A. D., Kretsch, N., Harden, K. P., & Crosnoe, R. (2014). Academic achievement as a moderator of genetic influences on alcohol use in adolescence. Developmental Psychology, 50, 1170–1178. Bennett, A. J., Lesch, K. P., Heils, A., Long, J. C., Lorenz, J. G., Shoaf, S. E., . . . Higley, J. D. (2002). Early experience and serotonin transporter gene variation interact to influence primate CNS function. Molecular Psychiatry, 7, 118–122. Bhatti, P., Sigurdson, A. J., Wang, S. S., Chen, J., Rothman, N., Hartge, P., . . . Landi, M. T. (2005). Genetic variation and willingness to participate in epidemiologic research: Data from three studies. Cancer Epidemiology Biomarkers & Prevention, 14, 2449–2453. doi: 10.1158/1055-9965.epi-05-0463 Bickeböller, H., Houwing-Duistermaat, J. J., Wang, X., & Yan, X. (2011). Dealing with high dimensionality for the identification of common and rare variants as main effects and for gene–environment interaction. Genetic Epidemiology, 35, S35–S40. Boardman, J. D. (2009). State-level moderation of genetic tendencies to smoke. American Journal of Public Health, 99, 480–486. Boardman, J. D., Saint Onge, J. M., Haberstick, B. C., Timberlake, D. S., & Hewitt, J. K. (2008). Do schools moderate the genetic determinants of smoking? Behavior Genetics, 38, 234–246. Bogdan, R., Agrawal, A., Gaffrey, M. S., Tillman, R., & Luby, J. L. (2014). Serotonin transporter-linked polymorphic region (5-HTTLPR) genotype and stressful life events interact to predict preschool-onset depression: A replication and developmental extension. Journal of Child Psychology and Psychiatry, 55, 448–457. doi: 10.1111 /jcpp. 12142 Boks, M. P., Schipper, M., Schubart, C. D., Sommer, I. E., Kahn, R. S., & Ophoff, R. A. (2007). Investigating gene environment interaction in complex diseases: Increasing power by selective sampling for environmental exposure. International Journal of Epidemiology, 36, 1363–1369. Boomsma, D., Busjahn, A., & Peltonen, L. (2002). Classical twin studies and beyond. Nature Reviews Genetics, 3, 872–882. Bos, G., Dekker, J. M., Feskens, E. J. M., Ocke, M. C., Nijpels, G., Stehouwer, C. D. A., . . . Jansen, H. (2005). Interactions of dietary fat intake and the hepatic lipase –480C→T polymorphism in determining hepatic lipase activity: the Hoorn Study. American Journal of Clinical Nutrition, 81, 911–915. Bosker, F. J., Hartman, C. A., Nolte, I. M., Prins, B. P., Terpstra, P., Posthuma, D., . . . Nolen, W. A. (2011). Poor replication of candidate genes for major depressive disorder using genome-wide association data. Molecular Psychiatry, 16, 516–532.

321

Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary–developmental theory of the origins and function of stress reactivity. Development and Psychopathology, 17, 271–301. Boyce, W. T., Sokolowski, M. B., & Robinson, G. E. (2012). Toward a new biology of social adversity. Proceedings of the National Academy of Sciences, 109, 17143–17148. doi: 10.1073/pnas.1121264109 Breslow, N. E., & Chatterjee, N. (1999). Design and analysis of two-phase studies with binary outcome applied to Wilms tumor prognosis. Applied Statistics, 48, 457–468. Breslow, N. E., Lumley, T., Ballantyne, C. M., Chambless, L. E., & Kulich, M. (2009). Using the whole cohort in the analysis of case-cohort data. American Journal of Epidemiology, 169, 1398–1405. Bronfenbrenner, U., & Ceci, S. J. (1994). Nature–nurture reconceptualized in developmental perspective: A bioecological model. Psychological Review, 101, 568–586. doi: 10.1037/0033-295X.101.4.568 Brookes, S. T., Whitley, E., Peters, T. J., Mulheran, P. A., Egger, M., & Smith, G. D. (2001). Subgroup analyses in randomised controlled trials: Quantifying the risks of false-positives and false-negatives. Health Technology Assessment, 5, 1–56. Bryant, A. L., Schulenberg, J., Bachman, J. G., O’Malley, P. M., & Johnston, L. D. (2000). Understanding the links among school misbehavior, academic achievement, and cigarette use: A national panel study of adolescents. Prevention Science, 1, 71– 87. doi: 10.1023/A:1010038130788 Buchanan, A. V., Weiss, K. M., & Fullerton, S. M. (2006). Dissecting complex disease: The quest for the Philosopher’s Stone? International Journal of Epidemiology, 35, 562–571. Burmeister, M., McInnis, M. G., & Zöllner, S. (2008). Psychiatric genetics: Progress amid controversy. Nature Reviews Genetics, 9, 527–540. Burton, P. R., Hansell, A. L., Fortier, I., Manolio, T. A., Khoury, M. J., Little, J., & Elliott, P. (2009). Size matters: Just how big is BIG? Quantifying realistic sample size requirements for human genome epidemiology. International Journal of Epidemiology, 38, 263–273. Burton, P. R., Tobin, M., & Hopper, J. (2005). Key concepts in genetic epidemiology. Lancet, 366, 941–951. Calati, R., Gressier, F., Balestri, M., & Serretti, A. (2013). Genetic modulation of borderline personality disorder: Systematic review and meta-analysis. Journal of Psychiatric Research, 47, 1275–1287. doi: 10.1016/j.jpsychires.2013.06.002 Capen, E. C., Clapp, R. V., & Campbell, W. M. (1971). Competitive bidding in high-risk situations. Journal of Petroleum Technology, 23, 641–653. Caporaso, N. E., Gu, F., Chatterjee, N., Sheng-Chih, J., Yu, K., Yeager, M., . . . Bergen, A. W. (2009). Genome-wide and candidate gene association study of cigarette smoking behaviors. PLoS One, 4, e4653. doi: 10.1371/journal.pone.0004653 Carlson, M. D., Mendle, J., & Harden, K. P. (2014). Early adverse environments and genetic influences on age at first sex: Evidence for gene × environment interaction. Developmental Psychology, 50, 1532–1542. Caspi, A., Hariri, A. R., Holmes, A., Uher, R., & Moffitt, T. E. (2010). Genetic sensitivity to the environment: The case of the serotonin transporter gene and its implications for studying complex diseases and traits. American Journal of Psychiatry, 167, 509–527. Caspi, A., & Moffitt, T. E. (2006). Gene–environment interactions in psychiatry: Joining forces with neuroscience. Nature Reviews Neuroscience, 7, 583–590. Caspi, A., Moffitt, T. E., Cannon, M., McClay, J., Murray, R., Harrington, H., . . . Craig, I. W. (2005). Moderation of the effect of adolescent-onset cannabis use on adult psychosis by a functional polymorphism in the catechol-o-methyltransferase gene: Longitudinal evidence of a gene × environment interaction. Biological Psychiatry, 57, 1117–1127. doi: 10.1016/j.biopsych.2005.01.026

322

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . Poulton, R. (2003). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301, 386–389. Chakravarti, A., & Little, P. (2003). Nature, nurture and human disease. Nature, 421, 412–414. Chen, S. H., Sun, J., Dimitrov, L., Turner, A. R., Adams, T. S., Meyers, D. A., . . . Hsu, F. C. (2008). A support vector machine approach for detecting gene-gene interaction. Genetic Epidemiology, 32, 152–167. Chen, X., Liu, C. T., Zhang, M., & Zhang, H. (2007). A forest-based approach to identifying gene and gene gene interactions. PNAS, 104, 19199–19203. Cicchetti, D., Rogosch, F. A., Gunnar, M. R., & Toth, S. L. (2010). The differential impacts of early physical and sexual abuse and internalizing problems on daytime cortisol rhythm in school-aged children. Child Development, 81, 252–269. doi: 10.1111/j .1467-8624.2009.01393.x Clayton, D., & McKeigue, P. M. (2001). Epidemiological methods for studying genes and environmental factors in complex diseases. Lancet, 358, 1356–1360. Colditz, G. A., Manson, J. E., & Hankinson, S. E. (1997). The nurses’ health study: 20-year contribution to the understanding of health among women. Journal of Women’s Health, 6, 49–62. Colditz, G. A., Sellers, T. A., & Trapido, E. (2006). Epidemiology— Identifying the causes and preventability of cancer? Nature Reviews Cancer, 6, 75–83. Coley, R. L., & Chase-Lansdale, P. L. (1998). Adolescent pregnancy and parenthood: Recent evidence and future directions. American Psychologist, 53, 152–166. doi: 10.1037/0003-066X.53.2.152 Colhoun, H. M., McKeigue, P. M., & Davey Smith, G. (2003). Problems of reporting genetic associations with complex outcomes. Lancet, 361, 865–872. Collins, A. L., Kim, Y., Sklar, P., O’Donovan, M. C., & Sullivan, P. F. (2012). Hypothesis-driven candidate genes for schizophrenia compared to genome-wide association results. Psychological Medicine, 42, 607–616. doi: doi:10.1017/S0033291711001607 Collins, F. S. (2004). The case for a US prospective cohort study of genes and environment. Nature, 429, 475–477. Collins, F. S., & Manolio, T. A. (2007). Necessary but not sufficient. Nature, 445, 259. Colman, I., & Jones, P. B. (2004). Birth cohort studies in psychiatry: Beginning at the beginning. Psychological Medicine, 34, 1375–1383. doi: doi:10.1017/S0033291704003277 Cologne, J. B., Sharp, G. B., Neriishi, K., Verkasalo, P. K., Land, C. E., & Nakachi, K. (2004). Improving the efficiency of nested case–control studies of interaction by selecting controls using counter matching on exposure. International Journal of Epidemiology, 33, 485–492. Cordell, H. J. (2009). Detecting gene–gene interactions that underlie human diseases. Nature Reviews Genetics, 10, 392–404. Cox, A., Dunning, A. M., Garcia-Closas, M., Balasubramanian, S., Reed, M. W. R., Pooley, K. A., . . . Easton, D. F. (2007). A common coding variant in CASP8 is associated with breast cancer risk. Nature Genetics, 39, 352–358. doi: http://www.nature.com/ng/journal/v39/n3/ suppinfo/ng1981_S1.html Cox, D. R. (1984). Interaction. International Statistical Review, 52, 1–31. Crosnoe, R. (2006). The connection between academic failure and adolescent drinking in secondary school. Sociology of Education, 79, 44–60. doi: doi:10.1177/003804070607900103 Crow, T. J. (2011). The missing genes: What happened to the heritability of psychiatric disorders? Molecular Psychiatry, 16, 362–364. Cui, J. S., Spurdle, A. B., Southey, M. C., Dite, G. S., Venter, D. J., McCredie, M. R. E., . . . Hopper, J. L. (2003). Regressive logistic

and proportional hazards disease models for within-family analyses of measured genotypes, with application to a CYP17 polymorphism and breast cancer. Genetic Epidemiology, 24, 161–172. doi: 10.1002/gepi.10222 Dabelea, D., Hanson, R. L., Lindsay, R. S., Pettitt, D. J., Imperatore, G., Gabir, M. M., . . . Knowler, W. C. (2000). Intrauterine exposure to diabetes conveys risks for type 2 diabetes and obesity: A study of discordant sibships. Diabetes, 49, 2208–2211. Dai, J. Y., LeBlanc, M., & Kooperberg, C. (2009). Semiparametric estimation exploiting covariate independence in two-phase randomized trials. Biometrics, 65, 178–187. Darroch, J. (1997). Biologic synergism and parallelism. American Journal of Epidemiology, 145, 661–668. DeBellis, M. D. (2001). Developmental traumatology: The psychobiological development of maltreated children and its implications for research, treatment, and policy. Development and Psychopathology, 13, 539–564. Decoster, J., van Os, J., Myin-Germeys, I., De Hert, M., & van Winkel, R. (2012). Genetic variation underlying psychosis-inducing effects of cannabis: critical review and future directions. Current Pharmaceutical Design, 18, 5015–5023. Dick, D. M., Agrawal, A., Schuckit, M. A., Bierut, L., Hinrichs, A., Fox, L., . . . Begleiter, H. (2006). Marital status, alcohol dependence, and GABRA2: Evidence for gene–environment correlation and interaction. Journal of Studies on Alcohol, 67, 185–194. Dick, D. M., Rose, R. J., Viken, R. J., Kaprio, J., & Koskenvuo, M. (2001). Exploring gene–environment interactions: Socioregional moderation of alcohol use. Journal of Abnormal Psychology, 110, 625–632. Dick, D. M., Viken, R., Purcell, S., Kaprio, J., Pulkkinen, L., & Rose, R. J. (2007). Parental monitoring moderates the importance of genetic and environmental influences on adolescent smoking. Journal of Abnormal Psychology, 116, 213–218. Didelez, V., & Sheehan, N. (2007). Mendelian randomization as an instrumental variable approach to causal inference. Statistical Methods in Medical Research, 16, 309–330. Doll, R. (2002). Proof of causality. Perspectives in Biology and Medicine, 45, 499–515. Doll, R., & Peto, R. (1981). The causes of cancer. Journal of the National Cancer Institute, 66, 1191–1308. Duman, E. A., & Canli, T. (2010). Social behavior and serotonin. In C. P. Müller & B. L. Jacobs (Eds.), Handbook of behavioral neuroscience (Vol. 21, pp. 449–456). New York, NY: Elsevier. Duncan, L. E., & Keller, M. C. (2011). A critical review of the first ten years of candidate gene-by-environment interaction research in psychiatry. American Journal of Psychiatry, 168, 1041–1049. Eaves, L. J. (2006). Genotype × environment interaction in psychopathology: Fact or artifact? Twin Research and Human Genetics, 9, 1–8. Eaves, L. J., & Erkanli, A. (2003). Markov chain Monte Carlo approaches to analysis of genetic and environmental components of human development change and GxE interaction. Behavior Genetics, 33, 279–300. Egger, M., Smith, G. D., & Phillips, A. N. (1997). Meta-analysis: Principles and procedures. British Medical Journal, 315, 1533–1537. doi: 10.1136/bmj.315.7121.1533 Eley, T. C., Sugden, K., Corsico, A., Gregory, A. M., Sham, P., McGuffin, P., . . . Craig, I. W. (2004). Gene–environment interaction analysis of serotonin system markers with adolescent depression. Molecular Psychiatry, 10, 908–915. Ellenberg, J. H., & Nelson, K. B. (1980). Sample selection and the natural history of disease. Studies of febrile seizures. JAMA, 243, 1337–1340. Elliott, J. G., & Grigorenko, E. L. (2014). The dyslexia debate. New York, NY: Cambridge University Press.

References Ellis, B. J., Essex, M. J., & Boyce, W. T. (2005). Biological sensitivity to context: II. Empirical explorations of an evolutionary–developmental theory. Development and Psychopathology, 17, 303–328. doi: 10.1017/0S0954579405050157 Ellis, B. J., Jackson, J. J., & Boyce, W. T. (2006). The stress response systems: Universality and adaptive individual differences. Developmental Review, 26, 175–212. Ellis, B. J., Schlomer, G. L., Tilley, E. H., & Butler, E. A. (2012). Impact of fathers on risky sexual behavior in daughters: A genetically and environmentally controlled sibling study. Development and Psychopathology, 24, 317–332. doi: 10.1017/S095457941100085X Engelman, C. D., Baurley, J. W., Chiu, Y.-F., Joubert, B. R., Lewinger, J. P., Maenner, M. J., . . . Gauderman, W. J. (2009). Detecting gene–environment interactions in genome-wide association data. Genetic Epidemiology, 33, S68–S73. doi: 10.1002/gepi.20475 Evans, D. E., & Rothbart, M. K. (2007). Developing a model for adult temperament. Journal of Research in Personality, 41, 868–888. Feinstein, A. R. (1985). Experimental requirements and scientific principles in case–control studies. Journal of Chronic Diseases, 38, 127–133. Feldman, M. W., & Lewontin, R. C. (1975). The heritability hang-up. Science, 190, 1163–1168. Fields, R. D. (2011). Imaging learning: the search for a memory trace. Neuroscientist, 17, 185–196. Fisher, R. A. (1918). The correlation between relatives on the supposition of Mendelian inheritance. Transactions of the Royal Society of Edinburgh, 52, 399–433. Fisher, R. A. (1925). Statistical methods for research workers. Edinburgh, UK: Oliver and Boyd Ltd. Fisher, R. A., & Mackenzie, W. A. (1923). Studies in crop variation. II. The manurial response of different potato varieties. Journal of Agricultural Science, 13, 311–320. Flint, J., & Munafò, M. R. (2007). The endophenotype concept in psychiatric genetics. Psychological Medicine, 37, 163–180. Flint, J., & Munafò, M. R. (2008). Forum: Interactions between gene and environment. Current Opinion in Psychiatry, 21, 315–317. Foppa, I., & Spiegelman, D. (1997). Power and sample size calculations for case–control studies of gene–environment interactions with a polytomous exposure variable. American Journal of Epidemiology, 146, 596–604. Foster, M. W., & Sharp, R. R. (2005). Will investments in largescale prospective cohorts and biobanks limit our ability to discover weaker, less common genetic and environmental contributors to complex diseases? Environmental Health Perspectives, 113, 119–122. Franke, A., McGovern, D. P. B., Barrett, J. C., Wang, K., Radford-Smith, G. L., Ahmad, T., . . . Parkes, M. (2010). Genome-wide meta-analysis increases to 71 the number of confirmed Crohn’s disease susceptibility loci. Nat Genet, 42, 1118–1125. doi: http://www.nature.com/ ng/journal/v42/n12/abs/ng.717.html#supplementary-information Fujisawa, K. K., Yamagata, S., Ozaki, K., & Ando, J. (2012). Hyperactivity/inattention problems moderate environmental but not genetic mediation between negative parenting and conduct problems. Journal of Abnormal Child Psychology, 40, 189–200. doi: 10.1007/s10802-011-9559-6 Gail, M., & Simon, R. (1985). Testing for qualitative interactions between treatment effects and patient subsets. Biometrics, 41, 361–372. Garcia-Closas, M., & Lubin, J. H. (1999). Power and sample size calculations in case–control studies of gene–environment interactions: Comments on different approaches. American Journal of Epidemiology, 149, 689–692. Gauderman, W. J. (2002). Sample size requirements for matched case–control studies of gene–environment interaction. Statistics in Medicine, 21, 35–50.

323

Gauderman, W. J., & Faucett, C. L. (1997). Detection of gene–environment interactions in joint segregation and linkage analysis. American Journal of Human Genetics, 61, 1189–1199. Gauderman, W. J., & Siegmund, K. D. (2001). Gene–environment interaction and affected sib pair linkage analysis. Human Heredity, 52, 34–46. Gauderman, W. J., Witter, J., & Thomas, D. (1999). Family-based association studies. Journal of the National Cancer Institute Monographs, 26, 31–39. Ghio, A. J., Carraway, M. S., & Madden, M. C. (2011). Composition of air pollution particles and oxidative stress in cells, tissues, and living systems. Journal of Toxicology and Environmental Health, Part B, 15, 1–21. doi: 10.1080/10937404.2012.632359 Gibson, E., Brazil, K., Coughlin, M. D., Emerson, C., Fournier, F., Schwartz, L., . . . Willison, D. J. (2008). Who’s minding the shop? The role of Canadian research ethics boards in the creation and uses of registries and biobanks. BMC Medical Ethics, 9, 17. Gilliland, F. D., McConnell, R., Peters, J., & Gong, H. J. (1999). A theoretical basis for investigating ambient air pollution and children’s respiratory health. Environmental Health perspectives, 107, 403–407. Godard, B., Ozdemir, V., Fortin, M., & Egalite, N. (2010). Ethnocultural community leaders’ views and perceptions on biobanks and population specific genomic research: a qualitative research study. Public Understanding of Science, 19, 469–485. Goldberg, M., & Zins, M. (2014). [Population-based cohorts. Example of the Gazel and Constances cohorts]. Bulletin de l Academie Nationale de Medecine, 197, 299–311; discussion 311–313. Gordis, L. (2000). Epidemiology. Philadelphia, PA: W.B. Saunders & Co. Goring, H. H., Terwilliger, J. D., & Blangero, J. (2001). Large upward bias in estimation of locus-specific effects from genomewide scans. American Journal of Human Genetics, 69, 1357–1369. Gottweis, H., Chen, H., & Starkbaum, J. (2011). Biobanks and the phantom public. Human Genetics, 130, 433–440. Grant, S. F., Thorleifsson, G., Reynisdottir, I., Benediktsson, R., Manolescu, A., Sainz, J., . . . Stefansson, K. (2006). Variant of transcription factor 7-like 2 (TCF7L2) gene confers risk of type 2 diabetes. Nature Genetics, 38, 320–323. Gray, R., & Wheatley, K. (1991). How to avoid bias when comparing bone marrow transplantation with chemotherapy. Bone Marrow Transplant, 7, 9–12. Greene, C. S., Penrod, N. M., Williams, S. M., & Moore, J. H. (2009). Failure to replicate a genetic association may provide important clues about genetic architecture. PLoS One, 4, e5639. Greenland, S. (2000). An introduction to instrumental variables for epidemiologists. International Journal of Epidemiology, 29, 1102. Gressier, F., Calati, R., Balestri, M., Marsano, A., Alberti, S., Antypa, N., & Serretti, A. (2013). The 5-HTTLPR polymorphism and posttraumatic stress disorder: A meta-analysis. Journal of Traumatic Stress, 26, 645–653. doi: 10.1002/jts.21855 Guo, S. W. (1998). Inflation of sibling recurrence-risk ratio, due to ascertainment bias and/or overreporting. American Journal of Human Genetics, 63, 252–258. Guttmacher, A. E., Collins, F. S., & Carmona, R. H. (2004). The family history—More important than ever. New England Journal of Medicine, 351, 2333–2336. Haga, S. B., & Beskow, L. M. (2008). Ethical, legal, and social implications of biobanks for genetics research. In D. C. Rao & C. C. Gu (Eds.), Advances in genetics (Vol. 60, pp. 505–544). New York, NY: Academic Press. Halme, L., Paavola-Sakki, P., Turunen, M. L., Färkkilä, M., & Kontula, K. (2006). Family and twin studies in inflammatory bowel

324

The Trilogy of GxE: Conceptualization, Operationalization, and Application

disease. World Journal of Gastroenterology, 12, 3668–3672. doi: 10.3748/wjg.v12.i23.3668 Hamza, T. H., Chen, H., Hill-Burns, E. M., Rhodes, S. L., Montimurro, J., Kay, D. M., . . . Payami, H. (2011). Genome-wide gene–environment study identifies glutamate receptor gene GRIN2A as a Parkinson’s disease modifier gene via interaction with coffee. PLoS Genetics, 7, e1002237. doi: 10.1371/journal.pgen.1002237 Hariri, A. R., & Holmes, A. (2006). Genetics of emotional regulation: The role of the serotonin transporter in neural function. TRENDS in Cognitive sciences, 10, 182–191. doi: 10.1016/j.tics.2006.02.011 Hariri, A. R., Mattay, V. S., Tessitore, A., Kolachana, B., Fera, F., Goldman, D., . . . Weinberger, D. R. (2002). Serotonin transporter genetic variation and the response of the human amygdala. Science, 297, 400–404. Harvey, N. C., Matthews, P., Collins, R., Cooper, C., & U.K. Biobank Musculoskeletal Advisory Group. (2013). Osteoporosis epidemiology in UK Biobank: A unique opportunity for international researchers. Osteoporosis International, 24, 2903–2905. Hebebrand, J., Scherag, A., Schimmelmann, B. G., & Hinney, A. (2010). Child and adolescent psychiatric genetics. European Child & Adolescent Psychiatry, 19, 259–279. doi: 10.1007/s00787-010-0091-y Hein, R., Beckmann, L., & Chang-Claude, J. (2008). Sample size requirements for indirect association studies of gene–environment interactions (GXE). Genetic Epidemiology, 32, 235–245. Helgason, A., Yngvadottir, B., Hrafnkelsson, B., Gulcher, J., & Stefansson, K. (2005). An Icelandic example of the impact of population structure on association studies. Nature Genetics, 37, 90–95. Hemminki, K., Lorenzo Bermejo, J., & Forsti, A. (2006). The balance between heritable and environmental aetiology of human disease. Nature Reviews Genetics, 7, 958–965. ˇ Henquet, C., Rosa, A., Delespaul, P., Papiol, S., Fananás, L., Van Os, J., & Myin-Germeys, I. (2009). COMT Val158Met moderation of cannabis-induced psychosis: A momentary assessment study of “switching on” hallucinations in the flow of daily life. Acta Psychiatrica Scandinavica, 119, 156–160. doi: 10.1111/j.1600-0447 .2008.01265.x Henquet, C., Rosa, A., Krabbendam, L., Papiol, S., Fananas, L., Drukker, M., . . . van Os, J. (2006). An Experimental study of Catechol-O-Methyltransferase Val158Met moderation of [Delta]-9-Tetrahydrocannabinol-induced effects on psychosis and cognition. Neuropsychopharmacology, 31, 2748–2757. Hewitt, J. K. (2012). Editorial policy on candidate gene association and candidate gene-by-environment interaction studies of complex traits. Behavior Genetics, 42, 1–2. doi: 10.1007/s10519-011-9504-z Hicks, B. M., South, S. C., DiRago, A. C., Iacono, W. G., & McGue, M. (2009). Environmental adversity and increasing genetic risk for externalizing disorders. Archives of General Psychiatry, 66, 640–648. Higginson, J. (1968). Present trends in cancer epidemiology. Proceedings. Canadian Cancer Conference, 8, 40–75. Hill, A. B. (1965). The environment and disease: Association or causation? Proceedings of the Royal Society of Medicine, 58, 295–300. Hill, A. V. S. (1999). Genetics and genomics of infectious disease susceptibility. British Medical Bulletin, 55, 401–413. Hindorff, L. A., Sethupathy, P., Junkins, H. A., Ramos, E. M., Mehta, J. P., Collins, F. S., & Manolio, T. A. (2009). Potential etiologic and functional implications of genome-wide association loci for human diseases and traits. PNAS, 106, 9362–9367. Hirschhorn, J. N. (2009). Genomewide association studies—Illuminating biologic pathways. New England Journal of Medicine, 360, 1699–1701. Hoffmann, T. J., Vansteelandt, S., Lange, C., Silverman, E. K., DeMeo, D. L., & Laird, N. M. (2011). Combining disease models to test

for gene–environment interaction in nuclear families. Biometrics, 67, 1260–1270. Hogben, L. (1932). Genetic principles in medicine and social science. New York, NY: Alfred A. Knopf. Hoh, J., Wille, A., & Ott, J. (2001). Trimming, weighting, and grouping SNPs in human case–control association studies. Genome Research, 11, 2115–2119. Holmans, P., Green, E. K., Pahwa, J. S., Ferreira, M. A. R., Purcell, S. M., Sklar, P., . . . Craddock, N. (2009). Gene ontology analysis of GWA study data sets provides insights into the biology of bipolar disorder. American Journal of Human Genetics, 85, 13–24. doi: 10.1016/j.ajhg.2009.05.011 Homberg, J. R. (2012). The stress-coping (mis)match hypothesis for nature × nurture interactions. Brain Research, 1432, 114–121. doi: 10.1016/j.brainres.2011.11.037 Homberg, J. R., & Lesch, K.-P. (2011). Looking on the bright side of serotonin transporter gene variation. Biological Psychiatry, 69, 513–519. doi: 10.1016/j.biopsych.2010.09.024 Hu, X. Z., Lipsky, R. H., Zhu, G., Akhtar, L. A., Taubman, J., Greenberg, B. D., . . . Goldman, D. (2006). Serotonin transporter promoter gain-of-function genotypes are linked to obsessive-compulsive disorder. American Journal of Human Genetics, 78, 815–826. Hull, J. G., Tedlie, J. C., & Lehn, D. A. (1992). Moderator variables in personality research: The problem of controlling for plausible alternatives. Personality and Social Psychology, 18, 115–117. Hunter, D. J. (2005). Gene–environment interactions in human diseases. Nature Reviews Genetics, 6, 287–298. Hwang, S. J., Beaty, T. H., Liang, K. Y., Coresh, J., & Khoury, M. J. (1994). Minimum sample size estimation to detect gene–environment interaction in case–control designs. American Journal of Epidemiology, 140, 1029–1037. IARC. (1990). Cancer: Causes, occurrence and control. Lyon, France: IARC: International Agency for Research on Cancer. IARC. (2004). Tobacco smoke and involuntary smoking. Lyon, France: IARC: International Agency for Research on Cancer. Insel, T. R., & Collins, F. S. (2003). Psychiatry in the genomics era. American Journal of Psychiatry, 160, 616–620. Ioannidis, J. P. (2005). Why most published research findings are false. PLoS Medicine, 2, e124. Ioannidis, J. P. (2006). Common genetic variants for breast cancer: 32 largely refuted candidates and larger prospects. Journal of the National Cancer Institute, 98, 1350–1353. Ioannidis, J. P. (2007). Non-replication and inconsistency in the genome-wide association setting. Human Heredity, 64, 203–213. Ioannidis, J. P., & Trikalinos, T. A. (2007). An exploratory test for an excess of significant findings. Clinical Trials, 4, 245–253. Ioannidis, J. P., Trikalinos, T. A., & Khoury, M. J. (2006). Implications of small effect sizes of individual genetic variants on the design and interpretation of genetic association studies of complex diseases. American Journal of Epidemiology, 164, 609–614. Jaffee, S. R., Price, T. S., & Reyes, T. M. (2013). Behavior genetics: past, present, future. Development & Psychopathology, 25, 1225–1242. Jamrozik, K., Weller, D. P., & Heller, R. F. (2005). Biobank: who’d bank on it? Medical Journal of Australia, 182, 56–57. Jensen, A. R. (1969). How much can we boost IQ and scholastic achievement? Harvard Educational Review, 39, 1–123. Jensen, A. R. (1973). Educability and group differences. New York, NY: Harper & Row. Johnston, C., Lahey, B. B., & Matthys, W. (2013). Editorial policy for candidate gene studies. Journal of Abnormal Child Psychology, 41, 511–514. doi: 10.1007/s10802-013-9741-0

References Julier, C., Hyer, R. N., Davies, J., Merlin, F., Soularue, P., Briant, L., . . . Lathrop, G. M. (1991). Insulin-IGF2 region on chromosome 11p encodes a gene implicated in HLA-DR4-dependent diabetes susceptibility. Nature, 354, 155–159. Kaiser, J. (2002). Biobanks. Population databases boom, from Iceland to the U.S. Science, 298, 1158–1161. Kaiser, J. (2003). Toxicology: Tying genetics to the risk of environmental diseases. Science, 300, 563. Kalil, A. C., Mattei, J., Florescu, D. F., Sun, J., & Kalil, R. S. (2010). Recommendations for the assessment and reporting of multivariable logistic regression in transplantation literature. American Journal of Transplantation, 10, 1686–1694. Kannel, W. B. (1995). Clinical misconceptions dispelled by epidemiological research. Circulation, 92, 3350–3360. Kannel, W. B. (2000). The Framingham Study: its 50-year legacy and future promise. Journal of Atherosclerosis and Thrombosis, 6, 60–66. Karg, K., Burmeister, M., Shedden, K., & Sen, S. (2011). The serotonin transporter promoter variant (5-HTTLPR), stress, and depression meta-analysis revisited: Evidence of genetic moderation. Archives of General Psychiatry, 68, 444–454. doi: 10.1001/archgenpsychiatry .2010.189 Kaufman, J., Gelernter, J., Kaffman, A., Caspi, A., & Moffitt, T. E. (2010). Arguable assumptions, debatable conclusions. Biological Psychiatry, 67(4), e19–e20. doi: 10.1016/j.biopsych.2009.07.041 Keller, M. C. (2014). Gene × environment interaction studies have not properly controlled for potential confounders: the problem and the (simple) solution. Biological Psychiatry, 75, 18–24. Kendler, K. S., Chen, X., Dick, D., Maes, H., Gillespie, N., Neale, M. C., & Riley, B. (2012). Recent advances in the genetic epidemiology and molecular genetics of substance use disorders. Nature Neuroscience, 15, 181–189. doi: 10.1038/nn.3018 Kendler, K. S., Gardner, C., & Dick, D. M. (2011). Predicting alcohol consumption in adolescence from alcohol-specific and general externalizing genetic risk factors, key environmental exposures and their interaction. Psychological Medicine, 41, 1507–1516. Kendler, K. S., & Gardner, C. O. (2010). Interpretation of interactions: guide for the perplexed. British Journal of Psychiatry, 197, 170–171. doi: 10.1192/bjp.bp. 110.081331 Khoury, M. J. (2004). The case for a global human genome epidemiology initiative. Nature Genetics, 36, 1027–1028. Khoury, M. J., & Wacholder, S. (2009). Invited commentary: From genome-wide association studies to gene–environment-wide interaction studies—Challenges and opportunities. American Journal of Epidemiology, 169, 227–230. Khoury, M. J., & Wagener, D. K. (1995). Epidemiological evaluation of the use of genetics to improve the predictive value of disease risk factors. American Journal of Human Genetics, 56, 835–844. Kiecolt, J., Aggen, S. H., & Kendler, K. S. (2013). Genetic and environmental influences on the relationship between mastery and alcohol dependence. Alcoholism: Clinical and Experimental Research, 37, 905–913. doi: 10.1111/acer.12058 Kilpatrick, D. G., Koenen, K. C., Ruggiero, K. J., Acierno, R., Galea, S., Resnick, H. S., . . . Gelernter, J. (2007a). The serotonin transporter genotype and social support and moderation of posttraumatic stress disorder and depression in hurricane-exposed adults. American Journal of Psychiatry, 164, 1693–1699. doi: 10.1176/appi.ajp. 2007.06122007 Kim, D.-K., Tolliver, T. J., Huang, S.-J., Martin, B. J., Andrews, A. M., Wichems, C., . . . Murphy, D. L. (2005). Altered serotonin synthesis, turnover and dynamic regulation in multiple brain regions of mice lacking the serotonin transporter. Neuropharmacology, 49, 798–810. doi: 10.1016/j.neuropharm.2005.08.010

325

Kim, H., Watkinson, J., & Anastassiou, D. (2011). Biomarker discovery using statistically significant gene sets. Journal of Computational Biology, 18, 1329–1338. Kiser, D., Steemers, B., Branchi, I., & Homberg, J. R. (2012). The reciprocal interaction between serotonin and social behaviour. Neuroscience & Biobehavioral Reviews, 36, 786–798. doi: 10.1016/j.neubiorev.2011.12.009 Knodt, A. (2012). Imputation of microsatellite markers with tag SNPs. Durham, NC: Duke University. Knol, M. J., van der Tweel, I., Grobbee, D. E., Numans, M. E., & Geerlings, M. I. (2007). Estimating interaction on an additive scale between continuous determinants in a logistic regression model. International Journal of Epidemiology, 36, 1111–1118. doi: 10.1093/ije/dym157 Ko, Y.-L., Hsu, L.-A., Hsu, K.-H., Ko, Y.-H., & Lee, Y.-S. (2004). The interactive effects of hepatic lipase gene promoter polymorphisms with sex and obesity on high-density-lipoprotein cholesterol levels in Taiwanese–Chinese. Atherosclerosis, 172, 135–142. doi: 10.1016/j.atherosclerosis.2003.09.013 Kochanska, G., Askan, N., & Joy, M. E. (2007). Children’s fearfulness as a moderator of parenting in early socialization: Two longitudinal studies. Developmental Psychology, 43, 222–237. Koenen, K. C., Aiello, A. E., Bakshis, E., Amstadter, A. B., Ruggiero, K. J., Acierno, R., . . . Galea, S. (2009a). Modification of the association between serotonin transporter genotype and risk of posttraumatic stress disorder in adults by county-level social environment. American Journal of Epidemiology, 169, 704–711. doi: 10.1093/aje/kwn397 Kolassa, I. T., Ertl, V., Eckart, C., Glockner, F., Kolassa, S., Papassotiropoulos, A., . . . Elbert, T. (2010a). Association study of trauma load and SLC6A4 promoter polymorphism in posttraumatic stress disorder: Evidence from survivors of the Rwandan genocide. Journal of Clinical Psychiatry, 71, 543–547. doi: 10.4088/JCP.08m04787blu Kolonel, L. N., Henderson, B. E., Hankin, J. H., Nomura, A. M., Wilkens, L. R., Pike, M. C., . . . Nagamine, F. S. (2000). A multiethnic cohort in Hawaii and Los Angeles: Baseline characteristics. American Journal of Epidemiology, 151, 346–357. Koopmans, J. R., Slutske, W. S., van Baal, G. C., & Boomsma, D. I. (1999). The influence of religion on alcohol use initiation: Evidence for genotype X environment interaction. Behavior Genetics, 29, 445–453. doi: 10.1023/A:1021679005623 Kosseim, P., Pullman, D., Perrot-Daley, A., Hodgkinson, K., Street, C., & Rahman, P. (2013). Privacy protection and public goods: Building a genetic database for health research in Newfoundland and Labrador. Journal of the American Medical Informatics Association, 20, 38–43. Krafka, J. (1920). The effect of temperature upon facet number in the bar-eyed mutant of Drosophila. Journal of General Physiology, 2, 409–464. Kraft, P., & Hunter, D. (2005). Integrating epidemiology and genetic association: the challenge of gene–environment interaction. Philosophical Transactions of the Royal Society B: Biological Sciences, 360, 1609–1616. Kraft, P., Yen, Y. C., Stram, D. O., Morrison, J., & Gauderman, W. J. (2007). Exploiting gene–environmental interaction to detect genetic associations. Human Heredity, 63, 111–119. Kriegebaum, C., Gutknecht, L., Schmitt, A., Lesch, K. P., & Reif, A. (2010). Serotonin now: Part 1. Neurobiology and developmental genetics. Fortschritte der Neurologie-Psychiatrie, 78, 319–331. Lachance, J., & Tishkoff, S. A. (2013). SNP ascertainment bias in population genetic analyses: why it is important, and how to correct it. Bioessays, 35, 780–786. Laird, N. M., & Lange, C. (2006). Family-based designs in the age of large-scale gene-association studies. Nature Genetics, 7, 385–394.

326

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Lander, E. S. (2011). Initial impact of the sequencing of the human genome. Nature, 470, 187–197. Langholz, B., & Goldstein, L. (1996). Risk set sampling in epidemiologic cohort studies. Statistical Science, 11, 35–53. Langholz, B., Rothman, N., Wacholder, S., & Thomas, D. C. (1999). Cohort studies for characterizing measured genes. Journal of the National Cancer Institute Monographs, 26, 39–42. Lasky-Su, J., Neale, B. M., Franke, B., Anney, R. J., Zhou, K., Maller, J. B., . . . Faraone, S. V. (2008). Genome-wide association scan of quantitative traits for attention deficit hyperactivity disorder identifies novel associations and confirms candidate gene associations. American Journal of Medical Genetics Part B, 147B, 1345–1354. Latvala, A., Castaneda, A. E., Perälä, J., Saarni, S. I., Aalto-Setälä, T., Lönnqvist, J., . . . Tuulio-Henriksson, A. (2009). Cognitive functioning in substance abuse and dependence: A population-based study of young adults. Addiction, 104, 1558–1568. doi: 10.1111/j.13600443.2009.02656.x Lawler, J. M., Hostinar, C. E., Mliner, S. B., & Gunnar, M. R. (2014). Disinhibited social engagement in postinstitutionalized children: differentiating normal from atypical behavior. Development and Psychopathology, 26, 451–464. doi: 10.1017/S0954579414000054 Le Marchand, L. (2005). The predominance of the environment over genes in cancer causation: implications for genetic epidemiology. Cancer Epidemiology, Biomarkers & Prevention, 14, 1037–1039. Le Marchand, L., & Wilkens, L. R. (2008). Design considerations for genomic association studies: importance of gene–environment interactions. Cancer Epidemiology, Biomarkers & Prevention, 17, 263–267. Leibowitz, H. M., Krueger, D. E., Maunder, L. R., Milton, R. C., Kini, M. M., Kahn, H. A., . . . Dawber, T. R. (1980). The Framingham Eye Study monograph: An ophthalmological and epidemiological study of cataract, glaucoma, diabetic retinopathy, macular degeneration, and visual acuity in a general population of 2631 adults, 1973–1975. Survey of Ophthalmology, 24, 335–610. Lesch, K. P., Bengel, D., Heils, A., Sabol, S. Z., Greenberg, B. D., Petri, S., . . . Murphy, D. L. (1996). Association of anxiety-related traits with a polymorphism in the serotonin transporter gene regulatory region. Science, 274, 1527 – 1531. Levenson, J. M., & Sweatt, J. D. (2005). Epigenetic mechanisms in memory formation. Nature Reviews Neuroscience, 6, 108–118. Lewis, S. J., Araya, R., Smith, G. D., Freathy, R. M., Gunnell, D., Palmer, T., & Munafó, M. (2011). Smoking is associated with, but does not cause, depressed mood in pregnancy—A Mendelian randomization study. PLoS One, 6, 1–5. doi: 10.1371/journal.pone.0021689 Lewontin, R. C. (1974). The analysis of variance and the analysis of causes. American Journal of Human Genetics, 26, 400–411. Li, R., Boerwinkle, E., Olshan, A. F., Chambless, L. E., Pankow, J. S., Tyroler, H. A., . . . Heiss, G. (2000). Glutathione S-transferase genotype as a susceptibility factor in smoking-related coronary heart disease. Atherosclerosis, 149, 451–462. doi: 10.1016/S0021-9150(99) 00483-9 Licinio, J. (2003). Molecular psychiatry: Moving away from onepolymorphism genetic association studies. Molecular Psychiatry, 8, 247. Little, J., Higgins, J. P., Ioannidis, J. P., Moher, D., Gagnon, F., von Elm, E., . . . Birkett, N. (2009). Strengthening the reporting of genetic association studies (STREGA): an extension of the STROBE statement. European Journal of Epidemiology, 24, 37–55. Liu, X., Fallin, M. D., & Kao, W. H. (2004). Genetic dissection methods: designs used for tests of gene–environment interaction. Current Opinion in Genetics & Development, 14, 241–245. Lloyd-Jones, D. M., Larson, M. G., Beiser, A., & Levy, D. (1999). Lifetime risk of developing coronary heart disease. Lancet, 353, 89–92.

Lobach, I., Carroll, R. J., Spinka, C., Gail, M. H., & Chatterjee, N. (2008). Haplotype-based regression analysis and inference of case–control studies with unphased genotypes and measurement ererrors in environmental exposures. Biometrics, 64, 673–684. Loeber, R., & Farrington, D. P. (1998). Serious and violent juvenile offenders: Risk factors and successful interventions. Thousand Oaks, CA: Sage. Lohmueller, K. E., Pearce, C. L., Pike, M. C., Lander, E. S., & Hirschhorn, J. N. (2003). Meta-analysis of genetic association studies supports a contribution of common variants to susceptibility to common disease. Nature Genetics, 33, 177–182. Lotrich, F. E., & Lenze, E. (2009). Gene–environment interactions and depression. Journal of the American Medical Association, 302, 1859–1860. Lu, A. T., Bakker, S., Janson, E., Cichon, S., Cantor, R. M., & Ophoff, R. A. (2012). Prediction of serotonin transporter promoter polymorphism genotypes from single nucleotide polymorphism arrays using machine learning methods. Psychiatric Genetics, 22, 182–188. Luan, J. A., Wong, M. Y., Day, N. E., & Wareham, N. J. (2001). Sample size determination for studies of gene–environment interaction. International Journal of Epidemiology, 30, 1035 –1040. Lush, J. L. (1937). Animal breeding plans. Ames, IA: Collegiate Press. Macleod, M. R., Michie, S., Roberts, I., Dirnagl, U., Chalmers, I., Ioannidis, J. P. A., . . . Glasziou, P. (2014). Biomedical research: Increasing value, reducing waste. Lancet, 383, 101–104. Maher, B. S. (2008). The case of the missing heritability. Nature, 456, 18–21. Manolio, T. A. (2003). Novel risk markers and clinical practice. New England Journal of Medicine, 349, 1587–1589. Manolio, T. A., Bailey-Wilson, J. E., & Collins, F. S. (2006). Genes, environment and the value of prospective cohort studies. Nature Reviews Genetics, 7, 812–820. Manolio, T. A., Brooks, L. D., & Collins, F. S. (2008). A HapMap harvest of insights into the genetics of common disease. Journal of Clinical Investigations, 118, 1590–1605. Mantel, N. (1973). Synthetic retrospective studies and related topics. Biometrics, 29, 479–486. Marchini, J., Donnelly, P., & Cardon, L. R. (2005). Genome-wide strategies for detecting multiple loci that influence complex diseases. Nature Genetics, 37, 413–417. Marshall, E. (1997). Whose DNA is it, anyway? Science, 278, 564–567. Mathews, T. A., Fedele, D. E., Coppelli, F. M., Avila, A. M., Murphy, D. L., & Andrews, A. M. (2004). Gene dose-dependent alterations in extraneuronal serotonin but not dopamine in mice with reduced serotonin transporter expression. Journal of Neuroscience Methods, 140, 169–181. doi: 10.1016/j.jneumeth.2004.05.017 Matullo, G., Berwick, M., & Vineis, P. (2005). Gene–environment interactions: how many false positives? Journal of the National Cancer Institute, 97, 550–551. McCaffery, J. M., Papandonatos, G. D., Lyons, M. J., Koenen, K. C., Tsuang, M. T., & Niaura, R. (2008). Educational attainment, smoking initiation and lifetime nicotine dependence among male Vietnam-era twins. Psychological Medicine, 38, 1287–1298. doi: 10.1017/S0033291707001882 McCarthy, M. I., & Hirschhorn, J. N. (2008). Genome-wide association studies: potential next steps on a genetic journey. Human Molecular Genetics, 17, R156–R165. McClellan, J., & King, M. C. (2010). Genetic heterogeneity in human disease. Cell, 141, 210–217. McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties of detecting interactions and moderator effects. Psychological Bulletin, 114, 376–390.

References McGrath, J. J., Mortensen, P. B., Visscher, P. M., & Wray, N. R. (2013). Where GWAS and epidemiology meet: Opportunities for the simultaneous study of genetic and environmental risk factors in schizophrenia. Schizophrenia Bulletin, 39, 955–959. doi: 10.1093/schbul/sbt108 McKinney, B. A., Reif, D. M., Ritchie, M. D., & Moore, J. H. (2006). Machine learning for detecting gene-gene interactions: a review. Applied Bioinformatics, 8, 77–88. Meehl, P. E. (1962). Schizotaxia, schizotypy, schizophrenia. American Psychologist, 17, 827–838. Miller, F. J., Schlosser, P. M., & Janszen, D. B. (2000). Haber’s rule: a special case in a family of curves relating concentration and duration of exposure to a fixed level of response for a given endpoint. Toxicology, 149, 21–34. doi: 10.1016/S0300-483X(00)00229-8 Mizushima, S., Tsuchida, K., & Yamori, Y. (1999). Preventive nutritional factors in epidemiology: interaction between sodium and calcium. Clinical & Experimental Pharmacology & Physiology, 26, 573–575. Modinos, G., Iyegbe, C., Prata, D., Rivera, M., Kempton, M. J., Valmaggia, L. R., . . . McGuire, P. (2013). Molecular genetic gene–environment studies using candidate genes in schizophrenia: A systematic review. Schizophrenia Research, 150, 356–365. doi: 10.1016/j.schres.2013.09.010 Moffitt, T. E. (2005). The new look of behavioral genetics in developmental psychopathology: gene–environment interplay in antisocial behaviors. Psychological Bulletin, 131, 533–554. Moffitt, T. E., Caspi, A., & Rutter, M. (2005). Strategy for investigating interactions between measured genes and measured environments. Archives of General Psychiatry, 62, 473–481. Molenaar, P. C. M., Boomsma, D. I., & Dolan, C. V. (1997). The detection of genotype-environment interaction in longitudinal genetic models. In M. C. LaBuda & E. L. Grigorenko (Eds.), On the way to individuality (pp. 53–70). New York, NY: Nova Science Publishers. Monroe, S. M., & Reid, M. W. (2008). Gene–environment interactions in depression research: Genetic polymorphisms and life-stress polyprocedures. Psychological Science, 19, 947–956. Monroe, S. M., & Simons, A. D. (1991). Diathesis–stress theories in the context of life stress research: Implications for the depressive disorders. Psychological Bulletin, 110, 406–425. Montañez, S., Owens, W. A., Gould, G. G., Murphy, D. L., & Daws, L. C. (2003). Exaggerated effect of fluvoxamine in heterozygote serotonin transporter knockout mice. Journal of Neurochemistry, 86, 210–219. doi: 10.1046/j.1471-4159.2003.01836.x Moore, J. H. (2003). The ubiquitous nature of epistasis in determining susceptibility to common human diseases. Human Heredity, 56, 73–82. Moore, J. H., & Williams, S. M. (2009). Epistasis and its implications for personal genetics. American Journal of Human Genetics, 85, 309–320. Mukherjee, B., Ahn, J., Gruber, S. B., Ghosh, M., & Chatterjee, N. (2010). Case–control studies of gene–environment interaction: Bayesian design and analysis. Biometrics, 66, 934–948. doi: 10.1111/j.1541-0420.2009.01357.x. Mukherjee, B., Ahn, J., Gruber, S. B., Rennert, G., Moreno, V., & Chatterjee, N. (2008). Tests for gene–environment interaction from case–control data: A novel study of type I error, power and designs. Genetic Epidemiology, 32, 615–626. Mukherjee, B., & Chatterjee, N. (2007). Exploiting gene–environment independence for analysis of case–control studies: An empirical-Bayes type shrinkage estimator to trade off between bias and efficiency. Biometrics, 64, 685–694. doi: 10.111/j.1541-0420.2007.00953.x Munafò, M. R., Brown, S. M., & Hariri, A. R. (2008). Serotonin transporter (5-HTTLPR) genotype and amygdala reactivity: A meta-analysis. Biological Psychiatry, 63, 852–857. Munafò, M. R., Clark, T. G., Moore, L. R., Payne, E., Walton, R., & Flint, J. (2003). Genetic polymorphisms and personality in healthy

327

adults: A systematic review and meta-analysis. Molecular Psychiatry, 8, 471–484. Munafò, M. R., Durrant, C., Lewis, G., & Flint, J. (2009). Gene X environment interactions at the serotonin transporter locus. Biological Psychiatry, 65, 211–219. Munafò, M. R., Durrant, C., Lewis, G., & Flint, J. (2010). Defining replication: A response to Kaufman and colleagues. Biological Psychiatry, 67(4), e21–e23. doi: 10.1016/j.biopsych.2009.09.035 Munafò, M. R., & Flint, J. (2009). Replication and heterogeneity in gene × environment interaction studies. International Journal of Neuropsychopharmacology, 12, 727–729. Munafò, M. R., Matheson, I. J., & Flint, J. (2007). Association of the DRD2 gene Taq1A polymorphism and alcoholism: A meta-analysis of case–control studies and evidence of publication bias. Molecular Psychiatry, 12, 454–461. Murcray, C. E., Lewinger, J. P., & Gauderman, W. J. (2009). Gene–environment interaction in genome-wide association studies. American Journal of Epidemiology, 169, 219–226. Murphy, D. L., Li, Q., Engel, S., Wichems, C., Andrews, A., Lesch, K. P., & Uhl, G. (2001). Genetic perspectives on the serotonin transporter. Brain Research Bulletin, 56, 487–494. Murrough, J. W., & Charney, D. S. (2011). The serotonin transporter and emotionality: Risk, resilience, and new therapeutic opportunities. Biological Psychiatry, 69, 510–512. doi: 10.1016/j.biopsych.2011.01 .019 Need, A. C., Ge, D., Weale, M. E., Maia, J., Feng, S., Heinzen, E. L., . . . Goldstein, D. B. (2009). A genome-wide investigation of SNPs and CNVs in schizophrenia. PLoS Genetics, 5, e1000373. doi: 10.1371/journal.pgen.1000373 Newman, A. B., Simonsick, E. M., Naydeck, B. L., Boudreau, R. M., Kritchevsky, S. B., Nevitt, M. C., . . . Harris, T. B. (2006). Association of long-distance corridor walk performance with mortality, cardiovascular disease, mobility limitation, and disability. JAMA, 295, 2018–2026. doi: 10.1001/jama.295.17.2018 Neyman, J. (1955). Statistics: Servant of all sciences. Science, 122, 401–406. Nieuwenhuijsen, M. J., Gómez-Perales, J. E., & Colvile, R. N. (2007). Levels of particulate air pollution, its elemental composition, determinants and health effects in metro systems. Atmospheric Environment, 41, 7995–8006. doi: 10.1016/j.atmosenv.2007.08.002 Ordovas, J. M., Corella, D., Demissie, S., Cupples, L. A., Couture, P., Coltell, O., . . . Tucker, K. L. (2002). Dietary fat intake determines the effect of a common polymorphism in the hepatic lipase gene promoter on high-density lipoprotein metabolism: Evidence of a strong dose effect in this gene-nutrient interaction in the Framingham Study. Circulation, 106, 2315–2321. doi: 10.1161/01.cir.0000036597 .52291.c9 Ottman, R. (1990). An epidemiologic approach to gene– environment interaction. Genetic Epidemiology, 7, 177–185. Ottman, R. (1996). Gene–environment interaction: Definitions and study designs. Preventive Medicine, 25, 764–770. Patsopoulos, N. A., Tatsioni, A., & Ioannidis, J. P. (2007). Claims of sex differences: an empirical assessment in genetic associations. JAMA, 298, 880–893. Peerbooms, O., Rutten, B. P. F., Collip, D., Lardinois, M., Lataster, T., Thewissen, V., . . . van Winkel, R. (2012). Evidence that interactive effects of COMT and MTHFR moderate psychotic response to environmental stress. Acta Psychiatrica Scandinavica, 125, 247–256. doi: 10.1111/j.1600-0447.2011.01806.x Pennington, B. F., McGrath, L. M., Rosenberg, J., Barnard, H., Smith, S. D., Willcutt, E. G., . . . Olson, R. K. (2009). Gene X environment interactions in reading disability and attention-deficit/hyperactivity disorder. Developmental Psychology, 45, 77–89.

328

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Piegorsch, W. W., Weinberg, C. R., & Taylor, J. (1994). Non hierarchical logistic models and case-only designs for assessing susceptibility in population-based case–control studies. Statistics in Medicine, 13, 153–162. Plomin, R., Haworth, C. M. A., Meaburn, E. L., Price, T. S., Wellcome Trust Case Control Consortium 2, & Davis, O. S. P. (2013). Common DNA markers can account for more than half of the genetic influence on cognitive abilities. Psychological Science. doi: 10.1177/0956797612457952 Pluess, M., & Belsky, J. (2013). Vantage sensitivity: Individual differences in response to positive experience. Psychological Bulletin, 139, 901–916. doi: 10.1037/a0030196 Ponder, B. (2001). Cancer genetics. Nature, 411, 336–341. Posner, M. I., & Rothbart, M. K. (2007). Research on attention networks as a model for the integration of psychological science. Annual Review of Psychology, 58, 1–23.

Rutter, M. (2006). Genes and behavior: Nature–nurture interplay explained. Malden, MA: Blackwell. Rutter, M., Moffitt, T. E., & Caspi, A. (2006). Gene–environment interplay and psychopathology: Multiple varieties but real effects. Journal of Child Psychology and Psychiatry, 47, 226–261. Rutter, M., & Pickles, A. (1991). Person–environment interactions: Concepts, mechanisms, and implications for data analysis. In R. Plomin & T. D. Wachs (Eds.), Conceptualization and measurement of organism–environment interaction (pp. 105–141). Washington, DC: American Psychological Association. Sackett, D. L. (1979). Bias in analytic research. Journal of Chronic Diseases, 32, 51–63. Sameroff, A. J. (1983). Developmental systems: Contexts and evolution. In P. Mussen (Ed.), Handbook of child psychology (Vol. 1, pp. 237–294). New York, NY: Wiley.

Pramanik, R., Thompson, H., Kistler, J. O., Wade, W. G., Galloway, J., Peakman, T., & Proctor, G. B. (2012). Effects of the UK Biobank collection protocol on potential biomarkers in saliva. International Journal of Epidemiology, 41, 1786–1797.

Sanders, A. R., Duan, J., Levinson, D. F., Shi, J., He, D., Hou, C., . . . Gejman, P. V. (2008). No significant association of 14 candidate genes with schizophrenia in a large European ancestry sample: Implications for psychiatric genetics American Journal of Psychiatry, 165, 497–506. doi: 10.1176/appi.ajp. 2007.07101573

Prentice, R. L. (1986). On the design of synthetic case–control studies. Biometrics, 42, 301–310.

Sarkar, S. (1999). From the reaktionsnorm to the adaptive norm: The norm of reaction, 1909–1960. Biology and Philosophy, 14, 235–252.

Purcell, S. (2002). Variance components models for gene–environment interaction in twin analysis. Twin Research, 5, 554–571.

Sayette, M. A., Griffin, K. M., & Sayers, W. M. (2010). Counterbalancing in smoking cue research: a critical analysis. Nicotine & Tobacco Research, 12, 1068–1079.

Relton, C. L., & Smith, G. D. (2012). Two-step epigenetic Mendelian randomization: a strategy for establishing the causal role of epigenetic processes in pathways to disease. International Journal of Epidemiology, 41, 161–176. doi: 10.1093/ije/dyr233 Riboli, E., & Kaaks, R. (1997). The EPIC Project: rationale and study design. European prospective investigation into cancer and nutrition. International Journal of Epidemiology, 26, S6–S14. Risch, N., Herrell, R., Lehner, T., Liang, K. Y., Eaves, L., Hoh, J., . . . Merikangas, K. R. (2009). Interaction between the serotonin transporter gene (5-HTTLPR), stressful life events, and risk of depression: A meta-analysis. JAMA, 301, 2462–2471. Ritchie, M. D., & Motsinger, A. A. (2005). Multifactor dimensionality reduction for detecting gene-gene and gene–environment interactions in pharmacogenomics studies. Pharmacogenomics, 6, 823–834. Rodgers, A., Ezzati, M., Vander Hoorn, S., Lopez, A. D., & Lin, R. B. (2004). Distribution of major health risks: Findings from the global burden of disease study. PLoS Medicine, 1, e27. Rose, G. (1992). The strategy of preventive medicine. Oxford, UK: Oxford University Press. Rosenberg, J., Pennington, B. F., Willcutt, E. G., & Olson, R. K. (2012). Gene by environment interactions influencing reading disability and the inattentive symptom dimension of attention deficit/hyperactivity disorder. Journal of Child Psychology & Psychiatry & Allied Disciplines, 53, 243–251. Rosenberg, N. A., Li, L. M., Ward, R., & Pritchard, J. K. (2003). Informativeness of genetic markers for inference of ancestry. American Journal of Human Genetics, 73, 1402–1422. Rothman, K. J. (1986). Interactions between causes. Boston, MA: Little, Brown.

Schaid, D. J. (1999). Case-parents design for gene–environment interaction. Genetic Epidemiology, 16, 261–273. Schaid, D. J., Olson, J. M., Gauderman, W. J., & Elston, R. C. (2003). Regression models for linkage: issues of traits, covariates, heterogeneity, and interaction. Human Heredity, 55, 86–96. Schlesselman, J. J. (1982). Case–control studies: Design, conduct, and analysis. New York, NY: Oxford University Press. Schulz-Heik, R. J., Rhee, S. H., Haberstick, B. C., Hewitt, J. K., Hopfer, C. J., Lessem, J. M., . . . Timberlake, D. S. (2007). Gene–environment correlations in childhood maltreatment and genetic mediation of the maltreatment–conduct problem relationship. Paper presented at the Behavioral Genetics Association, Amsterdam, The Netherlands. Searle, S., & Blackwell, J. M. (1999). Evidence for a functional repeat polymorphism in the promoter of the human NRAMP1 gene that correlates with autoimmune versus infectious disease susceptibility. Journal of Medical Genetics, 36, 295–299. Sebastiani, P., Ramoni, M. F., Nolan, V., Baldwin, C. T., & Steinberg, M. H. (2005). Genetic dissection and prognostic modeling of overt stroke in sickle cell anemia. Nature Genetics, 37, 435–440. Shin, J. H., Infante-Rivard, C., Graham, J., & McNeney, B. (2012). Adjusting for spurious gene-by-environment interaction using case-parent triads. Statistical Applications in Genetics & Molecular Biology, 11(2), Article 7. Silberberg, J. S., Wlodarczyk, J., Fryer, J., Ray, C. D., & Hensley, M. J. (1998). Correction for biases in a population-based study of family history and coronary heart disease. The Newcastle family history study I. American Journal of Epidemiology, 147, 1123–1132.

Rothman, K. J., & Greenland, S. (2005). Causation and causal inference in epidemiology. American Journal of Public Health, 95, S144–S150.

Singh, N. K., Repsilber, D., Liebscher, V., Taher, L., & Fuellen, G. (2013). Identifying genes relevant to specific biological conditions in time course microarray experiments. PLoS One, 8, e76561.

Rothman, K. J., Greenland, S., & Walker, A. M. (1980). Concepts of interaction. American Journal of Epidemiology, 112, 467–470.

Slatkin, M. (2009). Epigenetic inheritance and the missing heritability problem. Genetics, 182, 845–850. doi: 10.1534/genetics.109.102798

Rowe, D. (2003). Assessing genotype–environment interactions and correlations in the postgenomic era. In R. Plomin, J. C. DeFries, I. W. Craig, & P. McGuffin (Eds.), Behavior genetics in the post-genomic era, pp. 71–86.Washington, DC: American Psychological Association.

Smith, G. D., & Ebrahim, S. (2003). “Mendelian randomization”: Can genetic epidemiology contribute to understanding environmental determinants of disease? International Journal of Epidemiology, 32, 1–22.

References Smith, G. D., Ebrahim, S., Lewis, S., Hansell, A. L., Palmer, L. J., & Burton, P. R. (2005). Genetic epidemiology and public health: hope, hype, and future prospects. Lancet, 366, 1484–1498. doi: 10.1016/S0140-6736(05)67601-5 Smith, P. G., & Day, N. E. (1984). The design of case–control studies: the influence of confounding and interaction effects. International Journal of Epidemiology, 13, 356–365. South, S. C., & Krueger, R. F. (2013). Marital satisfaction and physical health: Evidence for an orchid effect. Psychological Science, 24, 373–378. doi: 10.1177/0956797612453116 Spira, A. (2014). [Large health cohorts in France]. Bulletin de l Academie Nationale de Medecine, 197, 315–328; discussion 328. St-Pierre, J., Miller-Felix, I., Paradis, M.-È., Bergeron, J., Lamarche, B., Després, J.-P., . . . Vohl, M.-C. (2003). Visceral obesity attenuates the effect of the hepatic lipase −514C>T polymorphism on plasma HDL-cholesterol levels in French-Canadian men. Molecular Genetics and Metabolism, 78, 31–36. doi: 10.1016/S1096-7192(02)00223-8 Stamler, J. (1991). Blood pressure and high blood pressure. Aspects of risk. Hypertension, 18, I95–I107. Sullivan, P. F., Lin, D., Tzeng, J. Y., van den Oord, E., Perkins, D., Stroup, T. S., . . . Close, S. L. (2008). Genomewide association for schizophrenia in the CATIE study: Results of stage 1. Molecular Psychiatry, 13, 570–584. doi: http://www.nature.com/mp/journal/v13/n6/suppinfo/ mp200825s1.html Suomi, S. (1997). Early determinants of behaviour. British Medical Bulletin, 53, 170–184. Surtees, P. G., Wainwright, N. W. J., Willis-Owen, S. A. G., Luben, R., Day, N. E., & Flint, J. (2006a). Social adversity, the serotonin transporter (5-HTTLPR) polymorphism and major depressive disorder. Biological Psychiatry, 59, 224–229. doi: 10.1016/j .biopsych.2005.07.014 Swanson, J. M. (2012). The UK Biobank and selection bias. Lancet, 380, 110. Tabery. (2007). Biometric and developmental gene–environment interactions: Looking back, moving forward. Development & Psychopathology, 19, 961–976. Tai, E. S., Corella, D., Deurenberg-Yap, M., Cutter, J., Chew, S. K., Tan, C. E., & Ordovas, J. M. (2003). Dietary fat interacts with the -514C>T polymorphism in the hepatic lipase gene promoter on plasma lipid profiles in a multiethnic Asian population: The 1998 Singapore national health survey. Journal of Nutrition, 133, 3399–3408. Taube, A. (1968). Matching in retrospective studies, sampling via the dependent variable. Acta Soc Med Upsaliensis, 73, 187–196. Taubes, G. (1995). Epidemiology faces its limits. Science, 269, 164–169. Taylor, J., Roehrig, A. D., Soden Hensler, B., Connor, C. M., & Schatschneider, C. (2010). Teacher quality moderates the genetic effects on early reading. Science, 328, 512–514. Terwilliger, J. D., & Weiss, K. M. (2003). Confounding, ascertainment bias, and the blind quest for a genetic “fountain of youth.” Annals of Medicine, 35, 532–544. Thomas, D. C. (1988). Exposure-time-response relationships with applications to cancer epidemiology. Annual Review of Public Health, 9, 451–482. Thomas, D. C. (2000). “Case-parents design for gene–environment interaction” by Schaid. Genetic Epidemiology, 19, 461–463. Thomas, D. C. (2010a). Gene–environment-wide association studies: Emerging approaches. Nature Reviews Genetics, 11, 259–272. Thomas, D. C. (2010b). Methods for investigating gene–environment interactions in candidate pathway and genome-wide association studies. Annual Review of Public Health, 21–36. doi: 10.1146/annurev .publhealth.012809.103619

329

Thomas, D. C., Stram, D., & Dwyer, J. (1993). Exposure measurement error: Influence on exposure–disease relationships and methods of correction. Annual Review of Public Health, 14, 69–93. Thompson, W. D. (1991). Effect modification and the limits of biological inference from epidemiologic data. Journal of Clinical Epidemiology, 44, 221–232. Thorgeirsson, T. E., Geller, F., Sulem, P., Rafnar, T., Wiste, A., Magnusson, K. P., . . . Stefansson, K. (2008). A variant associated with nicotine dependence, lung cancer and peripheral arterial disease. Nature, 452(7187), 638–642. doi: http://www.nature.com/nature/journal/v452/ n7187/suppinfo/nature06846_S1.html Thornton, A., & Lee, P. (2000). Publication bias in meta-analysis: its causes and consequences. Journal of Clinical Epidemiology, 53, 207–216. Timberlake, D. S., Hopfer, C. J., Rhee, S. H., Friedman, N. P., Haberstick, B. C., Lessem, J. M., & Hewitt, J. K. (2007). College attendance and its effect on drinking behaviors in a longitudinal study of adolescents. Alcoholism: Clinical & Experimental Research, 31, 1020–1030. Timberlake, D. S., Rhee, S. H., Haberstick, B. C., Hopfer, C., Ehringer, M., & Lessem, J. M. (2006). The moderating effects of religiosity on the genetic and environmental determinants of smoking initiation. Nicotine & Tobacco Research, 8, 123–133. Troyer, D. A., Mubiru, J., Leach, R. J., & Naylor, S. L. (2004). Promise and challenge: markers of prostate cancer detection, diagnosis and prognosis. Disease Markers, 20, 117–128. Tsai, A. W., Cushman, M., Rosamond, W. D., Heckbert, S. R., Tracy, R. P., Aleksic, N., & Folsom, A. R. (2002). Coagulation factors, inflammation markers, and venous thromboembolism: The longitudinal investigation of thromboembolism etiology (LITE). American Journal of Medicine, 113, 636–642. doi: 10.1016/S0002-9343(02)01345-1 Tucker-Drob, E. M., & Harden, K. P. (2012). Intellectual interest mediates gene X socioeconomic status interaction on adolescent academic achievement. Child Development, 83, 743–757. Turkheimer, E., Haley, A., Waldron, M., D’Onofrio, B., & Gottesman, I. I. (2003). Socioeconomic status modifies heritability of IQ in young children. Psychological Science, 14, 623–628. doi: 10.1046/j.0956-7976.2003.psci_1475.x Uher, R., & McGuffin, P. (2008). The moderation by the serotonin transporter gene of environmental adversity in the aetiology of mental illness: review and methodological analysis. Molecular Psychiatry, 13, 131–146. Ul-Haq, Z., Mackay, D. F., Martin, D., Smith, D. J., Gill, J. M., Nicholl, B. I., . . . Pell, J. P. (2014). Heaviness, health and happiness: A cross-sectional study of 163066 UK Biobank participants. Journal of Epidemiology & Community Health, 68, 340–348. van den Oord, E. (1999). Method to detect genotype–environment interactions for quantitative trait loci in association studies. American Journal of Epidemiology, 150, 1179–1187. van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., Belsky, J., Beach, S., Brody, G., Dodge, K. A., . . . Scott, S. (2011). Gene-by-environment experiments: a new approach to finding the missing heritability. Nature Reviews Genetics, 12, 881–882. van Os, J., & Sham, P. (2003). Gene–environment interaction and correlation in schizophrenia. In R. Murray, P. B. Jones, E. Susser, J. v. Os, & M. Cannon (Eds.), The epidemiology of schizophrenia (pp. 235–253). Cambridge, UK: Cambridge University Press. Vandenbroucke, J. P., Koster, T., Briet, E., Reitsma, P. H., Bertina, R. M., & Rosendaal, F. R. (1994). Increased risk of venous thrombosis in oral-contraceptive users who are carriers of factor V Leiden mutation. Lancet, 344, 1453–1457. VanderWeele, T. J., Hernández-Díaz, S., & Hernán, M. A. (2010). Case-only gene–environment interaction studies: when does

330

The Trilogy of GxE: Conceptualization, Operationalization, and Application

association imply mechanistic interaction? Genetic Epidemiology, 34, 327–334. doi: 10.1002/gepi.20484 Vineis, P. (2004). A self-fulfilling prophecy: Are we underestimating the role of environment in gene–environment interaction research? International Journal of Epidemiology, 33, 945–946. Vinkhuyzen, A. A., Wray, N. R., Yang, J., Goddard, M. E., & Visscher, P. M. (2013). Estimation and partition of heritability in human populations using whole-genome analysis methods. Annual Review of Genetics, 47, 75–95. Vogelstein, B., & Kinzler, K. (2002). The genetic basis of human cancer. New York, NY: McGraw–Hill. Vreeke, G.-J. (2000). Nature, nurture and the future of the analysis of variance. Human Development, 43, 32–45. Wacholder, S., Chanock, S., Garcia-Closas, M., El Ghormli, L., & Rothman, N. (2004). Assessing the probability that a positive report is false: An approach for molecular epidemiology studies. Journal of the National Cancer Institute, 96, 434–442. Wacholder, S., Silverman, D. T., McLaughlin, J. K., & Mandel, J. S. (1992). Selection of controls in case–control studies. III. Design options. American Journal of Epidemiology, 135, 1042–1050. Waddington, C. H. (1957). The strategy of the genes: A discussion of some aspects of theoretical biology. London, UK: Allen & Unwin. Waddington, C. H. (1975). The evolution of an evolutionist. Ithaca, NY: Cornell University Press. Waldron, M., Heath, A. C., Turkheimer, E. N., Emery, R. E., Nelson, E., Bucholz, K. K., . . . Martin, N. G. (2008). Childhood sexual abuse moderates genetic influences on age at first consensual sexual intercourse in women. Behavior Genetics, 38, 1–10. doi: 10.1007/s10519007-9176-x Wallace, H. M. (2006). A model of gene-gene and gene–environment interactions and its implications for targeting environmental interventions by genotype. Theoretical Biology and Medical Modelling, 3, 35. doi: 10.1186/1742-4682-3-35 Wang, S. S., Fridinger, F., Sheedy, K. M., & Khoury, M. J. (2001). Public attitudes regarding the donation and storage of blood specimens for genetic research. Community Genetics, 4, 18–26. Wang, W. Y., Barratt, B. J., Clayton, D. G., & Todd, J. A. (2005). Genome-wide association studies: theoretical and practical concerns. Nature Reviews Genetics, 6, 109–118. Weaver, V. M., Buckley, T. J., & Groopman, J. D. (1998). Approaches to environmental exposure assessment in children. Environmental Health Perspectives, 106, 827–832. Webster, P. (2008). Biobanks. Canada launches massive study of adult cancer precursors. Science, 320, 1572–1573. Weisbrot, D. (2012). The ethical, legal and social implications of umbilical cord blood banking: learning important lessons from the protection of human genetic information. Journal of Law & Medicine, 19, 525–549. White, J. E. (1982). A two stage design for the study of the relationship between a rare exposure and a rare disease. American Journal of Epidemiology, 115, 119–128. Wichmann, H. E. (2005). Genetic epidemiology in Germany—From biobanking to genetic statistics. Methods of Information in Medicine, 44, 584–589. Wichmann, H. E., & Gieger, C. (2007). [Biobanks]. Bundesgesundheitsblatt, Gesundheitsforschung, Gesundheitsschutz, 50, 192–199. Willett, W. C. (2002). Balancing life-style and genomics research for disease prevention. Science, 296, 695–698. Willett, W. C., Blot, W. J., Colditz, G. A., Folsom, A. R., Henderson, B. E., & Stampfer, M. J. (2007). Merging and emerging cohorts: Not worth the wait. Nature, 445, 257–258.

Winickoff, D. E. (2001). Biosamples, genomics, and human rights: context and content of Iceland’s Biobanks Act. Journal of Biolaw & Business, 4, 11–17. Wolf, M., & Krause, J. (2014). Why personality differences matter for social functioning and social structure. Trends in Ecology & Evolution, 29, 306–308. Wolf, M., van Doorn, G. S., & Weissing, F. J. (2008). Evolutionary emergence of responsive and unresponsive personalities. PNAS, 105, 15825–15830. Women’s Health Initiative Study Group. (1998). Design of the Women’s Health Initiative clinical trial and observational study. Controlled Clinical Trials, 19, 61–109. Wong, M. Y., Day, N. E., Luan, J. A., & Wareham, N. J. (2004). Estimation of magnitude in gene–environment interactions in the presence of measurement error. Statistics in Medicine, 23, 987–998. doi: 10.1002/sim.1662 Wu, C., Zhang, H., Liu, X., DeWan, A., Dubrow, R., Ying, Z., . . . Hoh, J. (2009). Detecting essential and removable interactions in genome-wide association studies. Statistics and Its Interface, 2, 161–170. Yang, Q., & Khoury, M. J. (1997). Evolving methods in genetic epidemiology. III. Gene–environment interaction in epidemiological research. Epidemiologic Reviews, 19, 33–43. Yang, Q., Khoury, M. J., & Flanders, W. D. (1997). Sample size requirements in case-only designs to detect gene–environment interaction. American Journal of Epidemiology, 146, 713–720. Yuasa, M., & Kishi, R. (2009). [Profile and issues of prospective birth cohort in Japan]. Nippon Eiseigaku Zasshi—Japanese Journal of Hygiene, 64, 774–781. Yzerbyt, V. Y., Muller, D., & Judd, C. M. (2004). Adjusting researchers’ approach to adjustment: On the use of covariates when testing interactions. Journal of Experimental Social Psychology, 40, 424–431. Zammit, S., Lewis, G., Dalman, C., & Allebeck, P. (2010). Examining interactions between risk factors for psychosis. British Journal of Psychiatry, 197, 207–211. Zeger, S. L., Liang, K. Y., & Albert, P. S. (1998). Models for longitudinal data: a generalized estimating equation approach. Biometrics, 44, 1049–1060. Zeggini, E., Weedon, M. N., Lindgren, C. M., Frayling, T. M., Elliott, K. S., Lango, H., . . . Hattersley, A. T. (2007). Replication of genome-wide association signals in UK samples reveals risk loci for type 2 diabetes. Science, 316, 1336–1341. doi: 10.1126/science.1142364 Zubin, J., & Spring, B. (1977). Vulnerability–A new view of schizophrenia. Journal of Abnormal Psychology, 86, 103–126. Zuckerman, M. (1999). Vulnerability to psychopathology: A biosocial model. Washington, DC: American Psychological Association. Zuk, O., Hechter, E., Sunyaev, S. R., & Lander, E. S. (2012). The mystery of missing heritability: Genetic interactions create phantom heritability. PNAS, 109, 1193–1198. doi: 10.1073/pnas.1119675109

APPENDIX GxE publications reviewed to create Figure 8.1 Aas, M., Djurovic, S., Athanasiu, L., Steen, N. E., Agartz, I., Lorentzen, S., . . . Melle, I. (2012). Serotonin transporter gene polymorphism, childhood trauma, and cognition in patients with psychotic disorders. Schizophrenia Bulletin, 38, 15–22. doi: 10.1093/schbul/sbr113 Aguilera, M., Arias, B., Wichers, M., Barrantes-Vidal, N., Moya, J., Villa, H., . . . Fananas, L. (2009). Early adversity and 5-HTT/BDNF genes: new evidence of gene–environment interactions on depressive symptoms in a general population. Psychological Medicine, 39, 1425–1432. doi: 10.1017/s0033291709005248

Appendix Akkermann, K., Kaasik, K., Kiive, E., Nordquist, N., Oreland, L., & Harro, J. (2012). The impact of adverse life events and the serotonin transporter gene promoter polymorphism on the development of eating disorder symptoms. Journal of Psychiatric Research, 46, 38–43. doi: 10.1016/j.jpsychires.2011.09.013 Alexander, N., Klucken, T., Koppe, G., Osinsky, R., Walter, B., Vaitl, D., . . . Hennig, J. (2012). Interaction of the serotonin transporter-linked polymorphic region and environmental adversity: Increased amygdala-hypothalamus connectivity as a potential mechanism linking neural and endocrine hyperreactivity. Biological Psychiatry, 72, 49–56. doi: 10.1016/j.biopsych.2012.01.030 Alexander, N., Kuepper, Y., Schmitz, A., Osinsky, R., Kozyra, E., & Hennig, J. (2009). Gene–environment interactions predict cortisol responses after acute stress: Implications for the etiology of depression. Psychoneuroendocrinology, 34, 1294–1303. doi: 10.1016/j.psyneuen.2009.03.017 Antypa, N., & Van der Does, A. J. W. (2010). Serotonin transporter gene, childhood emotional abuse and cognitive vulnerability to depression. Genes, Brain & Behavior, 9, 615–620. Araya, R., Hu, X., Heron, J., Enoch, M. A., Evans, J., Lewis, G., . . . Goldman, D. (2009). Effects of stressful life events, maternal depression and 5-HTTLPR genotype on emotional symptoms in pre-adolescent children. American Journal of Medical Genetics. Neuropsychiatric Genetics, 150B, 670–682. doi: 10.1002/ajmg.b.30888 Armbruster, D., Moser, D. A., Strobel, A., Hensch, T., Kirschbaum, C., Lesch, K.-P., & Brocke, B. (2009). Serotonin transporter gene variation and stressful life events impact processing of fear and anxiety. International Journal of Neuropsychopharmacology, 12, 393–401. doi: 10.1017/s1461145708009565 Armeli, S., Conner, T. S., Covault, J., Tennen, H., & Kranzler, H. R. (2008). A serotonin transporter gene polymorphism (5-HTTLPR), drinking-to-cope motivation, and negative life events among college students. Journal of Studies on Alcohol and Drugs, 69, 814–823. Artero, S., Touchon, J., Dupuy, A.-M., Malafosse, A., & Ritchie, K. (2011). War exposure, 5-HTTLPR genotype and lifetime risk of depression. British Journal of Psychiatry, 199, 43–48. doi: 10.1192/bjp.bp. 110.087924 Åslund, C., Leppert, J., Comasco, E., Nordquist, N., Oreland, L., & Nilsson, K. W. (2009). Impact of the interaction between the 5HTTLPR polymorphism and maltreatment on adolescent depression. A population-based study. Behavior Genetics, 39, 524–531. doi: 10.1007/s10519-009-9285-9 Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2008). Oxytocin receptor (OXTR) and serotonin transporter (5-HTT) genes associated with observed parenting. Social Cognitive and Affective Neuroscience, 3, 128–134. doi: 10.1093/scan/nsn004 Barry, R. A., Kochanska, G., & Philibert, R. A. (2008). GxE interaction in the organization of attachment: Mothers’ responsiveness as a moderator of children’s genotypes. Journal of Child Psychology and Psychiatry, 49, 1313–1320. doi: 10.1111/j.1469-7610.2008.01935.x Beaver, K. M., & Belsky, J. (2012). Gene–environment interaction and the intergenerational transmission of parenting: Testing the differential-susceptibility hypothesis. Psychiatric Quarterly, 83, 29–40. doi: 10.1007/s11126-011-9180-4 Beaver, K. M., Ratchford, M., & Ferguson, C. J. (2009). Evidence of genetic and environmental effects on the development of low self-control. Criminal Justice and Behavior, 36, 1158–1172. doi: 10.1177/0093854809342859 Beaver, K. M., Vaughn, M. G., Wright, J. P., & DeLisi, M. (2012). An interaction between perceived stress and 5HTTLPR genotype in the prediction of stable depressive symptomatology. American Journal of Orthopsychiatry, 82, 260–266. doi: 10.1111/j.1939-0025.2012.01148.x

331

Benjet, C., Thompson, R. J., & Gotlib, I. H. (2010). 5-HTTLPR moderates the effect of relational peer victimization on depressive symptoms in adolescent girls. Journal of Child Psychology and Psychiatry, 51, 173–179. doi: 10.1111/j.1469-7610.2009.02149.x Blaya, C., Salum, G. A., Moorjani, P., Seganfredo, A. C., Heldt, E., Leistner-Segal, S., . . . Manfro, G. G. (2010). Panic disorder and serotonergic genes (SLC6A4, HTR1A and HTR2A): Association and interaction with childhood trauma and parenting. Neuroscience Letters, 485, 11–15. doi: 10.1016/j.neulet.2010.08.042 Brezo, J., Bureau, A., Mérette, C., Jomphe, V., Barker, E. D., Vitaro, F., . . . Turecki, G. (2010). Differences and similarities in the serotonergic diathesis for suicide attempts and mood disorders: A 22-year longitudinal gene–environment study. Molecular Psychiatry, 15, 831–843. doi: 10.1038/mp. 2009.19 Brody, G. H., Beach, S. R., Philibert, R. A., Chen, Y. F., & Murry, V. M. (2009a). Prevention effects moderate the association of 5-HTTLPR and youth risk behavior initiation: gene × environment hypotheses tested via a randomized prevention design. Child Development, 80, 645–661. doi: 10.1111/j.1467-8624.2009.01288.x Brody, G. H., Beach, S. R. H., Chen, Y.-F., Obasi, E., Philibert, R. A., Kogan, S. M., & Simons, R. L. (2011). Perceived discrimination, serotonin transporter linked polymorphic region status, and the development of conduct problems. Development and Psychopathology, 23, 617–627. doi: 10.1017/s0954579411000046 Brody, G. H., Beach, S. R. H., Philibert, R. A., Chen, Y., Lei, M. K., Murry, V. M., & Brown, A. C. (2009b). Parenting moderates a genetic vulnerability factor in longitudinal increases in youths’ substance use. Journal of Consulting and Clinical Psychology, 77, 1–11. doi: 10.1037/a0012996 Brummett, B. H., Boyle, S. H., Siegler, I. C., Kuhn, C. M., Ashley-Koch, A., Jonassaint, C. R., . . . Williams, R. B. (2008). Effects of environmental stress and gender on associations among symptoms of depression and the serotonin transporter gene linked polymorphic region (5-HTTLPR). Behavior Genetics, 38, 34–43. doi: 10.1007/s10519-0079172-1 Brummett, B. H., Krystal, A. D., Ashley-Koch, A., Kuhn, C. M., Zuchner, S., Siegler, I. C., . . . Williams, R. B. (2007). Sleep quality varies as a function of 5-HTTLPR genotype and stress. Psychosomatic Medicine, 69, 621–624. doi: 10.1097/PSY.0b013e31814b8de6 Bryant, R. A., Felmingham, K. L., Falconer, E. M., Pe Benito, L., Dobson-Stone, C., Pierce, K. D., & Schofield, P. R. (2010). Preliminary evidence of the short allele of the serotonin transporter gene predicting poor response to cognitive behavior therapy in posttraumatic stress disorder. Biological Psychiatry, 67, 1217–1219. doi: 10.1016/j.biopsych.2010.03.016 Bukh, J. D., Bock, C., Vinberg, M., Werge, T., Gether, U., & Kessing, L. V. (2009). Interaction between genetic polymorphisms and stressful life events in first episode depression. Journal of Affective Disorders, 119, 107–115. doi: 10.1016/j.jad.2009.02.023 Burkhouse, K. L., Gibb, B. E., Coles, M. E., Knopik, V. S., & McGeary, J. E. (2011). Serotonin transporter genotype moderates the link between children’s reports of overprotective parenting and their behavioral inhibition. Journal of Abnormal Child Psychology, 39, 783–790. doi: 10.1007/s10802-011-9526-2 Carli, V., Mandelli, L., Zaninotto, L., Roy, A., Recchia, L., Stoppia, L., . . . Serretti, A. (2011). A protective genetic variant for adverse environments? The role of childhood traumas and serotonin transporter gene on resilience and depressive severity in a high-risk population. European Psychiatry, 26, 471–478. doi: 10.1016/j.eurpsy.2011.04.008 Carver, C. S., Johnson, S. L., Joormann, J., Lemoult, J., & Cuccaro, M. L. (2011). Childhood adversity interacts separately with 5-HTTLPR and BDNF to predict lifetime depression diagnosis. Journal of Affective Disorders, 132, 89–93. doi: 10.1016/j.jad.2011.02.001

332

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . Poulton, R. (2003b). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301, 386–389. doi: 10.1126/science.1083968 Cervilla, J. A., Molina, E., Rivera, M., Torres-González, F., Bellón, J. A., Moreno, B., . . . Gutiérrez, B. (2007). The risk for depression conferred by stressful life events is modified by variation at the serotonin transporter 5HTTLPR genotype: Evidence from the Spanish PREDICT-Gene cohort. Molecular Psychiatry, 12, 748–755. doi: 10.1038/sj.mp. 4001981 Chipman, P., Jorm, A. F., Prior, M., Sanson, A., Smart, D., Tan, X., & Easteal, S. (2007). No interaction between the serotonin transporter polymorphism (5-HTTLPR) and childhood adversity or recent stressful life events on symptoms of depression: results from two community surveys. American Journal of Medical Genetics. Neuropsychiatric Genetics, 144B, 561–565. doi: 10.1002/ajmg.b.30480 Chorbov, V. M., Lobos, E. A., Todorov, A. A., Heath, A. C., Botteron, K. N., & Todd, R. D. (2007). Relationship of 5-HTTLPR genotypes and depression risk in the presence of trauma in a female twin sample. American Journal of Medical Genetics. Neuropsychiatric Genetics, 144B, 830–833. doi: 10.1002/ajmg.b.30534 Cicchetti, D., Rogosch, F. A., & Oshri, A. (2011). Interactive effects of corticotropin releasing hormone receptor 1, serotonin transporter linked polymorphic region, and child maltreatment on diurnal cortisol regulation and internalizing symptomatology. Development and Psychopathology, 23, 1125–1138. doi: 10.1017/s0954579411000599 Cicchetti, D., Rogosch, F. A., Sturge-Apple, M., & Toth, S. L. (2010). Interaction of child maltreatment and 5-HTT polymorphisms: Suicidal ideation among children from low-SES backgrounds. Journal of Pediatric Psychiatry, 35, 536–546. doi: 10.1093/jpepsy/jsp078 Cicchetti, D., Rogosch, F. A., & Sturge-Apple, M. L. (2007). Interactions of child maltreatment and serotonin transporter and monoamine oxidase A polymorphisms: Depressive symptomatology among adolescents from low socioeconomic status backgrounds. Development and Psychopathology, 19, 1161–1180. doi: 10.1017/s0954579407000600 Cicchetti, D., Rogosch, F. A., & Thibodeau, E. L. (2012). The effects of child maltreatment on early signs of antisocial behavior: Genetic moderation by tryptophan hydroxylase, serotonin transporter, and monoamine oxidase A genes. Development and Psychopathology, 24, 907–928. doi: 10.1017/s0954579412000442 Cicchetti, D., Rogosch, F. A., & Toth, S. L. (2011). The effects of child maltreatment and polymorphisms of the serotonin transporter and dopamine D4 receptor genes on infant attachment and intervention efficacy. Development and Psychopathology, 23, 357–372. doi: 10.1017/s0954579411000113 Clasen, P. C., Wells, T. T., Knopik, V. S., McGeary, J. E., & Beevers, C. G. (2011). 5-HTTLPR and BDNF Val66Met polymorphisms moderate effects of stress on rumination. Genes, Brain & Behavior, 10, 740–746. doi: 10.1111/j.1601-183X.2011.00715.x Comasco, E., Sylvén, S. M., Papadopoulos, F. C., Sundström-Poromaa, I., Oreland, L., & Skalkidou, A. (2011). Postpartum depression symptoms: A case–control study on monoaminergic functional polymorphisms and environmental stressors. Psychiatric Genetics, 21, 19–28. doi: 10.1097/YPG.0b013e328341a3c1 Conway, C. C., Hammen, C., Brennan, P. A., Lind, P. A., & Najman, J. M. (2010). Interaction of chronic stress with serotonin transporter and catechol-O-methyltransferase polymorphisms in predicting youth depression. Depression and Anxiety, 27, 737–745. doi: 10.1002/ da.20715 Conway, C. C., Hammen, C., Espejo, E. P., Wray, N. R., Najman, J. M., & Brennan, P. A. (2012a). Appraisals of stressful life events as a genetically-linked mechanism in the stress–depression relationship.

Cognitive Therapy and Research, 36, 338–347. doi: 10.1007/s10608011-9368-9 Conway, C. C., Keenan-Miller, D., Hammen, C., Lind, P. A., Najman, J. M., & Brennan, P. A. (2012b). Coaction of stress and serotonin transporter genotype in predicting aggression at the transition to adulthood. Journal of Clinical Child & Adolescent Psychology, 41, 53–63. doi: 10.1080/15374416.2012.632351 Covault, J., Tennen, H., Armeli, S., Conner, T. S., Herman, A. I., Cillessen, A. H., & Kranzler, H. R. (2007). Interactive effects of the serotonin transporter 5-HTTLPR polymorphism and stressful life events on college student drinking and drug use. Biological Psychiatry, 61, 609–616. doi: 10.1016/j.biopsych.2006.05.018 Coventry, W. L., James, M. R., Eaves, L. J., Gordon, S. D., Gillespie, N. A., Ryan, L., . . . Wray, N. R. (2010). Do 5HTTLPR and stress interact in risk for depression and suicidality? Item response analyses of a large sample. American Journal of Medical Genetics. Neuropsychiatric Genetics, 153B, 757–765. doi: 10.1002/ajmg.b.31044 Cutuli, J. J., Raby, K. L., Cicchetti, D., Englund, M. M., & Egeland, B. (2013). Contributions of maltreatment and serotonin transporter genotype to depression in childhood, adolescence, and early adulthood. Journal of Affective Disorders, 149, 30–37. doi: 10.1016/j.jad.2012.08.011 De Pradier, M., Gorwood, P., Beaufils, B., Ades, J., & Dubertret, C. (2010). Influence of the serotonin transporter gene polymorphism, cannabis and childhood sexual abuse on phenotype of bipolar disorder: A preliminary study. European Psychiatry, 25, 323–327. doi: 10.1016/j.eurpsy.2009.10.002 Douglas, K., Chan, G., Gelernter, J., Arias, A. J., Anton, R. F., Poling, J., . . . Kranzler, H. R. (2011). 5-HTTLPR as a potential moderator of the effects of adverse childhood experiences on risk of antisocial personality disorder. Psychiatric Genetics, 21, 240–248. doi: 10.1097/YPG.0b013e3283457c15 Drury, S. S., Gleason, M. M., Theall, K. P., Smyke, A. T., Nelson, C. A., Fox, N. A., & Zeanah, C. H. (2012). Genetic sensitivity to the caregiving context: The influence of 5HTTLPR and BDNF. Physiology & Behavior, 106, 728–735. Du, Y., & Wan, Y. J. Y. (2009). The interaction of reward genes with environmental factors in contribution to alcoholism in Mexican Americans. Alcoholism: Clinical and Experimental Research, 33, 2103–2112. doi: 10.1111/j.1530-0277.2009.01050.x Eley, T. C., Sugden, K., Corsico, A., Gregory, A. M., Sham, P., McGuffin, P., . . . Craig, I. W. (2004b). Gene–environment interaction analysis of serotonin system markers with adolescent depression. Molecular Psychiatry, 9, 908–915. doi: 10.1038/sj.mp. 4001546 Everaerd, D., Gerritsen, L., Rijpkema, M., Frodl, T., van Oostrom, I., Franke, B., . . . Tendolkar, I. (2012). Sex modulates the interactive effect of the serotonin transporter gene polymorphism and childhood adversity on hippocampal volume. Neuropsychopharmacology, 37, 1848–1855. doi: 10.1038/npp. 2012.32 Fergusson, D. M., Horwood, L. J., Miller, A. L., & Kennedy, M. A. (2011). Life stress, 5-HTTLPR and mental disorder: Findings from a 30-year longitudinal study. British Journal of Psychiatry, 198, 129–135. doi: 10.1192/bjp.bp. 110.085993 Fisher, H. L., Cohen-Woods, S., Hosang, G. M., Korszun, A., Owen, M., Craddock, N., . . . Uher, R. (2013). Interaction between specific forms of childhood maltreatment and the serotonin transporter gene (5-HTT) in recurrent depressive disorder. Journal of Affective Disorders, 145, 136–141. doi: 10.1016/j.jad.2012.05.032 Fisher, H. L., Cohen-Woods, S., Hosang, G. M., Uher, R., Powell-Smith, G., Keers, R., . . . McGuffin, P. (2012). Stressful life events and the serotonin transporter gene (5-HTT) in recurrent clinical depression. Journal of Affective Disorders, 136, 189–193. doi: 10.1016/j.jad.2011.09.016

Appendix Fox, N. A., Nichols, K. E., Henderson, H. A., Rubin, K., Schmidt, L., Hamer, D., . . . Pine, D. S. (2005). Evidence for a gene–environment interaction in predicting behavioral inhibition in middle childhood. Psychological Science, 16, 921–926. doi: 10.1111/j.1467-9280.2005. 01637.x Fredericks, C. A., Drabant, E. M., Edge, M. D., Tillie, J. M., Hallmayer, J., Ramel, W., . . . Dhabhar, F. S. (2010). Healthy young women with serotonin transporter SS polymorphism show a pro-inflammatory bias under resting and stress conditions. Brain, Behavior and Immunity, 24, 350–357. doi: 10.1016/j.bbi.2009.10.014 Frigerio, A., Ceppi, E., Rusconi, M., Giorda, R., Raggi, M. E., & Fearon, P. (2009). The role played by the interaction between genetic factors and attachment in the stress response in infancy. Journal of Child Psychology and Psychiatry, 50, 1513–1522. doi: 10.1111/j.1469-7610. 2009.02126.x

333

Goodyer, I. M., Bacon, A., Ban, M., Croudace, T., & Herbert, J. (2009). Serotonin transporter genotype, morning cortisol and subsequent depression in adolescents. British Journal of Psychiatry, 195, 39–45. doi: 10.1192/bjp.bp. 108.054775 Gotlib, I. H., Joormann, J., Minor, K. L., & Hallmayer, J. (2008). HPA axis reactivity: A mechanism underlying the associations among 5-HTTLPR, stress, and depression. Biological Psychiatry, 63, 847–851. doi: 10.1016/j.biopsych.2007.10.008 Grabe, H. J., Lange, M., Wolff, B., Völzke, H., Lucht, M., Freyberger, H. J., . . . Cascorbi, I. (2005). Mental and physical distress is modulated by a polymorphism in the 5-HT transporter gene interacting with social stressors and chronic disease burden. Molecular Psychiatry, 10, 220–224. doi: 10.1038/sj.mp. 4001555

Frodl, T., Reinhold, E., Koutsouleris, N., Donohoe, G., Bondy, B., Reiser, M., . . . Meisenzahl, E. M. (2010). Childhood stress, serotonin transporter gene and brain structures in major depression. Neuropsychopharmacology, 35, 1383–1390. doi: 10.1038/npp. 2010.8

Grabe, H. J., Schwahn, C., Mahler, J., Appel, K., Schulz, A., Spitzer, C., . . . Völzke, H. (2012a). Genetic epistasis between the brain-derived neurotrophic factor Val66Met polymorphism and the 5-HTT promoter polymorphism moderates the susceptibility to depressive disorders after childhood abuse. Progress in Neuro-Psychopharmacology & Biological Psychiatry, 36, 264–270. doi: 10.1016/j.pnpbp. 2011.09.010

Gibb, B. E., Benas, J. S., Grassia, M., & McGeary, J. (2009). Children’s attentional biases and 5-HTTLPR genotype: Potential mechanisms linking mother and child depression. Journal of Clinical Child and Adolescent Psychology, 38, 415–426. doi: 10.1080/15374410902851705

Grabe, H. J., Schwahn, C., Mahler, J., Schulz, A., Spitzer, C., Fenske, K., . . . Freyberger, H. J. (2012b). Moderation of adult depression by the serotonin transporter promoter variant. American Journal of Medical Genetics. Neuropsychiatric Genetics, 3, 298–309.

Gibb, B. E., Johnson, A. L., Benas, J. S., Uhrlass, D. J., Knopik, V. S., & McGeary, J. E. (2011). Children’s 5-HTTLPR genotype moderates the link between maternal criticism and attentional biases specifically for facial displays of anger. Cognition & Emotion, 25, 1104–1120. doi: 10.1080/02699931.2010.508267

Grabe, H. J., Spitzer, C., Schwahn, C., Marcinek, A., Frahnow, A., Barnow, S., . . . Rosskopf, D. (2009). Serotonin transporter gene (SLC6A4) promoter polymorphisms and the susceptibility to posttraumatic stress disorder in the general population. American Journal of Psychiatry, 166, 926–933. doi: 10.1176/appi.ajp. 2009.08101542

Gibb, B. E., McGeary, J. E., Beevers, C. G., & Miller, I. W. (2006). Serotonin transporter (5-HTTLPR) genotype, childhood abuse, and suicide attempts in adult psychiatric inpatients. Suicide and Life-Threatening Behavior, 36, 687–693. doi: 10.1521/suli.2006.36.6. 687

Grassi, L., Rossi, E., Cobianchi, M., Aguiari, L., Capozzo, M., Martinis, E., . . . Giraldi, T. (2010). Depression and serotonin transporter (5-HTTLPR) polymorphism in breast cancer patients. Journal of Affective Disorders, 124, 346–350. doi: 10.1016/j.jad.2009.12.022

Gibb, B. E., Uhrlass, D. J., Grassia, M., Benas, J. S., & McGeary, J. (2009b). Children’s inferential styles, 5-HTTLPR genotype, and maternal expressed emotion-criticism: An integrated model for the intergenerational transmission of depression. Journal of Abnormal Psychology, 118, 734–745. doi: 10.1037/a0016765 Gibbons, F. X., Roberts, M. E., Gerrard, M., Li, Z., Beach, S., R. H., Simons, R. L. , . . . Philibert, R. A. (2012). The impact of stress on the life history strategies of African American adolescents: Cognitions, genetic moderation, and the role of discrimination. Developmental Psychology, 48, 722–739. doi: 10.1037/a0026599 Gilissen, R., Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & Linting, M. (2008). Electrodermal reactivity during the Trier social stress test for children: Interaction between the serotonin transporter polymorphism and children’s attachment representation. Developmental Psychobiology, 50, 615–625. doi: 10.1002/dev.20314 Gillespie, N. A., Whitfield, J. B., Williams, B., Heath, A. C., & Martin, N. G. (2005). The relationship between stressful life events, the serotonin transporter (5-HTTLPR) genotype and major depression. Psychological Medicine, 35, 101–111. doi: 10.1017/s0033291704002727 Gillihan, S. J., Rao, H., Brennan, L., Wang, D. J. J., Detre, J. A., Sankoorikal, G. M. V., . . . Farah, M. J. (2011). Serotonin transporter genotype modulates the association between depressive symptoms and amygdala activity among psychiatrically healthy adults. Psychiatry Research: Neuroimaging, 193, 161–167. doi: 10.1016/j.pscychresns.2011.03.003 Goldman, N., Glei, D. A., Lin, Y.-H., & Weinstein, M. (2010). The serotonin transporter polymorphism (5-HTTLPR): Allelic variation and links with depressive symptoms. Depression and Anxiety, 27, 260–269. doi: 10.1002/da.20660

Gunthert, K. C., Conner, T. S., Armeli, S., Tennen, H., Covault, J., & Kranzler, H. R. (2007). Serotonin transporter gene polymorphism (5-HTTLPR) and anxiety reactivity in daily life: A daily process approach to gene–environment interaction. Psychosomatic Medicine, 69, 762–768. doi: 10.1097/PSY.0b013e318157ad42 Hammen, C., Brennan, P. A., Keenan-Miller, D., Hazel, N. A., & Najman, J. M. (2010). Chronic and acute stress, gender, and serotonin transporter gene environment interactions predicting depression symptoms in youth. Journal of Child Psychology and Psychiatry, 51, 180–187. doi: 10.1111/j.1469-7610.2009.02177.x Hankin, B. L., Jenness, J., Abela, J. R. Z., & Smolen, A. (2011). Interaction of 5-HTTLPR and idiographic stressors predicts prospective depressive symptoms specifically among youth in a multiwave design. Journal of Clinical Child and Adolescent Psychology, 40, 572–585. doi: 10.1080/15374416.2011.581613 Hariri, A. R., Drabant, E. M., Munoz, K. E., Kolachana, B. S., Mattay, V. S., Egan, M. F., & Weinberger, D. R. (2005). A susceptibility gene for affective disorders and the response of the human amygdala. Archives of General Psychiatry, 62, 146–152. doi: 10.1001/archpsyc.62.2.146 Hermann, A., Kupper, Y., Schmitz, A., Walter, B., Vaitl, D., Hennig, J., . . . Tabbert, K. (2012). Functional gene polymorphisms in the serotonin system and traumatic life events modulate the neural basis of fear acquisition and extinction. PLoS One, 7, e44352. doi: 10.1371 /journal.pone.0044352 Ivorra, J. L., Sanjuan, J., Jover, M., Carot, J. M., de Frutos, R., & Molto, M. D. (2010). Gene–environment interaction of child temperament. Journal of Developmental and Behavioral Pediatrics, 31, 545–554. Jacob, C. P., Nguyen, T. T., Dempfle, A., Heine, M., Windemuth-Kieselbach, C., Baumann, K., . . . Reif, A. (2010). A gene– environment investigation on personality traits in two independent

334

The Trilogy of GxE: Conceptualization, Operationalization, and Application

clinical sets of adult patients with personality disorder and attention deficit/hyperactive disorder. European Archives of Psychiatry and Clinical Neuroscience, 260, 317–326. doi: 10.1007/s00406-0090079-0 Jacobs, N., Kenis, G., Peeters, F., Derom, C., Vlietinck, R., & Van Os, J. (2006). Stress-related negative affectivity and genetically altered serotonin transporter function: Evidence of synergism in shaping risk of depression. Archives of General Psychiatry, 63, 989–996. doi: 10.1001/archpsyc.63.9.989 Jacobs, R. H., Pine, D. S., Schoeny, M. E., Henry, D. B., Gollan, J. K., Moy, G., . . . Wakschlag, L. S. (2011). Maternal depressive history, teen 5HTTLPR genotype, and the processing of emotional faces: Exploring mechanisms of risk. Behaviour Research and Therapy, 49, 80–84. doi: 10.1016/j.brat.2010.10.004 Jenness, J. L., Hankin, B. L., Abela, J. R. Z., Young, J. F., & Smolen, A. (2011). Chronic family stress interacts with 5-HTTLPR to predict prospective depressive symptoms among youth. Depression and Anxiety, 28, 1074–1080. doi: 10.1002/da.20904 Johnson, A. L., Gibb, B. E., & McGeary, J. (2010). Reports of childhood physical abuse, 5-HTTLPR genotype, and women’s attentional biases for angry faces. Cognitive Therapy and Research, 34, 380–387. doi: 10.1007/s10608-009-9269-3 Jonassaint, C. R., Ashley-Koch, A., Whitfield, K. E., Hoyle, R. H., Richman, L. S., Siegler, I. C., . . . Williams, R. B. (2012). The serotonin transporter gene polymorphism (5HTTLPR) moderates the effect of adolescent environmental conditions on self-esteem in young adulthood: A structural equation modeling approach. Biological Psychology, 91, 111–119. doi: 10.1016/j.biopsycho.2012.05.004 Karwautz, A. F. K., Wagner, G., Waldherr, K., Nader, I. W., Fernandez-Aranda, F., Estivill, X., . . . Treasure, J. L. (2011). Gene– environment interaction in anorexia nervosa: Relevance of non-shared environment and the serotonin transporter gene. Molecular Psychiatry, 16, 590–592. doi: 10.1038/mp. 2010.125 Kaufman, J., Yang, B.-Z., Douglas-Palumberi, H., Crouse-Artus, M., Lipschitz, D., Krystal, J. H., & Gelernter, J. (2007). Genetic and environmental predictors of early alcohol use. Biological Psychiatry, 61, 1228–1234. doi: 10.1016/j.biopsych.2006.06.039 Kaufman, J., Yang, B.-Z., Douglas-Palumberi, H., Grasso, D., Lipschitz, D., Houshyar, S., . . . Gelernter, J. (2006). Brain-derived neurotrophic factor-5-HHTLPR gene interactions and environmental modifiers of depression in children. Biological Psychiatry, 59, 673–680. doi: 10.1016/j.biopsych.2005.10.026 Kaufman, J., Yang, B. Z., Douglas-Palumberi, H., Houshyar, S., Lipschitz, D., Krystal, J. H., & Gelernter, J. (2004). Social supports and serotonin transporter gene moderate depression in maltreated children. Proceedings of the National Academy of Sciences, 101, 17316–17321. doi: 10.1073/pnas.0404376101 Keers, R., Uher, R., Huezo-Diaz, P., Smith, R., Jaffee, S., Rietschel, M., . . . Aitchison, K. J. (2011). Interaction between serotonin transporter gene variants and life events predicts response to antidepressants in the GENDEP project. Pharmacogenomics Journal, 11, 138–145. doi: 10.1038/tpj.2010.14 Kendler, K. S., Kuhn, J. W., Vittum, J., Prescott, C. A., & Riley, B. (2005). The interaction of stressful life events and a serotonin transporter polymorphism in the trediction of episodes of major depression. Archives of General Psychiatry, 62, 529–535. doi: 10.1001/archpsyc.62.5.529 Kilpatrick, D. G., Koenen, K. C., Ruggiero, K. J., Acierno, R., Galea, S., Resnick, H. S., . . . Gelernter, J. (2007b). The serotonin transporter genotype and social support and moderation of posttraumatic stress disorder and depression in hurricane-exposed adults. American Journal of Psychiatry, 164, 1693–1699. doi: 10.1176/appi.ajp.2007. 06122007

Kim, J.-M., Stewart, R., Kim, S.-W., Yang, S.-J., Shin, I.-S., Kim, Y.-H., & Yoon, J.-S. (2007). Interactions between life stressors and susceptibility genes (5-HTTLPR and BDNF) on depression in Korean elders. Biological Psychiatry, 62, 423–428. doi: 10.1016/j.biopsych.2006.11.020 Klauke, B., Deckert, J., Reif, A., Pauli, P., Zwanzger, P., Baumann, C., . . . Domschke, K. (2011). Serotonin transporter gene and childhood trauma—A G × E effect on anxiety sensitivity. Depression and Anxiety, 28, 1048–1057. doi: 10.1002/da.20840 Kochanska, G., Kim, S., Barry, R. A., & Philibert, R. A. (2011). Children’s genotypes interact with maternal responsive care in predicting children’s competence: Diathesis–stress or differential susceptibility? Development and Psychopathology, 23, 605–616. doi: 10.1017/s09545 79411000071 Kochanska, G., Philibert, R. A., & Barry, R. A. (2009). Interplay of genes and early mother-child relationship in the development of self-regulation from toddler to preschool age. Journal of Child Psychology and Psychiatry, 50, 1331–1338. doi: 10.1111/j.1469-7610.2008. 02050.x Koenen, K. C., Aiello, A. E., Bakshis, E., Amstadter, A. B., Ruggiero, K. J., Acierno, R., . . . Galea, S. (2009b). Modification of the association between serotonin transporter genotype and risk of posttraumatic stress disorder in adults by county-level social environment. American Journal of Epidemiology, 169, 704–711. doi: 10.1093/aje/kwn397 Kohen, R., Cain, K. C., Buzaitis, A., Johnson, V., Becker, K. J., Teri, L., . . . Mitchell, P. H. (2011). Response to psychosocial treatment in poststroke depression is associated with serotonin transporter polymorphisms. Stroke, 42, 2068–2070. Kolassa, I.-T., Ertl, V., Eckart, C., Glöckner, F., Kolassa, S., Papassotiropoulos, A., . . . Elbert, T. (2010b). Association study of trauma load and SLC6A4 promoter polymorphism in posttraumatic stress disorder: Evidence from survivors of the Rwandan genocide. Journal of Clinical Psychiatry, 71, 543–547. doi: 10.4088/JCP.08m04787blu Kuepper, Y., Wielpuetz, C., Alexander, N., Mueller, E., Grant, P., & Hennig, J. (2012). 5-HTTLPR S-allele: A genetic plasticity factor regarding the effects of life events on personality? Genes, Brain & Behavior, 11, 643–650. Kumsta, R., Stevens, S., Brookes, K., Schlotz, W., Castle, J., Beckett, C., . . . Sonuga-Barke, E. J. (2010). 5HTT genotype moderates the influence of early institutional deprivation on emotional problems in adolescence: Evidence from the English and Romanian Adoptee (ERA) Study. Journal of Child Psychology and Psychiatry, 51, 755–762. doi: 10.1111/j.1469-7610.2010.02249.x Langley, K., Turic, D., Rice, F., Holmans, P., van den Bree, M. B., Craddock, N., . . . Thapar, A. (2008). Testing for gene × environment interaction effects in attention deficit hyperactivity disorder and associated antisocial behavior. American Journal of Medical Genetics. Neuropsychiatric Genetics, 147B, 49–53. doi: 10.1002/ajmg.b.30571 Laucht, M., Treutlein, J., Blomeyer, D., Buchmann, A. F., Schmid, B., Becker, K. J., . . . Banaschewski, T. (2009a). Interaction between the 5-HTTLPR serotonin transporter polymorphism and environmental adversity for mood and anxiety psychopathology: Evidence from a high-risk community sample of young adults. International Journal of Neuropsychopharmacology, 12, 737–747. doi: 10.1017/s146114570 8009875 Laucht, M., Treutlein, J., Schmid, B., Blomeyer, D., Becker, K. J., Buchmann, A. F., . . . Banaschewski, T. (2009b). Impact of psychosocial adversity on alcohol intake in young adults: Moderation by the LL genotype of the serotonin transporter polymorphism. Biological Psychiatry, 66, 102–109. doi: 10.1016/j.biopsych.2009.02.010 Lazary, J. (2010). Serotonin transporter gene and negative life change events are associated with depressive phenotype. Neuropsychopharmacologia Hungarica, 12, 347–354.

Appendix Lazary, J., Lazary, A., Gonda, X., Benko, A., Molnar, E., Juhasz, G., & Bagdy, G. (2008). New evidence for the association of the serotonin transporter gene (SLC6A4) haplotypes, threatening life events, and depressive phenotype. Biological Psychiatry, 64, 498–504. doi: 10.1016/j.biopsych.2008.03.030 Lemogne, C., Bergouignan, L., Boni, C., Gorwood, P., Pélissolo, A., & Fossati, P. (2009). Genetics and personality affect visual perspective in autobiographical memory. Consciousness and Cognition: An International Journal, 18, 823–830. doi: 10.1016/j.concog.2009.04.002 Lemogne, C., Gorwood, P., Boni, C., Pessiglione, M., Lehéricy, S., & Fossati, P. (2011). Cognitive appraisal and life stress moderate the effects of the 5-HTTLPR polymorphism on amygdala reactivity. Human Brain Mapping, 32, 1856–1867. doi: 10.1002/hbm.21150 Lerer, E., Kanyas, K., Karni, O., Ebstein, R. P., & Lerer, B. (2006). Why do young women smoke? II. Role of traumatic life experience, psychological characteristics and serotonergic genes. Molecular Psychiatry, 11, 771–781. doi: 10.1038/sj.mp. 4001855 Li, J. J., & Lee, S. S. (2010). Latent class analysis of antisocial behavior: Interaction of serotonin transporter genotype and maltreatment. Journal of Abnormal Child Psychology, 38, 789–801. doi: 10.1007/s10802010-9409-y Luijk, M. P. C. M., Roisman, G. I., Haltigan, J. D., Tiemeier, H., Booth-LaForce, C., van IJzendoorn, M. H., . . . BakermansKranenburg, M. J. (2011). Dopaminergic, serotonergic, and oxytonergic candidate genes associated with infant attachment security and disorganization? In search of main and interaction effects. Journal of Child Psychology and Psychiatry, 52, 1295–1307. doi: 10.1111/j.1469 -7610.2011.02440.x Mandelli, L., Antypa, N., Nearchou, F. A., Vaiopoulos, C., Stefanis, C. N., Serretti, A., & Stefanis, N. C. (2012). The role of serotonergic genes and environmental stress on the development of depressive symptoms and neuroticism. Journal of Affective Disorders, 142, 82–89. doi: 10.1016/j.jad.2012.03.047 Manuck, S. B., Bleil, M. E., Petersen, K. L., Flory, J. D., Mann, J. J., Ferrell, R. E., & Muldoon, M. F. (2005). The socio-economic status of communities predicts variation in brain serotonergic responsivity. Psychological Medicine, 35, 519–528. Manuck, S. B., Flory, J. D., Ferrell, R. E., & Muldoon, M. F. (2004). Socio-economic status covaries with central nervous system serotonergic responsivity as a function of allelic variation in the serotonin transporter gene-linked polymorphic region. Psychoneuroendocrinology, 29, 651–668. doi: 10.1016/s0306-4530(03)00094-5 Markus, C. R., & De Raedt, R. (2011). Differential effects of 5-HTTLPR genotypes on inhibition of negative emotional information following acute stress exposure and tryptophan challenge. Neuropsychopharmacology, 36, 819–826. doi: 10.1038/npp. 2010.221 Mata, J., & Gotlib, I. H. (2011). 5-HTTLPR moderates the relation between changes in depressive and bulimic symptoms in adolescent girls: A longitudinal study. International Journal of Eating Disorders, 44, 383–388. doi: 10.1002/eat.20850 May, M. E., Lightfoot, D. A., Srour, A., Kowalchuk, R. K., & Kennedy, C. H. (2010). Association between serotonin transporter polymorphisms and problem behavior in adult males with intellectual disabilities. Brain Research, 1357, 97–103. doi: 10.1016/j.brainres.2010.08 .052 Mehta, D., Quast, C., Fasching, P. A., Seifert, A., Voigt, F., Beckmann, M. W., . . . Goecke, T. W. (2012). The 5-HTTLPR polymorphism modulates the influence on environmental stressors on peripartum depression symptoms. Journal of Affective Disorders, 136, 1192–1197. doi: 10.1016/j.jad.2011.11.042 Meng, X., Kou, C., Shi, J., Yu, Y., & Huang, Y. (2011). Susceptibility genes, social environmental risk factors and their interactions

335

in internalizing disorders among mainland Chinese undergraduates. Journal of Affective Disorders, 132, 254–259. doi: 10.1016/j.jad.2011 .01.005 Mercer, K. B., Orcutt, H. K., Quinn, J. F., Fitzgerald, C. A., Conneely, K. N., Barfield, R. T., . . . Ressler, K. J. (2012). Acute and posttraumatic stress symptoms in a prospective gene × environment study of a university campus shooting. Archives of General Psychiatry, 69, 89–97. doi: 10.1001/archgenpsychiatry.2011.109 Middeldorp, C. M., de Geus, E. J., Willemsen, G., Hottenga, J. J., Slagboom, P. E., & Boomsma, D. I. (2010). The serotonin transporter gene length polymorphism (5-HTTLPR) and life events: No evidence for an interaction effect on neuroticism and anxious depressive symptoms. Twin Research and Human Genetics, 13, 544–549. doi: 10.1375/twin.13.6.544 Mileva-Seitz, V., Kennedy, J., Atkinson, L., Steiner, M., Levitan, R., Matthews, S. G., . . . Fleming, A. S. (2011). Serotonin transporter allelic variation in mothers predicts maternal sensitivity, behavior and attitudes toward 6-month-old infants. Genes, Brain & Behavior, 10, 325–333. doi: 10.1111/j.1601-183X.2010.00671.x Mingione, C. J., Heffner, J. L., Blom, T. J., & Anthenelli, R. M. (2012). Childhood adversity, serotonin transporter (5-HTTLPR) genotype, and risk for cigarette smoking and nicotine dependence in alcohol dependent adults. Drug and Alcohol Dependence, 123, 201–206. doi: 10.1016/j.drugalcdep. 2011.11.013 Minnix, J. A., Robinson, J. D., Lam, C. Y., Carter, B. L., Foreman, J. E., Vandenbergh, D. J., . . . Cinciripini, P. M. (2011). The serotonin transporter gene and startle response during nicotine deprivation. Biological Psychology, 86, 1–8. doi: 10.1016/j.biopsycho.2010.09.013 Mitchell, C., Notterman, D., Brooks-Gunn, J., Hobcraft, J., Garfinkel, I., Jaeger, K., . . . McLanahan, S. (2011). Role of mother’s genes and environment in postpartum depression. Proceedings of the National Academy of Sciences, 108, 8189–8193. doi: 10.1073/pnas.1014129108 Mueller, A., Armbruster, D., Moser, D. A., Canli, T., Lesch, K.-P., Brocke, B., & Kirschbaum, C. (2011). Interaction of serotonin transporter gene-linked polymorphic region and stressful life events predicts cortisol stress response. Neuropsychopharmacology, 36, 1332–1339. doi: 10.1038/npp. 2011.11 Nederhof, E., Bouma, E. M., Oldehinkel, A. J., & Ormel, J. (2010a). Interaction between childhood adversity, brain-derived neurotrophic factor val/met and serotonin transporter promoter polymorphism on depression: the TRAILS study. Biological Psychiatry, 68, 209–212. doi: 10.1016/j.biopsych.2010.04.006 Nederhof, E., Bouma, E. M. C., Riese, H., Laceulle, O. M., Ormel, J., & Oldehinkel, A. J. (2010b). Evidence for plasticity genotypes in a gene–gene–environment interaction: The TRAILS study. Genes, Brain & Behavior, 9, 968–973. doi: 10.1111/j.1601-183X.2010.00637.x Nijmeijer, J. S., Hartman, C. A., Rommelse, N. N. J., Altink, M. E., Buschgens, C. J. M., Fliers, E. A., . . . Hoekstra, P. J. (2010). Perinatal risk factors interacting with catechol O-methyltransferase and the serotonin transporter gene predict ASD symptoms in children with ADHD. Journal of Child Psychology and Psychiatry, 51, 1242–1250. doi: 10.1111/j.1469-7610.2010.02277.x Nikolas, M., Friderici, K. H., Waldman, I., Jernigan, K. A., & Nigg, J. T. (2010). Gene × environment interactions for ADHD: Synergistic effect of 5HTTLPR genotype and youth appraisals of inter-parental conflict. Behavioral and Brain Functions, 6, 23. doi: 10.1186/1744-9081-6-23 Nikolova, Y., Bogdan, R., & Pizzagalli, D. A. (2011). Perception of a naturalistic stressor interacts with 5-HTTLPR/rs25531 genotype and gender to impact reward responsiveness. Neuropsychobiology, 65, 45–54. doi: 10.1159/000329105 Nilsson, K. W., Oreland, L., Kronstrand, R., & Leppert, J. (2009). Smoking as a product of gene–environment interaction. Upsala Journal of Medical Science, 114, 100–107. doi: 10.1080/03009730902833406

336

The Trilogy of GxE: Conceptualization, Operationalization, and Application

Nobile, M., Giorda, R., Marino, C., Carlet, O., Pastore, V., Vanzin, L., . . . Battaglia, M. (2007). Socioeconomic status mediates the genetic contribution of the dopamine receptor D4 and serotonin transporter linked promoter region polymorphisms to externalization in preadolescence. Development and Psychopathology, 19, 1147–1160. doi: 10.1017/s0954579407000594 Nobile, M., Rusconi, M., Bellina, M., Marino, C., Giorda, R., Carlet, O., . . . Battaglia, M. (2009). The influence of family structure, the TPH2 G-703T and the 5-HTTLPR serotonergic genes upon affective problems in children aged 10–14 years. Journal of Child Psychology and Psychiatry, 50, 317–325. doi: 10.1111/j.1469-7610.2008.01958.x Paaver, M., Kurrikoff, T., Nordquist, N., Oreland, L., & Harro, J. (2008). The effect of 5-HTT gene promoter polymorphism on impulsivity depends on family relations in girls. Progress in Neuropsychopharmacology & Biological Psychiatry, 32, 1263–1268. doi: 10.1016/j.pnpbp. 2008.03.021 Pauli-Pott, U., Friedl, S., Hinney, A., & Hebebrand, J. (2009). Serotonin transporter gene polymorphism (5-HTTLPR), environmental conditions, and developing negative emotionality and fear in early childhood. Journal of Neural Transmission, 116, 503–512. doi: 10.1007/ s00702-008-0171-z Peirson, B. R. “Richard Woltereck (1877–1944).” Embryo Project Encyclopedia (2012). Peyrot, W. J., Middeldorp, C. M., Jansen, R., Smit, J. H., de Geus, E. J. C., Hottenga, J.-J., . . . Penninx, B. W. J. H. (2013). Strong effects of environmental factors on prevalence and course of major depressive disorder are not moderated by 5-HTTLPR polymorphisms in a large Dutch sample. Journal of Affective Disorders, 146, 91–99. doi: 10.1016/j.jad. 2012.08.044 Pluess, M., Belsky, J., Way, B. M., & Taylor, S. E. (2010). 5-HTTLPR moderates effects of current life events on neuroticism: Differential susceptibility to environmental influences. Progress in NeuroPsychopharmacology & Biological Psychiatry, 34, 1070–1074. doi: 10.1016/j.pnpbp.2010.05.028 Pluess, M., Velders, F. P., Belsky, J., van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., Jaddoe, V. W. V., . . . Tiemeier, H. (2011). Serotonin transporter polymorphism moderates effects of prenatal maternal anxiety on infant negative emotionality. Biological Psychiatry, 69, 520–525. doi: 10.1016/j.biopsych.2010.10.006 Power, T., Stewart, R., Ancelin, M.-L., Jaussent, I., Malafosse, A., & Ritchie, K. (2010). 5-HTTLPR genotype, stressful life events and late-life depression: No evidence of interaction in a French population. Neurobiology of Aging, 31, 886–887. doi: 10.1016/j.neurobiolaging .2008.06.006 Quinn, C. R., Dobson-Stone, C., Outhred, T., Harris, A., & Kemp, A. H. (2012). The contribution of BDNF and 5-HTT polymorphisms and early life stress to the heterogeneity of major depressive disorder: A preliminary study. Australian and New Zealand Journal of Psychiatry, 46, 55–63. doi: 10.1177/0004867411430878 Raby, K. L., Cicchetti, D., Carlson, E. A., Cutuli, J. J., Englund, M. M., & Egeland, B. (2012). Genetic and caregiving-based contributions to infant attachment: Unique associations with distress reactivity and attachment security. Psychological Science, 23, 1016–1023. doi: 10.1177/0956797612438265 Racine, S. E., Culbert, K. M., Larson, C. L., & Klump, K. L. (2009). The possible influence of impulsivity and dietary restraint on associations between serotonin genes and binge eating. Journal of Psychiatric Research, 43, 1278–1286. doi: 10.1016/j.jpsychires.2009.05.002 Reif, A., Rösler, M., Freitag, C. M., Schneider, M., Eujen, A., Kissling, C., . . . Retz, W. (2007). Nature and nurture predispose to violent behavior: Serotonergic genes and adverse childhood environment. Neuropsychopharmacology, 32, 2375–2383. doi: 10.1038/sj.npp.1301 359

Ressler, K. J., Bradley, B., Mercer, K. B., Deveau, T. C., Smith, A. K., Gillespie, C. F., . . . Binder, E. B. (2010). Polymorphisms in CRHR1 and the serotonin transporter loci: gene × gene × environment interactions on depressive symptoms. American Journal of Medical Genetics. Neuropsychiatric Genetics, 153B, 812–824. doi: 10.1002/ajmg.b.31052 Rethorst, C. D., Landers, D. M., Nagoshi, C. T., & Ross, J. T. D. (2011). The association of 5-HTTLPR genotype and depressive symptoms is moderated by physical activity. Journal of Psychiatric Research, 45, 185–189. doi: 10.1016/j.jpsychires.2010.05.007 Retz, W., Freitag, C. M., Retz-Junginger, P., Wenzler, D., Schneider, M., Kissling, C., . . . Rosler, M. (2008). A functional serotonin transporter promoter gene polymorphism increases ADHD symptoms in delinquents: Interaction with adverse childhood environment. Psychiatry Research, 158, 123–131. doi: 10.1016/j.psychres.2007.05.004 Ritchie, K., Jaussent, I., Stewart, R., Dupuy, A. M., Courtet, P., Ancelin, M. L., & Malafosse, A. (2009). Association of adverse childhood environment and 5-HTTLPR Genotype with late-life depression. Journal of Clinical Psychiatry, 70, 1281–1288. doi: 10.4088/JCP.08m04510 Roy, A., Hu, X. Z., Janal, M. N., & Goldman, D. (2007). Interaction between childhood trauma and serotonin transporter gene variation in suicide. Neuropsychopharmacology, 32, 2046–2052. doi: 10.1038/ sj.npp. 1301331 Sadeh, N., Javdani, S., Jackson, J. J., Reynolds, E. K., Potenza, M. N., Gelernter, J., . . . Verona, E. (2010). Serotonin transporter gene associations with psychopathic traits in youth vary as a function of socioeconomic resources. Journal of Abnormal Psychology, 119, 604–609. doi: 10.1037/a0019709 Saiz, P. A., Garcia-Portilla, M. P., Herrero, R., Arango, C., Corcoran, P., Morales, B., . . . Bobes, J. (2010). Interactions between functional serotonergic polymorphisms and demographic factors influence personality traits in healthy Spanish Caucasians. Psychiatric Genetics, 20, 171–178. doi: 10.1097/YPG.0b013e32833a20b9 Scheid, J. M., Holzman, C. B., Jones, N., Friderici, K. H., Jernigan, K. A., Symonds, L. L., . . . Fisher, R. (2011). Life stressors and 5-HTTLPR interaction in relation to midpregnancy depressive symptoms among African-American women. Psychiatric Genetics, 21, 271–280. doi: 10.1097/YPG.0b013e32834603e8 Scheid, J. M., Holzman, C. B., Jones, N., Friderici, K. H., Nummy, K. A., Symonds, L. L., . . . Fisher, R. (2007). Depressive symptoms in mid-pregnancy, lifetime stressors and the 5-HTTLPR genotype. Genes, Brain & Behavior, 6, 453–464. doi: 10.1111/j.1601-183X. 2006.00272.x Shinozaki, G., Romanowicz, M., Kung, S., & Mrazek, D. A. (2011). A new interaction between SLC6A4 variation and child abuse is associated with resting heart rate. Depression and Anxiety, 28, 227–233. Sjöberg, R. L., Nilsson, K. W., Nordquist, N., Öhrvik, J., Leppert, J., Lindström, L., & Oreland, L. (2006). Development of depression: Sex and the interaction between environment and a promoter polymorphism of the serotonin transporter gene. International Journal of Neuropsychopharmacology, 9, 443–449. doi: 10.1017/s1461145705005936 Sonuga-Barke, E. J. S., Oades, R. D., Psychogiou, L., Chen, W., Franke, B., Buitelaar, J. K., . . . Faraone, S. V. (2009). Dopamine and serotonin transporter genotypes moderate sensitivity to maternal expressed emotion: The case of conduct and emotional problems in attention deficit hyperactivity disorder. Journal of Child Psychology and Psychiatry, 50, 1052–1063. doi: 10.1111/j.1469-7610.2009.02095.x Spangler, G., Johann, M., Ronai, Z., & Zimmermann, P. (2009). Genetic and environmental influence on attachment disorganization. Journal of Child Psychology and Psychiatry, 50, 952–961. doi: 10.1111/ j.1469-7610.2008.02054.x Steiger, H., Richardson, J., Joober, R., Gauvin, L., Israel, M., Bruce, K. R., . . . Young, S. N. (2007). The 5HTTLPR polymorphism, prior

Appendix maltreatment and dramatic-erratic personality manifestations in women with bulimic syndromes. Journal of Psychiatry & Neuroscience, 32, 354–362. Stein, M. B., Schork, N. J., & Gelernter, J. (2008). Gene-by-environment (serotonin transporter and childhood maltreatment) interaction for anxiety sensitivity, an intermediate phenotype for anxiety disorders. Neuropsychopharmacology, 33, 312–319. doi: 10.1038/sj.npp. 1301422 Stoltenberg, S. F., Anderson, C., Nag, P., & Anagnopoulos, C. (2012a). Association between the serotonin transporter triallelic genotype and eating problems is moderated by the experience of childhood trauma in women. International Journal of Eating Disorders, 45, 492–500. doi: 10.1002/eat.20976 Stoltenberg, S. F., Christ, C. C., & Highland, K. B. (2012b). Serotonin system gene polymorphisms are associated with impulsivity in a context dependent manner. Progress in Neuropsychopharmacology & Biological Psychiatry, 39, 182–191. doi: 10.1016/j.pnpbp. 2012.06.012 Sugden, K., Arseneault, L., Harrington, H., Moffitt, T. E., Williams, B., & Caspi, A. (2010). Serotonin transporter gene moderates the development of emotional problems among children following bullying victimization. Journal of the American Academy of Child & Adolescent Psychiatry, 49, 830–840. doi: 10.1016/j.jaac.2010.01.024 Surtees, P. G., Wainwright, N. W., Willis-Owen, S. A., Luben, R., Day, N. E., & Flint, J. (2006b). Social adversity, the serotonin transporter (5-HTTLPR) polymorphism and major depressive disorder. Biological Psychiatry, 59, 224–229. doi: 10.1016/j.biopsych.2005.07.014 Taylor, S. E., Way, B. M., Welch, W. T., Hilmert, C. J., Lehman, B. J., & Eisenberger, N. I. (2006). Early family environment, current adversity, the serotonin transporter promoter polymorphism, and depressive symptomatology. Biological Psychiatry, 60, 671–676. doi: 10.1016/j.biopsych.2006.04.019 Thomason, M. E., Henry, M. L., Hamilton, J. P., Joormann, J., Pine, D. S., Ernst, M., . . . Gotlib, I. H. (2010). Neural and behavioral responses to threatening emotion faces in children as a function of the short allele of the serotonin transporter gene. Biological Psychology, 85, 38–44. doi: 10.1016/j.biopsycho.2010.04.009 Tiemeier, H., Velders, F. P., Szekely, E., Roza, S. J., Dieleman, G., Jaddoe, V. W. V., . . . Verhulst, F. C. (2012). The generation R study: A review of design, findings to date, and a study of the 5-HTTLPR by environmental interaction from fetal life onward. Journal of the American Academy of Child & Adolescent Psychiatry, 51, 1119–1135. doi: 10.1016/j.jaac.2012.08.021 Uddin, M., de los Santos, R., Bakshis, E., Cheng, C., & Aiello, A. E. (2011). Building conditions, 5-HTTLPR genotype, and depressive symptoms in adolescent males and females. Journal of Adolescent Health, 49, 379–385. doi: 10.1016/j.jadohealth.2011.01.013 Uddin, M., Koenen, K. C., de los Santos, R., Bakshis, E., Aiello, A. E., & Galea, S. (2010). Gender differences in the genetic and environmental determinants of adolescent depression. Depression and Anxiety, 27, 658–666. doi: 10.1002/da.20692 Uher, R., Caspi, A., Houts, R., Sugden, K., Williams, B., Poulton, R., & Moffitt, T. E. (2011). Serotonin transporter gene moderates childhood maltreatment’s effects on persistent but not single-episode depression: Replications and implications for resolving inconsistent results. Journal of Affective Disorders, 135, 56–65. doi: 10.1016/j.jad.2011.03.010 van Roekel, E., Engels, R., Verhagen, M., Goossens, L., & Scholte, R. H. J. (2011). Parental depressive feelings, parental support, and the serotonin transporter gene as predictors of adolescent depressive feelings: A latent growth curve analysis. Journal of Youth and Adolescence, 40, 453–462. doi: 10.1007/s10964-010-9560-3 van Roekel, E., Scholte, R. H., Verhagen, M., Goossens, L., & Engels, R. C. (2010). Loneliness in adolescence: gene × environment interactions

337

involving the serotonin transporter gene. Journal of Child Psychology and Psychiatry, 51, 747–754. doi: 10.1111/j.1469-7610.2010.02225.x Vang, F. J., Lindström, M., Sunnqvist, C., Bah-Rösman, J., Johanson, A., & Träskman-Bendz, L. (2009). Life-time adversities, reported thirteen years after a suicide attempt: Relationship to recovery, 5HTTLPR genotype, and past and present morbidity. Archives of Suicide Research, 13, 214–229. doi: 10.1080/13811110903044328 Vaske, J., Newsome, J., & Wright, J. P. (2012). Interaction of serotonin transporter linked polymorphic region and childhood neglect on criminal behavior and substance use for males and females. Development and Psychopathology, 24, 181–193. doi: 10.1017/s0954579411000769 Vaughn, M. G., DeLisi, M., Beaver, K. M., & Wright, J. P. (2009). DAT1 and 5HTT are associated with pathological criminal behavior in a nationally representative sample of youth. Criminal Justice and Behavior, 36, 1113–1124. doi: 10.1177/0093854809342839 Verona, E., Joiner, T. E., Johnson, F., & Bender, T. W. (2006). Gender specific gene–environment interactions on laboratory-assessed aggression. Biological Psychology, 71, 33–41. doi: 10.1016/j.biopsycho .2005.02.001 Verschoor, E., & Markus, C. R. (2011). Affective and neuroendocrine stress reactivity to an academic examination: influence of the 5-HTTLPR genotype and trait neuroticism. Biological Psychology, 87, 439–449. doi: 10.1016/j.biopsycho.2011.06.001 Verschoor, E., & Markus, C. R. (2011b). Effects of acute psychosocial stress exposure on endocrine and affective reactivity in college students differing in the 5-HTTLPR genotype and trait neuroticism. Stress, 14, 407–419. Wagner, S., Baskaya, Ô., Lieb, K., Dahmen, N., & Tadi´c, A. (2009). The 5-HTTLPR polymorphism modulates the association of serious life events (SLE) and impulsivity in patients with borderline personality disorder. Journal of Psychiatric Research, 43, 1067–1072. doi: 10.1016/j.jpsychires.2009.03.004 Walsh, N. D., Dalgleish, T., Dunn, V. J., Abbott, R., St Clair, M. C., Owens, M., . . . Goodyer, I. M. (2012). 5-HTTLPR–environment interplay and its effects on neural reactivity in adolescents. NeuroImage, 63(3), 1670–1680. doi: 10.1016/j.neuroimage.2012.07.067 Way, B. M., & Taylor, S. E. (2010). The serotonin transporter promoter polymorphism is associated with cortisol response to psychosocial stress. Biological Psychiatry, 67, 487–492. doi: 10.1016/j.biopsych. 2009.10.021 Way, B. M., & Taylor, S. E. (2011). A polymorphism in the serotonin transporter gene moderates cardiovascular reactivity to psychosocial stress. Psychosomatic medicine, 73, 310–317. doi: 10.1097/PSY. 0b013e31821195ed Wichers, M., Kenis, G., Jacobs, N., Mengelers, R., Derom, C., Vlietinck, R., & van Os, J. (2008). The BDNF Val(66)Met × 5-HTTLPR × child adversity interaction and depressive symptoms: An attempt at replication. American Journal of Medical Genetics. Neuropsychiatric Genetics, 147B, 120–123. doi: 10.1002/ajmg.b.30576 Wilhelm, K., Mitchell, P. B., Niven, H., Finch, A., Wedgwood, L., Scimone, A., . . . Schofield, P. R. (2006). Life events, first depression onset and the serotonin transporter gene. The British Journal of Psychiatry, 188, 210–215. doi: 10.1192/bjp.bp. 105.009522 Williams, R. B., Marchuk, D. A., Siegler, I. C., Barefoot, J. C., Helms, M. J., Brummett, B. H., . . . Schanberg, S. M. (2008). Childhood socioeconomic status and serotonin transporter gene polymorphism enhance cardiovascular reactivity to mental stress. Psychosomatic medicine, 70, 32–39. doi: 10.1097/PSY.0b013e31815f66c3 Xie, P., Kranzler, H. R., Poling, J., Stein, M. B., Anton, R. F., Brady, K., . . . Gelernter, J. (2009). Interactive effect of stressful life events and the serotonin transporter 5-HTTLPR genotype on posttraumatic stress disorder diagnosis in 2 independent populations. Archives

338

The Trilogy of GxE: Conceptualization, Operationalization, and Application

of General Psychiatry, 66, 1201–1209. doi: 10.1001/archgenpsychiatry.2009.153 Zalsman, G., Huang, Y.-y., Oquendo, M. A., Burke, A. K., Hu, X.-z., Brent, D. A., . . . Mann, J. J. (2006). Association of a triallelic serotonin transporter gene promoter region (5-HTTLPR) polymorphism with stressful life events and severity of depression. The American Journal of Psychiatry, 163, 1588–1593. doi: 10.1176/appi.ajp. 163.9.1588 Zannas, A. S., McQuoid, D. R., Steffens, D. C., Chrousos, G. P., & Taylor, W. D. (2012). Stressful life events, perceived stress, and 12-month course of geriatric depression: direct effects and moderation by the 5-HTTLPR and COMT Val158Met polymorphisms. Stress, 15, 425–434. doi: 10.3109/10253890.2011.634263

Zavos, H. M., Wong, C. C., Barclay, N. L., Keers, R., Mill, J., Rijsdijk, F. V., . . . Eley, T. C. (2012). Anxiety sensitivity in adolescence and young adulthood: The role of stressful. Depression and Anxiety, 29, 400–408. Zhang, K., Xu, Q., Xu, Y., Yang, H., Luo, J., Sun, Y., . . . Shen, Y. (2009). The combined effects of the 5-HTTLPR and 5-HTR1A genes modulates the relationship between negative life events and major depressive disorder in a Chinese population. Journal of Affective Disorders, 114, 224–231. doi: 10.1016/j.jad.2008.07.012 Zimmermann, P., Mohr, C., & Spangler, G. (2009). Genetic and attachment influences on adolescents’ regulation of autonomy and aggressiveness. Journal of Child Psychology and Psychiatry, 50, 1339–1347. doi: 10.1111/j.1469-7610.2009.02158.x

CHAPTER 9

Genetics and Family Systems: Articulation and Disarticulation DAVID REISS

GENETICS AND THE SOCIAL SCIENCES: A HISTORICAL SKETCH 339 THE TWIN METHOD AND THE EQUAL ENVIRONMENTS ASSUMPTION 340 THE ADOPTION METHOD AND SELECTIVE PLACEMENT 342 QUANTITATIVE GENETICS AND FAMILY THEORY: NONSHARED ENVIRONMENT, GENETIC INFLUENCES ON ENVIRONMENTAL MEASURES, AND REINTERPRETING ENVIRONMENTAL EFFECTS 344 The Nonshared and Shared Environment 345 Genetic Influences on Measures of the Family Environment 348 MOLECULAR GENETICS AND THE FAMILY 352 Cautions in Interpreting Molecular Data 353

SPECIFYING THEORIES OF FAMILY PROCESS AND INDIVIDUAL DIFFERENCES 356 Differences Among Children in Their Response to Parenting 356 Differences Among Children in Evoking Parental Responses 357 Individual Differences and Marital Processes 358 Differences Among Families in Response to Family-Level Interventions 359 Beyond Individual Differences: The Prospects for a Developmental Biology of Families 359 HAS GENETICS ALTERED THE WAY WE THINK ABOUT THE FAMILY AS A SOCIAL SYSTEM? 363 The State of the Evidence 363 Family Research Listening to Molecular Genetics 365 Family Research Listening to Quantitative Genetics 367 SUMMARY 369 REFERENCES 370

This chapter focuses on the rapprochement between genetics and the social sciences in the study of the family. This rapprochement began nearly a century ago and has accelerated in the last decade. This rapprochement has revised our concepts of family development, of how its subsystems influence one another and how the family shapes development of its child and adult members. This rapprochement extends into family-level interventions designed to treat individual and relational disorders or for preventing them. However, despite nearly a century of intellectual transaction between social and genetic approaches to the family system there remains troublesome disarticulation between the two perspectives. Moreover, there is notable disarticulation within the domain of genetic analyses between quantitative (twin and adoption methods) and molecular approaches.

genetics and social science. It is now possible to recognize significant turning points in a century-long history of combat, truce, and alliance between these two essential approaches to the study of human development. The purpose of this review is to outline how this emerging rapprochement of social and genetic analysis has reshaped our understanding of the family and to suggest the practical significance of this revised view. We also suggest that rapprochement has helped reshape our view of genetics and, while not a primary focus of this chapter, some of these changes will be noted in order to gain a better understanding that—in the rapprochement of two lines of work—both the science of genetics and of family systems has been irreversibly altered. As we note later, the systematic application of genetic principles to understand emotions and behavior extends back millennia. However, we lack a published history of behavioral genetics, as this field of inquiry is now termed, although there have been preliminary efforts (Gottesman, 2008; McGue, 2010). Moreover, there is no systematic effort to review the history of the transactions between

GENETICS AND THE SOCIAL SCIENCES: A HISTORICAL SKETCH The rapprochement achieved so far emerges from a long, and sometimes troubled, history of the connection between 339

340

Genetics and Family Systems

behavioral genetics and social science. In this vacuum several assumptions have arisen; among the most important of them is a view that social science is the older, softer, and less coherent discipline and that its inferences about families and human development must be significantly revised in the light of the much newer discipline of genetics. In this chapter, we will review areas where genetic studies require major revisions in thinking about the family. However, a systematic history of the interplay of genetics and social science might offer a different perspective: both the formal study of behavioral genetics and of social science have their roots in the late nineteenth century and each prodded the other to change during the ensuing decades. The ancients had a firm practical grasp of both genetics and family systems. Darwin in Variations in Plants and animals under domestication (Darwin, 1920 (1868) documented the acumen of the ancients in cross breeding. He cited the thirtieth chapter of Genesis, where Jacob exacts revenge on his father-in-law by the selective breeding of Lavan’s sheep (and became wealthy by keeping the best). Further, after Darwin praises the keen observations of and implicit knowledge of heredity by British animal and plant breeders, he cites Plato, Alexander the Great and Virgil for equal acumen. The Book of Genesis also provides evidence of the keenest observations of the interplay between family process and social convention. It would be hard to find in current accounts a shrewder appreciation of family social systems than the story of Tamara’s successful seduction of her father-in-law, an account interleaved between two halves of the sage account of the interplay between parent–child and sibling relationships in the story of Jacob, Joseph, and his brothers. At what point does keen observation and intuitive understanding of lawfulness of biological and social processes shade into science? While this question may give way to endless, and some might argue fruitless, philosophical debate there seems little doubt that the exacting experiments with peas by Gregor Mendel marks a starting point for scientific genetics. Mendel presented his results to the Natural History Society of Brunn in 1865. The society published his paper the following year (Mendel, 1948 (1866)) in their journal to which a large number of European scholarly libraries subscribed. Further, Mendel was said to have distributed many reprints of his paper and drew attention to it in scientific correspondence (Hunter, 2000). However, it was not until 1900 that Mendel’s work was recognized and highlighted by the German botanist, Carl Correns (Correns, 1950). By this time a searching and rigorous social science was well underway as exemplified

by the remarkable study of the social factors in suicide by Emil Durkheim (Durkheim, 1951 (1897)). With the same rigor that Mendel sought to illustrate the independent assortment of heritable traits of the pea plant so did Durkheim painstakingly attempt an isolation of factors of social cohesion—in religion, family life, and political stability—that prevented suicide. As we now understand the analysis of covariance, confounding, and counterfactuals, Durkheim understood that testing his thesis required accounting for the potentially confounding role of mental illness, alcoholism, heredity, season, and geographical location. Well before these historically decisive publications by Corrans and Durkheim, Francis Galton, with his strong interest in both heredity and life circumstance, published the first paper on the use of presumptively identical twins to identify sibling specific experiences in life that led to differences in their competence in adult life (Galton, 1876), a phenomenon that now goes by the awkward name nonshared environment and is investigated using the same strategy.

THE TWIN METHOD AND THE EQUAL ENVIRONMENTS ASSUMPTION The current age of behavioral genetics, however, had a very clear starting point: the discovery of the twin method. As Rende, Plomin, and Vandenburg (1990) pointed out, the first use of the twin method—as we know it today—was by the Stanford graduate student Curtis Merriman. His was the first of many major advances in behavioral genetics to be developed by a graduate student. Merriman published his thesis in 1924 to argue first, that there were two types of twins: fraternal and duplicative. Then he demonstrated that the correlation on intelligence tests between twins was higher for the latter. Since there were no agreed on criteria for identifying duplicative twins he knowingly used a short cut that blunted his analysis. Opposite-sex twins must be fraternal and among same-sex twins roughly half are duplicative. Hence, a comparison of same-sex versus opposite-sex twins should provide a lower bound for estimating genetic influence on intelligence. Rende and his colleagues also underscored an even more direct use of the twin method by the German dermatologist Herman Siemens who, in 1924, directly compared concordance rates and correlations between identical (one egg as he termed them) and fraternal (two-egg) twins. Siemens understandably focused on skin conditions but also reported on children’s mental capacities and school performance. His 1927 paper anticipates modern-day diagnosis of twinships

The Twin Method and the Equal Environments Assumption

specifying physical characteristics that are likely to be identical in one-egg twins (Siemens, 1927). In the two decades following Merriman’s (1924) and Siemens’ (1927) path breaking advances the twin method was applied to a broad range of human variability. Perhaps none were as controversial as those of Franz Kallman, who estimated high heritability of schizophrenia from comparing concordance rates of monozygotic (MZ) and dizygotic (DZ) twins (Kallmann, 1946). Conflict between the study of families as social systems and the study of genetics reached a fever point in the ensuing years because family systems researchers proposed an entirely alternate theory of the origins of schizophrenia as a disorder shaped by abnormal interactions in the family. The most influential and enduring social formulation was that of Margaret Singer and Lyman Wynne, who developed a family process concept of transactional thought disorder. Parents’ communication style, particularly their vague, amorphous, and shifting descriptions of ambiguous stimuli, could induce thought disorder and ultimately schizophrenia in their children. These parental patterns could be readily observed in the parents’ responses to Rorschach blots. Singer and Wynne could, using blind coding of parental communication deviance and child’s thought disturbance, successfully identify parents with a schizophrenic offspring (e.g., Singer & Wynne, 1965). Their argument for the social transmission of schizophrenia was indirectly buttressed by a stream of attacks on the twin method, mainly by family clinicians and researchers (Jackson, 1960), criticism that continues (Joseph, 1998, 2001). Critics compiled evidence that identical twins were treated much more similarly by their parents than were fraternal twins and hence differences of between-twin correlations between the two groups of twins reflected environmental rather than genetic influences. As strident as this debate became, it also set the stage for an extended process of scientific conflict resolution since both sides, the geneticists and the environmentalists, could agree that the validity of key assumptions of the twin method were not only critical but also testable. The fundamental assumption of the twin method was that the environments (social and physical) of DZ twins were as correlated as those of MZ twins: the equal environments assumption. Thus, an early order of business was the resolution of debate about the equal environments assumption of the twin method, an endeavor that reflects equally the work of social scientists and geneticists. The first step towards exploring the equal environments assumption was a more precise definition of this assumption. Clearly, family researchers were correct about the special treatment

341

of MZ twins: they were more frequently dressed alike, had names that went together such as Jane and June and were more likely to be confused by family members and strangers alike. But if estimates were of, say, the heritability of schizophrenia, could it be argued that clothes, first names, or even being confused with one’s sib are plausible etiologies of schizophrenia? Moreover, MZ twins might have more correlated environments because they are genetically similar (Eaves, Foley, & Silberg, 2003). For example, babies with irritable temperaments evoke annoyance or withdrawal of their parents. If this temperament is heritable then evoked reactions of parents will be more similar for MZ than DZ twins because the former are identical. Thus, the definition of the equal environments assumption focused on the equality of correlated environments plausibly linked to the behavior under investigation that do not arise as a consequence of the twins’ genotype. Stated thusly, the equal environments assumption is not easy to test by straightforward computations in traditional twins studies. Two strategies, however, have demonstrated its validity. The first has been analysis of twin samples where twins or their parents or both have misperceived their zygosity. Usually, it is MZ twins that are misperceived as DZ though the reverse sometimes happens. Early studies made efforts to correctly diagnose zygosity through fingerprinting (Goodman & Stevenson, 1989) or blood typing (Scarr, Scarf, & Weinberg, 1980) and recent studies have added more elaborate genetic testing (Conley, Rauscher, Dawes, Magnusson, & Siegal, 2013). All studies published thus far, using this approach, have shown that twin misperception does not materially affect heritability estimates: MZ-DZ differences are the same or greater in genetically diagnosed twins as they when the distinction is made only by the parents’ or the twins’ perceptions. A second strategy provides even more persuasive evidence: the analysis of twins reared apart. In two separate samples (Bouchard, Lykken, McGue, Segal, & Tellegen, 1990; Pedersen, Plomin, McClearn, & Friberg, 1988; Plomin, Pedersen, McClearn, Nesselroade, & Bergeman, 1988) correlations between MZ twins reared apart were equal to, or nearly so, for many but not all measures of temperament, personality, cognitive measures, and a range of physical measures such as height and blood pressure. These important confirmations of the equal environments assumption should have allayed the major concerns of social scientists studying the families although it needs testing continuously as the twin method is used in new domains of study where the equal environments assumption might not hold. Moreover, other assumptions lie beneath the standard computations of the twin method

342

Genetics and Family Systems

that need constant examination including the assumed absence of notable interaction of the environment with a subjects entire genotype and no genetically-based mate selection in the parents of the twins.

THE ADOPTION METHOD AND SELECTIVE PLACEMENT There are two different designs that capitalize on the natural experiment of adoption. Both involve comparing a characteristic of the parent, or the environment shaped by the parent, with characteristics of the child. In one design, adoptive parents and their children are compared and this comparison is contrasted with identical comparisons between a second group: parents rearing their biological children. A second, more difficult design, recruits birth parents who have placed their child for adoption at birth or very soon thereafter. Comparisons between the birth parents and the child are contrasted with those between the adoptive parent and their child. While the child may be placed at birth, the biological mother cares for her child for 9 months of pregnancy so that potential effects of prenatal rearing must—where possible—be ruled out. A recent design, to be discussed later, capitalizes on the new technology of in vitro fertilization. Mother-child comparisons are made between two groups: one where the mother’s own egg is fertilized and then implanted in her womb and another where a donor’s egg is implanted (Rice et al., 2009). The first use of the adoption design was published only three years after Merriman and Siemens published the first use of twin studies. In 1927, Barbara Stoddard Burks, a psychologist and student of Lewis Terman, published an elegant paper comparing adoptive parents’ intelligence scores with their adopted children and contrasted that with a closely matched group of biological parents raising their own children (Burks, 1927). This remarkable report recognized that biological parents provided both environmental and genetic influences whereas adoptive parents only environmental. Where parent–child correlations for the biological group exceeded those of the adoptive group, she reasoned, the difference represented the role of genes in parent–child transmission. Moreover, she recognized that parental intelligence was a limited measure of environmental influence on children and thus included a culture index of the home that included parental education, interests, and the size of their library—a measure that anticipated current and widely used measures of the home environment (Caldwell & Bradley, 1978). Finally,

Burks recognized that valid inferences drawn from the adoption design assume that the adoption process does not selectively place children in order to match characteristics of the birth parents—say intelligence, interests, or occupational background—with those of the adopting parents. The selective placement assumption is logically equivalent to the equal environments assumption; if it is violated results will seriously overestimate genetic effects. In presenting the results of this prescient paper, Burks exploited a special advantage of the adoption method: the capacity to directly assess the environmental effects of parental characteristics and of the home environment on child development. She found evidence that both genes and environment contributed to variation in child IQ. Two teams independently carried out the first use of the adoption method for the study of the transmission of psychopathology from parents to children. Leonard Heston, who at the time of his study was a psychiatric resident, reported a comparison of the adopted away offspring of a group of mothers with schizophrenia, hospitalized in Oregon, and a matched control group born to mothers without major mental illness. The children were placed at or near birth in a foundling home or with a family; the latter included some who were relatives of the mother with schizophrenia (a potential confound in an otherwise remarkable effort by an investigator taking his first step in research). Heston reported a preponderance of psychiatric problems in the offspring of schizophrenic mothers (Heston, 1966). At the time Heston was tracking the offspring of psychiatric patients after their Oregon hospitalization, Seymour Kety, Paul Wender, David Rosenthal, and Fini Schulsinger were tracking adopted offspring of schizophrenic female patients in Denmark and also both the biological and adopted relatives of the adopted offspring who became schizophrenic. They reported a preponderance of psychopathology in the offspring of the schizophrenic parents (Rosenthal et al., 1968) as well as in the biological but not in the adoptive relatives of adopted offspring who developed schizophrenia (Kety, Rosenthal, Wender, & Schulsinger, 1968). These data were presented at an epochal meeting in Dorado Beach, Puerto Rico where Heston reported additional data from his study showing no effect of the rearing conditions of his adoptees—institutional care versus family care—on the psychopathology he measured (Heston & Denney, 1968). The cumulated evidence argued for an important role of genetics in the parent-to-child transmission of schizophrenia, entirely circumventing the weaknesses of the twin method. However, these investigators also concluded that that there was little evidence that rearing

The Adoption Method and Selective Placement

played any role in schizophrenia. Heston showed no effect of the extreme variation between foundling-reared and family-reared adoptees, and the Kety group found no excess of schizophrenia and related disorders in the adoptive parents. The impact of these presentations and their conclusions were particularly dramatic since they directly followed a summary at Dorado Beach by Lyman Wynne of his many studies showing an association of communication deviance in parents with their offspring’s schizophrenia (Wynne, 1968). Although Wynne was circumspect about the etiological significance of his work, the overwhelming impact at the time was that genetics were important and the social origins of this disorder were, at the very least, unproven. However there were serious flaws—clear now after more than 40 years—in the arguments of both camps. The Danish and Oregon teams did not recognize that institutional versus family rearing—the key comparison in the Heston report—might not be the decisive environmental contributor to schizophrenia. (The effects of institutional rearing on child development were clarified in a much later adoption study by Michael Rutter and his colleagues (Rutter & O’Connor, 2004) where the effects were not psychotic disorders but on cognitive development and on promiscuous attachment to strangers.) Similarly, the diagnosis of schizophrenia in a rearing parent—the focus of the Rosenthal group—was not necessarily an index of a home environment specifically pathogenic for schizophrenia. Both teams missed the opportunity to do a simple cross tabulation to detect the possibility that a poor rearing environment might enhance the impact of genetic liability—however crudely measured—although the small samples were vastly underpowered to test for statistical interactions. For their part the Wynne group underestimated the importance of the genetic relatedness of the parents and children they studied. While they viewed parental Rorschach responses from an interpersonal perspective their measures might reflect only the behavioral expression of the same genes in the parents that, transmitted to their children, enhance their liability for schizophrenia. Now called passive gene–environment correlation, this form of parent–child transmission is a major counterfactual to all studies seeking to establish a causal link between family process and child development where parents and children are genetically related. Interestingly, the adoption design provides a uniquely direct method for estimating the importance of passive gene–environment correlation. For example, using the same adoption design as Barbara Stoddard Burks seven decades later, McGue and his colleagues

343

reported an association between parental alcohol dependence and impulsive and delinquent behavior in their offspring. However, this association held only for biological parents rearing their own offspring and not for adoptive parents and their children. The findings suggested parental alcoholism was not an environmental risk factor for offspring behavioral problems but that genetic factors influencing alcoholism in parents were passively transmitted to children where those same genetic factors impaired the offspring’s behavioral self-control (King et al., 2009). There were important epilogues to this dramatic Dorado Beach confrontation, a confrontation that may be among the most pivotal moments in the history of developmental psychopathology. Lyman Wynne partnered with the Finnish psychiatrist Pekka Tienari to replicate and extend the Danish and Oregon studies. Tienari tirelessly tracked children of Finnish schizophrenic mothers placed for adoption at an early age and developed a closely matched control group of adoptees whose biological parents were not psychotic. The ultimate sample was almost five times larger than the Danish and Oregon studies and included measures of communication deviance as well as other direct measures of family process. Assembling the sample took years, and it was not until 30 years after the Dorado Beach conference that the first major results were published. Parental communication deviance did not have a main effect on schizophrenia, but it did have decisive moderating effect on the genetic liability for thought disorder in the adopted offspring. At high levels of communication deviance offspring with genetic liability showed more thought disorder than did offspring without genetic liability, but at low levels of communication deviance the offspring with genetic liability showed less thought disorder than controls (Wahlberg, Wynne, Oja, Keskitalo, et al., 1997). Wynne and his colleagues recognized this pattern as one formulated by Kendler and Eaves 9 years before; Kendler and Eaves (1986) termed it genetic control of sensitivity to the environment. Individuals with this sensitivity genotype would show greater adaptation that those without the genotype under favorable environmental conditions but do worse under unfavorable conditions. This form of gene–environment interaction was reintroduced by Jay Belsky and colleagues with the term differential susceptibility (Belsky, Bakermans-Kranenburg, & Van IJzendoorn, 2007; Belsky & Pluess, 2009) apparently without recognizing either the original contribution of Kendler and Eaves or the findings of Wynne and Tienari. Perhaps because Belsky and his colleagues were better known in developmental psychology their papers became the basis for a wide range of research on differential susceptibility

344

Genetics and Family Systems

gene–environment research in developmental psychology and psychopathology. The Dorado Beach conference had a second long-term epilogue, one that led to a major disarticulation of genetic and family process studies that rivaled that of the 1960s. The data from the Danish and Oregon studies were, legitimately, very encouraging to research on the genetics of schizophrenia. After all this was an entirely different method than the twin study that was still under a cloud of methodological concern. Although a review of subsequent work on the genetics of schizophrenia is well beyond the scope of this chapter, it is important to recognize that many subsequent twin studies established that the genetic factors contributed well over half the variance for the schizophrenia diagnosis, that the search for specific genes—selected in advance of study as candidates—was unsatisfying and that a current focus is on genome-wide association studies (GWAS) using samples that exceed 30,000 individuals (for a review of the current status of this field along with a commentary on family and other environmental effects see Iyegbe, Campbell, Butler, Ajnakina, & Sham, 2014). Improved methods of genotyping have identified over 100 gene variants that are reliably associated with the schizophrenia phenotype; all these methods assume that important genes in the pathogenesis of schizophrenia enhance liability to illness and not, as initially suggested by Kendler and Eaves and demonstrated in the finding of Wynne and colleagues, that important genes might enhance sensitivity to environment. Comprehensive reviews of the etiology of schizophrenia acknowledge the large role of a broad range of prenatal and postnatal environmental social factors associated with schizophrenia. Some mention the work of Wynne (Tandon, Keshavan, & Nasrallah, 2008), and some do not (Iyegbe et al., 2014). To encompass the probability that environmental factors might additively or interactively contribute to GWAS analyses in understanding the pathogenesis of schizophrenia, GWAS researchers have proposed polygenic risk scores to summate the genetic component of such multifactorial inquiry. However, these scores reflect only genotypes identified as enhancing disease liability. The fundamental disarticulation here is that family processes researchers—beginning with Wynne and continuing with Belsky and those family researchers influenced by his work—have called attention to a functional significance of genotypic differences that, ironically, was first delineated by two behavioral geneticists, that is, the function of increasing sensitivity to environmental differences. Distressingly, this idea has been pushed to the margins in the current mammoth and hugely expensive search for vulnerability genes.

Leaving aside this very substantial disarticulation, let us summarize what the adoption design has contributed to the emerging articulation of genetic and family systems studies. First, by verifying findings on genetic influence of twin studies it has foundational status in all the more recent work on the genetics of psychopathology. Second, it has permitted more accurate estimates of the role of parenting and the family environment in the development of offspring. The original Burks study illustrates the delineation of family factors that have a substantial effect on child competence controlling for passive gene–environment correlation. The much later adoption work of McGue and colleagues illustrated the equally important weakening of a popular environmental theory positing parental alcoholism as an environmental risk factor for emerging impulsive disorders in offspring (King et al., 2009). Third, the adoption design provided a distinctive, direct way to demonstrate how the family environment moderates genetic risk for psychopathology. And finally, the adoption design provided a particularly graphic illustration of the role of genes in enhancing sensitivity to the environment; in doing so it offered an enticing and provocative challenge to mainline genetic search for liability genes to which the great majority of genetics research on psychopathology has not responded.

QUANTITATIVE GENETICS AND FAMILY THEORY: NONSHARED ENVIRONMENT, GENETIC INFLUENCES ON ENVIRONMENTAL MEASURES, AND REINTERPRETING ENVIRONMENTAL EFFECTS We have already seen that the assumptions of the twin method were being rigorously explored by 1980, and as early as 1968, the adoption method was buttressing the conclusions of twin studies on the heritability of psychiatry’s most severe disorder, schizophrenia. These two techniques, because they estimate the relative weights of all genetic influences and all environmental influences, are referred to as quantitative genetics, and in the 1980s the results of their use began to challenge, in some cases severely so, theories about families and their influence on child development. Three notable challenges will be reviewed here: the concepts of nonshared and shared environmental influences, genetic influences on measures of the family environment, and the role of genes in explaining the association of variables measuring the family and those measuring the psychological functioning and development of its members.

Quantitative Genetics and Family Theory

The Nonshared and Shared Environment As early as 1932 a curious pattern of findings began to emerge in twin studies. In a remarkably well-done twin study, Harold Carter, an assistant professor in a quantitatively sophisticated department of education at Berkeley, published a twin study of personality. He found, for example, that correlation for neurotic tendencies was .62 for identical and .36 for fraternal twins after correcting for measurement unreliability (very small in his studies) and range restriction. Ordinarily, we obtain a crude estimate of heritability by subtracting these correlations and then doubling the difference (since the difference between MZ and DZ twins covers only half of the possible range of genetic differences between sibs). Here we get the value of 56% suggesting the variance attributable to not-yet-identified or anonymous genotypic differences in this and comparable samples. That leaves 46% attributable to anonymous environmental differences. Since these values are already corrected for measurement error, 1–.62 is a direct estimate of the environment that makes siblings different, the awkwardly termed nonshared environment. In this case the value is 38%. Since the total variance attributable to environmental causes is 42% this means that .38/.42, or 90%, of the environmental effects are those that make siblings within the same family different. Many subsequent studies found similar patterns but John Loehlin and Robert Nichols were the first to emphasize how challenging these findings were to most family theories (Loehlin & Nichols, 1976). The major paper in the evolution of this line of reasoning was a review by Robert Plomin and Denise Daniels in 1987 (Plomin & Daniels, 1987). They emphasized that, as in the early Carter paper, evidence for the singular importance of the nonshared environment has been lying in plain sight in twin studies of personality, cognitive capabilities, and psychopathology. They also summarized the results of several adoption studies where the sample included families with two children adopted from different biological parents. If family attributes common to children in the same family influence child’s temperament, personality or cognitive abilities then genetically unrelated siblings—adopted very early in childhood—should show correlations with each other at well above chance levels and these correlations should increase the longer they are exposed to these conditions. In fact, in these special samples, between-sibling correlations were small, often statistically insignificant and became even smaller the longer adopted siblings were members of the same adopting family. If measurement error is taken into account, differences

345

between MZ twins are arguably the most direct way of assessing nonshared environmental influences. However, by definition, twins are identical in age. Thus, small differences between them—either in non-heritable physical features, temperament or environmental experiences, may be used by them and their parents to differentiate one MZ twin from the other and lead to an uncommonly high influence on their differences in their developmental trajectories. Thus, the MZ twin method may overestimate effects of the nonshared environment. Two studies using ordinary siblings support this idea: sibling correlations on several measures of adjustment tend to decrease as age differences between siblings decrease (Feinberg & Hetherington, 2000; Schachter, Gilutz, Shore, & Adler, 1978). Thus, adoption designs provide an important check here on twin findings as they did in the genetics of schizophrenia. That is, correlations among genetically unrelated siblings provides a direct measure of the shared environment; adding this to estimate of heritability derived from comparisons of MZ and DZ twins provides an alternate approach to estimating the nonshared environment. Years after the influential Plomin and Daniels review, Reiss and colleagues derived estimates of the nonshared environment on height, weight and vocabulary in adolescents using full, half and step siblings and correcting for between-sibling age differences; if anything estimates of the magnitude of nonshared effects were greater than those using just MZ and DZ twins in the same sample providing further support for the estimates by Plomin and Daniels (Reiss, Neiderhiser, Hetherington, & Plomin, 2000). For studies of family systems there were three important sequelae of the landmark Plomin and Daniels paper. First, twin studies of individual differences in both behavioral and medical measures have been consistent, for the most part, in showing that environments that make siblings different in the same family outweigh, at a single time of observation, the importance of between family differences that would render siblings similar (see a recent summary by Plomin, 2011). A second sequel to the Plomin review were efforts to identify specific environmental factors that might differ between siblings and be associated with outcomes for children, adolescents or adults. Researchers have frequently used the MZ twin difference measure although limitations in this design, as noted, surround these results with caution. If a difference in the environmental measure is associated with a difference in a measure of adjustment this suggest a nonshared effect. For example, Pike and her colleagues found—in a cross sectional MZ difference study—that differences in parental negativity directed at the adolescent

346

Genetics and Family Systems

were associated with differences in both depression and antisocial behavior (Pike, Reiss, Hetherington, & Plomin, 1996). More convincing, of course, are longitudinal studies that ether control for measures of adjustment in the first wave or directly employ cross lag analyses. Controlling for child antisocial behavior at age 5, Caspi and colleagues reported an association between maternal expressed emotion at age 5 and child antisocial behavior at age 7 (Caspi et al., 2004). Likewise, using a longitudinal MZ twins design, with 4,508 twins, Viding and colleagues successfully predicted differences in 12-year-old children’s conduct problems but not child callousness from differences in negative parental discipline 5 years previously. The associations were small, accounting for about 1% of the variance. More complex findings using the MZ twin method have been frequent. For example, Guimond and colleagues predicted differences in kindergartners, social reticence from differences in maternal and paternal over protectiveness and hostility at age 30 months (Guimond et al., 2011). The results were complex: main effects were found only for boys, and only girls showed an interaction of parenting differences scores with paternal depression. To add to the complexity, in a multipanel study Charlotte Cecil and her colleagues found reciprocal relationships between harsh parenting and self-control in children followed from age 3 to 12 (Cecil, Barker, Jaffee, & Viding, 2012). Child effects on parenting were at least as prominent as the reverse. Indeed, some studies conclude that nonshared parental environments are a consequence of sibling differences in adjustment (Vitaro et al., 2011). The MZ difference method has been used to study nonshared environments beyond the family of procreation. For example, Lichtenstein studied twins discordant for recent loss of a spouse to show that this later life differential spousal loss was associated with differentia mortality rates (Lichtenstein, Gatz, & Berg, 1998). Vitaro and colleagues showed that differential experience of aggression in friends anticipated differential aggressiveness in the MZ twins (Vitaro et al., 2011). A second approach to exploring the role of specific nonshared experiences, multivariate genetic modeling, begins with an observed association between an environmental variables and a specific outcome (e.g., a contemporaneous association between maternal negative parenting and adolescent depression of .57). Multivariate genetic modeling allows a decomposition of this covariance into a portion that can be attributable to (1) the influence of the children’s genotypic differences both on negative behavior of their parents toward them (a topic to which we will return) and on the child’s depression; (2) the shared environment common to parent and child measures; and

(3) the nonshared environment common to both. Reiss and his colleagues conducted the most comprehensive survey to date, using this method, of the role of specific nonshared factors. They used this technique to study parents, siblings, peers, and other nonfamily influences on both maladjustment and the positive development of adolescents and combined data from observer, child and parent reports (Reiss, Neiderhiser, Hetherington, & Plomin, 2000). Of the 86 contemporaneous associations examined, 29 showed a statistically significant role for the nonshared environment, mostly for maternal and paternal negative parenting and its association with both psychopathology and positive outcomes such sociability, social responsibility and self-worth. In the case of maternal negativity and adolescent depression, 10% of the covariance was attributable to the nonshared environment. Cross-lagged analyses, across three years of adolescence, showed a more diminished role for the nonshared environment almost certainly because this class of influences tends to be unstable across time in children and adolescents (Blonigen, Carlson, Hicks, Krueger, & Iacono, 2008; Ganiban, Saudino, Ulbricht, Neiderhiser, & Reiss, 2008; Larsson, Larsson, & Lichtenstein, 2004; Reiss et al., 2000). Meta-analyses of siblings from longitudinal twin and adoption studies by Tucker-Drob and Briley (Briley & Tucker-Drob; Tucker-Drob & Briley, 2014) suggest that for cognition stable genetic influences may peak or even decline in the transition to adulthood. In contrast, stable nonshared influences continue to rise across development into old age. For personality, stable genetic influences tend to decline as stable nonshared influences rise in the transition to adulthood. In general, as indivduals age stable nonshared influences become at least as important as stable genetic influences. Subsequent to the study of Reiss and his colleagues a number of investigators used multivariate models to evaluate the role of specific nonshared environments, again with a focus on parenting and child development. Most of these have focused on negative parent–child relationships and the development of poor self-control, impulsiveness, and aggression in children (Burt, McGue, Krueger, & Iacono, 2005; Jaffee et al., 2004; Knafo & Plomin, 2006; Larsson, Viding, Rijsdijk, & Plomin, 2008). The sibling specific experience of peer relationships and its relationship has also been tested in this paradigm (Button et al., 2007). In one case with sibling-specific peer experiences playing the preeminent role (Bullock, Deater-Deckard, & Leve, 2006), general genetic factors played the predominant role in most of the reported associations but, to a lesser degree, so did nonshared environment.

Quantitative Genetics and Family Theory

Perhaps the most notable success in exploring the nonshared environment is in adult rather than child development. In 1998, an influential paper by the criminologist John Laub and colleagues reported on a 25-year longitudinal study following boys at high risk for criminal behavior. They found that a broad range of childhood characteristics could not predict continuity of or desistance from adult criminal behavior but that men’s attachment to a spouse had a strong protective effect (Laub, Nagin, & Sampson, 1998). This echoed findings from 14 years before when Michael Rutter and colleagues reported the apparent salutary effects of a favorable relationship with a spouse on adult women who had been institutionalized as children (Rutter & Quinton, 1984). Each study examined only one child in a family and used standard procedures to ask whether the effects observed could be attributed to the effect of the marital relationship or whether they reflected hidden strengths of the high-risk individuals that enabled them to form these bonds. Genetically informed designs permit us to revisit this critical issue of selection versus influence from a novel perspective. For example we may follow the marriages of sibling pairs. If one twin has a happy marriage and the other does not, is there an effect of this difference on differences in their adjustment and is this effect the same for MZ and DZ twins? If both conditions are met findings favor an effect of marriage over an effect a heritable factor in the twins that favors a happy marriage. Erica Spotts and her colleagues reported the first cross sectional studies linking marital satisfaction to both depression and positive well-being in adult twin women an effect of marriage although, lacking longitudinal data, the effect might have been the other way around (Spotts, Neiderhiser, Ganiban, et al., 2004; Spotts, Pederson, et al., 2005). However, a prospective study reported by Burt and colleagues (2010) strengthened inference of a causal effect of marriage. Later in this chapter we will return to the significance of the nonshared environment for the study of adult development. A third sequel to Plomin and Daniel’s (1987) paper is the increasing number of reports on the importance of the shared environment. These reports have softened the blunt conclusions of the Plomin and Daniels review. These newer studies delineate environments shared by siblings that are associated with their psychological development and adaptation. These include reports on support from relatives in adult subjects (Agrawal, Jacobson, Prescott, & Kendler, 2002), types of peer groups chosen by adolescents (Tarantino et al., 2014), early substance abuse (Walden, McGue, lacono, Burt, & Elkins, 2004), adolescent delinquency (Burt, McGue, Krueger, & Iacono, 2007),

347

conduct disorder, attention-deficit/hyperactivity disorder (ADHD), and oppositional disorder in adolescents as well as their co-occurrence (Burt, McGue, Krueger, & Iacono, 2005), cognitive performance in early childhood (Petrill & Deater-Deckard, 2004), and for measures of autonomy and sociability in adolescence (Reiss et al., 2000). An adoption study, published well after the Plomin and Daniels review, took a second look at the comparison of full biological sibling and genetically unrelated adopted adolescent siblings (Buchanan, McGue, Keyes, & Iacono, 2009). Across six measures shared environment accounted for about 20% of the total variance. In a recent publication, Alexandra Burt has clarified why this approach provides a more direct and unbiased estimate of shared environment (Burt, 2014) Results showed notable correlations for adopted adolescent siblings not only for externalizing problems, consistent with more recent twin studies cited above, but also for IQ. This contrasts with the finding from Scarr and Weinberg (Scarr & Weinberg, 1978) of near zero correlations for the same group of siblings, a finding crucial to the argument of Plomin and Daniels. The Scarr and Weinberg sample were approximately 5 years older, suggesting the possibility that the influence of environments siblings share wanes as offspring age. Perhaps the most interesting result to date in this quest for shared environmental effects is the unheralded role of sibling relationships. Insofar as sibling relationships are reciprocal, the relationship constitutes a shared environment for each sibling. If one sibling’s aggression towards a co-sibling is fully reciprocated then both share an aggressive sibling environment. If between-family differences in this shared environment are associated with measures of adjustment for each sibling, correcting for genetic effects, then we can estimate a shared environmental influence on sibling adjustment, especially with the use of longitudinal data. Twin data provide creditable evidence for this effect but so do adoption data. For example, McGue and colleagues—in an adoption design—found that parental alcoholism problems were related to adolescent offspring alcohol use only when comparing biological parents and their offspring but was absent for adoptive parents and their adopted children; in sharp contrast sibling effects held in both groups (McGue & Sharma, 1995). Indeed, between-family shared sibling environments have regularly been associated with the development of aggression and of substance use in each sibling, controlling for genetic effects, in many twin studies (Natsuaki, Ge, Reiss, & Neiderhiser, 2009; Slomkowski, Rende, Novak, Lloyd-Richardson, & Niaura, 2005). A particularly striking set of findings comes

348

Genetics and Family Systems

from the comprehensive study of adolescent twins by Reiss, Plomin, and Hetherington. To meet concurrent standards of family research they combined reports by both parents and the siblings themselves and observed ratings of videotaped sibling interaction in their own home to assess both warm and supportive relationship patterns as well those of conflict, hostility and physical abuse. They reported a total of 18 associations between these sibling measures and measures of adolescent psychiatric symptoms as well as positive attributes such as sociability, social responsibility and engagement in school and related activities. In 13 of these associations the predominant source of covariance was shared environment. This suggested that reciprocal relationships among siblings might constitute a shared environment crucial to the development of both of them. We will return to this finding in the next section as a prime illustration of unanticipated yield for family studies of data from a genetically informed design. Genetic Influences on Measures of the Family Environment In 1981 David Rowe was a young psychology assistant professor at Oberlin College. Although trained in behavioral genetics, Robert Plomin supervised his dissertation, and he became fascinated by the reports of family systems research. In particular, he was intrigued by reports that children’s perceptions of how they were treated by their parents might be as predictive of their development as are observational measures of their parenting. He saw the twin method as a way of understanding this effect. If twin siblings were highly correlated in their perception of parenting, and there was little difference, between MZ and DZ twins then it might be the case that perceptions did reflect some aspect of parenting behavior as it actually occurred. Indeed, this is what Rowe found for most of his measures of perceived parenting save one: the twin perception of rejection by their fathers. Here, MZ twins perceived their father’s rejection or acceptance much more similarly than did DZ twins even when their greater time spent together was accounted for. This suggested that the perception may have arisen more from heritable features of the twins themselves rather than from their parents. Rowe seemed little interested in this apparently aberrant finding (until later in his career) and family researchers even less so. At the time Rowe published his study most family researchers were using direct observations of family interaction as a gold standard for measuring family process. For example, the meticulous work of Mishler and Waxler (1968) and Gerald Patterson (1982) and his colleagues were

major influences on this standard. Thus, a report claiming genetic influence on subjective impressions of children would have carried little weight even if it had been more definitive (and had been read at all by family researchers). Thus a pivotal finding was that of Judy Dunn, the British developmental psychologist, and Robert Plomin; in their study biological parents rearing their own toddlers treated them more similarly than did adoptive parents treat their two children adopted from different birth parents (Dunn & Plomin, 1986). While there are many interpretations possible of this finding, two studies published 10 years later and in the same issue of Developmental Psychology and including meticulous observational coding of parent–child interaction, heavily weighted a genetic explanation. Xiaojia Ge and his colleagues (1996), in a small (45) sample of adoptive parents and their children at midadolescence, reported that the harsh parenting of both adoptive mothers and of fathers could be predicted by severe, hospitalizable externalizing psychopathology of the child’s birth parents and this effect was mediated by antisocial behavior of the adolescents. Alison Pike (1996) and her colleagues, using data from the Nonshared Environment in Adolescent Development study, used a much larger sample of twins, full sibs, half sibs, and unrelated sibs totaling 719 pairs of sibs and their families unselected for psychopathology in parents or children. They reported the same phenomenon: genetic factors that influenced negative parenting also influenced not only antisocial behavior but also depressive symptoms in a group of somewhat younger adolescents. Most importantly, these common genetic factors accounted for most of the correlation between the measures of parenting and of adolescent adjustment. Reiss and his colleagues (2000), using data from the same study in the most comprehensive investigation to date of the role of genes in explaining covariance between family variables and offspring development, studied the association of a range of parenting, marital, sibling variables and both psychopathology symptoms as well cognitive achievement, sociability, autonomy and self worth in a sample of 719 pairs of twins, full sibs, half sibs and unrelated sibs. Meticulous coding of observed interaction in the family’s homes was combined with their self-report of relationships quality. Genetic factors common to parenting and child development and marital conflict about the child and development accounted for a substantial and often most of the covariance between the family and the offspring variable. These data might have been regarded as uncovering one of the most astonishing and unacknowledged confounds in the history of developmental psychopathology.

Quantitative Genetics and Family Theory

Alternatively, it might simply reflect the relatively noncontroversial heritability not only of psychopathological symptoms but also of personal strengths such as self-esteem and autonomy. Thus, genetic influence on the covariance may reflect simply the reaction of parents to a child that has either developed a full-blown psychiatric syndrome or a very well developed prosocial and self-confident personality. Indeed, Richard Q. Bell (1968), in his highly influential paper on child effects on parenting, used genetic data of this kind as part of his argument. Analyses carried out by Neiderhiser and other colleagues in the Reiss group suggested a different scenario: heritable features of less well developed behaviors—perhaps temperaments—evoke parental reactions which, as a consequence of this evocative gene environment correlation, go on to accelerate the transformation of simple temperaments into syndromes of pathology or psychological health. Were that the usual case it would revive the notion of genetic expression outside the skin first formulated by the geneticist Kenneth Kendler (2001). That is, genetic information proceeds through the usual channels (DNA → RNA → protein → neural processes → behavior), but the next step is not simply a prodrome of a syndrome of adaptation or maladaptation. The next step it is a social evocative behavior and the development of an adaptive or maladaptive syndrome does not proceed unless evocation actually occurs (see Reiss & Leve, 2007, for a full exploration of this idea). Follow-up twin studies partially confirmed the plausibility of this novel idea about genetic expression (see, e.g., Burt et al., 2005; Larsson et al., 2008; Tucker-Drob & Harden, 2012). However it has fallen to a currently ongoing prospective adoption design to delve more precisely into the precise sequencing and developmental timing of this outside-the-skin process. Data are drawn from the Early Growth and Development Study that is following, across infancy and childhood, over 500 adoption units consisting of a birth mother, two adoptive parents and a child placed for adoption at birth; birth fathers were also recruited for many of these families. In this study, for example, Harold and colleagues traced the genetic liability transmitted to four year olds by having birth mothers high on ADHD symptoms. By 4 years of age these children showed increases in impulsive behavior that evoked hostility towards the child from their adoptive mothers that appeared to amplify the impulsive behavioral style into ADHD symptoms age 6 (Harold et al., 2013). A similar sequence has been mapped from birth mother temperamental disengagement to social withdrawal by her placed child at 27 months to hostile responses by the

349

adoptive mother to disrupted peer behavior at age 4 (Elam et al., 2014). An intriguing conditional sequences was mapped by Fearon (2014) in this same research group: birth parent externalizing pathology to subjective feelings of maternal distress and incompetence by child age 9 months that persisted at least another 18 months at which time the sequence had led to child conduct problems. However this sequence occurred only in the presence of marital distress between the adoptive parents. To put the matter more graphically, marital satisfaction blocked this avenue of outside the skin expression of genetic influence as securely as methylation of a gene promoter site. As we will see, molecular genetic studies are also beginning to map out similar sequences beginning with a known gene variant and extending through an evoked response from a parent and continuing with the evolution of a pattern of behavioral maladjustment. However, the unparalleled advantage of an adoption design is that these sequences start with behavioral syndromes in the parents of already established adaptive significance. For example, the Fearon report begins the sequence with variation in birth mother’s externalizing problems including clinically diagnostic levels of antisocial personality disorders and substance abuse. We can then ask how are the genetic risks for these highly comorbid conditions first manifested in infancy, controlling for the effects of prenatal exposure to drugs, alcohol and smoke and the effects of birth mother’s psychopathology on self care and obstetrical complications. The answer is important here: by 9 months these genetic risks are first expressed in dysphoric feelings in the adoptive mothers about their functioning as a mother. But these initial effects are already being offset by satisfying marriages. What heritable features in the child evoked this maternal dysphoria? Fearon and colleagues diligently searched among the usual suspects, temperamental features of the child, without success. We will return to this informative failure in a moment. While these informative sequences of evocative gene–environment correlation are just being mapped out many studies have established the substantial effects of genetic factors on a broad range of environmental measures including peer relationships in childhood and adolescence and social support in adults (see Kendler & Baker, 2007, for a review). Many studies report substantial genetic effects on the probability of becoming married, achieving marital satisfaction, and getting divorced. (Jerskey et al., 2010; Jocklin, McGue, & Lykken, 1996; Johnson, McGue, Krueger, & Bouchard, 2004; Lykken, 2002; McGue & Lykken, 1992; Spotts, Lichtenstein, et al.,

350

Genetics and Family Systems

2005; Spotts, Neiderhiser, Towers, et al., 2004; Spotts, Prescott, & Kendler, 2006; Trumbetta, Markowitz, & Gottesman, 2007). These results reflect not only genetic influences of a spouse’s own behavior within a marriage but also evocative effects of one spouse’s heritable qualities on the marital satisfaction of the co-spouse. In toto, many of the heritable characteristics that have evocative effects on marital and parent–child relationships can be attributed in part to heritable features of temperament and personality but across many reports these are accounting for, at best, a modest proportion of these effects (see the previously cited review, Jaffee & Price, 2007). While beyond the scope of this chapter, these data suggest a need to refine and reformulate our understanding of heritable features of children and adults that evoke responses from others. Beyond patterns of self-regulation, cognitive processing, organization of habits and sociability—encompassed in current schemes of temperament and personality—we will certainly want to consider in children genetically influenced individual differences such as differences in their subtler social cognitions, their physical appearance, their smell and their tactile sensitivity. Returning specifically to the parent–child relationship, a recent meta-analysis by Ashlea Klahr and S. Alex Burt (2014) reviewing no fewer than 56 twin and adoption studies confirms substantial genetic effects on the parent–child relationship. Not only do they report robust evocative effects of heritable attributes of the child, but also there are notable genetic effects of the parents’ genotypes on parenting behavior. David Rowe, (1981) the first to report genetic influences on how adolescents perceived their parents, might be astonished to see this extraordinary confirmation of his first, very tentative observation; regrettably his creative and penetrating work in behavioral genetics was cut short by his very untimely death. Taken together, these findings of genetic influences on measures of the family environment constitute the starkest challenge to business as usual in both typical family studies and for most genetic studies as well. They make clear that models of family influence on offspring development are seriously misspecified without genetic information. First, despite Bell’s (1968) challenge to the field over 45 years ago child effects in parent–child relationships are almost certainly underestimated. Indeed, inferences drawn from classical longitudinal studies and their consequent cross-lagged correlations between child factors such as temperament and parental factors such as hostile parenting are seriously weakened by these data. In this example, we can note that child temperament is only partial heritable; individual differences in temperament

also reflect environmental influence including parenting. Child temperament at 3 months, for example, might be a proxy for variation in parenting at 1 month. In contrast, genotypically influenced behavior is tied to the child and evocative genotype environment correlation in this context is much more securely child based; children’s genotypes are what they bring to human relationships and although genetic expression—within the cell nucleus and outside the skin—is sensitive to social influence, genotype is completely resistant to it. Likewise, the effects of a parent’s own genotype on his or her parenting, an inference we draw from twins who are parents, poses another challenge to traditional family research. Parents may treat children harshly in part because of their own genotype. When their genotypes are transmitted to their offspring the genotype of the child may be expressed as antisocial behavior. This passive gene–environment correlation, a phenomenon we have already noted in describing earlier studies of the transmission of schizophrenia, can masquerade as a direct social effect of parenting. Two genetically informed research designs directly address this possibility. The first are adoption designs that correlate a characteristic of a parent, (one that is thought to have a direct social impact on the child) with a measure of adjustment in the children and then compare these two correlation for birth parents and the children they are raising with adoptive parents and their adopted child. This is precisely the design of the first adoption study by Barbara Stoddard Burks. A second design focuses on children of twins, particularly twins who are discordant for the parental characteristic. For example, parental alcoholism is thought to have disruptive effect on parenting and on household routines and through these social mechanisms to increase the liability of impulsive behavior and alcohol use in offspring. If this were the case then the correlations between parents and children should be the same in biological and adoptive families. To extend this example to children of twins, let us suppose we have sample of twins who are mothers and where there is a substantial number of these twin mothers who are discordant for alcoholism. In the case of MZ twins the child’s aunt (the identical twin sister of the child’s mother) is a genetic carbon copy of the child’s mother. Thus, if the parental influence is entirely genetic then the child’s liability should be the same no matter whether it his mother or his aunt who drinks; the liability for the child—in the case of the genetic transmission—should be less if his mother is a DZ twin. If the liability is social transmitted then the liability should only be among children of the drinking twin. As a recent review of children of twin studies (McAdams

Quantitative Genetics and Family Theory

et al., 2014) has shown, the pattern is quite distinct for parental depression and alcohol use. Social mechanisms are involved in the transmission of depression from parents to offspring although the reverse influence, the social effect of child psychopathology on the parent cannot be ruled out. However, the association of parental alcoholism with offspring alcohol use as well as with other forms of impulsive behavior is via genetic mechanisms with little evidence of the role of social processes. This corresponds to findings from adoption studies where parent–child associations of parental depression and offspring depressions (and other problems as well) are found almost equally in biological and adoptive families (Bauer et al., 2007; Marmorstein, Iacono, & McGue, 2012; Tully, Iacono, & McGue, 2008). The refrain must now be familiar: the verification of an important finding in quantitative genetics by two radically different procedures: twin and adoption. As the McAdams et al. (2014) review illustrates, the children of twin design has been used for more direct measures of family process including marital conflict and parenting styles with evidence for environmental effects and passive gene–environment correlation depending on the family measure and assessment of the offspring. A more recent elaboration of the children of twins design—incorporating a second sample of children who are twins in the same analytic model—can now distinguish between three mechanisms of transmission: social, passive gene–environment correlation and evocative gene–environment correlation. Among other results these studies have confirmed the importance of evocative gene–environment correlation in the association of parental negativity and externalizing in adolescent offspring (Marceau et al., 2013; Narusyte et al., 2011). A simple solution for controlling for passive gene– environment correlation—for estimates of parental social influence—is to study parents genetically unrelated to their children: adoptees. In this brief history of the interface of social and genetic research we encounter once again the pivotal role of adoption studies. Although excruciatingly difficult to design and execute, in comparison to twins studies, they are the most direct approach to controlling for the effects of passive gene–environment correlation. We have reviewed an impressive body of data converging on two major ideas. In the shorter term the association between parenting and child development or between the parents’ marital satisfaction and child development may be attributable to genetic factors common to both. In the longer term, the transmission of parental characteristics to children—including those that directly shape the child’s social environment—may influence child

351

development solely because the genes that influence the parentally-shaped environment also, when transmitted to the child, shape the child’s behavior. Need every study of family life now require a twin or adoption design? We think not, but we will return to this topic briefly in the coda to this chapter. Before concluding this section let us return to siblings for a final time to consider a paradox. By adolescence most genetically informed studies suggest that a great deal, if not most, of the covariance between many measures of parenting and measures of adolescent adjustment can be accounted can be accounted for by genes common to both. The reverse, several studies suggest, is true for sibling relationships. The covariance of measures of sibling adjustment with carefully measured qualities of the environment they provide for each other is attributable, in most cases, to shared environment. To put the matter another way for many family variables studied thus far, parents are important to adolescent development substantially because they amplify, as a consequence of evocative gene–environment correlation, heritable characteristics in their offspring. However, the environments created by siblings for themselves—through sustained patterns of interactional reciprocity—appear to have a main effect on adolescent adjustment independent of their genotype. Are genetic data suggesting that siblings are the tyrants of their families—for better or for worse? Not only are adolescents, by virtue of their genotype shaping their parents’ behavior toward them, but also they are providing an influential world for themselves. While many parents of adolescents might agree with such a view a single report by Danielle Bussell and her colleagues on the Nonshared Environment in Adolescent Development project suggest a different scenario. She proposed that multivariate genetic modeling can be applied to covariance between measures of two family subsystems. Thus, the covariance between parent– child subsystems and the sibling subsystem can be decomposed into that attributable to common genes (the same genetic factors in the child that evoke parent–child hostility also evoke sibling hostility), to shared environment (between family differences in parenting style are associated with between family differences in sibling relationships independent of child genotype) or nonshared (the parenting hostility uniquely directed at one sibling is associated with that siblings hostility toward the co-sib independent of genotype of the child and of between-family differences in parenting). In examining both positive and negative interaction of mothers with adolescents and the sibs with each other Bussell found that it was the shared environmental components of the parenting–sibling relationship

352

Genetics and Family Systems

association that was, by far, the most important (Bussell et al., 1999). Given the remarkable stability of the quality of parent–child relationships across adolescence (Reiss et al., 2000), it is reasonable to suppose that the quality of mother–child relationship measured at child age 14 reflects not only current status of that relationship but also its recent past. If that assumption is correct it might be said that the sibling relationship bears a memory of between family differences in parenting style and becomes an agency of that difference in emerging adolescence. We linger over these data because this commemorative function of sibling relationships has been little studied in classical family research but is an arresting example of possible new vistas on the family opened up by genetically informed design—new unless you count the Book of Genesis and its clear portrayals of the interleaving of parent–child and sibling relationships in the case of Rebecca, Jacob and Esau or, as noted, repeated in a second generation of the same family between Jacob, Joseph and his brothers. MOLECULAR GENETICS AND THE FAMILY How molecular genetics entered the field of family research. Thus far we have reviewed studies where the knowledge of genetics is inferred from correlations between individuals of known genetic relatedness. However, none of those studies identifies which of the 30,000 genes in the genome, or which noncoding regions, play a role in these effects. The lessons to be applied from this line of quantitative genetics to research on the human family are clear, challenging, and—for the most part—unheeded. The tools of quantitative genetics have been developed over decades; its inventors were psychologists, physicians and educators who wanted answers to practical problems. The assumptions behind the methods were vigorously investigated; different methods were used to arrive at comparable estimates and much of the work was conducted well under the radar of the broad reaches of biomedical science or public acclaim. Just the reverse has been true of molecular genetics. Francis Crick, a physicist by training, and James Watson a biologist with an early interest in molecular genetics, collaborated to identify the structure of DNA in a highly competitive and conspicuous race with Linus Pauling among others seeking the same goal. Their one page publication in Nature (Watson, 1952) attracted worldwide attention as the discovery of the secret of life. Nine years later they became Nobel Laureates. Skipping quickly over the following decades of this history we come, 31 years later, to a raucous meeting of the biotechnology firm Cetus. There, Kary Mullis presented a poster describing

his work, in collaboration with other Cetus scientists, on the double-probe technique that was to be the foundation of the polymerase chain reaction (PCR). Mullis, a tempestuous figure throughout his life, had to be separated from a physical brawl at that meeting. In contrast to the unpredictable temper of its prime developer, PCR emerged as a fully reliable tool for genotyping and spread rapidly throughout the biological sciences. Like his molecular genetic forebears a generation earlier, Mullis was awarded the Nobel exactly nine years after his first report of his findings. His trip to Stockholm also included a brush with the Swedish police (Rabinow, 1996) and left a continuing controversy over whether he deserved the award. PCR is fundamental to most of the molecular genetic studies published in the last decade relevant to understanding families. Subsequent phases of the story of molecular genetics hardly needs retelling but once again a monumental and widely publicized competition between the Celera Corporation and the US Federal Government led to the almost simultaneous announcement of the complete sequencing of the human genome, an accomplishment that has rapidly led to huge investments in searching the genome for both common and rare variants of genes that might be related to human disease including psychiatric disorders. These genome wide association studies, requiring mammoth samples, have been applied to social variables broadly considered (see a review and critique, Chabris et al., 2013) and even to a directly measured family variables (Butcher & Plomin, 2008). When a version of this chapter is written a decade hence, this tool may have yielded insight for the study of the family, but not yet. Not long after it was developed, PCR was applied to the understanding of behavior. A working premise was that a single gene variant might lead to a single, well-specified disorder. Plomin, Owen, and McGuffin (1994) reviewed this work to highlight its utility in a restricted number of syndromes, mainly those of severe intellectual deficiency. They drew attention to an alternate conception of the relationship between genes and disorders, quantitative trait loci, where many genes operating in concert were necessary to influence the emergence of complex behavioral syndromes. They proposed that the severity or distinctiveness of a genetically influenced syndrome might be related to the number of these gene variants possessed by an individual, that some of the genes may not be necessary for development of a syndrome and that somewhat different gene sets may lead to the same disorder. Several developments led to a major revision of this view. A new idea was that genetic variation might not be a necessary or contributory cause of complex human behavioral syndromes but rather that it rendered individuals

Molecular Genetics and the Family

more or less sensitive to environmental variation. Quantitative genetics made the first contributions to this idea. As we have seen, this idea was formalized by Kendler and Eaves in 1986 and was demonstrated clearly in the work of Wahlberg (1997) and his colleagues on schizophrenia. Kendler and colleagues, using a population-based twin sample, showed that identical non-depressed twins of depressed women were more sensitive to stress, and subsequent depression, than were fraternal twin siblings of depressed co-twins (Kendler, Kessler, Walters, MacLean, et al., 1995). Long before, Cadoret observed that genetic liability for antisocial behavior rendered children more sensitive to adverse rearing conditions in their adoptive families (Cadoret, 1982; Cadoret & Cain, 1981; Cadoret, Cain, & Crowe, 1983). At the other end of the spectrum, animal research—often in mice—was able to deploy powerful knockout and transgenic methods to cause or identify specific genetic deficit and reverse these with highly targeted pharmaceuticals (e.g., Cases et al., 1995); these more highly controlled models suggested clearer links between single genes and more complex behavior such as aggression. Finally, variations in single genes seemed to have reproducible effects on the structure and function of the human brain as revealed by constantly improving brain-imaging techniques (Hariri et al., 2002). These lines of evidence prompted Stephen Suomi (working with rhesus macaques) and Avshalom Caspi and his colleagues (working with a large, longitudinal human study) to select particular genes for scrutiny. Suomi was the first to report on the association of a single gene variant with heightened sensitivity of infant macaques to variations in their early rearing environment. They focused on the statistical interaction of two variants of the rhesus version of the serotonin transporter gene with adverse early rearing environments, mother rearing or nursery rearing. Macaques with the inefficient gene variant showed the clearest effects of early rearing experience on early-appearing adaptive behavior, particularly activity patterns (Champoux et al., 2002). A the same time Caspi and colleagues reported children with the less efficient version of an x-linked gene, MAOA (similar to the gene well explored in animal studies), and who were exposed to maltreatment, developed significant aggressive behaviors (Caspi et al., 2002). A year later the Caspi group reported similar findings for the serotonin transporter gene; in this second report, young adults with the inefficient form of the serotonin transporter gene showed the effects of both childhood maltreatment and recent stress on the evolution of depression (Caspi et al., 2003). These findings attracted immediate and broad attention across all of biological science and in the public press as had other reports in the

353

brief history of molecular genetics. They also attracted the attention of many behavioral scientists because these findings were drawn from the meticulous longitudinal Dunedin study, named for its place of origin on the South Island of New Zealand. The study had already yielded major findings in a group of almost 1,000 children followed from early childhood, and as this chapter is written, most of these children have been followed well into adulthood. All three of these highly influential papers had three major features in common: the main effects of environmental adversity were very substantial; the main effects of the gene variants were tiny or non-existent; and the effects of the interaction were between the two. Strongly influenced by this paradigm, most subsequent work has examined single gene variants, analyzed singly or in very small numbers combined into polygenic risk scores, as indices of individual differences in response to environmental variation. That is, almost all subsequent work has focused on genetically influenced receptive capacities of individuals in families. In contrast the robust evidence for quantitative genetics of very substantial genetic effects on effector capacities, the capacities of individuals to select and influence their own environments, has received much less attention. Moreover, taken together, these papers had a persistent influence on the study of parent child relationships—to focus on individual differences in the sensitivity the offspring reactions to parents rather than on differences in the parents’ sensitivity to children. The Caspi 2003 report focused on stress in adult life although it contained a generous section on child maltreatment that reinforced an underlying theme in the Suomi 2002 and Caspi 2002 papers adding weight, almost certainly inadvertently, a picture of individuals as responders to rather that creators of their own environments. To be clear, molecular gene–environment interaction studies have probed many aspects of healthy and pathological adult development but have rarely focused on adults as parents. Thus the great majority of subsequent reports relevant to this chapter on the family have, collectively, focused on genetic effects on individual differences among children that accounted for their varying response to their rearing conditions in contrast to a parallel focus on genetic differences among parents that account for differences in their response to their children although we will review some studies in this latter category. We will also review the very small number of molecular genetic studies that explored effector capacities in children. Cautions in Interpreting Molecular Data Despite these limits on the new paradigm, clearer in retrospect, the interest in these early reports was well founded.

354

Genetics and Family Systems

All three were built on prior findings in quantitative genetics but also, and of immediate interest, on well controlled laboratory studies of animal and human neurobiology. Thus, there was a promise of bringing studies of the family into a unifying field of developmental human biology. Indeed, in retrospect, it is now clear that the introduction of molecular genetics into the study of the family has had entirely different dynamic than its older cousin, quantitative genetics. The former emerged slowly across half a century and with great resistance of researchers on social processes; many of its major lessons remain unlearned in family research. Since the Barr et al., 2003 papers, genotyping samples of families has been so rapidly included in the repertoire of family research that those from other disciplines—including geneticists—have called for caution (among many examples see Plomin, 2013). Some of these cautions seem well founded and we begin our review of selected, illustrative reports of this work with three of them in mind. First, several reviews—using standard approaches to meta-analysis or the analysis of publication bias—have raised concerns about the reproducibility of some of these initial findings, particularly the moderation by the serotonin gene variants on the relationship of stress to depression in adults (Duncan & Keller, 2011; Karg, Burmeister, Shedden, & Sen, 2011; Munafo, Durrant, Lewis, & Flint, 2009; Risch et al., 2009). A particularly striking failure to replicate comes from the improbable circumstance that a second, large 30-year longitudinal study of children—the Christchurch Health and Development Study—is also ongoing on the South Island of New Zealand (Fergusson, Horwood, Miller, & Kennedy, 2011) whose population of just over a million seems to include an inordinate number of developmental researchers. However, using broader criteria for reproducibility several reviews find evidence for replicability of the Caspi report on the serotonin transporter gene (Karg et al., 2011; Mak, Kong, Mak, Sharma, & Ho, 2012; van IJzendoorn, Belsky, & Bakermans-Kranenburg, 2012). Caspi’s 2002 paper on the interaction of the less efficient MAOA gene variant and child maltreatment on the development of aggression in boys has been confirmed by a very recent meta-analysis of 20 studies; importantly, this analysis found no difference in findings between smaller or larger samples or between earlier or later publications addressing, at least, some concerns about publication bias (Byrd & Manuck, 2014). Equally important, successful replication was limited to child maltreatment and not other childhood adversities. Included among the confirmatory finding is indeed one

from the other South Island longitudinal study (Fergusson, Boden, Horwood, Miller, & Kennedy, 2011). Kary Mullis’s PCR gave social scientists a tool they never had before: a reliable, inexpensive ascertainment of differences among individuals, differences that preceded all else in their prenatal and postnatal social and physical world. Moreover, PCR can be used years after a study had concluded (if only a cheek swab or saliva sample can be obtained). This unique ascertainment has enabled an astonishing number of attempted reproductions of the original studies in little more than a decade. However, energized by these recent reports, investigators have examined many other gene variants—some of that work is reviewed here—with a broad variety of environmental variation and developmental outcomes. Many of these reports come from multivariate, multi-panel studies that genotyped subjects for many gene variants thus allowing investigators a wide latitude in associations they choose to study and report. Often these reports contain data only about one of these variants leaving skeptical readers to wonder whether null findings with other variants have gone unreported. Thus, particularly in this case, the empty file drawer procedure—usually regarded as a safeguard in meta-analysis against the findings of unpublished papers— cannot offer protection against unpublished analyses (Duncan & Keller, 2011). Caspi and colleagues have presented a strong case for an alternate strategy to meta-analysis for reviewing evidence in this field: reviewing studies in animals and humans that are theoretically linked and may help explain, the moderation of environmental effects by specific gene variants (Caspi, Hariri, Holmes, Uher, & Moffitt, 2010). The fundamental idea here is that the foundational papers were exploring a theory of inherited sensitivity to stress rather than a specific hypothesis about the pathogenesis of depression. Thus, for example, a comprehensive review should include—as theirs did—experimental studies of the moderating effect of the gene variant on cortisol responses to standard stress stimuli in a laboratory setting. We cautiously explore this strategy in our current selective review. Our interest is to explore both possible as well as established use of molecular genetic tools to better understand family process. Finally, we note that questions about the reproducibility of the foundational papers in this field—or of those that follow—is not unique to molecular genetic studies of the family. Nor have many, if any, studies in the recent history of behavioral sciences attracted such persistent efforts at reproducing the original findings. More generally, reproducibility of findings in all branches of science remains

Molecular Genetics and the Family

a major concern. A survey by Begley and Ellis (2012) noted serious problems of replication problems in cancer treatment research and a special issue of Psychological Science raised concerns about replicability in psychology considered broadly (e.g., Pashler & Wagenmakers, 2012). As an ironic twist, the Begley and Ellis survey is—because of its lack of detail and precision—very likely itself not replicable. A second problem in evaluating this rapidly emerging field is the high variability with which investigators explore genetic confounds. For example, population stratification is a threat to all gene association studies—for main effects and for interactions. Race is a crude marker of population stratification and there are examples of gene–environment interaction findings that do not generalize across race that serve as a reminder of the potential importance of this source of confounding (Widom & Brzustowicz, 2006). Within a study sample, unidentified subgroups may have different genetic ancestries. As a consequence, they may differ both in the frequency of the gene variants being investigated and in the behavioral measure of interest thus producing a spurious gene–behavior correlation. Likewise, where environmental variation is attributable to a blood relative (such as variation in parenting) passive gene–environment correlation may masquerade as a gene–environment interaction. It is a rare and commendable paper that examines these potential confounds (see, e.g., Sulik et al., 2012) for a test of population stratification and Kim-Cohen and colleagues for ascertainment of passive gene–environment correlation (Kim-Cohen et al., 2006). Further, investigators rarely specify in advance whether they are expecting the genetic effect to be additive (where the effect of the interaction increases from individuals who have neither deficient gene variant to those who have one to those who have two) or nonadditive, and if the latter, whether the effect should be dominant or recessive. A particularly intractable problem is the possibility that two or more genes are at work completely mimicking gene–environment interaction: the measured gene, thought to be moderating an environmental effect, may be interacting with unmeasured genes that have notable influence on the same environmental variance. In their 2003 paper on the serotonin transporter gene, Caspi and colleagues addressed this problem by varying the timing of variables measured but only animals studies that experimentally vary rearing conditions, as in the Suomi studies, or human studies that experimentally vary the environment through an experiment (e.g. the laboratory induction of stress by Way, Taylor, and their colleagues (Way, & Taylor, 2010;

355

Way & Taylor, 2011) or a randomized trial (e.g., Brody, Beach, Philibert, Chen, & Murry, 2009) fully addressed this issue. Indeed, virtually none of the papers we will review even address this concern, let alone test for it. Reviews of the literature in this area, either meta-analyses or conceptual, that we have already cited fail to examine papers for their adherence to this genetic toilette. A final problem with evaluating this literature is its heavy dependence on statistical interaction although, unlike the three foundational papers, genetic main effects have been reported and we will survey some of these. Some prospects (Reiss, Leve, & Neiderhiser, 2013) and problems (e.g., Eaves, 2006) of interpreting statistical interactions have been reviewed elsewhere. The first set of problems is inherent in the statistics of interactions and, because of their wide coverage elsewhere, need only be mentioned here; these include statistical power, scaling of variables and testing for regions of significance. Of particular interest here is the problem of controlling for confounds in interaction testing. For example, Keller has suggested that most gene–environment interaction testing has failed to test for confound environment interactions or confounder–gene interactions and hence ruling out confounds has been insecure at best (Keller, 2014). A second problem is deriving mechanistic interpretation from patterns of interaction. For example, Widaman and colleagues presented an ingenious approach to both identifying crossing points in statistical interactions and attaching confidence intervals to these points. This rigorous approach can distinguish ordinal from disordinal interactions, with well-specified assumptions and limits. Widaman and colleagues argued that disordinal interactions, rigorously analyzed, favor a mechanistic theory termed differential susceptibility, a theory noted above in relationship to gene–environment interactions in schizophrenia. Ordinal interactions, according to this view, favor the original Caspi idea of inherited sensitivity. However, as Reiss and colleagues note (Reiss et al., 2013), disordinal interaction can also indicate a quite different mechanism of goodness of fit: some individuals benefit for environment A but do much worse in environment B whereas, in other individuals, the reverse is true. The Early Growth and Development Study, using an adoption design, nicely illustrated this phenomenon: toddlers at risk for externalizing disorders (birth mothers had externalizing problems) develop favorably when parents structure assigned tasks in contrast to children at low risk for these disorders who do well when task are unstructured but respond poorly to high levels of parental structuring

356

Genetics and Family Systems

(Leve et al., 2009). As for ordinal interactions, they may be due to effective social compensation for deficits associated with genotypic differences as well as to those genotypic effect on positive adjustment that are only manifest under favorable social environments (e.g. the heritability of intelligence is most apparent in favorable economic circumstances (Bates, Lewis, & Weiss, 2013; Turkheimer, Haley, Waldron, D’Onofrio, & Gottesman, 2003). There is a growing consensus that experimental studies are crucial to advancing causal reasoning beyond statistical interactions however carefully the latter are accomplished. Thus, in viewing over a decade of work following from the initial Suomi and Caspi papers we exercise both enthusiasm about the enlarging prospects for placing the study of the human family with a broader science of developmental biology but also remain cautious until a literature accumulates where genotyping is planned from the start of the adequately powered study, where well-conceived hypotheses provide substantial prior probabilities, where statistical and genetic artifacts are reasonably ruled out and detailed mechanistic explanations are sought in planned comparisons across studies, methods and species and in valid, experimental models. An exhaustive review of studies published thus far is beyond the scope of this chapter; as noted, we have selected studies more to illustrate the potential of molecular genetics techniques for studying the family, particularly in better specifying theories of individual differences in both responding to and evoking family process.

SPECIFYING THEORIES OF FAMILY PROCESS AND INDIVIDUAL DIFFERENCES A core objective for systematic observation of families, certainly since the time of the redaction of the Book of Genesis, has been a better understanding of six processes within the family: variations in how children respond to very adverse parenting as well as its more ordinary variation; variations in how children evoke parental responses; factors inherent in parents that shape differences among them in sensitivity and style; variations inherent in adults that shape their marital relationships; difference among families in their response to stress; how family experience in one generation shapes the family lives of offspring when, as adults they constitute their own families years; and differences among families in their response to family-level therapeutic and preventive interventions to improve mental health. As noted, the great majority of molecular genetic studies address the first of these six questions. Some of the

remaining questions have been addressed by no more than one or two studies. Differences Among Children in Their Response to Parenting Molecular genetic techniques have been applied to understanding individual variation in children in their response to parenting and, in far fewer instances, individual differences in parents in response to children. As noted, Caspi et al. (2002) reported that boys with the less variant of the MAOA, x-linked gene explain variations in response to maltreatment, much of it by their parents, as expressed in their developing their own antisocial behavior. This is the most securely replicated finding in the molecular genetic investigation of family social process. Cicchetti and colleagues (Cicchetti, Rogosch, & Thibodeau, 2012) explored other moderating genes regulating serotonergic brain systems that might account for individual variation in children’s response; they replicated the MAOA finding and found that the inefficient variant of serotonin transporter as well as the inefficient variant of tryptophan hydroxylase gene had a moderating relationship as well. In general, these findings suggested that children with less efficient genotypes inherited a sensitivity to maltreatment. This group also explored genes regulating other systems and with other outcomes. For example, they reported that a gene variant regulating corticotrophin releasing hormone receptors in the HPA system moderated the impact of maltreatment on diurnal cortisol variation with an interaction suggesting inherited sensitivity to stress (Cicchetti, Rogosch, & Oshri, 2011) and a gene variant regulating the oxytocin system moderated the association of maltreatment with diminished perceived social support from friends and family (Hostinar, Cicchetti, & Rogosch, 2014). The GG variant of this gene was associated with children’s increased sensitivity to the stress of maltreatment although the literature on this variant is unclear whether it is expected to have protective or sensitizing properties. A recent paper from this group is of special interest; a gene variant regulating receptors of acetylcholine moderated the effects of maltreatment on personality, particularly neuroticism. Unlike, other studies of maltreatment these results suggested the gene variant enhanced neuroticism in maltreated children but diminished it in nonmaltreated controls (Grazioplene, Deyoung, Rogosch, & Cicchetti, 2013), an example that arguably reflects differential susceptibility Molecular genetic techniques have also been applied to study variation among children in their response to less extreme variation of parenting than the stark distinction

Specifying Theories of Family Process and Individual Differences

between maltreatment and no maltreatment. A few examples, from a rapidly growing literature, illustrate the possibilities here. Starting during fetal development inefficient gene variants of the serotonin transporter are reported to influence individual differences in fetal response following exposure to maternal anxiety in utero as manifested in postpartum infant negative emotionality (Pluess et al., 2011) and the effects of maternal postnatal anxiety on infant irritability (Ivorra et al., 2010). The former interaction pattern suggested differential sensitivity to stress but the latter, given a main effect of the genotypic difference, suggested that nonanxious mothers could offset the apparent negative effect of the inefficient variant. Pluess and colleagues (2011) reported a similar study with a pattern suggesting inherited sensitivity to stress. Sulik and colleagues (2012) reported on variation in parenting quality and aggression and conduct problems in toddlers; the interaction in this report revolved around the efficient form of the serotonin transporter as marked by variation in two locations in the same gene. Surprisingly, children with the efficient form of the gene benefited from favorable parenting but suffered from unfavorable parenting; a pattern suggesting differential susceptibility. Kochanska and colleagues have followed approximately 100 children from toddlerhood to nearly school age; when the children were about 4 the deficient form of the serotonin transporter gene was sensitive to differences in security of attachment between the mother and the child. On a test battery assessing facets of self control (Kochanska, Philibert, & Barry, 2009) secure attachment offset the disadvantage associated with the deficient genotype. A similar pattern emerged a year later in the children’s development. This time a responsive maternal style offset deficits in school and social competence (Kochanska, Kim, Barry, & Philibert, 2011). However, for measures of children’s moral competence, children with less efficient variant showed performance that was better than those without if they had highly responsive mothers and worse if they did not, an example of differential susceptibility. Differences Among Children in Evoking Parental Responses As quantitative genetics has shown, genotypic variation in children not only influences variation in their receptiveness or sensitivity to parenting but also plays a major role in evoking parental response. As noted in the discussion of quantitative genetics, genetic studies add a significant dimension to the study of child effects. For example, family research has accumulated evidence that irritable children who display negative emotions frequently evoke equally

357

negative response from parents (e.g., for infants see van den Boom & Hoeksma, 1994). Molecular genetics has taken an early, hesitant start in identifying specific genotypic difference that might account for these evocative effects. Mills-Koonce and his colleagues (2007) appear to be the first to explore this area examining the role of an infrequent variant of a gene regulating a dopamine receptor because of its association with child temperament. Following from reports of van den Boom and Hoeksma (1994) and others, we might expect such genetically influenced temperament to be disruptive of mother–infant relationship, a finding confirmed by Mills-Koonce et al. Unfortunately, they tested the effect of this variant after ruling out the effects of child mood and temperament rather than examining their mediating role in the effect of this genotype. However, working among a research group with a primary focus on the family as a social system, Mills-Koonce et al. added an innovation: genotyping the mothers as well as children raising the possibility of studying an interpersonal gene by gene interaction. In this analysis such an interaction was not found, but the less frequent gene variant in the child was associated with less sensitive parenting, an important documentation that the genotype of one member of the family might affect the behavior of another. Elizabeth Hayden and her colleagues (2010) reproduced this finding in a larger sample of older children. In this study, the child’s negative affect was explored as a mediator of the genetic effect; it did made a marginal contribution to mediating the association of gene variant and evoked parenting. The dopamine systems in the brain, presumably those most involved in emotion regulation, were explored using another gene variant—this one regulating dopamine transport from intercellular space to intra-neuronal locations. Again, because of its association with potentially troubling child temperament, Hayden and colleagues (2013) explored its role in evoking negative parenting; they reported its apparent influence on both parental hostility and lack of guidance; in these findings child negative affect played a more distinctive mediating role. Copeland and colleagues (2011), working in a group focusing on psychiatric epidemiology, studied a gene variant regulating the endogenous opioid system in positive parenting; in animal studies a comparable gene variant is important in influencing attachment behavior. Copeland reported that children with the variant that enhances opioid receptor efficiency elicited more positive parenting for those parents that had a history of psychiatric problems; these findings resonate with an idea about genotypic variation in children that might actively heal potentially problematic parent–child relationships.

358

Genetics and Family Systems

Differences Inherent Among Parents That Shape Their Response to Children Molecular genetics techniques have also been used to explore the role of genetic variation in parents. Much of this work has been stimulated by an extended series of animal studies on the role of the neuropeptides vasopressin and oxytocin (see Donaldson & Young, 2008, for a review). These nearly identical peptides are produced in the hypothalamus, transferred to the pituitary, and then released into the peripheral circulation where they have a number of regulatory roles. They are also transferred to brain regions where their effects on social behavior in animals have been intensely studied. Perhaps the most influential animal models have been contrasts among monogamous voles, as compared to promiscuous voles, where vasopressin plays a distinctive role not only in male pair bonding but also in male parental behavior, whereas oxytocin plays an approximately comparable role in females. Working in the context of family process, attachment, and early intervention research group, BakermansKranenburg and van IJzendoorn (2008) were the first to report the association of one of several variants of the gene controlling the oxytocin receptor and observed sensitivity of parenting. Four years later Feldman and her group reported a comparable finding but with a different variant of the same gene (Feldman et al., 2012). In another study Mileva-Seitz and colleagues (2013) reported the association of a variant in a gene regulating oxytocin synthesis to maternal vocalizations towards her infant but not in maternal sensitivity. Genes regulating the oxytocin system also are associated with an apparent buffering effect. Using the same gene variant as in the original Bakermans-Kranenburg and van IJzendoorn study (2008), Sturge-Apple and her colleagues found that it rendered the mother more sensitive to conflict with her husband (Sturge-Apple, Cicchetti, Davies, & Suor, 2012). Those with the sensitive variant were less sensitive and less warm with their toddlers when there was high marital conflict but more so when there was low conflict. Notably in this study, as in the Mileva-Seitz 2011 study, there were no main effects on maternal parenting of this gene variant. As another example, Apter-Levy and colleagues, in the Feldman research group, reported a buffering effect for the same variant on the oxytocin receptor gene that, in their previous study, was associated with sensitive parenting (Apter-Levy, Feldman, Vakart, Ebstein, & Feldman, 2013). In the new study, they focused on the effects of sustained mothers’ depression on the development of psychopathology of

their school-aged offspring. The variant associated with maternal sensitive parenting in the first study here was associated with a protective effect on offspring development. Genes regulating other neural systems have also been explored both for their direct effects and buffering effects on maternal parenting. These include genes regulating dopamine (Fortuna et al., 2011; Kaitz et al., 2010; Lee et al., 2010), serotonin (Bakermans-Kranenburg & van IJzendoorn, 2008; Mileva-Seitz et al., 2011), and vasopressin (Bisceglia et al., 2012). Individual Differences and Marital Processes As noted, quantitative genetics has illuminated both environmental and genetic process related to marriage; for the latter there appear to be notable genetic influences on whether individuals become married and then divorced or stay single. Moreover, genetic influences have been reported for satisfaction in marital relationships with evidence for evocative effects; the genotype of one spouse is associated with the marital satisfaction of the other spouse. There has, however, been scant attention to the molecular bases of these findings. In the very first attempts investigators have used two strategies. The first approach was strongly influenced by findings in the prairie voles on the role of oxytocin and vasopressin in pair bonding. Two reports came from the same Swedish twin study. For men, the presumably more efficient variant of the gene regulating vasopressin receptors in the husbands were associated with their wives’ marital satisfaction. For women, variations in their oxytocin receptor influenced, but to a lesser extent, their husbands’ marital satisfaction (Walum et al., 2008, 2012). As expected, these were main effects though small in magnitude. A second strategy follows from many others we have cited: genotypic variation was explored for its association to individual’s sensitivity to the marital context. In two studies, the less efficient allele of the serotonin transporter gene was associated with enhanced sensitivity of the spouse to the affect of the other marital partner (Haase et al., 2013; Schoebi, Way, Karney, & Bradbury, 2012). Finally, the mother’s marital history has been studied as a varying environment in the development of ADHD in boys across a broad age range. The same gene regulating the D2 dopamine receptor, that played a role in other studies cited here had opposite effects depending on whether mothers had an adverse marital history (divorces, remarriages etc.) or not (Waldman, 2007). It is possible that this maternal history was a proxy for serious personality disturbances and the negative parenting that often accompanies them.

Specifying Theories of Family Process and Individual Differences

Individual Differences in Family Functioning in Response to External Stress We can cite but only example here on differences among mothers in their vulnerability to postpartum depression in response to either economic adversity or advantage. Mitchell and colleagues report that two less efficient variants of the serotonin transporter gene are associated with heightened risk for postpartum depression under adverse economic circumstance—indexed by mother’s educational attainment—but reduced risk under favorable circumstances (Mitchell et al., 2011). Differences Among Adults in the Role of Childhood Experiences in Shaping Their Adult Family Roles At least three studies have examined intergenerational effects in families: do individual offspring differ in the degree to which harsh or favorable parenting influences their behavior, years later, with a spouse or romantic partner? Initial evidence suggests that several gene variants enhance the transmission from positive parenting to subsequent positive partner relationships and also favor the transmission of negative parenting to unfavorable relationship with subsequent partners (Beaver & Belsky, 2012; Masarik et al., 2014; Simons et al., 2013). Differences Among Families in Response to Family-Level Interventions Finally, molecular genetic techniques have been applied to understand individual differences in response to family-oriented behavioral interventions. In doing so, these studies follow in the tradition of pharmacogenetic studies that investigate how genetic variation may forecast response to specific pharmacologic agents. In applying these techniques to behavioral intervention it is important to distinguish two questions. First, does a specified genetic variation forecast an individual’s responsiveness to a therapeutic agent or, second, does it offset a current or future pathological condition associated with the genetic variation? In both cases, individuals with the predictive genotype will show a greater response to an intervention but for distinct different reasons (see Reiss et al., 2013 for a more complete explanation). This distinction in mechanism has great practical significance. If a genotypic responsiveness mechanism is established by repeated, adequately powered randomized trials then genotyping might become a rational basis for recommending particular treatments only to those individuals who have a responsive genotype. In contrast, where a data on a behavioral

359

treatment clearly demonstrates a compensatory mechanism then that treatment could be recommended to individuals even though their syndrome has a clear association with (and is probably influenced by) their distinctive genotypic profile. Investigators experienced in family intervention techniques have reported small-sample studies of brief parent training interventions for externalizing behavior (Bakermans-Kranenburg, van IJzendoorn, Pijlman, Mesman, & Juffer, 2008) and ADHD (van den Hoofdakker et al., 2012). The former showed evidence that a variant of a dopamine receptor gene forecast greater responsiveness of children to the intervention; in the latter graphical presentation of the data (not analyzed statistically) suggested treatment had a compensatory effect for a more severe form of ADHD associated with a dopamine transport gene variant. In both cases genotyping was completed only on a subset of subjects after the trial was completed with a consequent loss of the protection from confounding ordinarily afforded by a randomized trial; thus, both reports serve as informative proof of concept’ studies. This reservation does not to apply to data from two large trials of a family-oriented program for the prevention of alcoholism in African American teenagers reported by Gene Brody and his group. The great majority of participants were genotyped and gene variants associated with several neural systems were evaluated for their capacity to predict alcohol use in teens that were not already affected. These sizable genotypic predictive effects appeared only in the control groups but not in the treatment groups providing one of the best sets of evidence we now have that behavioral interventions can offset the behavioral risks associated with genotypic differences (Brody, Chen, & Beach, 2013). Yet to be resolved in these studies are the surprisingly large main effects of single genes given very much smaller single gene effects noted in GWAS studies. Perhaps a combination of poverty and racism that face these families operate as moderator, not measurable in this study, enhancing single gene effects. Or it could be a winner’s curse. Beyond Individual Differences: The Prospects for a Developmental Biology of Families Within an astonishingly brief period of time, the PCR technique first reported by the flamboyant and controversial Nobelist Kary Mullins has spread rapidly among researchers on family process. The foundational papers by Suomi and colleagues and Caspi and colleagues are rightly credited with making a link between a molecular technique

360

Genetics and Family Systems

and the social process of the family. The reports we have reviewed here constitute only an illustrative sample of this notable spread. In introducing this work we have already outlined reasons why it should be read with caution. Nonetheless, the exuberance of this work is based on much more than the new tools it affords for specifying unheralded sources of individual differences in the responses of family members to one another. Beyond individual differences these techniques, linked with other rapidly developing tools of developmental biology, promise a more mechanistic understanding of the development of family relationships and of its constituent members. However, to reach these goals, two major steps are in order now. First is the careful testing of longitudinal process models of family and individual development, and second is a more strategic use of experimental modeling to test the assumptions and explore mechanisms behind the associations studies that we have reviewed and related reports we did not. A full account of these approaches is beyond the scope of this chapter, yet a failure to mention them unfairly narrows the scope of our appraisal of what molecular genetics may contribute to the study of the family. Longitudinal Process Models Process models are distinct form the individual difference models we reviewed in the last section. They are designed to investigate the specific gene–environment interplay, in mechanistic fashion, that unfolds as development proceeds through well-specified periods of time. They begin to address the question not of whether but of how genotypic variation influences family process and individual development. The best illustrations of this approach come from the application of molecular genetic techniques to understand evocative gene–environment correlations. For example, in the same report, but using two different samples, Walum and colleagues reported that same oxytocin receptor gene variant that influenced the marital satisfaction of spouses of married women at midlife also influenced variation in the quality of the social relationships for girls at ages 8–9, as rated by their parents (Walum et al., 2012). A tempting speculation here is that the oxytocin receptor variant expressed itself earlier in positive social relations that widened and deepened the social skills of the developing woman, affording her a happy marriage years later. In the parlance of genetics relational skills of adolescent girls might be an intermediate phenotype in the causal path from genetically influenced variation in the oxytocin receptor and marital satisfaction. Or as Kendler and Neale (2010) pointed out, the genotypic influence in

social relationships of young girls might simply be an early indicator of gene variant that improves the prospect of a happy marriage but has no causal role. Moreover, to sketch a longitudinal process model, we need to make a shaky assumption that the two different samples reported by Walum and his colleagues are equivalent. In another instance, and again piecing together results from samples, two separate research teams reported the evocative effect of a gene regulating brain opioid receptors. In childhood this variant is associated with positive parent–child relationships—especially where parents have prior behavioral problems (Copeland et al., 2011). Later in development, this same variant protects against the development of substance abuse in the transition from adolescence to young adulthood in males, an effect mediated by the adolescent’s participation in alcohol using peer groups (Chassin et al., 2012). Again, it is tempting speculate that the opioid receptor variation, because it evoked positive parent–child relationships under difficult circumstances protects the developing boy so that at adolescence, he avoids alcohol-using peers and thus avoids an alcoholabusing early adulthood. Such a sequence would be consistent with the quantitative genetic study of Neiderhiser and colleagues that reported a common genetic influence on evoked parenting, delinquency of peers and substance abuse by early adulthood (Neiderhiser, Marceau, & Reiss, 2013). However, inferences about developmental processes of this kind require longitudinal data within a single sample. An intriguing example is provided by Propper and her colleagues (Propper, Shanahan, Russo, & Mills-Koonce, 2012) who studied a gene variant controlling the D4 dopamine receptor in girls. This variant was associated with several measure of these girls’ academic achievement in first grade as mediated by its apparent evocation of engaged parenting when these girls were toddlers. Although statistically challenging, process models can be explored not only for genetic main effects, as in evocative gene–environment correlation, but for gene–environment interactions. For example, Davies and Cicchetti (2014) explored a process model to explain the interaction of maternal unresponsiveness to their 2-year-olds with a variant of the serotonin transporter gene on the development of externalizing two years later. This interaction was mediated by the aggressive responses of the child towards the mother at age 3, suggesting that this reaction to mothers’ negative parenting is a developmental way station on the road to maladjustment. However, the child’s gene variant here that was sensitive to parental negativity was the so-called long or efficient form of the transporter gene that, in most studies, has been

Specifying Theories of Family Process and Individual Differences

shown to be the insensitive variant. As at many points in the molecular genetic literature it hard to know if this is a novel finding or a failure to support previous findings on the role of the short variant in stress sensitivity. Longitudinal process models will be immeasurably aided by a detailed knowledge of the timing in development when gene–environment interplay occurs. For example, Daniel Choe and Daniel Shaw and their colleagues reported the interaction of the inefficient variant of the MAOA gene with harsh parenting on adult antisocial behavior in a longitudinal design that had followed these men from infancy: those with low efficiency variant were particularly vulnerable at observations made between ages 1.5 to 5 (Choe, Shaw, Hyde, & Forbes, 2014). Experiments Passive cross sectional and longitudinal studies, that is, those that have no planned and randomized interventions, are insufficient to understand mechanisms underlying observed main effects and interaction of gene variants including those we have already reviewed. All independent variables in these studies are subject to confounding. For example, natural variation in parenting may be a proxy for genetically influenced variation in parental personality; those personality genes when transmitted to a child may influence that child’s adjustment—an instance of confounding passive gene–environment correlation. Or the natural variation in parenting may reflect the evocative effect of children’s heritable characteristics that are either not measured in the study or are, as yet, unknown. Thus an apparent gene–environment interaction is here confounded by gene–gene interaction. Also, causal pathways may be too complex to unravel in passive longitudinal studies. In the example above, many variables can affect a parent’s engaged response above and beyond the influence of the D4 dopamine receptor gene variant. Although never used in genetic studies, the pseudo-family design—where parents and children are randomly matched in a laboratory setting—can help to isolate the effect on a parent of a child’s heritable characteristics (Reiss & Oliveri, 1983). Likewise, many processes critical to understanding the genetically influenced biology of family processes, such as functioning of the brain’s cortex, cannot be visualized in the settings where most passive longitudinal studies unfold. Thus, the major objectives of experiments are effective control of independent variables free of confounds, simplification of complex causal networks, and visualization of otherwise occult biological process. Experimental procedures have a long history in the study of family relationships (Howe & Reiss, 1993). The Strange

361

Situation is an experiment holding constant mother–child reunions while focusing on variations in children’s reactions. Experiments have sought to experimentally vary parental response through standardized instructions to parents to feign affect (Hornik, Risenhoover, & Gunnar, 1987) or hold infant temperament constant by the use of standardized simulated babies (Rutherford, Goldberg, Luyten, Bridgett, & Mayes, 2013). Where visualization of evoked potential is required to identify cortical response involved in parent–child relationships, computer simulations of reunion experience can be programmed so that the precise moment of simulated reunion can be related to the timing of cortical responses that follow just a few milliseconds afterwards (White, Wu, Borelli, Mayes, & Crowley, 2013). The ecological validity of these experimental settings are always open to questions as are the degree to which variables manipulated in experiments successfully simulate more ordinary occurring experience in families. However, experimental models involving genotyping have one advantage over other models of complex family processes: the genotyped variable is absolutely constant across settings. To a large extent this is true even when animal models are used in experimental probes of mechanisms of gene-family process interplay. While even in close human relatives, such as Rhesus macaques, there are some differences in gene structure and variation the analogies in structure and function provide high levels of security that, in these animal experiments, genetic variation effectively models that in humans. As an example of an experiment exploring a molecular genetic influence on family process, Peltola and colleagues (2014) provided women with brief exposure to infant and adult faces and measured the evoked cortical potentials immediately following these exposures. Their results suggested that the same gene variant associated with sensitive mothering in a passive design also led to more rapid and intense cortical responses to infant but not adult faces displaying intense emotions. Mothers, in contrast to women without children, showed more rapid and intense reactions to high emotion in the infant pictures but the effect of the gene variant on cortical response held for both mothers and nonmother controls. These results, not yet replicated, suggest that the gene variant is associated with a trait of emotional responsiveness to infants independent of the substantial neuroendocrine and psychological transformations that attend pregnancy, parturition and early child rearing. There is a broad repertoire of experimental approaches available for linking the available findings on molecular genetics to a to the emergent science of developmental

362

Genetics and Family Systems

biology. Experimental studies of humans, such as those of Peltola and colleagues (2014), have been abundant; among many see the work of Baldwin Way, Shelly Taylor, and their colleagues (Moons, Way, & Taylor, 2014; Way & Taylor, 2010, 2011). Some investigators have argued that randomized, controlled trials might also be considered as experiments (see, e.g., van IJzendoorn & Bakermans-Kranenburg, 2012) since randomized interventions that change family interaction patterns do, as in more traditional experiments, experimentally control family environment assuming the randomly assigned treatment condition is effective. However, treatments that are designed to remediate or prevent serious clinical problems are invariably complex with many components and many undetected effects that may contribute to their efficacy. Howe, Reiss, and Yuh (2002) termed these meditational confounds. While techniques for componential and meditational analyses are well developed in psychotherapy research their combined complexity renders randomized controlled clinical trials a poor way to model naturally occurring, well-specified variation in the environment as occurs in ordinary development. Experiments designed to unravel mechanisms have a fourth objective, especially in the investigation of gene–environment interaction. As detailed elsewhere (Reiss et al., 2013), experiments seek to uncover the genetic main effect accounting for the interaction. Recall that in their foundational 2003 report by Caspi and colleagues found no main effect of the less efficient variant of the serotonin gene on depression but rather an interaction between stress and the variant. Caspi postulated a genetic main effect on stress sensitivity and selected the serotonin transporter gene in part because of an already published brain imaging experiment where, in a small sample, Hariri and colleagues (2002) reported that this variant had a sizable main effect on amygdale function in response to experimental manipulation of stressful images. It is tempting to place these findings into a hypothesized longitudinal process model: children born with one or two copies of the inefficient gene variant have amygdale that are excessively responsive to the stress of early maltreatment and when they become adults, the gene variant sensitizes them to a range of life stressors. This main effect of a genetic variant on amygdale function then interacts with stress during the life course to lead to depression (and perhaps other adverse outcomes). In this scenario, amygdale function would be a genetically influenced intermediate phenotype in a longitudinal process model. Indeed, Caspi’s foundational publication has stimulated well over 30 subsequent experiments where brain imaging

studies of individuals with and those without the inefficient gene variant are placed in a magnet for brain imaging in ways roughly comparable to the original Hariri report. We could not identify another experiment that came close to this one for rigorous attempts at replicating a result on a main genetic effect on a process that might mediate a molecular gene–environment interaction. A recent meta-analysis of these publications suggest, however, that this “candidate” intermediate phenotype may not be able to play its proposed role in the interplay among the gene variant, stress, and development (Murphy et al., 2013). The effect size in this analysis, though replicating in direction the original Hariri study, was far smaller than the original with an excess of positive reports over those expected given the statistical power of the publications in the meta analysis (suggesting a reporting bias). Perhaps most pertinent, a major source of heterogeneity among studies was age of subject; contrary to what might be expected for an early-appearing intermediate phenotype effect size increased in direct proportion to the increases in subjects’ age. However, the few studies of children and adolescents were not included in the review and its overall estimates included unpublished as well as published data. Just as certain patterns of gene–environment interaction suggest an inherited sensitivity to stress so have some patterns suggested—as we noted—an inherited differential susceptibility to the environment for better or worse. Experiments able to define these mechanisms and test the putative plasticity genes for the main effects on these mechanisms. Van IJzendoorn and Bakermans-Kranenburg (2012) recently reviewed evidence of these experiments, but, small in number, they are all randomized clinical trials that, as noted, have limited probative value. As noted, animal experiments also play an essential role in delineating developmental mechanisms. However, not to exceed the bounds of this chapter we mention them only briefly. Recall, they helped shape the hypotheses of the foundational Caspi papers. In animal experimental models, subjects can be randomly assigned to adverse environments. For example, infants of mothers likely to be abusive may be cross-fostered to mothers who are highly unlikely to be and vice versa (see Maestripieri, Lindell, & Higley, 2007). Rodent models allow for a range of experiments unthinkable in humans, and in many primates including inbreeding, transgenic, knockout, and knockdown experiments. Finally, it is now essential to link passive human studies of gene–environment interplay to in vitro experiments of two kinds. The first of these two are continuing in vitro experiments to establish the transcriptional efficiency of

Has Genetics Altered the Way We Think About the Family as a Social System?

gene variants employed in human studies (Flint & Munafo, 2013). For example, among the studies we have reviewed on gene variants regulating the oxytocin receptor, no fewer than six have been reported across the different studies with investigators each pursing different variants. At the level of biological efficacy are all these gene variants equivalent in effect? Are each a separable indicator of the status of oxytocin receptors, and if so, shouldn’t they all be used in future studies to test the role of this receptor in family process? A strategy of this kind may be thought of as the backward utility of in vitro experiments. However, in vitro experiments may have forward utility as well. For example, Steven Cole and his colleagues identified a variant of a gene regulating the proinflammatory interleukin 6 by determining the effects of norepinephrine on the expression of the gene that regulates interleukin 6’s circulating level; norepinephrine, a reliable product of stress-induced activation of the sympathetic nervous system, served in an in vitro model of the role of stress on the interleukin gene. Bringing this model forward, Cole and colleagues tested the role of this gene variation on its role in moderating the effect of depression on mortality (Cole et al., 2010). Although this study is not, strictly speaking, relevant to family studies it an important proof of concept for the forward utility of in vitro experiments in identifying gene variants relevant to the stress of family life HAS GENETICS ALTERED THE WAY WE THINK ABOUT THE FAMILY AS A SOCIAL SYSTEM? The State of the Evidence This chapter has attempted to convey both the promise of genetic methods in the study of the family but also methodological challenges and controversies that these methods have raised. There is little question that the use of these methods have changed our thinking about families but before summarizing these changes it is important to inventory the security with which we can draw major inferences from the data we have reviewed. Quantitative genetics is built on twin and adoption methods that began their development almost 90 years ago. They became highly relevant to the study of the family, particularly on mechanisms by which psychopathology is transmitted from parent to offspring, by the 1960s when the fateful Dorado Beach conference witnessed a confrontation between social and genetic explanations. There has been a vigorous program of investigation of many but not all the assumptions behind these methods, and as our review here indicates, many of the major findings

363

in the application of these methods to the family have been reproduced. However, uneasy lies the head that is complacent about these methods. Model testing approaches to analyzing twin data still rest on assumptions that are probably untenable unless specifically tested each time they are applied. For example, assortative mating of the twins’ parents (the selection of one another on the basis of similar, heritable traits such as alcohol use or intelligence) can underestimate heritability of traits since this pattern of mating will inflate DZ correlations but not MZ correlations. As we noted, Burt (2014) clearly showed that the twin method, particularly the model testing methods used to analyze its data, may underestimate shared environmental effects and thus underestimate the importance of between-family differences. Not to batter the twin method too severely, but we have also noted its potential problems in overestimating the nonshared environment and also have presented data to argue against that concern. The adoption method has often come to the rescue of the twin method either by providing a corrective for estimates derived from the twin method or by circumventing its assumptions. Assortative mating in adoption studies leads to overestimating heritability; moreover, the adoption method better estimates the shared environment—especially in adoption designs where siblings born to different birth parents are adopted into the same family. As we have seen, the adoption method entirely circumvents the equal environments assumption; that is why it not only had a major impact at Dorado Beach but also established beyond doubt the substantial genetic contribution to the pathogenesis of schizophrenia and thus its use by Kety, Rosenthal, Wender, Schulsinger, and Heston emerged as a foundation of biological psychiatry 1968. However, the adoption method has its own headaches. For example, as noted, the birth mother does rear her offspring for nine months: from conception to birth. While it is relatively easy to rule out the confounding effects of her use of prescription and nonprescription drugs during pregnancy the science of fetal environment is growing rapidly. Birth’s mother self care, psychological status and exposure to second had smoke and toxins may confound inferences about the role of genetics in the transmission of maternal characteristics to offspring development and these are much harder to track either prospectively or retrospectively. As noted, a new version of the adoption design includes a group of mothers who adopt a fertilized egg shortly after conception thus permitting a cleaner separation of genetic and fetal environmental influence. However, the adoption-at-conception design cannot

364

Genetics and Family Systems

readily distinguish prenatal from postnatal environmental effects and the number of women who, having endured the substantial medical stress of egg implantation procedures, go on to smoke or use drugs during pregnancy is extremely small (Rice et al., 2009) reducing the power of this design to distinguish genetic from exposure effects on fetal development. Recently, the adoption-at-conception and the adoption-at-birth design have been combined (Gaysina et al., 2013) to provide a much clearer delineation of genetic, prenatal, and postnatal influences of smoking on the development of conduct problems. Criticism has also been leveled at the conventional adoption design because adoptive families are typically middle class and hence there is a restricted range of environmental adversity and, critics have argued, a diminished opportunity to estimate the effects of harsh environment on child development (Stoolmiller, 1999). However, a searching empirical test of this possibility did not support this concern (McGue et al., 2007). Overall, the corpus of work in the field of quantitative genetics is substantial, increasingly well designed, and the major findings have been frequently reproduced. Our historical review has clarified how essential it has been to ferret out all the assumptions behind these methods, to test them rigorously and to have two entirely separate methods, each with its own improvements, to estimate both genetic and environmental influences at work in the human family. With some notable exceptions, these claims cannot be made for molecular genetic approaches. After all, its history is barely a tenth the age of its senior partner but its use has become much more widespread. We have introduced our selective review with cautious about the findings thus far. Indeed, in some minds, these cautious are sizable leading to recommendations that this line of research is of limited value (Flint & Munafo, 2013; Plomin, 2013); further, some investigators in this domain are encountering considerable difficulty in gaining funding for their work. Work in this field needs challenge and may thrive on it. Funding continuing investigation and encouraging students to enter this field should remain a high priority. Nonetheless, investigators in this field—and to some extent editors of journals that publish their reports—have often made matters more difficult and opened it to unnecessary criticism. For example, it is rare that papers reporting results of the interplay of specific gene variants and social variables in the family include an independent replication. Given the relative ease of genotyping (in contrast, for example, to developing a twin sample) such a practice would enhance the likelihood of reproducing results in follow on studies—particularly when a new combination

of gene variant, moderating environment and outcome are reported from the first time. A recent report provides an example. Petra Zimmermann and her colleagues (2011) at the Max Planck Institute of Munich combined forces with Avshalom Caspi and Terrie Moffitt and their colleagues at Duke. They drew on two longitudinal studies to report independent and confirmatory analyses of the role of variants of a gene regulating the corticoid receptor in the HPA axis feedback system on sensitivity to stress as indexed by the development of first-episode depression. As we noted, critical assumptions behind the methods of molecular genetics as applied to families have remained largely unexamined. We have already noted many of these and need only re-state three of these assumptions here: (1) no passive gene–environment correlation for those studies of the effects of parenting; (2) no interaction between the measured gene of interest and an unmeasured gene or genes that influence the environmental variable; and (3) no confounding population stratification. Finally, the path between quantitative and molecular genetics needs smoothing. Quantitative genetics now has excellent tools for identifying family process variables that moderate genetic influences on developmental psychopathology. Using those moderating variables in molecular genetic studies adds explanatory power to the quantitative analyses and is an important source of heightened prior probability (and hence replicability) for molecular analysis. The work of Danielle Dick (2011) illustrates this nicely. Using a twin design, she identified parental monitoring as a moderator of genetic influences on adolescent substance use; high levels of monitoring reduced genetic influence, a finding she extended using specific gene variants (see her account of this strategy). In another domain, molecular genetic analyses seems little concerned with heritable evocative effects that quantitative genetics has illustrated repeatedly not only in children and adolescents but in the marriages of adults. These have become more compelling as recent adoption research has illustrated more clearly their probable role in pathogenesis. Moreover, molecular genetics is not digging into an enormous domain for the study of gene–environment interaction: variations in the sensitivity of the social environment that moderate these heritable evocative effects. Finally, it seems odd that molecular genetic tools have not been deployed to investigate the most striking findings of all in the quantitative genetics of the family, the finding that genetic influences on family variables (e.g., parent–child conflict) overlap with those on individual measures of adjustment (e.g., antisocial behavior) and in notable instances account for the majority of the observed covariance between the two. Twelve years

Has Genetics Altered the Way We Think About the Family as a Social System?

after Suomi and Caspi introduced molecular genetic tools to study the family we have little insight into what genes might be involved in these associations. Thus, a fair reading of nearly a century of work applying genetic approaches to the study of the family can conclude: quantitative genetics speaks to the study of family process with the voice of authority, whereas molecular genetics speaks, at present, with the softer voice of promise. What are these voices saying? Family Research Listening to Molecular Genetics We turn first to molecular genetic analyses. Here we summarize three areas of family research where molecular genetics can be illuminating. These examples are offered with a clear purpose. As already indicated, the attractiveness of this approach to analysis is the promise of placing the study of the human family into a secure location in the rapidly advancing domain of developmental biology. Thus, it is more than high time for molecular genetics to move beyond the almost promiscuous search for more variants connected to more environments connected to more outcomes although continued promiscuity in the search of fresh liaison with the truth is still defensible. Child Maltreatment Even our selective review makes it clear that many filaments of molecular genetic analyses of family converge on this major public health problem. Molecular genetics has added a much clearer delineation of individual differences in response to this severe and sustained stress in childhood. Most literature focuses on the genetically moderated impact of maltreatment on behavioral disorders in children, adolescent, and adults. However, from a public health perspective, the long-term impact of child maltreatment on adult medical health may be its most important scourge. Precisely, what is the risk of adult mental and physical illness attributable to child maltreatment? The computation of attributable risk is a secure public health indicator because it allows a calculation of the reduction in disease incidence and prevalence that one might expect if the risk factor were eliminated or fully offset by intervention. Such estimates are just being published. For example, Kessler and colleagues estimate the risk for a broad spectrum of adolescent and adult psychiatric disorders attributable to early adversity, mostly maltreatment, is an astonishing 30% (Kessler et al., 2010). This estimate may be revised downward as future studies correct for passive gene–environment correlation and retrospective bias but estimates might be revised upward

365

in a subset of individuals if their vulnerability genes are included in the epidemiological inquiry. We do not have gene–environment studies for adult medical illness associated with child maltreatment comparable to those for child and adolescent psychopathology. However, we do not lack fastidious longitudinal studies documenting this risk (e.g., Danese et al., 2009; Widom, Czaja, Bentley, & Johnson, 2012; Wilson & Widom, 2011). Nor do we lack studies that explore the mechanisms by which this risk persists into adulthood (e.g., see Miller, Chen, & Parker, 2011) for a review chronic inflammation occasioned by the stress of early maltreatment and (McCrory, De Brito, & Viding, 2012) for persistent effects on brain structure and function). Data thus far on molecular genetic vulnerability to child maltreatment augurs well for this vital extension to adult medical illness and a trickle of data on this domain is beginning to appear (Brummett et al., 2008; Xie et al., 2013; Zhao, Bremner, Goldberg, Quyyumi, & Vaccarino, 2013). One might think that identifying gene variants that increased children’s sensitivity to maltreatment would be the most obvious public health benefit of current research and perhaps it will lead at some point to fitting remediation to specific children by virtue of their genotype. However, it hard to imagine that current efforts to reduce the prevalence of child maltreatment or to reduce the time that children are exposed to it or to remediate its potential impact on the children would become more focused or even efficient with further molecular genetic study. Would anyone now deny a child the benefit of the nurse visitation programs or child–parent psychotherapy programs, programs that seem to have an overall beneficial effect in maltreatment (Cicchetti & Banny, 2014), on the basis of their genotype? Genotyping provides fresh insights into mechanisms that might account for the impact of maltreatment on children but has no obvious impact yet on therapeutic innovation. The reverse is true for the long-term effects of child maltreatment on adult medical illness that may become manifest decades after the maltreatment. Even if we had a thoroughly enlightened international policy on child maltreatment and its amelioration it is likely that many adults, seriously maltreated in childhood, would reach adulthood with no obvious signs of medical illness apparent but at substantial risk of future malady. It is here where an understanding of genotypes and liability can have huge implications for selective preventive interventions. Developmental Family Neuroscience It is fair to say that major advances in social neuroscience rest substantially on laboratory experiments where

366

Genetics and Family Systems

complex social stimuli are modeled by a broad array of auditory and visual stimuli, briefly presented to just one member per family and correlated with various cortical recordings, brain imaging and blood sampling techniques. Social neuroscience becomes more social when comparable recordings are made in simulated or actual interactional games (see Todorov, Fiske, & Prentice, 2011, for review of these approaches). Social neuroscience moves into the family when these techniques test the moderation of the presence of family members on neural responses to standard stimuli (e.g., Coan, Schaefer, & Davidson, 2006; Conner et al., 2012) and validated against the quality and durability of important intimate relationships (e.g., Eisenberger et al., 2011). Many of the molecular genetics papers we have reviewed offer a tantalizing vision of a substantial enlargement of this field to focus on the development of the family as a system deploying the technology of social neuroscience but guided by genotyping. Genotyping can be regarded as a probe of relatively stable individual differences in the operation of neural systems although Plomin et al. (1994) cautioned against the simple notion of one gene being associated specifically with one disease. A comparable caution is warranted here: the neural systems of the brain are complex and interrelated and variations in a single gene may be imprecise indicators of stable individual differences in their function. Bearing these caution in mind, let us imagine a sequence suggested by these papers. The sequence can start with mating and marriage here. Quantitative genetic data suggest genetic influences on becoming married (Jerskey et al., 2010); it is conceivable that specific genes may be identified that regulate this starting point in the assembly of a family; it is relevant that Walum and colleagues (2008) showed an association of very small effect of a variant of the gene regulating vasopressin receptor in men on whether couples with adolescent children were married or cohabiting. That is, where a man possessed one variant the couple remained together without the commitment of marriage and if he possessed the other variant they were married. It is conceivable that members of a couple pick each other because of similarities in genotypes although such an investigation may confound population stratification effects with choice effects. Nonetheless, once married small effects of gene variants—vasopressin for men and oxytocin receptor for women—have been reported (Walum et al., 2008, 2012). Moreover, parental gene variants may promote the parent–child relationship through its enhancement of sensitive parenting (Bakermans-Kranenburg & van IJzendoorn, 2008; Feldman et al., 2012). Children’s genes

may also promote this relationship through eliciting or evoking positive parenting and (Hayden et al., 2013; Hayden et al., 2010; Mills-Koonce et al., 2007; Propper et al., 2012). Moreover, there are gene variants in parents and in children that might be said to have a protective effect on parent–child relationships. For example, genotypic variants may protect mothers post partum from depression despite external difficulties (Mitchell et al., 2011); they may protect parental function from marital problems (Sturge-Apple et al., 2012) and protect the parent–child relationship from the mother’s own depression (Apter-Levy et al., 2013). We have described one report of a child gene-variant regulating the endogenous opiate system that appears to protect a positive parent–child relationship when parental psychiatric problems might otherwise perturb it (Copeland et al., 2011). Finally, there appear gene variants that protect relationships in a subsequent generation in the family from the transmission of interactional difficulties in the prior generation (Beaver & Belsky, 2012; Masarik et al., 2014; Simons et al., 2013). Even from these initial fragments it is possible to sketch, with the lightest and most tentative hand, a developmental sequence of between-family difference in genetically influenced promoting and protective processes unfolding across two generations in the development of the family. When a sequel to this chapter is written, perhaps 10 years hence, it seems likely that many of these specific findings will have evaporated as attempted replications may fail or improved genetic and statistical toilette is satisfied. Nonetheless, the framework suggested here—a focus on the developmental stages of the family—may survive as a way of integrating newer, more solid findings, framing more solid hypotheses and developing better methods. As for the last, Mills-Koonce et al. (2007) and Choe et al. (2014) provided a look at the future of the role of genetics in a developmental family neuroscience; the former used the family as the unit of genotyping with genetic data collected from at least two members in each family and the latter presages a more precise timing of when, in the development of family relationships, genetic effects have their most significant influence. The framework we propose here can keep molecular genetics focused on the family and on how these molecular techniques may enlarge our knowledge of how the family unfolds over time and across generations and how it shapes and is shaped by each of its members. Fitting Family Interventions to the Family The field is indebted to the early publications in this field for their perspicacity, three of them reviewed here. However, it

Has Genetics Altered the Way We Think About the Family as a Social System?

is time to move on to other strategies to take full advantage of the molecular genetic investigation of the family. As we noted, clinically efficacious interventions are a poor way to model naturally occurring variation in family process. Moreover, post hoc genotyping of already completed trials—even where the genotyping is nearly complete for all subjects—can produce misleading results, especially where there is a genetic main effect on the liability or pathology to which the intervention is addressed. In such an instance, the apparent sensitivity of individuals (or families) with a presumably sensitive genotype under study cannot be separated from the ceiling effect of the treatment. The apparent sensitivity of the genotype may just reflect a familiar clinical truth: sicker patients get more out of treatment than those who are less so. In fitting a specific treatment to a particular family genotype there is no substitute for a prospectively designed randomized trial, planned from the start, with blocking on the genotype in question such that those with and without the genotype are equally sick or equally likely to become so. Since genotype is a blocking variable, all subjects must be genotyped before randomization with careful attention to whatever sampling bias such a requirement may inadvertently produce. Although contemporary techniques such as urn randomization might help it should not be necessary. The theory of differential susceptibility marshals empirical evidence buttressed by evolutionary reasoning to posit a set of genes that confer plasticity on individuals (and perhaps their families). The scientific task of fitting family interventions to family genotypes provides a test, in the most practical terms possible, of the mettle of this theory. Because of their for-better-or-for-worse quality, these genotypes should be untroubled by main effects but, if the theory is true, should predict more sizable treatment effects, which should become clearly apparent in a blocked design. During an early stage of a new branch of science inquisitive exploration, even playful exploration without rigorously derived prior probabilities, should be encouraged. The exuberance with which family researchers have embraced PCR has, in some instances, exemplified this productive vein in science. Fitting a treatment to a person or to a family with serious troubles or with liability for future troubles, on the other hand, is a serious business. Suppose genotype A proves sensitive to the treatment but genotype B does not; replicated results from well-designed clinical trials would rationally lead to recommending against the treatment for genotype B. A family-based use of molecular genetics—as we have reviewed it here—offers notable potential for this fitting process. However this evolving research paradigm should conform to well-established

367

practices for other clinical trials. These include placing the measures including genotyping, and including expected outcomes in the lockbox of clinicaltrials.gov. It is entirely conceivable that such a study is already on the drawing board if not already under way. Family Research Listening to Quantitative Genetics Common Genes and Covariances We have reviewed abundant evidence that there are substantial genetic influences on most measures of family relationships including parent–child, marital, and even sibling relationships. Equally abundant evidence suggests that many of these genetic influences are shared with those that influence such important measures of individual adjustment such as depression, antisocial behavior, social responsibility, autonomy, and cognitive motivation and achievement. These are evident in relatively short periods of developmental time where, for example, negative parenting might be assessed during early adolescence and child delinquency in later adolescence. These effects of common genes are also apparent across generations as in the case of parental alcoholism and offspring impulsive behavior. Every model of family influence prior to the use of genetically informative designs is potentially misspecified. While we have already belabored the point this circumstance is also the case for molecular gene–environment interactions where presumed environmental effect on adjustment could be attributable, in whole or in part to evocative gene–environment correlation for genes not measured in the study (or unknown and hence unmeasurable) or to passive gene–environment correlation. We re-emphasize here that genetic studies have effectively delineated family subsystems that appear to have a decisive impact on the development of its members by dint of social, not genetic, mechanisms. In our review, marital and siblings systems stand out here but the genetically informed literature also clarifies where parent–child relationships are associated with child adjustment for environmental, not genetic, reasons. Must every study, then, be a twin or an adoption study? No. But family researchers from here on out must think genetically (Reiss, 2010) and read the genetic literature pertinent to the problem they are studying. Further, in considering interpretations of their findings they need to include the possibility of genetic influences on covariances they report. However, genetically informed studies need to be included somewhere in an extensive program of research. We think twin and adoption studies should be accessible resources for the field of family studies and

368

Genetics and Family Systems

research funding should be available for their establishment and maintenance just as government funding has helped distribute high density EEG and MRI facilities, each piece of equipment available for the use of many nearby researchers. However, genetically informed designs can be used even in the pilot phases of a family research project. Let us suppose, for example, that a researcher has a novel idea about a parent–child interaction pattern than increases the child’s liability for anxiety disorders. The researcher might consider trying out this novel idea on adoptive families in the initial study. Recruiting adoptive families is much easier if they do not have to be linked to the birth parents of the adopted child. If our investigator gets a positive finding on the first go-around then the investigator has identified a potentially salient attribute of parent–child relationships and already knows, using this design, that its covariance with child symptoms cannot be attributed to passive gene–environment correlation. Another option for this investigator is to use a sibling design; it allows for an immediate distinction: are the observed effects due to differences among families (either genetic or shared environment) or differences within families (nonshared or genetic). There are other compelling reason for prioritizing sibling designs including the important effects of sibling socialization and for studies of adult development, a topic to which we return below. Evocative Gene–Environment Correlation and the Pathogenesis of Heritable Disorders We have reviewed several quantitative genetic (Burt et al., 2005; Larsson et al., 2008; Neiderhiser, Reiss, Hetherington, & Plomin, 1999; Reiss et al., 2000; Tucker-Drob & Harden, 2012) and molecular genetic studies (Propper et al., 2012) that suggest as a novel mechanism by which genetic influence is expressed in behavioral disorders. A heritable disposition in the child evokes a response from a parent that, in turn, amplifies the child’s disposition that evokes an even more vigorous parental response. These reciprocal effects may cycle back and forth until a clear pathological outcome is noted in the child (such outcomes from these reciprocal processes are also possible in the parents; see Gross, Shaw, & Moilanen, 2008). All of this process can be hidden in estimates of the heritability of a disorder as determined by, for example, twin designs or even newer methods of estimating heritability based on correlated patterns of genotypic similarity between pairs of genetically unrelated individuals who are similar in any measured phenotype (Yang, Lee, Goddard, & Visscher, 2011). All these heritability estimates say to us that genotypic differences among individuals lead to phenotypic

differences—somehow and some way. As noted, the way might be entirely through conventional pathways from DNA to RNA to proteins relevant for fetal, child, or adult brain function (or other organs relevant for individual differences in behavior such as the adrenal gland). Or there could be an outside-the-skin mechanisms. Here the intermediate behavioral phenotypes are heritable characteristics that evoke distinctive social responses. That is, these heritable characteristics are likely to induce the same reactions in other family members across families so that they are detectable as a reliable and heritable social signals. We have cited two papers from the same adoption study that illustrate how this reciprocal process may work for the early the development of conduct problems (Fearon et al., in press) and for ADHD symptoms (Harold et al., 2013). The Fearon paper illustrated how between family differences might blunt this process: families with happy marriages somehow cancels mother’s role in this reciprocal process with salutary effects on the child’s development of self-control. But within-family differences may also matter. For example, the heritable properties that elicit warmth in mothers are uncorrelated with those that evoke warmth in fathers (Reiss et al., 2000); to be detectable these differences in response—shaped by either role or gender of the parent—must generalize across families (at least in the sample studied) to be detectable. But it seems likely that families may react to a range of aversive behaviors in children—each determined by a different gene set—in comparable ways: each elicits a negative response. The gene-variants we reviewed in molecular studies of evocative processes in children suggested that effects of many different genes do converge on negative parenting. To follow the logic here it is the response of the family to a diverse set of heritable signals in the child that plays a decisive role in the outside the skin mechanism of the expression of genetic effects in behavioral syndromes. The family may make fine distinctions—as in the case of mother or father differences in warm responses to their adolescent offspring or lump many genetically distinct effects on child’s characteristics together as functionally equivalent in odiousness or attractiveness. John Loehlin presented a method for tracking how the family may lump or discriminate among these heritable genetic signals (Loehlin, Neiderhiser, & Reiss, 2005). In a reanalysis of the data from the nonshared study by Reiss and his colleagues, he used a genetically informed factor analysis to suggest that the same genetic factors that elicit warmth from parents suppress conflict negativity (or vice versa). That is, so far as the family’s response is concerned

Summary

all the heritable features leading to either warmth or negativity shape a continuous dimension of family reaction with warmth on the high end and negativity on the low end and this dimension of reaction—initially influenced entirely by the child’s genotype but shaped by the family’s reaction—is, in turn, distinctively related to child’s autonomy and negatively related to the child depression and antisocial behavior. However, the child factors that elicit monitoring and control are uncorrelated with those eliciting the continuous dimension of warmth-negativity. This dimension, with high monitoring and control at the upper end is distinctively related to sociability and cognitive achievement in the adolescents; these positive attributes are unrelated to the genetically elicited warmth-negativity dimension. While highly preliminary these data do suggest that, where families are integral to the outside the skin mechanism of gene expression, they may recode the genetic information that resides in the heritable characteristics of the child. Siblings and Adult Development Adult development is as important to the field of developmental psychopathology as is child development. An emerging pattern of data from quantitative genetics seems to have much to say about these processes. While specific nonshared influences have been identified in early development their magnitude has not only be less than expected from the twin and adoption studies reviewed by Plomin and Daniels (1987). Moreover, in early development these influences were transitory and their shifts, even over short periods of time, accounted for much of the instability of measures of child adjustment over time. The picture looks quite different as individuals age. The nonshared influences become stable and efforts to identify them have successfully centered on marriage although sibling differences in occupational experience, exogenously caused health problems or even residential location are also candidates for large-scale effects. A major base for this interest in nonshared environment in adult development comes from the meta-analyses of Briley and TuckerDrob already cited (Briley & Tucker-Drob; Tucker-Drob & Briley, 2014) which clarifies the increasing importance of stable nonshared influences, across development, in two large domains of personal functioning: cognition and personality. Currently, one of the most astonishing initiatives in behavioral science is the U.S. Health and Retirement Study that follows a large number of individuals from midlife until advanced old age or death. This seminal study has spawned many other comparable surveys in

369

Europe, Asia, the Middle East, Africa, and Latin America. All of them are searching for ways to characterize patterns of aging and for the influences that shape those patterns over time. Quantitative genetics, though early in its exploration of this domain, provides an impudent suggestion for these very large studies. The fundamental questions about aging may not be estimates of factors accounting for differences among individuals. Rather the question about trajectories of aging might be reframed: what factors account for difference and similarities between siblings in the same family over the life course? However, the quantitative genetics literature we have reviewed suggests optimism for the discovery of the potential sources of these differences—above and beyond marriage. Except in the rare case of MZ twins (they are likely to be present in only 1 of approximately 80 cases in a random sample of aging individuals) these differences could be attributable to genetic differences among ordinary siblings. Or they could be due to critical differential experiences such as a good marriage for one or a bad marriage or no marriage for the other. We think a sibling difference in life course is a productive approach to adult development but rigorous estimates of the role of genetic differences among siblings will depend on oversampling twins. Several large-scale survey studies provide precedent (e.g., Rowe, Jacobson, & Van den Oord, 1999; South & Krueger, 2008). There is, of course, another reason for studies of midlife and aging to sample subjects by sibling pairs. We have already reviewed data suggesting that the reciprocal environments that siblings build for themselves appears to have a sustained influence on the development of each at least through the transition to adulthood. In their reviews of sibling studies, Tucker-Drob and Briley found that stable, shared environmental influences persisted across the life span (Tucker-Drob & Briley, 2014). Do reciprocal sibling relationships play a significant role in the persistence of these stable shared environmental influences? Do these reciprocal sibling environments still bear the imprint of much earlier parent–child relationships?

SUMMARY We have sketched a history of rapprochement between genetics and the social sciences in the study of the family. There are many domains in the study of the family that have been inalterably changed by this rapprochement: a heightened interest in differences between families in the sensitivity of members to one another, early steps in

370

Genetics and Family Systems

embedding the study of family processes in a genuine transactional neuroscience, the prospect of a better fit between families and family-level interventions design to reduce problems they already have or to prevent them from occurring in the first place, a clearer recognition that genetic relatedness among family members influences observed associations between characteristics of parents and of their children, a new etiological model of developmental psychopathology focusing on the role in the expression of genetic influences on psychopathology through reciprocal relationships of genetically influenced evocative behavior in the child and the response of the parent, and finally, a perspective on sibling-based sampling and analysis in the study of adult development and aging. We hope geneticists are also listening to the results of this rapprochement to better appreciate the circumstances under which specific gene variants are expressed or not expressed and to include in their investigations the outside-the-skin mechanisms reviewed here.

REFERENCES Agrawal, A., Jacobson, K., Prescott, C., & Kendler, K. (2002). A twin study of sex differences in social support. Psychological Medicine, 32(7), 1155–1164. doi: 10.1017/S0033291702006281 Apter-Levy, Y., Feldman, M., Vakart, A., Ebstein, R. P., & Feldman, R. (2013). Impact of maternal depression across the first 6 years of life on the child’s mental health, social engagement, and empathy: The moderating role of oxytocin. American Journal of Psychiatry, 170(10), 1161–1168. doi: 10.1176/appi.ajp. 2013.12121597 Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2008). Oxytocin receptor (OXTR) and serotonin transporter (5-HTT) genes associated with observed parenting. Social Cognitive and Affective Neuroscience, 3(2), 128–134. doi: 10.1093/scan/nsn004 Bakermans-Kranenburg, M. J., Van IJzendoorn, M. H., Pijlman, F. T., Mesman, J., & Juffer, F. (2008). Experimental evidence for differential susceptibility: Dopamine D4 receptor polymorphism (DRD4 VNTR) moderates intervention effects on toddlers’ externalizing behavior in a randomized controlled trial. Developmental Psychology, 44(1), 293–300. doi: 10.1037/0012–1649.44.1.293 Barr, C. S., Newman, T. K., Becker, M. L., Champoux, M., Lesch, K. P., Suomi, S. J., . . . Higley, J. (2003). Serotonin transporter gene variation is associated with alcohol sensitivity in rhesus macaques exposed to early-life stress. Alcoholism: Clinical and Experimental Research, 27(5), 812–817. Bates, T. C., Lewis, G. J., & Weiss, A. (2013). Childhood socioeconomic status amplifies genetic effects on adult intelligence. Psychological Science, 24(10), 2111–2116. doi: 10.1177/0956797613488394 Bauer, L. O., Covault, J., Harel, O., Das, S., Gelernter, J., Anton, R., & Kranzler, H. R. (2007). Variation in GABRA2 predicts drinking behavior in Project MATCH subjects. Alcoholism: Clinical and Experimental Research, 31(11), 1780–1787. Beaver, K. M., & Belsky, J. (2012). Gene–environment interaction and the intergenerational transmission of parenting: Testing the differential-susceptibility hypothesis. Psychiatric Quarterly, 83(1), 29–40. doi: 10.1007/s11126–011–9180–4

Begley, C. G., & Ellis, L. M. (2012). Drug development: Raise standards for preclinical cancer research. Nature, 483(7391), 531–533. doi: 10.1038/483531a Bell, R. Q. (1968). A reinterpretation of the direction of effects in studies of socialization. Psychological Review, 75, 81–95. doi:http://dx.doi.org /10.1037/h0025583 Belsky, J., Bakermans-Kranenburg, M. J., & Van IJzendoorn, M. H. (2007). For better and for worse: Differential susceptibility to environmental influences. Current Directions in Psychological Science, 16(6), 300–304. doi: 10.1111/j.1467–8721.2007.00525.x Belsky, J., & Pluess, M. (2009). Beyond diathesis stress: Differential susceptibility to environmental influences. Psychological Bulletin, 135(6), 885–908. Bisceglia, R., Jenkins, J. M., Wigg, K. G., O’Connor, T. G., Moran, G., & Barr, C. L. (2012). Arginine vasopressin 1a receptor gene and maternal behavior: evidence of association and moderation. Genes Brain and Behavior, 11(3), 262–268. doi: 10.1111/j.1601–183X.2012.00769.x Blonigen, D. M., Carlson, M. D., Hicks, B. M., Krueger, R. F., & Iacono, W. G. (2008). Stability and change in personality traits from late adolescence to early adulthood: A longitudinal twin study. Journal of Personality, 76(2), 229–266. doi: 10.1111/j.1467–6494.2007.00485.x Bouchard, T. J., Jr.,, Lykken, D. T., McGue, M., Segal, N. L., & Tellegen, A. (1990). Sources of human psychological differences: The Minnesota study of twins reared apart. Science, 250(4978), 223–228. doi: 10.1126/science.2218526 Briley, D. A., & Tucker-Drob, E. M. (2014). Genetic and environmental continuity in personality development: A meta-analysis. Psychological Bulletin, 140, 1303–1331. 10.1037/a0037091 Brody, G. H., Beach, S. R., Philibert, R. A., Chen, Y. F., & Murry, V. M. (2009). Prevention effects moderate the association of 5-HTTLPR and youth risk behavior initiation: gene × environment hypotheses tested via a randomized prevention design. Child Development, 80(3), 645–661. doi: 10.1111/j.1467–8624.2009.01288.x Brody, G. H., Chen, Y. F., & Beach, S. R. (2013). Differential susceptibility to prevention: GABAergic, dopaminergic, and multilocus effects. Journal of Child Psychology and Psychiatry, 54(8), 863–871. doi: 10.1111/jcpp. 12042 Brummett, B. H., Boyle, S. H., Siegler, I. C., Züchner, S., Ashley-Koch, A., & Williams, R. B. (2008). Lipid levels are associated with a regulatory polymorphism of the monoamine oxidase—A gene promoter (MAOA-uVNTR). Medical Science Monitor: International Medical Journal of Experimental and Clinical Research, 14(2), CR57–CR61. Buchanan, J. P., McGue, M., Keyes, M., & Iacono, W. G. (2009). Are there shared environmental influences on adolescent behavior? Evidence from a study of adoptive siblings. Behavior Genetics, 39(5), 532–540. doi: 10.1007/s10519–009–9283-y Bullock, B. M., Deater-Deckard, K., & Leve, L. D. (2006). Deviant peer affiliation and problem behavior: a test of genetic and environmental influences. Journal of Abnormal Child Psychology, 34(1), 27–39. Burks, B. S. (1927). Foster parent–foster child comparisons as evidence upon the nature–nurture problem. Proceedings of the National Academy of Sciences USA, 13(12), 846–848. Burt, S., McGue, M., Krueger, R. F., & Iacono, W. G. (2005). How are parent–child conflict and childhood externalizing symptoms related over time? Results from a genetically informative cross-lagged study. Development and Psychopathology, 17(1), 145–165. doi: 10.1017/ S095457940505008X Burt, S. A. (2014). Research review: the shared environment as a key source of variability in child and adolescent psychopathology. Journal of Child Psychology and Psychiatry, 55(4), 304–312. doi: 10.1111/ jcpp. 12173

References Burt, S. A., Donnellan, M. B., Humbad, M. N., Hicks, B. M., McGue, M., & Iacono, W. G. (2010). Does marriage inhibit antisocial behavior?: An examination of selection vs causation via a longitudinal twin design. Archives of General Psychiatry, 67(12), 1309–1315. doi: 10.1001/archgenpsychiatry.2010.159 Burt, S. A., McGue, M., Krueger, R. F., & Iacono, W. G. (2005). Sources of covariation among the child-externalizing disorders: informant effects and the shared environment. Psychological Medicine, 35(8), 1133–1144. doi: 10.1017/S0033291705004770 Burt, S. A., McGue, M., Krueger, R. F., & Iacono, W. G. (2007). Environmental contributions to adolescent delinquency: A fresh look at the shared environment. Journal of Abnormal Child Psychology, 35(5), 787–800. doi: 10.1007/s10802–007–9135–2 Bussell, D. A., Neiderhiser, J. M., Pike, A., Plomin, R., Simmens, S., Howe, G. W., . . . Reiss, D. (1999). Adolescents’ relationships to siblings and mothers: A multivariate genetic analysis. Developmental Psychology, 35(5), 1248–1259. Butcher, L. M., & Plomin, R. (2008). The nature of nurture: A genomewide association scan for family chaos. Behavior Genetics, 38(4), 361–371. doi: 10.1007/s10519–008–9198-z Button, T. M., Corley, R. P., Rhee, S. H., Hewitt, J. K., Young, S. E., & Stallings, M. C. (2007). Delinquent peer affiliation and conduct problems: A twin study. Journal of Abnormal Psychology, 116(3), 554–564. Byrd, A. L., & Manuck, S. B. (2014). MAOA, childhood maltreatment, and antisocial behavior: Meta-analysis of a gene– environment interaction. Biological Psychiatry, 75(1), 9–17. doi: 10.1016/j.biopsych.2013.05.004 Cadoret, R. J. (1982). Genotype–environment interaction in antisocial behavior. Psychological Medicine, 12(2), 235–239. Cadoret, R. J., & Cain, C. A. (1981). Genotype–environmental interaction in antisocial behavior. Psychological Medicine, 12, 235–239. Cadoret, R. J., Cain, C. A., & Crowe, R. R. (1983). Evidence for gene– environment interaction in the development of adolescent antisocial behavior. Behavior Genetics, 13(3), 301–310. Caldwell, B. M., & Bradley, R. H. (1978). Home observation for measurement of the environment. Little Rock, AR: University of Arkansas. Cases, O., Self, I., Grimsby, J., Gaspar, P., Chen, K., Pournin, S., . . . De Maeyer, E. (1995). Aggressive behavior and altered amounts of brain serotonin and norepinephrine in mice lacking MAOA. Science, 268(5218), 1763–1766. Caspi, A., Hariri, A. R., Holmes, A., Uher, R., & Moffitt, T. E. (2010). Genetic sensitivity to the environment: the case of the serotonin transporter gene and its implications for studying complex diseases and traits. American Journal of Psychiatry, 167(5), 509–527. doi: 10.1176/appi.ajp. 2010.09101452 Caspi, A., McClay, J., Moffitt, T., Mill, J., Martin, J., Craig, I. W., . . . Poulton, R. (2002). Role of genotype in the cycle of violence in maltreated children. Science, 297(5582), 851–854. doi: 10.1126/science.1072290 Caspi, A., Moffitt, T. E., Morgan, J., Rutter, M., Taylor, A., Arseneault, L., . . . Polo-Tomas, M. (2004). Maternal expressed emotion predicts children’s antisocial behavior problems: Using monozygotic-twin differences to identify environmental effects on behavioral development. Developmental Psychology, 40(2), 149–161. doi: 10.1037/0012–1649.40.2.149 Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . Poulton, R. (2003). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301(5631), 386–389. doi: 10.1126/science.1083968

371

Cecil, C. A., Barker, E. D., Jaffee, S. R., & Viding, E. (2012). Association between maladaptive parenting and child self-control over time: Cross-lagged study using a monozygotic twin difference design. British Journal of Psychiatry, 201(4), 291–297. doi: 10.1192/bjp.bp. 111.107581 Chabris, C. F., Lee, J. J., Benjamin, D. J., Beauchamp, J. P., Glaeser, E. L., Borst, G., . . . Laibson, D. I. (2013). Why it is hard to find genes associated with social science traits: Theoretical and empirical considerations. American Journal of Public Health, 103 Suppl 1, S152–166. doi: 10.2105/AJPH.2013.301327 Champoux, M., Bennett, A., Shannon, C., Higley, J., Lesch, K., & Suomi, S. (2002). Serotonin transporter gene polymorphism, differential early rearing, and behavior in rhesus monkey neonates. Molecular Psychiatry, 7(10), 1058–1063. Chassin, L., Lee, M. R., Cho, Y. I., Wang, F. L., Agrawal, A., Sher, K. J., & Lynskey, M. T. (2012). Testing multiple levels of influence in the intergenerational transmission of alcohol disorders from a developmental perspective: The example of alcohol use promoting peers and mu-opioid receptor M1 variation. Development and Psychopathology, 24(3), 953–967. doi: 10.1017/S0954579412000478 Choe, D. E., Shaw, D. S., Hyde, L. W., & Forbes, E. E. (2014). Interactions Between Monoamine Oxidase A and Punitive Discipline in African American and Caucasian Men’s Antisocial Behavior. Clinical Psychological Science, 2(5), 591–601. doi: 10.1177/2167702613518046 Cicchetti, D., & Banny, A. (2014). A developmental psychopathology perspective on child maltreatment. In, D. Cicchetti, Handbook of developmental psychopathology (pp. 723–741). Springer. Cicchetti, D., Rogosch, F. A., & Oshri, A. (2011). Interactive effects of corticotropin releasing hormone receptor 1, serotonin transporter linked polymorphic region, and child maltreatment on diurnal cortisol regulation and internalizing symptomatology. Development and Psychopathology, 23(4), 1125–1138. doi: 10.1017/S0954579411000599 Cicchetti, D., Rogosch, F. A., & Thibodeau, E. L. (2012). The effects of child maltreatment on early signs of antisocial behavior: Genetic moderation by tryptophan hydroxylase, serotonin transporter, and monoamine oxidase A genes. Development and Psychopathology, 24(3), 907–928. doi: 10.1017/S0954579412000442 Coan, J. A., Schaefer, H. S., & Davidson, R. J. (2006). Lending a hand: Social regulation of the neural response to threat. Psychological Science, 17(12), 1032–1039. Cole, S. W., Arevalo, J. M., Takahashi, R., Sloan, E. K., Lutgendorf, S. K., Sood, A. K., . . . Seeman, T. E. (2010). Computational identification of gene–social environment interaction at the human IL6 locus. Proceedings of the National Academy of Sciences U S A, 107(12), 5681–5686. doi:10.1073/pnas.0911515107 Conley, D., Rauscher, E., Dawes, C., Magnusson, P. K., & Siegal, M. L. (2013). Heritability and the equal environments assumption: Evidence from multiple samples of misclassified twins. Behavior Genetics, 43(5), 415–426. doi: 10.1007/s10519–013–9602–1 Conner, O. L., Siegle, G. J., McFarland, A. M., Silk, J. S., Ladouceur, C. D., Dahl, R. E., . . . Ryan, N. D. (2012). Mom-it helps when you’re right here! Attenuation of neural stress markers in anxious youths whose caregivers are present during fMRI. PLoS One, 7(12), e50680. doi: 10.1371/journal.pone.0050680 Copeland, W. E., Sun, H., Costello, E. J., Angold, A., Heilig, M. A., & Barr, C. S. (2011). Child mu-opioid receptor gene variant influences parent–child relations. Neuropsychopharmacology, 36(6), 1165–1170. doi: 10.1038/npp. 2010.251 Correns, C. (1950). G. Mendel’s law concerning the behavior of progeny of varietal hybrids. Genetics, 35(5), 33–41.

372

Genetics and Family Systems

Danese, A., Moffitt, T. E., Harrington, H., Milne, B. J., Polanczyk, G., Pariante, C. M., . . . Caspi, A. (2009). Adverse childhood experiences and adult risk factors for age-related disease: depression, inflammation, and clustering of metabolic risk markers. Archives of Pediatrics and Adolescent Medicine, 163(12), 1135–1143. doi: 10.1001/archpediatrics.2009.214 Darwin, C. (1920 (1868)). The variation of animals and plants under domestication. New York, NY: D. Appelton. Davies, P. T., & Cicchetti, D. (2014). How and why does the 5-HTTLPR gene moderate associations between maternal unresponsiveness and children’s disruptive problems? Child Development, 85(2), 484–500. doi: 10.1111/cdev.12148 Dick, D. M. (2011). Gene–environment interaction in psychological traits and disorders. Annual Review of Clinical Psychology, 7, 383–409. Donaldson, Z. R., & Young, L. J. (2008). Oxytocin, vasopressin, and the neurogenetics of sociality. Science, 322(5903), 900–904. Duncan, L. E., & Keller, M. C. (2011). A critical review of the first 10 years of candidate gene-by-environment interaction research in psychiatry. Am J Psychiatry, 168(10), 1041–1049. doi: 10.1176/appi .ajp. 2011.11020191 Dunn, J., & Plomin, R. (1986). Determinants of maternal behaviour towards 3-year-old siblings. British Journal of Developmental Psychology, 4(2), 127–137. Durkheim, E. (1951). Suicide. A study in sociology (J. A. t. Spaulding & G. t. Simpson, Eds.). Glencoe, IL: Free Press. (Original work published 1897.) Eaves, L. (2006). Genotype × environment interaction in psychopathology: Fact or artifact? Twin Research and Human Genetics, 9(1), 1–8. Eaves, L., Foley, D., & Silberg, J. (2003). Has the “equal environments” assumption been tested in twin studies? Twin Research, 6(6), 486–489. Eisenberger, N. I., Master, S. L., Inagaki, T. K., Taylor, S. E., Shirinyan, D., Lieberman, M. D., & Naliboff, B. D. (2011). Attachment figures activate a safety signal-related neural region and reduce pain experience. Proceedings of the National Academy of Sciences U S A, 108(28), 11721–11726. doi: 10.1073/pnas.1108239108 Elam, K. K., Harold, G. T., Neiderhiser, J. M., Reiss, D., Shaw, D. S., Natsuaki, M. N., . . . Leve, L. D. (2014). Adoptive parent hostility and children’s peer behavior problems: Examining the role of genetically informed child attributes on adoptive parent behavior. Developmental Psychology, 50(5), 1543–1552. doi: 10.1037/a0035470 Fearon, R. M. P., Reiss, D., Leve, L. D., Shaw, D., Ganiban, J., & Neiderhiser, J. (in press). Child-evoked maternal negativity from 9 to 27 months: evidence of gene–environment correlation and its moderation by marital distress. Development and Psychopathology. Feinberg, M. E., & Hetherington, E. (2000). Sibling differentiation in adolescence: Implications for behavioral genetic theory. Child Development, 71(6), 1512–1524. doi: 10.1111/1467–8624.00243 Feldman, R., Zagoory-Sharon, O., Weisman, O., Schneiderman, I., Gordon, I., Maoz, R., . . . Ebstein, R. P. (2012). Sensitive parenting is associated with plasma oxytocin and polymorphisms in the OXTR and CD38 genes. Biological Psychiatry, 72(3), 175–181. doi: 10.1016/j.biopsych.2011.12.025 Fergusson, D. M., Boden, J. M., Horwood, L. J., Miller, A. L., & Kennedy, M. A. (2011). MAOA, abuse exposure and antisocial behaviour: 30-year longitudinal study. British Journal of Psychiatry, 198(6), 457–463. doi: 10.1192/bjp.bp. 110.086991 Fergusson, D. M., Horwood, L. J., Miller, A. L., & Kennedy, M. A. (2011). Life stress, 5-HTTLPR and mental disorder: findings from a 30-year longitudinal study. British Journal of Psychiatry, 198(2), 129–135. doi: 10.1192/bjp.bp. 110.085993

Flint, J., & Munafo, M. R. (2013). Candidate and non-candidate genes in behavior genetics. Current Opinion in Neurobiology, 23(1), 57–61. doi: 10.1016/j.conb.2012.07.005 Fortuna, K., van IJzendoorn, M. H., Mankuta, D., Kaitz, M., Avinun, R., Ebstein, R. P., & Knafo, A. (2011). Differential genetic susceptibility to child risk at birth in predicting observed maternal behavior. PLoS One, 6(5), e19765. doi: 10.1371/journal.pone.0019765 Galton, F. (1876). The history of twins as a criterion of the relative powers of nature and nurture. Journal of the Anthropologic Institute of Great Britain and Ireland, 5, 391–406. Ganiban, J. M., Saudino, K. J., Ulbricht, J., Neiderhiser, J. M., & Reiss, D. (2008). Stability and change in temperament during adolescence. Journal of Personality and Social Psychology, 95(1), 222–236. doi: 10.1037/0022–3514.95.1.222 Gaysina, D., Fergusson, D. M., Leve, L. D., Horwood, J., Reiss, D., Shaw, D. S., . . . Harold, G. T. (2013). Maternal smoking during pregnancy and offspring conduct problems: Evidence from 3 independent genetically sensitive research designs. JAMA Psychiatry. doi: 10.1001/jamapsychiatry.2013.127 Ge, X., Conger, R. D., Cadoret, R. J., Neiderhiser, J. M., Yates, W., Troughton, E., & Stewart, M. A. (1996). The developmental interface between nature and nurture: A mutual influence model of child antisocial behavior and parent behaviors. Developmental Psychology, 32(4), 574–589. Goodman, R., & Stevenson, J. (1989). A twin study of hyperactivity—II. The aetiological role of genes, family relationships and perinatal adversity. Journal of Child Psychology and Psychiatry, 30(5), 691–709. Gottesman, I. I. (2008). Milestones in the history of behavioral genetics: Participant observer. Acta Psychologica Sinica, 40(10), 1042–1050. doi: 10.3724/SP.J.1041.2008.01042 Grazioplene, R. G., Deyoung, C. G., Rogosch, F. A., & Cicchetti, D. (2013). A novel differential susceptibility gene: CHRNA4 and moderation of the effect of maltreatment on child personality. Journal of Child Psychology and Psychiatry, 54(8), 872–880. doi: 10.1111/ jcpp. 12031 Gross, H. E., Shaw, D. S., & Moilanen, K. L. (2008). Reciprocal associations between boys’ externalizing problems and mothers’ depressive symptoms. Journal of Abnormal Child Psychology, 36(5), 1–17. Guimond, F. A., Brendgen, M., Forget-Dubois, N., Dionne, G., Vitaro, F., Tremblay, R. E., & Boivin, M. (2011). Associations of mother’s and father’s parenting practices with children’s observed social reticence in a competitive situation: a monozygotic twin difference study. Journal of Abnormal Child Psychology, 40(3), 391–402. doi: 10.1007/s10802–011–9573–8 Haase, C. M., Saslow, L. R., Bloch, L., Saturn, S. R., Casey, J. J., Seider, B. H., . . . Levenson, R. W. (2013). The 5-HTTLPR polymorphism in the serotonin transporter gene moderates the association between emotional behavior and changes in marital satisfaction over time. Emotion, 13(6), 1068–1079. doi: 10.1037/a0033761 Hariri, A. R., Mattay, V. S., Tessitore, A., Kolachana, B., Fera, F., Goldman, D., . . . Weinberger, D. R. (2002). Serotonin transporter genetic variation and the response of the human amygdala. Science, 297(5580), 400–403. doi: 10.1126/science.1071829 Harold, G. T., Leve, L. D., Barrett, D., Elam, K., Neiderhiser, J. M., Natsuaki, M. N., . . . Thapar, A. (2013). Biological and rearing mother influences on child ADHD symptoms: revisiting the developmental interface between nature and nurture. Journal of Child Psychology and Psychiatry, 54(10), 1038–1046. doi: 10.1111/jcpp. 12100 Hayden, E. P., Hanna, B., Sheikh, H. I., Laptook, R. S., Kim, J., Singh, S. M., & Klein, D. N. (2013). Child dopamine active transporter 1 genotype and parenting: Evidence for evocative gene–environment

References correlations. Development and Psychopathology, 25(1), 163–173. doi: 10.1017/S0954579412000971 Hayden, E. P., Klein, D. N., Dougherty, L. R., Olino, T. M., Laptook, R. S., Dyson, M. W., . . . Singh, S. M. (2010). The dopamine D2 receptor gene and depressive and anxious symptoms in childhood: Associations and evidence for gene–environment correlation and gene–environment interaction. Psychiatric Genetics, 20(6), 304–310. doi: 10.1097/YPG.0b013e32833adccb Hayden, E. P., Klein, D. N., Dougherty, L. R., Olino, T. M., Laptook, R. S., Dyson, M. W., . . . Singh, S. M. (2010). The dopamine D2 receptor gene and depressive and anxious symptoms in childhood: Associations and evidence for gene–environment correlation and gene–environment interaction. Psychiatric Genetics, 20(6), 304–310. doi: 10.1097/YPG.0b013e32833adccb Heston, L. L. (1966). Psychiatric disorders in foster home reared children of schizophrenic mothers. British Journal of Psychiatry, 112(489), 819–825. doi: 10.1192/bjp. 112.489.819 Heston, L. L., & Denney, D. (1968). Interaction between early life experience and biological factors in schizophrenia. In D. Rosenthal & S. S. Kety (Eds.), The transmission of schizophrenia. Oxford, UK: Pergamon Press. Hornik, R., Risenhoover, N., & Gunnar, M. (1987). The effects of maternal positive, neutral, and negative affective communications on infant responses to new toys. Child Development, 937–944. Hostinar, C. E., Cicchetti, D., & Rogosch, F. A. (2014). Oxytocin receptor gene polymorphism, perceived social support, and psychological symptoms in maltreated adolescents. Development and Psychopathology, 26(2), 465–477. doi: 10.1017/S0954579414000066 Howe, G. W., & Reiss, D. (1993). Simulation and experimentation in family research. In, Pauline Boss, Sourcebook of family theories and methods (pp. 303–321). New York, NY: Springer. Howe, G. W., Reiss, D., & Yuh, J. (2002). Can prevention trials test theories of etiology? Development and Psychopathology, 14(4), 673–694. doi: 10.1017/S0954579402004029 Hunter, G. K. (2000). Vital forces: The discovery of the molecular basis of life. London, UK: Academic Press. Ivorra, J. L., Sanjuan, J., Jover, M., Carot, J. M., de Frutos, R., & Molto, M. D. (2010). Gene–environment interaction of child temperament. Journal of Developmental and Behavioral Pediatrics, 31(7), 545–554. doi: 10.1097/dbp. 0b013e3181ee4072 Iyegbe, C., Campbell, D., Butler, A., Ajnakina, O., & Sham, P. (2014). The emerging molecular architecture of schizophrenia, polygenic risk scores and the clinical implications for GxE research. Social Psychiatry and Psychiatric Epidemiology, 49(2), 169–182. doi: 10.1007/s00127–014–0823–2 Jackson, D. (1960). A critique of the literature on the genetics of schizophrenia. In D. Jackson (Ed.), The etiology of schizophrenia (pp. 37–90). New York, NY: Basic Books. Jaffee, S. R., Caspi, A., Moffitt, T. E., Polo-Tomas, M., Price, T. S., & Taylor, A. (2004). The limits of child effects: Evidence for genetically mediated child effects on corporal punishment but not on physical maltreatment. Developmental Psychology, 40(6), 1047–1058. doi: 10.1037/0012–1649.40.6.1047 Jaffee, S. R., & Price, T. S. (2007). Gene–environment correlations: A review of the evidence and implications for prevention of mental illness. Molecular Psychiatry, 12(5), 432–442. doi: 10.1038/sj.mp .4001950 Jerskey, B. A., Panizzon, M. S., Jacobson, K. C., Neale, M. C., Grant, M. D., Schultz, M., . . . Lyons, M. J. (2010). Marriage and divorce: A genetic perspective. Personality and Individual Differences, 49(5), 473–478. doi: 10.1016/j.paid.2010.05.007

373

Jocklin, V., McGue, M., & Lykken, D. T. (1996). Personality and divorce: A genetic analysis. Journal of Personality & Social Psychology, 71(2), 288–299. Johnson, W., McGue, M., Krueger, R. F., & Bouchard, T. J., Jr., (2004). Marriage and personality: A genetic analysis. Journal of Personality and Social Psychology, 86(2), 285–294. doi: 10.1037/0022–3514.86.2.285 Joseph, J. (1998). The equal environment assumption of the classical twin method: A critical analysis. Journal of Mind & Behavior, 19(3), 325–358. Joseph, J. (2001). Don Jackson’s “a critique of the literature on the genetics of schizophrenia”: A reappraisal after 40 years. Genetic, Social, and General Psychology Monographs, 127, 27–57. Kaitz, M., Shalev, I., Sapir, N., Devor, N., Samet, Y., Mankuta, D., & Ebstein, R. P. (2010). Mothers’ dopamine receptor polymorphism modulates the relation between infant fussiness and sensitive parenting. Developmental Psychobiology, 52(2), 149–157. doi: 10.1002/dev.20423 Kallmann, F. J. (1946). The genetic theory of schizophrenia; an analysis of 691 schizophrenic twin index families. American Journal of Psychiatry, 103, 309–322. Karg, K., Burmeister, M., Shedden, K., & Sen, S. (2011). The serotonin transporter promoter variant (5-HTTLPR), stress, and depression meta-analysis revisited: Evidence of genetic moderation. Archives of General Psychiatry, 68(5), 444–454. doi: 10.1001/archgenpsychiatry.2010.189 Keller, M. C. (2014). Gene × environment interaction studies have not properly controlled for potential confounders: the problem and the (simple) solution. Biological Psychiatry, 75(1), 18–24. doi: 10.1016/j.biopsych.2013.09.006 Kendler, K., & Neale, M. (2010). Endophenotype: A conceptual analysis. Molecular Psychiatry, 15(8), 789–797. doi: 10.1038/mp. 2010.8 Kendler, K. S. (2001). Twin studies of psychiatric illness: an update. Archives of General Psychiatry, 58(11), 1005–1014. doi:10.1001/archpsyc.58.11.1005. Kendler, K. S., & Baker, J. H. (2007). Genetic influences on measures of the environment: A systematic review. Psychological Medicine, 37(5), 615–626. doi: 10.1017/S0033291706009524 Kendler, K. S., & Eaves, L. J. (1986). Models for the joint effect of genotype and environment on liability to psychiatric illness. American Journal of Psychiatry, 143(3), 279–289. Kendler, K. S., Kessler, R. C., Walters, E. E., MacLean, C., & et al. (1995). Stressful life events, genetic liability, and onset of an episode of major depression in women. American Journal of Psychiatry, 152(6), 833–842. Kessler, R. C., McLaughlin, K. A., Green, J. G., Gruber, M. J., Sampson, N. A., Zaslavsky, A. M., . . . Williams, D. R. (2010). Childhood adversities and adult psychopathology in the WHO World Mental Health Surveys. British Journal of Psychiatry, 197(5), 378–385. doi: 10.1192/bjp.bp. 110.080499 Kety, S. S., Rosenthal, D., Wender, P. H., & Schulsinger, F. (1968). The types and prevalence of mental illness in the biological and adoptive families of adopted schizophrenics. Journal of Psychiatric Research, 6, 345–362. doi: 10.1016/0022–3956(68)90026–5 Kim-Cohen, J., Caspi, A., Taylor, A., Williams, B., Newcombe, R., Craig, I., & Moffitt, T. (2006). MAOA, maltreatment, and gene– environment interaction predicting children’s mental health: New evidence and a meta-analysis. Molecular Psychiatry, 11(10), 903–913. doi: 10.1038/sj.mp. 4001851 King, S. M., Keyes, M., Malone, S. M., Elkins, I., Legrand, L. N., Iacono, W. G., & McGue, M. (2009). Parental alcohol dependence

374

Genetics and Family Systems

and the transmission of adolescent behavioral disinhibition: A study of adoptive and non-adoptive families. Addiction, 104(4), 578–586. doi: 10.1111/j.1360–0443.2008.02469.x Klahr, A. M., & Burt, S. A. (2014). Elucidating the etiology of individual differences in parenting: A meta-analysis of behavioral genetic research. Psychological Bulletin, 140(2), 544–586. doi: 10.1037/ a0034205 Knafo, A., & Plomin, R. (2006). Parental discipline and affection and children’s prosocial behavior: Genetic and environmental links. Journal of Personality and Social Psychology, 90(1), 147–164. doi: 10.1037/ 0022–3514.90.1.147 Kochanska, G., Kim, S., Barry, R. A., & Philibert, R. A. (2011). Children’s genotypes interact with maternal responsive care in predicting children’s competence: Diathesis–stress or differential susceptibility? Development and Psychopathology, 23(2), 605–616. doi: 10.1017/ S0954579411000071 Kochanska, G., Philibert, R. A., & Barry, R. A. (2009). Interplay of genes and early mother–child relationship in the development of self-regulation from toddler to preschool age. Journal of Child Psychology and Psychiatry, 50(11), 1331–1338. doi: 10.1111/j.1469 –7610.2008.02050.x Larsson, H., Viding, E., Rijsdijk, F. V., & Plomin, R. (2008). Relationships between parental negativity and childhood antisocial behavior over time: A bidirectional effects model in a longitudinal genetically informative design. Journal of Abnormal Child Psychology, 36(5), 633–645. doi: 10.1007/s10802–007–9151–2 Larsson, J.-O., Larsson, H., & Lichtenstein, P. (2004). Genetic and environmental contributions to stability and change of ADHD symptoms between 8 and 13 years of age: A longitudinal twin study. Journal of the American Academy of Child & Adolescent Psychiatry, 43(10), 1267–1275. doi: 10.1097/01.chi.0000135622.05219.bf Laub, J. H., Nagin, D. S., & Sampson, R. J. (1998). Trajectories of change in criminal offending: Good marriages and the desistance process. American Sociological Review, 63(2), 225–238. doi: http://www .jstor.org/stable/2657324 Lee, S., Chronis-Tuscano, A., Keenan, K., Pelham, W., Loney, J., Van Hulle, C., . . . Lahey, B. (2010). Association of maternal dopamine transporter genotype with negative parenting: Evidence for gene × environment interaction with child disruptive behavior. Molecular Psychiatry, 15(5), 548–558. doi: 10.1038/mp. 2008.102 Leve, L. D., Harold, G. T., Ge, X., Neiderhiser, J. M., Shaw, D., Scaramella, L. V., & Reiss, D. (2009). Structured parenting of toddlers at high versus low genetic risk: two pathways to child problems. Journal of the American Academy of Child & Adolescent Psychiatry, 48(11), 1102–1109. doi: 10.1097/CHI.0b013e3181b8bfc0 Lichtenstein, P., Gatz, M., & Berg, S. (1998). A twin study of mortality after spousal bereavement. Psychological Medicine, 28(3), 635–643. Loehlin, J. C., Neiderhiser, J. M., & Reiss, D. (2005). Genetic and environmental components of adolescent adjustment and parental behavior: A multivariate analysis. Child Development, 76(5), 1104–1115. doi: 10.1111/j.1467–8624.2005.00900.x Loehlin, J. C., & Nichols, R. C. (1976). Heredity, environment, & personality: A study of 850 sets of twins. Austin, TX: University of Texas Press. Lykken, D. T. (2002). How relationships begin and end: A genetic perspective In A. L. Vangelisti, H. T. Reis, Mary Anne Fitzpatrick (Eds.), Stability and change in relationships Advances in personal relationships (pp. 83–102). New York, NY: Cambridge University Press. Maestripieri, D., Lindell, S. G., & Higley, J. (2007). Intergenerational transmission of maternal behavior in rhesus macaques and its underlying mechanisms. Developmental Psychobiology, 49(2), 165–171. doi: 10.1002/dev.20200

Mak, K. K., Kong, W. Y., Mak, A., Sharma, V. K., & Ho, R. C. (2012). Polymorphisms of the serotonin transporter gene and post-stroke depression: A meta-analysis. Journal of Neurology, Neurosurgery & Psychiatry, 84(3), 322–328. doi: 10.1136/jnnp-2012–303791 Marceau, K., Horwitz, B. N., Narusyte, J., Ganiban, J. M., Spotts, E. L., Reiss, D., & Neiderhiser, J. M. (2013). Gene–environment correlation underlying the association between parental negativity and adolescent externalizing problems. Child Development, 84(6), 2031–2046. doi: 10.1111/cdev.12094 Marmorstein, N. R., Iacono, W. G., & McGue, M. (2012). Associations between substance use disorders and major depression in parents and late adolescent-emerging adult offspring: An adoption study. Addiction, 107(11), 1965–1973. doi: 10.1111/j.1360–0443.2012 .03934.x Masarik, A. S., Conger, R. D., Donnellan, M. B., Stallings, M. C., Martin, M. J., Schofield, T. J., . . . Widaman, K. F. (2014). For better and for worse: genes and parenting interact to predict future behavior in romantic relationships. Journal of Family Psychology, 28(3), 357–367. doi: 10.1037/a0036818 McAdams, T. A., Neiderhiser, J. M., Rijsdijk, F. V., Narusyte, J., Lichtenstein, P., & Eley, T. C. (2014). Accounting for genetic and environmental confounds in associations between parent and child characteristics: A systematic review of children-of-twins studies. Psychological Bulletin, 140(4), 1138–1173. doi: 10.1037/a0036416 McCrory, E., De Brito, S. A., & Viding, E. (2012). The link between child abuse and psychopathology: A review of neurobiological and genetic research. Journal of the Royal Society of Medicine, 105(4), 151–156. doi: 10.1258/jrsm.2011.110222 McGue, M. (2010). The end of behavioral genetics? Behavior Genetics, 40, 284–296. McGue, M., Keyes, M., Sharma, A., Elkins, I., Legrand, L., Johnson, W., & Iacono, W. G. (2007). The environments of adopted and non-adopted youth: Evidence on range restriction from the sibling interaction and behavior study (SIBS). Behavior Genetics, 37(3), 449–462. doi: 10.1007/s10519–007–9142–7 McGue, M., & Lykken, D. T. (1992). Genetic influence on risk of divorce. Psychological Science, 3(6), 368–373. doi: 10.1111/j.1467 –9280.1992.tb00049.x McGue, M., & Sharma, A. (1995). Parent and sibling influences on adolescent alcohol use and misuse: Evidence from a U.S. adoption cohort. Journal of Studies on Alcohol, 57(1), 8–18. Mendel, G. (1948 (1866)). Experiments in plant hybridization. Cambridge, MA: Harvard University Press. Merriman, C. (1924). The intellectual resemblance of twins. Psychological Monographs: General and Applied, 33(5), i–57. Mileva-Seitz, V., Kennedy, J., Atkinson, L., Steiner, M., Levitan, R., Matthews, S. G., . . . Fleming, A. S. (2011). Serotonin transporter allelic variation in mothers predicts maternal sensitivity, behavior and attitudes toward 6-month-old infants. Genes Brain and Behavior, 10(3), 325–333. doi: 10.1111/j.1601–183X.2010.00671.x Mileva-Seitz, V., Steiner, M., Atkinson, L., Meaney, M. J., Levitan, R., Kennedy, J. L., . . . Fleming, A. S. (2013). Interaction between oxytocin genotypes and early experience predicts quality of mothering and postpartum mood. PLoS One, 8(4), e61443. doi: 10.1371/journal.pone.0061443 Miller, G. E., Chen, E., & Parker, K. J. (2011). Psychological stress in childhood and susceptibility to the chronic diseases of aging: Moving toward a model of behavioral and biological mechanisms. Psychological Bulletin, 137(6), 959–997. doi: 10.1037/a0024768 Mills-Koonce, W. R., Propper, C. B., Gariepy, J.-L., Blair, C., Garrett-Peters, P., & Cox, M. J. (2007). Bidirectional genetic and environmental influences on mother and child behavior: The family

References system as the unit of analyses. Development and Psychopathology, 19(4), 1073–1087. doi: 10.1017/S0954579407000545 Mishler, E. G., & Waxler, N. E. (1968). Interaction in families: An experimental study of family process and schizophrenia. New York, NY: John Wiley. Mitchell, C., Notterman, D., Brooks-Gunn, J., Hobcraft, J., Garfinkel, I., Jaeger, K., . . . McLanahan, S. (2011). Role of mother’s genes and environment in postpartum depression. PNAS Proceedings of the National Academy of Sciences of the United States of America, 108(20), 8189–8193. doi: 10.1073/pnas.1014129108 Moons, W. G., Way, B. M., & Taylor, S. E. (2014). Oxytocin and vasopressin receptor polymorphisms interact with circulating neuropeptides to predict human emotional reactions to stress. Emotion, 14(3), 562–572. doi: 10.1037/a0035503 Munafo, M. R., Durrant, C., Lewis, G., & Flint, J. (2009). Gene × environment interactions at the serotonin transporter locus. Biological Psychiatry, 65(3), 211–219. doi: 10.1016/j.biopsych.2008.06.009 Murphy, S. E., Norbury, R., Godlewska, B. R., Cowen, P. J., Mannie, Z. M., Harmer, C. J., & Munafo, M. R. (2013). The effect of the serotonin transporter polymorphism (5-HTTLPR) on amygdala function: A meta-analysis. Molecular Psychiatry, 18(4), 512–520. doi: 10.1038/mp. 2012.19 Narusyte, J., Neiderhiser, J. M., Andershed, A. K., D’Onofrio, B. M., Reiss, D., Spotts, E., . . . Lichtenstein, P. (2011). Parental criticism and externalizing behavior problems in adolescents: The role of environment and genotype–environment correlation. Journal of Abnormal Psychology, 120, 365–376. doi: 10.1037/a0021815 Natsuaki, M. N., Ge, X., Reiss, D., & Neiderhiser, J. M. (2009). Aggressive behavior between siblings and the development of externalizing problems: Evidence from a genetically sensitive study. Developmental Psychology, 45(4), 1009–1018. doi: 10.1037/a0015698 Neiderhiser, J. M., Marceau, K., & Reiss, D. (2013). Four factors for the initiation of substance use by young adulthood: A 10-year follow-up twin and sibling study of marital conflict, monitoring, siblings, and peers. Development and Psychopathology, 25(1), 133–149. doi: 10.1017/s0954579412000958 Neiderhiser, J. M., Reiss, D., Hetherington, E. M., & Plomin, R. (1999). Relationships between parenting and adolescent adjustment over time: Genetic and environmental contributions. Developmental Psychology, 35(3), 680–692. doi.apa.org/journals/dev/35/3/680.pdf Pashler, H., & Wagenmakers, E. J. (2012). Editors’ introduction to the special section on replicability in psychological science: A crisis of confidence? Perspectives on Psychological Science, 7(6), 528–530. doi: 10.1177/1745691612465253 Patterson, G. (1982). Coercive family process: A social learning approach. Eugene, OR: Castalia. Pedersen, N. L., Plomin, R., McClearn, G. E., & Friberg, L. (1988). Neuroticism, extraversion, and related traits in adult twins reared apart and reared together. Journal of Personality and Social Psychology, 55, 95–957. Peltola, M. J., Yrttiaho, S., Puura, K., Proverbio, A. M., Mononen, N., Lehtimaki, T., & Leppanen, J. M. (2014). Motherhood and oxytocin receptor genetic variation are associated with selective changes in electrocortical responses to infant facial expressions. Emotion, 14(3), 469–477. doi: 10.1037/a0035959 Petrill, S. A., & Deater-Deckard, K. (2004). Task orientation, parental warmth and SES account for a significant proportion of the shared environmental variance in general cognitive ability in early childhood: Evidence from a twin study. Developmental Science, 7(1), 25–32. doi: 10.1111/j.1467–7687.2004.00319.x Pike, A., Reiss, D., Hetherington, E., & Plomin, R. (1996). Using MZ differences in the search for nonshared environmental effects.

375

Journal of Child Psychology & Psychiatry & Allied Disciplines, 37(6), 695–704. Plomin, R. (2011). Commentary: Why are children in the same family so different? Non-shared environment three decades later. International Journal of Epidemiology, 40(3), 582–592. doi: 10.1093/ije/dyq144 Plomin, R. (2013). Commentary: missing heritability, polygenic scores, and gene–environment correlation. Journal of Child Psychology and Psychiatry, 54(10), 1147–1149. doi: 10.1111/jcpp. 12128 Plomin, R., & Daniels, D. (1987). Why are children in the same family so different from one another? Behavioral and Brain Sciences, 10, 1–16. Plomin, R., Owen, M. J., & McGuffin, P. (1994). The genetic basis of complex human behaviors. Science, 264(5166), 1733–1739. Plomin, R., Pedersen, N. L., McClearn, G. E., Nesselroade, J. R., & Bergeman, C. S. (1988). EAS temperaments during the last half of the life span: Twins reared apart and twins reared together. Psychology and Aging, 3(1), 43–50. doi: 10.1037/0882–7974.3.1.43 Pluess, M., Velders, F. P., Belsky, J., van, I. M. H., BakermansKranenburg, M. J., Jaddoe, V. W., . . . Tiemeier, H. (2011). Serotonin transporter polymorphism moderates effects of prenatal maternal anxiety on infant negative emotionality. Biological Psychiatry, 69(6), 520–525. doi:10.1016/j.biopsych.2010.10.006 Propper, C. B., Shanahan, M. J., Russo, R., & Mills-Koonce, W. R. (2012). Evocative gene-parenting correlations and academic performance at first grade: An exploratory study. Development and Psychopathology, 24(4), 1265–1282. doi: 10.1017/S0954579412000697 Rabinow, P. (1996). Making PCR. A story of biotechnology. Chicago, IL: University of Chicago Press. Reiss, D. (2010). Genetic thinking in the study of social relationships. Perspectives on Psychological Science, 5(5), 502–515. doi: 10.1177/1745691610383516 Reiss, D., & Leve, L. D. (2007). Genetic expression outside the skin: Clues to mechanisms of genotype × environment interaction. Development and Psychopathology, 19(4), 1005–1027. doi: 10.1017/S0954579407000508 Reiss, D., Leve, L. D., & Neiderhiser, J. M. (2013). How genes and the social environment moderate each other. American Journal of Public Health, 103 Suppl 1, S111–121. doi: 10.2105/ajph.2013.301408 Reiss, D., Neiderhiser, J., Hetherington, E. M., & Plomin, R. (2000). The relationship code: Deciphering genetic and social patterns in adolescent development. Cambridge, MA: Harvard University Press. Reiss, D., & Oliveri, M. E. (1983). Sensory experience and family process: Perceptual styles tend to run in but not necessarily run families. Family Process, 22(3), 289–308. doi: 10.1111/j.1545–5300.1983.00289.x Rende, R. D., Plomin, R., & Vandenberg, S. G. (1990). Who discovered the twin method? Behavioral Genetics, 20(2), 277–285. Rice, F., Harold, G. T., Boivin, J., Hay, D. F., van den Bree, M., & Thapar, A. (2009). Disentangling prenatal and inherited influences in humans with an experimental design. PNAS Proceedings of the National Academy of Sciences of the United States of America, 106(7), 2464–2467. doi: 10.1073/pnas.0808798106 Risch, N., Herrell, R., Lehner, T., Liang, K.-Y., Eaves, L., Hoh, J., . . . Merikangas, K. R. (2009). Interaction between the serotonin transporter gene (5-HTTLPR), stressful life events, and risk of depression: A meta-analysis. JAMA 301, 2462–2471 doi: 10.1001/jama.2009.878 Rosenthal, D., Wender, P. H., Kety, S. S., Schulsinger, F., Welner, J., & Ostergaard, L. (1968). Schizophrenics’ offspring reared in adoptive homes. In D. Rosenthal & S. S. Kety (Eds.), The transmission of schizophrenia. Oxford, UK: Pergamon Press. Rowe, D. C. (1981). Environmental and genetic influences on dimensions of perceived parenting: A twin study. Developmental Psychology, 17(2), 203–208.

376

Genetics and Family Systems

Genetic and environmental influences on vocabulary IQ: parental education level as moderator. (1999). Rowe, D. C., Jacobson, K. C., & Van den Oord, E. J. Child Development, 70(5), 1151–1162. Rutherford, H. J., Goldberg, B., Luyten, P., Bridgett, D. J., & Mayes, L. C. (2013). Parental reflective functioning is associated with tolerance of infant distress but not general distress: Evidence for a specific relationship using a simulated baby paradigm. Infant Behavior & Development, 36(4), 635–641. doi: 10.1016/j.infbeh.2013.06.008

Spotts, E. L., Prescott, C., & Kendler, K. (2006). Examining the origins of gender differences in marital quality: A behavior genetic analysis. Journal of Family Psychology, 20(4), 605–613. doi: 10.1037/ 0893–3200.20.4.605 Stoolmiller, M. (1999). Implications of the restricted range of family environments for estimates of heritability and nonshared environment in behavior-genetic adoption studies. Psychological Bulletin, 125(4), 392–409. doi: 10.1037/0033–2909.125.4.392

Rutter, M., & O’Connor, T. G. (2004). Are there biological programming effects for psychological development? Findings from a study of romanian adoptees. Developmental Psychology, 40(1), 81–94. doi: 10.1037/0012–1649.40.1.81

Sturge-Apple, M. L., Cicchetti, D., Davies, P. T., & Suor, J. H. (2012). Differential susceptibility in spillover between interparental conflict and maternal parenting practices: Evidence for OXTR and 5-HTT genes. Journal of Family Psychology, 26(3), 431–442. doi: 10.1037/a0028302

Rutter, M., & Quinton, D. (1984). Long-term follow-up of women institutionalized in childhood: Factors promoting good functioning in adult life. British Journal of Developmental Psychology, 2(3), 191–204. doi: 10.1111/j.2044–835X.1984.tb00925.x

Sulik, M. J., Eisenberg, N., Lemery-Chalfant, K., Spinrad, T. L., Silva, K. M., Eggum, N. D., . . . Verrelli, B. C. (2012). Interactions between serotonin transporter gene haplotypes and quality of mothers’ parenting predict the development of children’s noncompliance. Developmental Psychology, 48(3), 740–754. doi: 10.1037/a0025938

Scarr, S., Scarf, E., & Weinberg, R. A. (1980). Perceived and actual similarities in biological and adoptive families: Does perceived similarity bias genetic inferences? Behavior Genetics, 10(5), 445–457. Scarr, S., & Weinberg, R. A. (1978). The influence of “family background” on intellectual attainment. American Sociological Review, 43, 674–692. Schachter, F. F., Gilutz, G., Shore, E., & Adler, M. (1978). Sibling deidentification judged by mothers: Cross-validation and developmental studies. Child Development, 49, 543–546. Schoebi, D., Way, B. M., Karney, B. R., & Bradbury, T. N. (2012). Genetic moderation of sensitivity to positive and negative affect in marriage. Emotion, 12, 208–212. doi: 10.1037/a0026067 Siemens, H. W. (1927). The diagnosis of identity in twins. Journal of Heredity, 18(5), 201–209. Simons, R. L., Simons, L. G., Lei, M.-K., Beach, S. R. H., Brody, G. H., Gibbons, F. X., & Philibert, R. A. (2013). Genetic moderation of the impact of parenting on hostility toward romantic partners. Journal of Marriage and Family, 75(2), 325–341. doi: 10.1111/jomf.12010 Singer, M. T., & Wynne, L. C. (1965). Thought disorder and family relations of schizophrenics: IV Results and implications. Archives of General Psychiatry, 12, 201–212. Slomkowski, C., Rende, R., Novak, S., Lloyd-Richardson, E., & Niaura, R. (2005). Sibling effects on smoking in adolescence: Evidence for social influence from a genetically informative design. Addiction, 100, 430–438. doi: 10.1111/j.1360–0443.2004.00965.x

Tandon, R., Keshavan, M. S., & Nasrallah, H. A. (2008). Schizophrenia, “just the facts” what we know in 2008. 2. Epidemiology and etiology. Schizophrenia Research, 102(1–3), 1–18. doi: 10.1016/j.schres .2008.04.011 Tarantino, N., Tully, E. C., Garcia, S. E., South, S., Iacono, W. G., & McGue, M. (2014). Genetic and environmental influences on affiliation with deviant peers during adolescence and early adulthood. Developmental Psychology, 50(3), 663–673. doi: 10.1037/a0034345 Todorov, A., Fiske, S., & Prentice, D. (Eds.). (2011). Social neuroscience: Toward understanding the underpinnings of the social mind. Oxford, UK: Oxford University Press. Trumbetta, S. L., Markowitz, E. M., & Gottesman, I. I. (2007). Marriage and genetic variation across the lifespan: Not a steady relationship? Behavior Genetics, 37(2), 362–375. doi: 10.1007/s10519–006–9132–1 Tucker-Drob, E. M., & Briley, D. A. (2014). Continuity of genetic and environmental influences on cognition across the life span: A meta-analysis of longitudinal twin and adoption studies. Psychological Bulletin, 140(4), 949–979. Tucker-Drob, E. M., & Harden, K. (2012). Early childhood cognitive development and parental cognitive stimulation: Evidence for reciprocal gene–environment transactions. Developmental Science, 15(2), 250–259. doi: 10.1111/j.1467–7687.2011.01121.x

South, S. C., & Krueger, R. F. (2008). Marital quality moderates genetic and environmental influences on the internalizing spectrum. Journal of Abnormal Psychology, 117(4), 826–837. doi: 10.1037/a0013499

Tully, E. C., Iacono, W. G., & McGue, M. (2008). An adoption study of parental depression as an environmental liability for adolescent depression and childhood disruptive disorders. American Journal of Psychiatry, 165(9), 1148–1154. doi: 10.1176/appi.ajp. 2008.07091438

Spotts, E. L., Lichtenstein, P., Pedersen, N., Neiderhiser, J. M., Hansson, K., Cederblad, M., & Reiss, D. (2005). Personality and marital satisfaction: A behavioural genetic analysis. European Journal of Personality, 19(3), 205–227. doi: 10.1002/per.545

Turkheimer, E., Haley, A., Waldron, M., D’Onofrio, B., & Gottesman, I. I. (2003). Socioeconomic status modifies heritability of IQ in young children. Psychological Science, 14(6), 623–628. doi: 10.1046/j.0956 –7976.2003.psci_1475.x

Spotts, E. L., Neiderhiser, J. M., Ganiban, J., Reiss, D., Lichtenstein, P., Hansson, K., . . . Pedersen, N. L. (2004). Accounting for depressive symptoms in women: A twin study of associations with interpersonal relationships. Journal of Affective Disorders, 82(1), 101–111. doi: 10.1016/j.jad.2003.10.005

van den Boom, D. C., & Hoeksma, J. B. (1994). The effect of infant irritability on mother–infant interaction: A growth-curve analysis. Developmental Psychology, 30(4), 581–590. doi: 10.1037/0012–1649 .30.4.581

Spotts, E. L., Neiderhiser, J. M., Towers, H., Hansson, K., Lichtenstein, P., Cederblad, M., . . . Reiss, D. (2004). Genetic and environmental influences on marital relationships. Journal of Family Psychology, 18(1), 107–119.

van den Hoofdakker, B. J., Nauta, M. H., Dijck-Brouwer, D. A., van der Veen-Mulders, L., Sytema, S., Emmelkamp, P. M., . . . Hoekstra, P. J. (2012). Dopamine transporter gene moderates response to behavioral parent training in children with ADHD: A pilot study. Developmental Psychology, 48(2), 567–574. doi: 10.1037/a0026564

Spotts, E. L., Pederson, N. L., Neiderhiser, J. M., Reiss, D., Lichtenstein, P., Hansson, K., & Cederblad, M. (2005). Genetic effects on women’s positive mental health: Do marital relationships and social support matter? Journal of Family Psychology, 19(3), 339–349. doi: 10.1037/ 0893–3200.19.3.339

van IJzendoorn, M. H., & Bakermans-Kranenburg, M. J. (2012). Differential susceptibility experiments: Going beyond correlational evidence: Comment on beyond mental health, differential susceptibility articles. Developmental Psychology, 48(3), 769–774. doi: 10.1037/a0027536

References van IJzendoorn, M. H., Belsky, J., & Bakermans-Kranenburg, M. J. (2012). Serotonin transporter genotype 5HTTLPR as a marker of differential susceptibility[quest] A meta-analysis of child and adolescent gene-by-environment studies. Translational Psychiatry, 2, e147. doi: 10.1038/tp. 2012.73 Vitaro, F., Brendgen, M., Boivin, M., Cantin, S., Dionne, G., Tremblay, R. E., . . . Perusse, D. (2011). A monozygotic twin difference study of friends’ aggression and children’s adjustment problems. Child Development, 82(2), 617–632. doi: 10.1111/j.1467–8624.2010.01570.x Wahlberg, K.-E., Wynne, L. C., Oja, H., Keskitalo, P. et al. (1997). Gene–environment interaction in vulnerability to schizophrenia: Findings from the Finnish family study of schizophrenia. American Journal of Psychiatry, 154(3), 355–362. Walden, B., McGue, M., lacono, W. G., Burt, S., & Elkins, I. (2004). Identifying shared environmental contributions to early substance use: The respective roles of peers and parents. Journal of Abnormal Psychology, 113(3), 440–450. doi: 10.1037/0021–843X.113.3.440 Waldman, I. D. (2007). Gene–environment interactions reexamined: Does mother’s marital stability interact with the dopamine receptor D2 gene in the etiology of childhood attention-deficit/hyperactivity disorder? Development and Psychopathology, 19(4), 1117–1128. doi: 10.1017/S0954579407000570 Walum, H., Lichtenstein, P., Neiderhiser, J. M., Reiss, D., Ganiban, J. M., Spotts, E. L., . . . Westberg, L. (2012). Variation in the oxytocin receptor gene is associated with pair-bonding and social behavior. Biological Psychiatry, 71(5), 419–426. doi: 10.1016/j.biopsych.2011.09.002 Walum, H., Westberg, L., Henningsson, S., Neiderhiser, J. M., Reiss, D., Igl, W., . . . Lichtenstein, P. (2008). Genetic variation in the vasopressin receptor 1a gene (AVPR1A) associates with pair-bonding behavior in humans. Proceedings of the National Academy of Sciences, 105(37), 14153–14156. doi: 10.1073/pnas.0803081105 Watson, J. D., Crick, F. H. C. (1952). Molecular structure of nuceliec acids. A structure for deoxyribose nucleic acid. Nature, 171(4356), 737–738. Way, B. M., & Taylor, S. E. (2010). The serotonin transporter promoter polymorphism is associated with cortisol response to psychosocial stress. Biological Psychiatry, 67(5), 487–492. doi: 10.1016/j .biopsych.2009.10.021 Way, B. M., & Taylor, S. E. (2011). A polymorphism in the serotonin transporter gene moderates cardiovascular reactivity to psychosocial

377

stress. Psychosomatic Medicine, 73(4), 310–317. doi: 10.1097/ PSY.0b013e31821195ed [doi] White, L. O., Wu, J., Borelli, J. L., Mayes, L. C., & Crowley, M. J. (2013). Play it again: neural responses to reunion with excluders predicted by attachment patterns. Developmental Science, 16(6), 850–863. Widom, C. S., & Brzustowicz, L. M. (2006). MAOA and the “cycle of violence:” Childhood abuse and neglect, MAOA genotype, and risk for violent and antisocial behavior. Biological Psychiatry, 60(7), 684–689. doi: 10.1016/j.biopsych.2006.03.039 Widom, C. S., Czaja, S. J., Bentley, T., & Johnson, M. S. (2012). A prospective investigation of physical health outcomes in abused and neglected children: new findings from a 30-year follow-up. American Journal of Public Health, 102(6), 1135–1144. doi: 10.2105/AJPH.2011.300636 Wilson, H. W., & Widom, C. S. (2011). Pathways from childhood abuse and neglect to HIV-risk sexual behavior in middle adulthood. Journal of Consulting and Clinical Psychology, 79(2), 236–246. doi: 10.1037/a0022915 Wynne, L. C. (1968). Methodological and conceptual issues in the study of schizophrenics and their families. In D. Rosenthal & S. S. Kety (Eds.), The transmission of schizophrenia, 185–200. Oxford, UK: Pergamon Press. Xie, B., Li, D., London, S., Palmer, P., Johnshon, C., Li, Y., . . . Swan, G. (2013). Gender difference in interactions between MAOA promoter uVNTR polymorphism and negative familial stressors on body mass index among Chinese adolescents. Pediatric Obesity, 9(5), e80–e90. doi: 10.1111/j.2047–6310.2013.00181.x Yang, J., Lee, S. H., Goddard, M. E., & Visscher, P. M. (2011). GCTA: A tool for genome-wide complex trait analysis. American Journal of Human Genetics, 88(1), 76–82. doi: 10.1016/j.ajhg.2010.11.011 Zhao, J., Bremner, J. D., Goldberg, J., Quyyumi, A. A., & Vaccarino, V. (2013). Monoamine oxidase a genotype, childhood trauma, and subclinical atherosclerosis: A twin study. Psychosomatic Medicine, 75(5), 471–477. doi: 10.1097/PSY.0b013e318292922a Zimmermann, P., Bruckl, T., Nocon, A., Pfister, H., Binder, E. B., Uhr, M., . . . Ising, M. (2011). Interaction of FKBP5 gene variants and adverse life events in predicting depression onset: Results from a 10-year prospective community study. American Journal of Psychiatry, 168(10), 1107–1116. doi: 10.1176/appi.ajp. 2011.10111577

CHAPTER 10

Molecular Genetics Methods for Developmental Scientists SERGEY A. KORNILOV and ELENA L. GRIGORENKO

INTRODUCTION 378 BASIC CONCEPTS 379 MOLECULAR GENETIC STUDY DESIGNS AND GENOTYPING 382 Linkage Studies and Short-Tandem Repeat Polymorphisms (STRPs) 382 Candidate Gene Association Studies and Single Nucleotide Polymorphisms (SNPs) 385 Genome-Wide Association Studies and High-Density SNP Microarray panels 392 Augmenting Genetic Association Studies Using Recent Statistical Developments: Mixed Models, Chip Heritabilities, Gene-Based Associations, and Meta-analytic Approaches 395

STRUCTURAL VARIATION IN THE HUMAN GENOME AND COPY NUMBER VARIANT (CNV) ASSOCIATION STUDIES 397 DNA SEQUENCING: FROM SANGER SEQUENCING TO HIGH-THROUGHPUT NEXT-GENERATION SEQUENCING TECHNOLOGIES 401 GENE EXPRESSION PROFILING 405 FUTURE DIRECTIONS 407 CONCLUSION 407 REFERENCES 408

INTRODUCTION

publication and number of published reports indexed by SCOPUS over the same period can be described with a point estimate of Pearson’s r = .91 (p < .000001)—a large effect size rarely seen by developmental scientists in their own data! Whether or not this accelerating increase in the amount of generated and published data will result in a better understanding of the complex relationships between human genetics and development partly depends on the accessibility of these data (in terms of both acquisition and interpretation) to its main consumers, that is, to developmental scientists. In 2012, the US National Institutes of Health launched an online educational initiative for behavioral and social scientists called Genetics and Social Science: Expanding Transdisciplinary Research (http://www.nchpeg.org/bssr/), aimed at providing accessible coverage of the core concepts and issues in the field of genetics. Nevertheless, in their own teaching and collaboration experience, the authors of this chapter have been faced with the virtual absence of a concise, integrative, state-of-the-art overview of the basic methodological issues and advances in human genetics as applied to the study of the development and manifestation of complex traits and common disorders in general and neurodevelopmental disorders in particular. Most of the published treatments of these topics and methods are relatively isolated in

The progressive biologization of developmental science in general and the science of developmental psychopathology in particular in the past several decades has been tightly linked to technological, methodological, and statistical advances in developmental biology and genetic epidemiology. Predictably, this biologization has been accompanied by a rapid increase in the number of multidisciplinary projects conducted and published by the field. For example, at the time of the writing of this chapter, a quick SCOPUS search limited to the subject area of psychology using a combination of search terms “genes”/“molecular genetics” AND “development” returned almost 3,000 results, with 240 reports published in 2013 alone—a sixfold increase compared with the total number of publications with the same keyword 20 years ago. Even more impressive, perhaps, is that the linear relationship between year of The writing of this chapter was supported by grants R01 DC007665, R21 HD070594, and P50 HD052120 from the US National Institutes of Health and grant No 14.Z50.31.0027 from the Government of the Russian Federation. Grantees undertaking such projects are encouraged to express freely their professional judgment. The chapter, therefore, does not necessarily reflect the position or policies of the aforementioned funding agencies, and no official endorsement should be inferred. 378

Basic Concepts

their scope and highly technical, thereby reducing their utility for interdisciplinary researchers that frequently do not have formal training in molecular genetics, genetic epidemiology, bioinformatics, and biostatistics. Therefore, the current chapter aims to fill this gap in the literature by providing an accessible and comprehensive overview of molecular genetics methods that would hopefully aid the new generation of developmental scientists in both digesting the rapidly growing number of reports on molecular genetic studies in the field and potentially serving as a starting point for designing their own projects. This chapter is organized as follows. First, we will provide a brief overview of the basic concepts of molecular cell biology, thus setting the necessary background for presenting molecular genetics methods. Then, we will review several types of genetic variation and corresponding genotyping methods and research designs used in genetic linkage and association studies (with an emphasis on microarray genotyping and genome-wide association studies). We will then turn to the discussion of the current state of affairs with respect to the growing appreciation of the role of the structural variation in the human genome in human development and its consequences for the debate between those who attribute complex trait variation in humans to frequent or common vs. rare genetic variants. We will then turn to DNA sequencing methods, starting with the original Sanger sequencing and following with the next-generation DNA sequencing methods. After a brief discussion of gene expression profiling methods that builds on the information presented in previous sections (i.e., microarray and sequencing platforms), we will conclude with a set of comments pertaining to current and future utilization of molecular methods in the developmental sciences.

BASIC CONCEPTS The focus of this chapter is on the methods used for acquiring information about the genetic make up of an individual (or a group of individuals) and study designs that are aimed at gaining knowledge about the etiology of a particular trait (typical and atypical) or a syndromic trait constellation. Before describing these methods, however, we need to establish what exactly the substrate for genetic information is, where it is located, and how it is transmitted and used to build a living organism. To minimize the technicality of this overview, we therefore refer the interested reader to other sources for more in-depth treatments of the

379

relevant issues (Alberts et al., 2002; Laird & Lange, 2011; Russell, 2009; Wu, Ma, & Casella, 2007; Ziegler, König, & Pahlke, 2010). With the exception of red blood cells (erythrocytes) that lack cell nucleus, every human body cell contains a full set of information about the primary structure of the genetic makeup of the individual. This information is stored in thread-like structures made of deoxyribonucleic acid (DNA) and other proteins. DNA, the main carrier of genetic information, is a large two-stranded molecule composed of a backbone that is formed by an alternating repeated pattern of a sugar group (deoxyribose) and a phosphate group, and organic nucleobases (or, simply, bases). The primary bases of the DNA are purines (guanine [G] and adenine [A]) and pyrimidines (thymine [T] and cytosine [C]). A structural unit of a single strand of DNA consists of one sugar group, one phosphate group, and a base, and is called a nucleotide. Recall, however, that DNA is a double-stranded molecule. The two strands are connected by a hydrogen bond between two opposing bases on the complementary strands comprising a DNA molecule: adenine always bonds with thymine (A-T rule), and cytosine always bonds with guanine (C-G rule). The specificity of these pairings is regulated by the dimensional1 properties of the DNA bases and is frequently referred to as complementary base pairing rule since one can infer the sequence of one strand if the other one is known. The resulting DNA molecule resembles a twisted ladder described as a double helix. Human DNA is structured into 23 chromosomes that function in cells. Each chromosome has a centromere and two arms, a short one (p arm) and a long one (q arm), that end with telomeres. Of these, 22 are autosomes consecutively labeled from 1 to 22 roughly according to their length: chromosome 1 is the longest with a length of ∼248.9 million base pairs (bp); 21 is the shortest with ∼46.7 million bp (http://www.ncbi.nlm.nih.gov/projects/genome /assembly/grc/human/data/). The remaining twenty-third chromosome is either an X or a Y sex chromosome, around 156 and 57.2 million bp long, respectively. Human sex cells (gametes—egg and sperm cells) contain a single set of 23 1 Purines have a two-ring structure and are larger (wider) than pyrimidines that only have one ring. The formation of stable hydrogen bonds between DNA bases depends on the distance between strands, the size of the bases, and their geometrical properties. Correspondingly, purine-purine, pyrimidine-pyrimidine, and noncomplementary purine-pyrimidine pairings locally disrupt the DNA structure leading to DNA instability and are disfavored.

380

Molecular Genetics Methods for Developmental Scientists

chromosomes and are termed to be haploid (k = 23). Most human somatic cells, however, are diploid (k = 46): they contain a double set of chromosomes, with one inherited from the mother and other one—the father, for the full diploid set of 44 autosomes and 2 sex chromosomes. Thus, while a haploid set of human DNA is approximately 3 billion bp long, the complete diploid set of genetic information is twice this length. The genetic information necessary to make an entire organism2 is stored in the linear sequence of DNA nucleotides and is used to guide protein synthesis. Proteins consist of polypeptide chains of repeating amino acids. Twenty different amino acids found in human proteins are encoded by triplets of DNA nucleotides called codons (e.g., the GUU sequence codes for amino acid called valine, while CGU codes for arginine). The three-letter code can be used to generate 64 different codons, but since there are only 20 amino acids that need to be encoded, the genetic code itself is considered degenerate (i.e., more than one codon can code for a particular amino acid). According to what is frequently referred to as the central dogma of molecular biology, protein synthesis follows a particular path, the DNA → RNA3 → protein pathway. The first step in this pathway is called transcription. While the DNA is located in the cell nucleus, protein synthesis takes place in the plasma. Correspondingly, the first step requires transcription of DNA into a single-stranded messenger RNA molecule (mRNA) that carries the information to the plasma. During transcription, the DNA sequence is read by RNA polymerase (RNAP), an enzyme that is used to produce mRNA4 —a template that will be used for the actual protein synthesis during the second 2

Note that genetic information encoded in the DNA molecule is not only sufficient to make an entire organism, but is also transmittable (from parent to offspring), replicable (in order to be transmittable), and exhibits significant variation (which can be mapped to at least some facets of phenotypic variance within species). We will turn to these points later in the chapter. 3 The RNA (ribonucleic acid) molecule, like DNA, is composed of a chain of nucleotides. However, it is typically a less stable single-stranded molecule. Unlike DNA that contains thymine as a complementary base to adenine, the complementary base to adenine in RNA is not thymine, but rather uracil (U), an unmethylated form of thymine. 4 mRNA is not the only kind of the RNA molecules resulting from transcription and not the only one RNA molecule that participates in translation. Other RNAs (i.e., ribosomal rRNA and transfer tRNA) also partake in protein synthesis as supporting or regulating elements.

phase called translation. This second step takes place in ribosomes in the cell plasma, and includes decoding the transcription-generated RNA to produce an amino acid or polypeptide chain that is then modified (folded) into the actual active functional protein. The full sequence of DNA is called the genome. For a long time, the researchers’ attention was focused on the amino-acid coding regions of the genome which contain codon sequences that initiate and terminate transcription, and whose transcripts undergo translation to produce a functional protein product—genes. These coding genic regions are separated by stretches of non-coding intergenic DNA. A gene can be formally defined as “a locatable region of genomic sequence, corresponding to a unit of inheritance, which is associated with regulatory regions, transcribed regions, and/or other functional significance region” (Pearson, 2006, p. 401). This recent definition underscores the shift from coding regions per se to regions that have at least some functional significance. At the time of the writing of this chapter, the latest human genome build (i.e., a public release of an assembly of contiguous reference DNA sequences in a chromosomal order that is representative of the human genome; the latest build has been released in 2014 and is called the Genome Reference Consortium human genome build 38 or GRCh38; http://www.ensembl.org) included 20,389 coding genes, accounting for about 1% of the total human genome length. However, the results of the Encyclopedia of DNA Elements (ENCODE; http://www.genome.gov/10005107), a public research project focused on identifying all functional elements in the genome, suggested that ∼80% of the genome might have a biochemical function (The ENCODE Project Consortium, 2012), and that the distinction between genic and intergenic DNA regions is not clear-cut, but rather complicated by such phenomena as the existence of previously ignored transcription of regulatory elements and the presence of transcribed but non-protein-coding RNAs and pseudogenes (Gerstein et al., 2007; Mattick, 2004). The transmission of genetic information from parents to their offspring involves the transfer of the genetic information into parental gametes during meiosis, and the fusion of parental gametes during zygote formation. Since meiosis is a special form of cell division that includes only one round of the duplication of genetic material but two rounds of cell division, each diploid progenitor cell gives rise to four haploid gametes. Thus, at the beginning of this process, a diploid progenitor cell contains a double set of

Basic Concepts

genetic material with each of the pair homologous chromosomes coming from one of the parents. Homologous chromosomes are similar in length, centromere location, and gene composition and position but are not necessarily fully identical as one is coming from the mother, and the other one—the father. After the DNA is duplicated, the cell contains four sets of the genetic material, and homologous chromosomes form connected pairs called bivalents. At this stage, it is possible for maternal and paternal chromosomes to exchange portions of genetic material during a process known as crossing over. After two rounds of meiotic division, four haploid gametes are formed. Chromosomal crossing over results in the recombination of maternal and paternal alleles (i.e., genetic variants) on the same chromosome and increases genetic variability among gametes. Importantly, the probability of a crossing-over occurring between two genomic segments at specific locations (loci) depends on the distance between these segments (i.e., segments and genes located closer to each other have a lower probability of being separated during crossing over than segments located further apart from each other—in the extreme case, on different chromosomes). This results in the deviation of the observed frequencies of haplotypes (combinations of alleles) from their expected frequencies in the population, the phenomenon referred to as linkage disequilibrium (LD). Two loci are considered linked if they are jointly transmitted from parent to offspring more often than under independent inheritance. Generally, genetic segments and genes that are located close together on the same chromosome and undergo recombination in less than 50% of meiotic events are considered to be linked, whereas genes and segments that are located on different chromosomes or at different ends of the same chromosome with crossing over happening in approximately 50% of meiotic events are considered to be inherited independently. We will return to the importance of the idea of linkage and linkage disequilibrium for molecular genetic studies of human traits in the next section. The possibility of inheriting different alleles on the maternal and paternal DNA strands at the same genetic locus highlights the presence of genetic variability across individuals. In general, although different individuals can be said to have the same set of genes, they usually (with the exception of monozygotic twins) do not share the same exact sequence of DNA nucleotides. In diploid cells, the combination of alleles at a particular locus defines the genotype at that locus (e.g., if an individual has inherited the A allele from the mother and the G allele from the

381

father, they have an AG genotype at that locus). If the two alleles are the same on both chromosomes, an individual is said to be homozygous at this locus, whereas different alleles form a heterozygous genotype. Variation in the human genome is a consequence of different types of mutation events that alter the base sequence composition of the DNA. These include large events during cell division that lead to deletions, insertions, duplications, translocations, and inversions of relatively large chromosomal segments and smaller events that typically occur during DNA replication, such as point mutations (i.e., substitution of one nucleotide for another). Generally, a locus is considered polymorphic if the mutant (minor or less frequent) allele is present in the referent population (e.g., individuals of European ancestry) at the frequency of at least 1% while the referent wild-type (major or more frequent) allele is found with the complementary frequency. The corresponding variation in the DNA sequence at that locus is called a polymorphism. The human genome contains several types of genetic variation and corresponding polymorphisms. These include the single nucleotide polymorphisms mentioned above (SNPs; the most prevalent type of genetic variation, with 97,535,033 validated SNPs reported in the most recent build of the public SNP database dbSNP 144. (http://www .ncbi.nlm.nih.gov/projects/SNP/snp_summary.cgi), short tandem repeats (repetitions of a pattern of one or more nucleotides adjacent to one another otherwise known as micro- and mini-satellites), short insertions and deletions of genetic material (so-called indels that generally range in size from 1 bp to 1,000 bp or 1 kilobase pair, kb), and copy number variants (deletions, duplications, translocations, inversions, and insertions of genetic material that are typically over 1 kb long). Prior to the technological advances in molecular genetics, only large (generally over 3 million bp or 3 megabases, Mb, in size) and rare types of structural variations that affected quantity or structure of chromosomes were identified, as they were detectable using microscopic techniques. Over the course of the past two decades, however, a large number of smaller, frequent, or common (>1% frequency) variants has been discovered and described. We will now turn to the most commonly employed study designs in molecular genetics of human traits, methods used to genotype the main types of variants described already, and inferential (including statistical) procedures that underlie the inference about the involvement of a particular genetic variant or a gene in the etiology of human traits.

382

Molecular Genetics Methods for Developmental Scientists

MOLECULAR GENETIC STUDY DESIGNS AND GENOTYPING Linkage Studies and Short-Tandem Repeat Polymorphisms (STRPs) The vast majority of complex human traits, from height (Silventoinen et al., 2003) and insomnia-related symptoms (Hublin, Partinen, Koskenvuo, & Kaprio, 2011) to child language development and literacy-related skills (Grigorenko, Ngorosho, Jukes, & Bundy, 2006; Kovas et al., 2005) to adolescent psychopathy (Tuvblad, Bezdjian, Raine, & Baker, 2014) to patterns of resting EEG connectivity (Schutte et al., 2013), exhibit significant heritability, with the contribution of genetic factors estimated in behavioral studies to generally range from 30% to 80% with a median estimate around 50%. Establishing familiality and heritability5 using behavioral genetics methods and designs such as twin studies, the analysis of familial trait aggregation (i.e., the extent to which the trait is overrepresented in families where there is at least one affected proband), and segregation (i.e., the establishment of the mode of inheritance or genetic model of the trait) are important steps in the discovery of the genetic architecture of a trait. These steps are typically followed by the application of molecular genetics methods to locate genomic regions and identify genetic variants contributing to the trait, with the goal of revealing main genetic factors that play a role in its development and manifestation. For so-called simple, Mendelian traits characterized by relatively low prevalence, Mendelian patterns of inheritance (autosomal dominant, autosomal recessive, and X-linked), and high penetrance (i.e., probability of observing the phenotype given the presence of the mutation) of single-locus genetic variants, the past three decades of research have 5 Familiality refers to the familial aggregation (clustering) of the trait. For example, developmental disorders of language are considered to be familial since prevalence of these disorders is higher in first-degree relatives of probands affected with the disorder than in families of unaffected children (Tallal et al., 2001; Tomblin, 1989). Heritability, on the other hand, refers to the proportion of the total trait variance attributable to genetic factors. In most cases, heritability is estimated as narrow-sense heritability that reflects “the degree to which the genes transmitted from parents determine the phenotype of their children” (Tenesa & Haley, 2013, p. 141). For example, heritability estimates from twin studies suggest that childhood depression symptoms are 30–80% heritable (i.e., 30–80% of the variance in these symptoms could be attributed to additive shared genetic effects). Every heritable trait is also familial, but not every familial trait is heritable.

been extremely fruitful with respect to the discovery of genes underlying these traits and accompanying disorders (Botstein & Risch, 2003). The molecular bases of more than 3,500 such rare disorders have been identified to date, accounting for approximately 50% of the estimated 7,000 monogenic diseases (Boycott, Vanstone, Bulman, & MacKenzie, 2013). Among them are such somatic disorders as cystic fibrosis (Online Mendelian Inheritance in Man, http://www.omim.org; MIM#219700; caused by mutations in the CFTR gene located on chromosome 7q316 ) and developmental disorders such as DiGeorge syndrome (MIM#188400; caused by a deletion of 1.5–3 Mb region of chromosome 22q11.2 that contains, among others, the TBX1 gene) or Charcot-Marie-Tooth syndrome (MIM#606482; caused by mutations in the DNM2 gene located on chromosome 19p13.2). The success of Mendelian gene discovery is largely dependent on the methodological advances that transformed statistical linkage analysis as a behavioral analysis tool to investigate familial segregation of traits into a molecular analysis tool (that involves positional cloning with molecular techniques) used to establish the location of the segments of the DNA shared by related and phenotypically similar individuals. This revolutionary change has been brought on by the technological advances that made the genotyping (i.e., evaluating the exact state) of polymorphic genetic markers (i.e., DNA sequences with known physical/chromosomal locations) accessible to researchers. Initially, these markers were restriction fragment length polymorphisms (RFLPs; highly locus-specific markers with alternative alleles associated with restriction fragments7 that differ in size), followed by short tandem repeat polymorphisms (STRPs). STRPs are represented 6 Each arm of a chromosome is divided into regions called cytogenetic bands. These regions correspond to the patterns of staining (light and dark areas) of chromosomal regions that can be detected under a microscope and are labeled consecutively starting from the centromere out toward the telomeres. Cytogenetic bands (e.g., 22q11) can be further subdivided into subbands (e.g., 22q11.1). 7 Restriction enzymes can recognize specific nucleotide sequences and cut DNA at specific locations called restriction sites. An RFLP probe hybridizes with (i.e., attaches itself to) one or more fragments of the cut (or digested) DNA samples separated by gel electrophoresis based on their size, producing a blotting pattern (e.g., visualized on an x-ray film) specific to a genotype at a particular locus. Mutations in the genome (e.g., SNPs, indels) change the relative positions of restriction sites and the length of DNA segments between them, making them amenable to genotyping through this method.

Molecular Genetic Study Designs and Genotyping

by repeating sequences8 of typically 1–6 nucleotides that are abundant in the human genome and, due to their elevated mutation rate, are highly polymorphic (e.g., with an average of 6–8 alleles estimated for European populations; Calafell, Shuster, Speed, Kidd, & Kidd, 1998). By themselves these polymorphisms mostly do not have functional significance as they are biased away from the protein-coding regions of DNA (Metzgar, Bytof, & Wills, 2000) due to potentially high disruptiveness of non-triplet (tri- and hexanucleotide, i.e., STRPs in multiples of three) STRPs that cause frameshift mutations (i.e., changes in the grouping of codons that affect which amino acids are coded for by the codons that occur after the mutation) and result in the alteration of protein structure and function. However, STRPs have been invaluable with respect to their utility in locating genomic regions affecting a variety of traits and remain the most frequently employed markers in linkage studies. Recall that linked genetic loci are transmitted together from parent to offspring more often than expected under independent inheritance. In so-called parametric9 or model-based linkage analysis used to study Mendelian traits, the investigator is interested in the recombination fraction between multiple loci genotyped in familial samples. The analysis of the segregation of these markers in pedigrees reveals their relative position on the genome and is used to map trait loci. Given the highly polymorphic nature of STRP markers and their relatively low genotyping error rate, it is possible to efficiently trace the transmission of DNA segments through the pedigrees. The co-inheritance of the same haplotype by individuals affected with a Mendelian disorder but not by the 8

Genotyping STRPs typically requires the polymerase chain reaction (PCR) amplification of the STRP region using primers that flank a particular STRP. The size of the amplified region depends on the number of repeats and, as in the case of RFLPs, gel electrophoresis can be used to separate DNA fragments of different size and establish the STRP genotype. Multiplexing (i.e., co-amplification of multiple STRPs in a single PCR reaction) facilitates genotyping of multiple STRPs at the same time, and in this case genotyping relies on using automated laser-induced fluorescence detection methods that label alleles with fluorescent dyes (fluorophores, same or different, depending on the overlap in allele size) attached to one of the primers for each of the STRP markers (for an overview of the current methods, see Guichoux et al., 2011). 9 Parametric linkage assumes that the models describing the trait and the genetic marker are known. Nonparametric linkage relies on fewer assumptions about these models and their precise specifications.

383

unaffected individuals (i.e., co-segregation of the phenotype and the haplotype) in extended pedigrees and nuclear families serves as primary evidence for the involvement of the particular genomic location in that phenotype. Parametric linkage studies specify several parameters to calculate the primary linkage statistic, logarithm of the odds (LOD) score, such as disease allele frequency and mode of inheritance, marker allele frequencies, and marker maps. The parametric linkage analysis itself is based on the maximization the LOD score10 function to provide an estimate of the recombination fraction between STRP markers and the disease locus (and, correspondingly, an estimate of the position of the disease locus). Linkage studies frequently follow a two-staged approach where the first scan is performed using multiple markers spaced 5–10 centimorgans (cM11 ) apart followed by a higher-density scan that is focused on the region for which initial evidence for linkage has been obtained. Given that the resolution of linkage studies is relatively low (1–10 cM), they might identify from a few to a few hundred genes predicted to be relevant for the trait and located within about 1 cM (or approximately 1 Mb) from the linkage site, depending on the pattern of the cross-overs in the region (Botstein & Risch, 2003). This two-staged approach, for example, has been successfully used to establish the linkage of a severe speech and language disorder to the locus labeled SPCH1 and located on chromosome 7q31, followed by the identification of FOXP2 as the causal gene for the autosomal-dominant Mendelian speech and language disorder in a single three-generational UK pedigree (Fisher, Vargha-Khadem, Watkins, Monaco, & Pembrey, 1998; Lai, Fisher, Hurst, Vargha-Khadem, & Monaco, 2001). Although there are several examples of successful linkage mapping of Mendelian disorders using single extended 10 The LOD score is based on the comparison of the likelihood of the null hypothesis (no linkage) to the likelihood of the alternative hypothesis of linkage given certain model parameters. Positive and large LOD scores provide evidence for linkage. An LOD score of 3 is equivalent to p = .0001 or the odds of linkage of a thousand to one that the two loci (or a marker and the disease) are linked and inherited together. Although this value has been used as the cut-off for a long time in linkage studies, more recent work points to the LOD value of 3.3 as being more appropriate for controlling type I error rates in genome-wide linkage studies with multiple markers distributed across the whole genome at a certain density level. 11 Centimorgan are units of genetic distance between chromosomal positions that are based on the evaluation of the frequency of crossing over (recombination) events between them (i.e., 1 cM equal 1% recombination, or, in physical distance, ∼1 Mb).

384

Molecular Genetics Methods for Developmental Scientists

pedigrees such as the KE family mentioned above (e.g., Asher, Morell, & Friedman, 1991; Sobreira et al., 2010), most linkage studies employ samples of multiple families to increase power to detect linkage. The statistical power of linkage (both parametric and non-parametric—see below) studies depends on several factors such as locus heterogeneity (i.e., more families are needed to detect signals that are coming from more than one locus involved in the etiology of the trait), recombination fraction between the trait and the marker locus (the further apart they are, the more families are required), correctness of the specification of allele frequency (and, for parametric linkage, other model parameters), diagnostic classification accuracy, presence of phenocopy (i.e., when a phenotype caused by nonhereditary factors matches or is very similar to a phenotype caused by genetic mutations), and genetic marker heterozygosity (or informativeness). Note that parametric linkage analyses that specify modes of inheritance are best suited for locating loci involved in single-gene Mendelian disorders, especially in cases when there are few ungenotyped individuals in a family and the specification of the phenotype is unambiguous. When the trait is polygenic, common, complex, or quantitative in nature, however, it typically does not conform to a specific mode of inheritance, and parametric LOD score analyses are ill suited for establishing linkage in this situation. Correspondingly, nonparametric linkage analyses methods have been developed to account for this uncertainty in model parameters, relax them, or even remove the necessity to specify them unambiguously (e.g., remove the mode of inheritance assumption and decrease variant penetrance since complex traits and disorders are commonly thought to involve multiple genes and variants of small(er) effect sizes). This approach assumes that affected related individuals should show an increased sharing of haplotypes that are Identical by Descent (IBD12 ) in regions close to the involved genes independent of the specific mode of inheritance; that is, phenotypically similar 12 Alleles are considered Identical By State (IBS) if they are represented by the same variant of a polymorphic locus. Alleles are considered IBD if they are not only IBS but also if they are inherited from a common ancestor. Two individuals (related or not) can share 0, 1, or 2 alleles IBS. While two unrelated individuals by definition share no alleles IBD, the expected proportion of shared IBD alleles in pairs of related individuals depends on the type of the relationship based on the principle of random segregation of gametes combined with the number of meiosis events that separate the two relatives in a pair (e.g., MZ twins are expected to share 2 alleles IBD at a given locus, while siblings are expected to show a distribution of 0 (25%), 1 (50%), and 2 (25%) alleles IBD).

related individuals should have similar alleles for a marker located close to the genes contributing to the trait. A common design for a nonparametric linkage study is to evaluate allele sharing in sibships, in particular, affected sib pairs (for binary traits). When the IBD status for alleles for a marker can be determined (i.e., parents were genotyped and their genotypes at a particular locus can distinguish between children’s IBS and IBD alleles), the observed proportion of sib pairs sharing 0, 1, or 2 IBD alleles can be compared with the expected proportions (see footnote 6) using a chi-square test or one-sided tests that compare proportion of sib pairs who share 2 alleles IBD with the expected proportion of 25% or compare average IBD sharing with the expected proportion of 50%. However, in the majority of cases, the IBD status cannot be determined unambiguously for either part of the sample or the whole sample (e.g., if the markers are not sufficiently polymorphic or the parents have not been genotyped). Several methods have been developed to allow for this uncertainty in linkage analysis. Among these are methods that take into account partially informative13 matings from extended families (Sandkuijl, 1989), and multipoint methods utilizing information from multiple genetic markers at the same time to estimate the IBD distribution at each marker (Kruglyak & Lander, 1995). The pairwise sib-pair method can be extended to include larger sibships, and alternative pairwise IBD sharing analysis methods were developed to perform linkage analyses with multiple affected relatives of different degrees of relatedness (Gudbjartsson, Jonasson, Frigge, & Kong, 2000; Kruglyak, Daly, Reeve-Daly, & Lander, 1996). So far we have mostly been discussing linkage in the context of binary traits such as disease status (affected vs. unaffected). However, many (if not most) traits of interest to developmental scientists are fundamentally continuous or quantitative in nature. These traits can be analyzed in quantitative trait linkage paradigms, and loci identified in studies that are based on these paradigms are frequently referred to as quantitative trait loci (QTLs). For sib-pair designs, early methods included regressing the squared difference between quantitative trait values of siblings on the estimated proportion of IBD alleles with the expectation that trait similarity in the presence of the linkage of this trait to a particular gene should be related to higher proportion of IBD alleles shared by siblings at the marker close to that gene (Haseman & Elston, 1972). 13

That is, when the parental genotype is not known but genotypes available for other family members can be used to calculate the probability distribution of the missing genotype.

Molecular Genetic Study Designs and Genotyping

More recent methods include extensions of the regression method that also incorporate additional information contained in mean-corrected cross-product of trait values (i.e., sibling covariance; Xu, Weiss, Xu, & Wei, 2000) and variance component methods that partition trait covariance between different relatives into the estimated IBD sharing component and the kinship component (Almasy & Blangero, 1998; Sham, Purcell, Cherny, & Abecasis, 2002). Keeping the sample size constant, variance component methods have the highest power when the number of pairs of relatives is maximized by studying fewer but larger families as opposed to more but smaller families (Williams & Blangero, 1999). Cost-efficient whole-genome linkage studies that employ 400–800 STRP markers (sufficient to cover the whole genome) replaced early linkage studies that used numbers of markers an order of magnitude smaller in 1990s. Since then, a multitude of positive findings for simple Mendelian disorders have been reported. More recently, less polymorphic but extremely high density markers, single nucleotide polymorphisms (SNPs; see next section), have also been proposed for linkage studies. Yet the utility of linkage analyses for complex traits remains severely limited by their low genomic resolution, power constraints due to large sample size requirements (Risch, 2000), and cost and validity considerations specific to family-based designs, especially in the context of complex developmental traits. Although several approaches to increase the power of linkage studies of complex traits have been suggested, such as multivariate analyses that capitalize on the simultaneous interrogation of related phenotypes typical of complex cognitive traits (Marlow et al., 2003), developmental phenotypes that limit the age band where a particular phenotyping method is efficient remain a problem for these designs. Moreover, traditional mapping methods that focus on single time-point quantitative phenotype measures neglect the developmental nature of these traits and the dynamic pattern of their genetic control (Wu & Lin, 2006). We mention this developmental limitation here but would like to note that it is not specific to linkage studies and can be extended to genetic association studies in general (see next section). In addition, it is worth noting that sample ascertainment for linkage studies is rarely population-based and, correspondingly, traits that are characterized by locus heterogeneity (i.e., when a trait is influenced by multiple mutations at different unrelated chromosomal loci) might produce family-specific linkage signals that complicate the interpretation of linkage findings with respect to the general population (Teare & Barrett, 2005).

385

Candidate Gene Association Studies and Single Nucleotide Polymorphisms (SNPs) In contrast to linkage studies, the goal of genetic association studies is to establish an association between the genotype at a particular locus and a particular trait. Although the distinction between linkage and association is subtle, it is crucial for understanding the differences between the two approaches for identifying the genetic bases of human traits. Genetic association studies assume that the same allele is associated with the trait in a similar fashion in the population, while linkage studies allow different alleles to be linked with the disorder or trait in different families, as long as the location of the genomic segment harboring these alleles shows evidence for linkage. Most genetic association studies are performed using DNA samples collected from unrelated individuals genotyped using markers that provide a much finer resolution than the RFLP or STRP markers traditionally used in linkage studies—single nucleotide polymorphisms (SNPs). SNPs are single base-pair substitution polymorphisms that are typically less polymorphic than STRPs (i.e., most SNPs only have two alleles; see the NCBI database of SNPs called dbSNP, http://www.ncbi.nlm.nih.gov/SNP/). They represent around 90% of the variation in DNA and are found at the frequency of approximately 1 SNP per 1 kb of the human genome. They are found at a slightly higher frequency in intergenic compared with genic DNA, and among those that occur within coding genome segments, slightly less than half are synonymous (i.e., the base pair substitution does not change the coded amino acid) with the rest being nonsynonymous (i.e., missense, resulting in identifiable alterations of protein structure or nonsense, when the protein is rendered nonfunctional due to truncation introduced by a premature stop codon). SNPs that are located outside of the coding regions might also have regulatory and transcription-altering functions, although for the majority of these markers their functional impact and regulatory potential is unknown. SNP genotyping and association analysis has become the leading tool for trait gene discovery due to their abundance in the human genome, low mutation rates, and simplicity of methods for SNP genotyping. We already briefly covered one such method: the RFLP genotyping touched upon in the previous section can actually be used for SNP genotyping when the restriction cut occurs for one allele but not the other, resulting in different segment lengths corresponding to each allele. The utility of this method is limited by its inapplicability to the majority of SNPs and the low degree of automatization of the process. There are several

386

Molecular Genetics Methods for Developmental Scientists

alternatives available to researchers, and the choice of the genotyping platform largely depends on the scale of the study (i.e., ranging from platforms that genotype one to several thousand to several million SNPs). Below we will first describe two genotyping platforms used to interrogate several to several thousand SNPs to set the stage for the further discussion of SNP genotyping platforms in the later sections of the chapter in the context of genome-wide association studies. One of the most frequently used methods for SNP genotyping in human association studies is allele-specific oligonucleotide hybridization (ASO). This method discriminates between SNP alleles by hybridizing complementary synthetic oligonucleotides (i.e., short single-stranded DNA molecules) to the target DNA segment. Allele discrimination in ASO genotyping relies on the labeling of allele-specific products with different fluorescent dye labels that are then read using fluorescent readout systems. For example, the TaqMan® assay (Life Technologies, Applied Biosystems) is based on the ASO method coupled with 5’ nuclease activity of Taq polymerase during the PCR reaction (Shen, Abdullah, & Wang, 2009). TaqMan probes are labeled with a fluorescent reporter dye on one end (5’) and a common nonfluorescent quencher on the other end (3’) that eliminates background fluorescence of the intact probe. Each TaqMan assay uses two allele-specific probes and a pair of PCR primers. The hybridization of the probe during PCR results in the separation of the dye and the quencher, promoting fluorescence. Allele detection in this case is based on the Forster resonance energy transfer (FRET) analysis. FRET technique measures changes in fluorescence during dye excitation due to the transfer of energy in the donor-acceptor complex. In the absence of the quencher, the fluorophore dye, excited at a particular wavelength, emits light while returning to its ground state, while the presence of the quencher permits the transfer of energy from the excited fluorophore to the quencher without the emission of light. SNP genotyping typically uses cyanine dyes Cy3 (dark pink) and Cy5 (blue) for allele discrimination. The simplicity of TaqMan Assay chemistry lends it to automation and millions of assays have been developed, but low SNP multiplexing (i.e., genotyping of more than one SNP in parallel in the same assay/reaction) and high cost of dual-labeled probes limits the application of this genotyping method to projects that recruit large samples that are then genotyped using a large number of markers. In contrast, Illumina GoldenGate Assay (Illumina, Inc), a highly multiplexing (i.e., simultaneously interrogating 96 to 3,072 SNPs) microarray platform, is based

on the allele-specific primer extension (ASPE) and ligation method (Fan et al., 2006; Shen et al., 2005). In broad strokes, the increasingly popular microarray technology relies on using a small solid support surface with attached immobilized target DNA sequences (probes) to which the complementary amplified and labeled sample DNA (i.e., the DNA that is being genotyped) hybridizes. The hybridization signal is then read and interpreted with respect to evidence for a particular genotype. Illumina’s microarray genotyping platforms are designed so that three oligonucleotides are synthesized for each SNP: two allele-specific oligonucleotides (ASOs) that are used to distinguish the SNP alleles, and a locus-specific oligonucleotide (LSO) that hybridizes 1 to 20 bases downstream from the ASO site. ASOs and LSOs contain target sequences for a set of universal PCR primers, and the LSOs also contain SNP-unique address sequences complementary to capture sequences attached to ultra-small silica beads. These silica beads are placed in microwells etched out of optical fibers or a silicon wafer. After the oligonucleotides have hybridized to the sample DNA, the ASO with a 3’ end base complementary to the base at the SNP site is extended by polymerase and ligated to the corresponding LSO, forming a PCR template. The templates are amplified using the set of primers common to all SNPs, of which two are labeled with either Cy3 or Cy5 fluorescent dyes (but each anneals to only one of the two ASO sequences), and the third one anneals to a LSO common sequence. These reactions generate allele-specific PCR products that hybridize (using LSO to recognize a particular SNP site) to the beads carrying LSO-complementary oligonucleotides. The resulting hybridization signal is composed of the ratio of Cy3 and Cy5 fluorescence and is used to determine the genotype at each SNP site (i.e., equal intensity corresponds to the heterozygote, while the ratios of 1:0 and 0:1 correspond to the two alternative homozygote states). Note that genotyping based on indirect fluorescent signals is essentially a bioinformatic inference problem, and, although generally robust, is not immune to errors and depends on such parameters as the quality of the DNA sample, technical errors during sample and library preparation and handling, the specifics of SNPs selected for genotyping, and the accuracy of the algorithms used for SNP calling. Both of the platforms described above are aimed at establishing the genotype (or allelic state) for each of the person in the study sample at each of the interrogated SNPs. These genotypes are represented by combinations of alleles that are labeled according to their frequency (typically the more frequent allele is considered the major

Molecular Genetic Study Designs and Genotyping

allele, and the less frequent one is labeled as the minor allele) in the current or the referent sample. Commercial genotyping platforms rely on prior knowledge about the sequence of the human genome and interrogate known sites of variation. Recall that a locus is considered polymorphic if the frequency of the minor allele (MAF) is at or above 1% in a particular population. However, it is certainly possible for an individual to carry private (i.e., found only in a particular family) or rare mutations that are missed by commercial platforms due to their low frequency (and novel mutations that have never been described and, correspondingly, could not have made it to the commercial SNP genotyping platforms). We will return to this point later in the chapter when talking about structural variants and rare variants identified using next-generation sequencing approaches. For now, we would like to concentrate on the main design considerations for candidate gene association studies. Case–Control Designs, Single SNP, and Haplotype Association Testing Strategies When the researcher is interested in the association between one (or several) SNPs and a binary or a quantitative trait, the simplest design strategy is to recruit either two ascertained groups of individuals that differ with respect to the binary trait (affected cases and unaffected controls) or a sample of individuals that exhibit significant variation in the quantitative trait. When the DNA from cases and controls is genotyped for a set of SNP markers of interest, straightforward association analyses can be employed to detect association signals separately for each marker. These include methods familiar to most behavioral scientists such as logistic regression that models log odds of disease as a linear function of the genotype/allele status and chi-square tests of association that test for the independence of disease status and genotype/allele frequencies with a standard 2 (cases + controls) × 3 (two homozygotes and the heterozygote) contingency table. It is now widely accepted that the contributions of genetic factors to disease risk for complex disorders are roughly additive with the heterozygotes’ risk estimated to be in-between that for the two homozygotes. Special statistical procedures have been developed to account for this additivity and improve statistical power compared with the general class of tests mentioned above. For example, the Cochrane-Armitage test (Armitage, 1955), also known as the proportion trend test, evaluates the hypothesis of zero slope for the line fitted to the three genotypic risk estimates—in this case the risk is estimated as the proportion of cases to the total sample of cases

387

of controls harboring 0, 1, or 2 risk alleles. The choice of the particular association testing strategy, however, is not straightforward, as for some associations the nature of the relationship deviates from assumptions behind the employed statistical tests—for example, if the relationship significantly deviates from the additivity assumption (as in the case of overdominance when the two homozygotes have similar risk estimates that are also different from that for a heterozygote), the Cochrane-Armitage test will have low to null power to detect the true association. There are several strategies for overcoming this limitation, including calculating test statistics for all possible models (i.e., additive, dominant, and recessive effects) and choosing the maximum test statistic, or using Bayesian procedures (Balding, 2006). For quantitative traits, association methods based on the extension of generalized linear model (GLM) approaches such as linear regression or analysis of variance (ANOVA) are frequently used to test for association under the assumption of additive genetic effects. These association tests typically require the distribution of the trait values to approximate normality for each genotype, and for each genotype group to have similar trait variances. In situations when the assumption of normality is violated, mathematical transformations of trait data such as log-transformation or quantile normalization can be used to approximate normal distribution. Score tests for binary outcomes like the Cochrane-Armitage test and linear regression approaches for quantitative traits can also incorporate covariates in the analysis, for example, to adjust for participants’ gender and age. One of the important concerns in genetic association studies is the presence of population stratification. Allele frequencies for a multitude of genetic variants vary both within and between different populations. Correspondingly, the detection of an association signal can be influenced (in both directions—by producing spurious associations and masking true association signals due to loss of statistical power) by the presence of differences in baseline allele frequencies between the populations from which cases and controls are sampled that have nothing to do with the trait of interest but rather with systematic differences in these populations’ ancestry. In fact, population stratification is one of the most widely acknowledged reasons for nonreplication of genetic associations (Cardon & Palmer, 2003). In the ideal case, population stratification should be controlled by careful sample matching between cases and controls, and several methods (e.g., matching on such variables as geographic proximity, physical characteristics, self-reported family ancestry) have

388

Molecular Genetics Methods for Developmental Scientists

been proposed to address this issue at the level of study design planning. Most frequently, however, the researchers supplement these procedures with statistical analyses of background genetic differences between cases and controls. These methods range from simply detecting background differences in genetic makeup between the populations by evaluating an independent set of multiple SNP markers that are unrelated to the markers of interest (Pritchard & Rosenberg, 1999) to more advanced methods such as structured association methods that assign individuals to population subclusters and perform association analyses within these clusters or use cluster membership as a covariate in the analyses (Pritchard, Stephens, & Donnelly, 2000), and genomic control methods that control for the inflation of test statistic caused by stratification (Devlin, Roeder, & Wasserman, 2001). When properly applied (e.g., by genotyping a recommended set of at least 25 additional SNPs) in situations when population stratification is not extreme, genomic control efficiently corrects for population stratification with minimal loss of statistical power (Dadd, Weale, & Lewis, 2009). Note that the methods mentioned in the previous paragraph require genotyping multiple SNPs at unlinked or independent loci. So far we have been discussing association testing for single SNPs. However, in the vast majority of cases, the researchers genotype more than one SNP (e.g., multiple SNPs within a single candidate gene or in a candidate region). Independent testing of multiple SNPs in the same study introduces two problems. First, testing multiple SNPs separately leads to the statistical problem of multiple testing, well known to behavioral scientists as inflating Type I error rate. Standard multiple comparisons correction procedures such as the Bonferroni correction are overly conservative given that SNPs are frequently not fully independent from each other (i.e., are in linkage disequilibrium) and other methods that adjust for multiple comparisons in association studies have been developed. We will discuss these methods in the later paragraphs in the context of the extreme version of this problem specific to genome-wide association studies. Second, such independent testing ignores the wealth of information conveyed by the combinations of SNPs reflected in haplotypes—specific combinations (or sequences) of alleles that occur together more frequently than expected by chance in regions of highly correlated genetic material with limited recombination. These regions are called linkage disequilibrium blocks and contain haplotypes consisting of SNPs that are in the high degree of linkage disequilibrium with each other due to shared ancestry (The International HapMap Consortium, 2005,

2007). Haplotype analyses reduce the burden of multiple testing by reducing the actual number of statistical tests performed. However, haplotype analyses can also capture complex interaction effects, such as when mutations that occur at different sites within the same gene on the same chromosome (in cis position as a haplotype) interact to create super alleles that have large multiplicative effects on the trait and disease risk (Drysdale, 2000; Tavtigian et al., 2001). When haplotype phase (i.e., whether and which alleles observed for adjacent SNPs for a given individual are located on the same chromosome because they were inherited together as a haplotype from the same parent) is known, haplotype association testing is relatively straightforward and relies on extensions of methods used for SNP allele testing (e.g., by comparing haplotype frequencies in cases vs. controls). In case–control studies, however, the information about haplotype phase is typically unavailable as genotyping platforms produce unphased observed genotypes for individuals in the sample, and their parental genotypes are also typically unknown. Correspondingly, haplotype analysis in case–control studies requires statistical haplotype inference and reconstruction using likelihood-based or Bayesian methods followed by haplotype association testing (Schaid, Rowland, Tines, Jacobson, & Poland, 2002; Stephens & Donnelly, 2003; Sun, Greenwood, & Neal, 2007). For example, in an iterative two-staged approach of Schaid and colleagues (2002), haplotypes are inferred by computing estimates of probabilities of all possible haplotype pairs consistent with each individual’s genotypes; these estimates are used as weights to update regression coefficients in association analysis in the GLM-based score testing framework that evaluates the presence of global (omnibus) haplotype effects and effects of specific haplotypes on binary, as well as quantitative traits. Family-based Designs Genetic association testing strategies described in the previous section rely on recruiting and genotyping samples of unrelated individuals. These case–control and case-cohort designs became the de facto standard in the field of complex trait genetics after linkage mapping fell out of favor. That does not mean, however, that studies of genetic underpinnings of complex traits exclusively use samples of unrelated individuals. A complementary class of study designs, called family-based designs, relies on using sets of genotyped relatives of various familial configurations to test for genetic associations between a trait and a set of SNP markers. These designs offer important advantages

Molecular Genetic Study Designs and Genotyping

over case–control studies: i.e., they are very robust against population stratification, simultaneously test for linkage and association, and offer unique strategies for battling the issues of multiple testing and genotyping quality control (Laird & Lange, 2006). One of the simplest family-based designs is that of family trios which requires genotyping an affected proband and both of their parents. A specific non-parametric analytic strategy, called the transmission disequilibrium test (TDT), has been developed for these designs and tests the null hypothesis of no linkage and no association by evaluating the observed and the expected (under Mendelian laws) number of transmitted alleles (Spielman, McGinnis, & Ewens, 1993). For each marker, the TDT only utilizes data from informative families with at least one heterozygous parent and evaluates whether a particular allele is excessively transmitted to affected offspring (i.e., by comparing the numbers of transmitted vs. nontransmitted alleles), thus providing evidence for both linkage and association of that allele and the trait. Given that TDT is nonparametric, it does not require the specification of the genetic model or even the distribution of the trait. It is, however, inapplicable when parental genotypes are missing or when the analysis is focused on larger (so-called general) pedigrees. The generalization of the trio-based TDT analytic strategy is realized in the widely used family-based association testing (FBAT) approach that permits using different configurations of relatives and missing data (e.g., nuclear or extended families with multiple affected and unaffected siblings that provide within-family controls), while preserving the nonparametric nature and robustness of the association testing found in TDT. FBAT represents a general class of family-based association tests that evaluate test statistics expressed as the covariance between an arbitrary function of offspring’s genotype (genotype transmissions) and an arbitrary function of offspring’s phenotype (a residual-trait deviation measure such as residuals from the sample mean offspring trait values); the details about the complex statistics underlying this method are presented elsewhere (Horvath, Xu, & Laird, 2001). Importantly, the FBAT approach and its extensions can not only handle different configurations of familial structures, but also work with covariates, are efficient with quantitative and time-to-onset traits, and permit multi-phenotype and multi-marker association testing (Lange, DeMeo, & Laird, 2002; Lange, DeMeo, Silverman, Weiss, & Laird, 2004; Lange, Silverman, Xu, Weiss, & Laird, 2003). Other extensions of the TDT test include the affectedfamily-based controls (AFBAC; Thomson, 1995) method

389

that uses data from all rather than just heterozygous parents (although it loses its immunity to population stratification) and the recently developed generalized family-based association test for dichotomous traits (GDT) that utilizes data from all possible discordant pairs of relatives in the family with resulting statistical power being higher than for other frequently employed tests of association (Chen, Manichaikul, & Rich, 2009). As noted above, one of the main advantages of family-based designs is their robustness to population stratification. Genotyping parents also facilitates establishing haplotype phase, crucial for haplotype analyses, and the analysis of imprinting or parent-of-origin and mother–child genotype interaction effects (Cordell, Barratt, & Clayton, 2004; Weinberg, 1999). Another advantage of these designs is specific to childhood disorders where recruiting appropriately matched controls is more difficult than the affected child’s parents (Laird & Lange, 2006). However, these designs have an important limitation. Although family-based designs are particularly efficient in establishing linkage and association for rare Mendelian disorders, for common disorders and complex traits their power is generally lower than that of case–control designs, while genotyping cost and recruitment time requirements are generally higher than that for case–control studies (McGinnis, Shifman, & Darvasi, 2002). Nevertheless, family-based designs might still experience a rise in popularity in the context of the analysis of next-generation sequencing data, and recent advances in analytic techniques (albeit parametric) permit using hybrid designs that utilize both population-based and family-based samples simultaneously (Zheng, Heagerty, Hsu, & Newcomb, 2010). Candidate Gene and SNP Selection Strategies So far we have been discussing genetic association studies in general, without reference to the specific genes or polymorphisms within them. However, candidate gene association studies typically focus on a more or less well-defined set of genes and SNPs within those genes that interest researchers. Several factors influence the selection of candidate genes and SNPs for a genetic association study, ranging from the number and nature of tested hypotheses to sample size to expected effect sizes (these issues will be covered in later sections). Multiple strategies are available for candidate gene selection, although assessing the candidacy potential of a particular gene with respect to a complex trait is an inherently imprecise and difficult task (Hattersley & McCarthy, 2005). When evidence for significant linkage to a particular genomic region

390

Molecular Genetics Methods for Developmental Scientists

has been obtained, the researchers can select a set of SNPs and genes that are located under the linkage signal peak. This positional strategy thus focuses on candidate susceptibility genes that account for linkage peaks obtained in prior genetic linkage scans. Although intuitively appealing, this strategy has significant limitations as identified and replicated linkage regions are frequently relatively large and, given the general power and resolution limitations of linkage studies, are likely to miss multiple candidate genes located in other regions. The positional strategy for candidate gene selection can be used in isolation or in conjunction with other strategies, including the statistical strategy that prioritizes genes implicated in the trait or the disorder (or a related trait and disorder) in previous association studies or meta-analytical investigations (see later sections of the chapter). Candidate genes might also be selected for association analysis using prior knowledge about the biological function of the protein encoded by the gene and the information about the expression of that gene in tissues relevant for the trait (e.g., in specific brain regions) at appropriate developmental periods. The utilization of the biological function information about the gene is an efficient strategy for candidate gene selection if and when the pathophysiology of the disease/disorder or the pathways involved in the development of the trait is known. This strategy can be augmented by using information from pharmacological studies that identify proteins implicated in the action of the diseaseor trait-modifying drug or animal model studies that identify candidate genes that influence a phenotype similar to the human phenotype of interest in the model animal. The subjectivity of candidate gene selection procedures, coupled with high non-replication rates of candidate gene association studies (Ioannidis, Ntzani, Trikalinos, & Contopoulos-Ioannidis, 2001), led to the development of a variety of computational tools that attempt to address or at least minimize problems caused by the lack of the background knowledge about the molecular underpinnings of complex traits (Zhu & Zhao, 2007). These in silico tools augment and integrate several of the identified strategies of candidate gene selection (sometimes also called candidate gene prioritization) with the goal of arriving at a multivariate strategy that would more likely result in an accurate candidate gene identification than an isolated use of just one strategy. These computational prioritization tools rely on the extraction, filtration, and complex statistical analysis and integration of heterogeneous data sets and resources available in a variety of public databases that are typically organized according to the principles of biological ontology (anatomical, cell/tissue, developmental,

gene, and phenotype ontology). Many of the existing gene prioritization methods rely on two sources of input: 1) a list of candidate genes for prioritization (from a single locus, multiple loci, or even the whole genome), and 2) prioritization criteria (e.g., involvement in a disease or a certain pathway) in the form of biological keywords or seed genes that have already been linked to the trait or related traits (Brunner & Oti, 2007). These tools and computational strategies generally produce lists of candidate genes by filtering the initial lists into smaller subsets (e.g., based on the combination of the information about gene function, association data from genome-wide association studies, protein–protein and pathway interactions) or by ranking the input genes (Moreau & Tranchevent, 2012). The limitation of the filtering approach is the relatively strict nature of the filtration strategies that sometimes lead to false negative results when the true candidate gene is filtered out as failing one or more of the filtration criteria. An alternative strategy is to rank the candidate genes from least to most likely to be involved in the etiology of the trait without the application of strict filters. These methods include text mining of publicly available sources (e.g., publication abstracts in MEDLINE), similarity profiling that identifies candidate genes based on their similarity to the known genes relevant to the trait or respective biological processes, and network analytic methods that rely on the application of methods such as random walks or diffusion kernel to networks (e.g., a true protein–protein interaction network or an integrated functional linkage network) to identify trait-relevant nodes in preestablished networks through which information is propagated. There are multiple recently published excellent reviews of these (and other) methods available to researchers, with several accessible tutorials on web-based tools the interested reader is welcome to consult (Moreau & Tranchevent, 2012; Patnala, Clements, & Batra, 2013; Tranchevent et al., 2011). However, the application of these methods in a situation when little or no prior knowledge about the etiology of the trait and the pathways involved in related developmental processes is available remains challenging, and only a subset of human genes have been sufficiently annotated to be informative in such analyses. Once the candidate gene or a set of candidate genes have been selected, the researchers typically face the issue of choosing polymorphisms within these genes. SNP selection for candidate gene studies, like the selection of candidate genes, can rely on the evaluation of the functional effects of SNPs. Recall that SNPs can be classified according to their functional class. A natural strategy is to select those SNPs within a gene that have identifiable consequences for the

Molecular Genetic Study Designs and Genotyping

structure and, correspondingly, the function of the protein products of that gene. In this case the evaluation of the association is performed in the context of the functional impact of the SNP on the biological pathway relevant for the trait. When there are multiple relatively frequent non-synonymous SNPs per gene, it is common to include several or all of these SNPs in the pool of candidate SNPs. The evaluation of the functional effects of SNPs might promote SNP selection by providing information regarding the degree to which a particular SNP is damaging obtained in experimental studies and analytically. Similarly to candidate gene prioritization frameworks, multiple in silico SNP effect analysis tools have been proposed, generally falling into two classes of sequence evaluation vs. structure evaluation algorithms (Mah, Low, & Lee, 2011). Sequence evaluation algorithms are more frequently used than structural algorithms since dimensional structural information is currently unavailable for most proteins, and evaluate the effects of SNPs on the function of the protein based on such parameters as the extent of change in amino acid residues (sequence homology) in evolutionarily conserved (and, therefore, assumed to have important functions) loci and their physiochemical similarity. Structural analyses, on the other hand, predict the functional effects of SNPs based on their impact on the three-dimensional protein structure, for example, by focusing on protein stability. Recently developed tools integrate sequence and structure information, as well as other prioritization-related data (such as network, pathway, and association information). Researchers might also want to include SNPs in the regulatory gene elements (e.g., promoters and TF14 binding sites), regions that can have long-distance effects on gene function (e.g., enhancer regions), and regions coding for microRNAs—small (∼22 nucleotides long) noncoding RNAs that are extensively involved in post-transcriptional gene expression regulation (Ameres & Zamore, 2013). All of these elements are involved in multiple mechanisms that govern gene expression (Ameres & Zamore, 2013; Riethoven, 2010). Therefore, SNPs can disrupt or alter gene expression by influencing these elements, sometimes in a cell-type and tissue-specific manner (Dimas et al., 2009). Recently, a class of elements called expression quantitative trait loci (eQTL) has been discovered by establishing associations between variants located in these elements that can either be proximal (cis) or distal (trans) with respect to the location of the gene with the actual 14 Transcription factors (TF) are proteins that initiate and regulate transcription (and, therefore, gene expression) by binding to specific DNA sequences.

391

levels of that gene’s expression (Dixon et al., 2007; Goring et al., 2007). Emerging evidence points to the pivotal role of eQTLs in complex traits—e.g., it has been found that SNPs associated with complex traits are much more likely to be located in eQTLs compared with SNPs located in other regions (Nicolae et al., 2010), and, correspondingly, point to the regulatory (rather than, for example, disruptive) nature of the variants associated with complex traits and common disorders. Note, however, that commercial SNP genotyping platforms rely on using known SNP sites with a certain frequency threshold (i.e., >1% MAF in the population) that also satisfy several additional constraints necessary for high-quality genotyping (e.g., the unambiguity of the surrounding sequence is pivotal for successful target hybridization), and genotyping the full set of all discovered SNPs within a gene is impractical both from a statistical viewpoint and for cost reasons (it is, however, becoming increasingly possible with the application of next-generation sequencing methods; see further chapter sections). Given that the human genome has a linkage disequilibrium-laden correlated structure candidate gene association studies frequently employ the tag SNP selection strategy. Thus, loci demonstrating significant levels of linkage disequilibrium can be evaluated for association with a trait using so-called common tag SNPs that are correlated with other SNPs in the region because of the shared history and therefore tag common SNP haplotypes, removing the necessity to genotype SNPs at all possible loci. The presence of linkage disequilibrium and the availability of common tag SNPs for genetic association studies implies that it is possible to detect an association between a tag SNP and a trait without the tag SNP having any direct functional relationship with the trait and without directly genotyping the causal variant. Therefore, an important distinction for genetic association studies is that between a direct association and an indirect association that arises due to the correlated structure of the genome and the corresponding redundancy in the information provided by closely located SNPs. Differentiating between the two classes of association phenomena is an important issue for genetic association studies and frequently relies on the extensive characterization of the functional impact of the SNP on a trait or further fine-mapping (e.g., by using next-generation DNA sequencing) of the region around the SNP for which association has been demonstrated. The selection of tag SNPs is possible in situations when the patterns of linkage disequilibrium between common variants at a particular locus are known for a given population. Fortunately, large-scale genome research

392

Molecular Genetics Methods for Developmental Scientists

projects such as the International HapMap Project that focuses on discovering common patterns in the variation of the DNA sequence in the human genome (The International HapMap Consortium, 2005, 2007) provide such high-resolution genome-wide data for samples of varying ancestries: e.g., over 90% of the common (MAF > 5%) SNPs in non-African populations and around 80% of the common SNPs in Yoruba populations are highly correlated with at least one SNP in HapMap-produced data set of over 3.7 million SNPs with an oversampling of nonsynonymous SNPs. Researchers are typically interested in obtaining the smallest possible set of tag SNPs with minimal loss of information based on pairwise correlations between SNPs or correlations of multiple SNPs indexing haplotypes. Some studies show that the reduction in the number of genotypes required to capture the common variation using tag SNPs is around 46–65% when the SNPs are chosen to capture all of the SNPs in the reference set, with a further reduction of 36–55% when the selection criteria are relaxed and without noticeable decreases in statistical power to detect true association signals, with even greater efficiency for multimarker methods (de Bakker et al., 2005). Correspondingly, several computational tools have been developed to optimize the process of tag SNP selection for candidate gene association studies using data produced by the International HapMap as well as other projects (Xu, Kaplan, & Taylor, 2007) for single as well as multiple populations simultaneously (Liu, Wang, & Wong, 2010; Patnala et al., 2013; Sicotte, Rider, Poland, Dhiman, & Kocher, 2011). However, although common tag SNPs can efficiently capture common variation in the human genome, they are much less efficient at predicting rare variants with MAF below 5% (Zeggini et al., 2005), and researchers should be aware of this limitation while designing candidate gene association studies. An example of a successful implementation of a candidate association strategy can be found in a recent report on the genetic bases of adolescent substance abuse by Trucco, Villafuerte, Heitzeg, Burmeister, and Zucker (2014). Trucco and colleagues examined the associations between GABRA2, a gene previously implicated in substance abuse and conduct disorder in adults, and rule breaking as a potential mechanism mediating the effects GABRA2 on substance and alcohol abuse risk in the developmental context. GABRA2 is characterized by a strong LD structure, resulting in two common haplotypes specific to this gene. These haplotypes have been tagged in the study by three SNPs genotyped using the TaqMan platform for each of the 518 participants. The results were consistent across

three SNPs and indicated that variation in GABRA2 was more strongly associated with rule breaking in mid–late but not early adolescence, with over 70% of the total effect of GABRA2 on substance abuse attributable to mediating rule breaking (which prospectively predicted substance abuse in the sample). This study represents not only a successful candidate association study that interrogated the candidate gene identified for a complex trait in adult samples but also, notably, explicitly modeled cascading developmental processes and their genetic regulation (i.e., using path modeling). Genome-Wide Association Studies and High-Density SNP Microarray panels The ability of common SNPs to tag genetic variants in linkage disequilibrium with them lies at the heart of the principle of genome-wide association studies (GWAS). In contrast to candidate gene association studies, GWAS studies are hypothesis-free in the sense that instead of depending on hit-or-miss candidate gene and variant selection, they are aimed at interrogating the whole genome. GWAS studies were made possible by the combination of advances in human genetics such as the completion of the Human Genome Project15 (International Human Genome Sequencing Consortium, 2001, 2004), the deposition of a large number of SNPs to public databases, the initiation of the HapMap projects, and the rapid development of commercially-available high-throughput microarray genotyping platforms by Illumina and Affymetrix in the mid 2000s. Initially, these platforms aimed at using the microarray approach (described in previous sections) to genotype approximately 100,000 SNPs with MAF > 5% per sample spread across the whole genome and tagging other variants in linkage disequilibrium with them. Current instantiations of Illumina’s and Affymetrix’s GWAS microarray panels include 10–50 times more SNPs than the number of SNPs used in the first published GWAS that evaluated associations between 116,204 SNPs and age-related macular degeneration in 96 cases in 50 controls (Klein et al., 2005). For example, Illumina currently offers a family of BeadChip-based microarray panels that evaluate from 715,000 SNPs per sample 15

Coordinated by the National Institutes of Health and the U.S. Department of Energy, the Human Genome Project was the international consortium-based research program aimed at establishing the precise genomic location and DNA sequence of human genes. Its completion in 2003, 13 years after the start of the project, marked the key milestone in the science of human genetics.

Molecular Genetic Study Designs and Genotyping

(HumanOmniExpress-24) to over 4.5 million SNPs per sample (HumanOmni5Exome). In addition to including more SNPs, some of these arrays also (1) capitalize on information obtained in the 1000 Genomes Project16 and large-scale sequencing projects, permitting genotyping of rare SNPs with MAF between 1% and 5% as opposed to only common SNPs with MAF >5% in previous generations of panels, (2) oversample exonic SNPs, boosting the power and resolution of GWAS studies to detect associations with SNPs located in coding gene regions, (3) include markers for other types of genetic variation (e.g., indels and CNVs; see later sections), and (4) allow for panel customization with the option of adding up to thousands of custom markers that specifically interest the researchers in specific studies. Recent studies comparing the efficiency, coverage, and cost–benefit ratios (Ha, Freytag, & Bickeboeller, 2014) for different genotyping platforms suggest that while genome coverage for these platforms is slightly lower than advertised by the manufacturers and varies by study population, their tagging efficiency is still relatively high, and techniques like the imputation of unobserved genotypes based on the markers genotyped on the microarray are likely to prove advantageous for GWAS studies. Several overviews on the design of GWAS studies focusing on the evaluation of the prospective sample size and power as related to researchers’ budgetary constraints, research questions and hypotheses, and choice of genotyping platform have been published and can be used as practical guides for choosing a particular genotyping platform (Ha et al., 2014; Nelson et al., 2013; Spencer, Su, Donnelly, & Marchini, 2009). Given the availability of these resources, we would like to instead focus on a set of design (and interpretation) issues beyond the choice of genotyping platform that are relatively specific to GWAS studies. Most GWAS studies adopt a case–control approach, while recruiting large samples of cases and controls that are genotyped using high-throughput platforms mentioned above. Since GWAS studies typically evaluate hundreds of thousands to millions of SNP markers, they are faced with the necessity to perform the corresponding large number of statistical tests that dramatically increases the Type I error rate. Traditionally, researchers have used a relatively 16 The 1000 Genomes Project launched in 2008 with the goal of arriving at a refined, detailed catalog and map of human genetic variation by discovering >95% of genetic variants of up to 1% MAF across the genome, estimating their population frequencies and linkage disequilibrium patterns by performing DNA sequencing of 1,000 human genomes.

393

arbitrary cutoff of p < 5 × 10– 7 put forward by the Wellcome Trust GWAS publications (The Wellcome Trust Case Control Consortium, 2007) or standard Bonferroni adjustments (𝛼 / number of genotyped SNPs) to declare an association statistically significant. However, Bonferroni adjustments are overly conservative in the context of the correlated structure of the human genome, and several alternatives that correct for non-independence between tested SNPs have been proposed. These include techniques that correct p-values based on the direct genotype-based calculation of the effective number of independent tests (Duggal, Gillanders, Holmes, & Bailey-Wilson, 2008; Gao, Becker, Becker, Starmer, & Province, 2010) and permutation testing (Han, Kang, & Eskin, 2009; Johnson et al., 2010). Performing and correcting for multiple comparisons in the GWAS context means a dramatic power reduction for detecting associations, especially when the signal to noise ratio is high. This issue is especially pronounced when searching for indirect genetic associations with complex, polygenic traits that exhibit significant allelic (i.e., when multiple different polymorphisms in the same gene are associated with the trait) and locus (i.e., when multiple genes and genomic loci are implicated) heterogeneity, and when phenotyping is imprecise. Although the aforementioned first published GWAS study included ∼150 individuals, currently performed GWAS studies generally employ much larger samples. For highly heritable polygenic traits with a strong genetic component such as height or Crohn’s disease susceptibility, effect sizes for common SNPs are generally small and vary in the range of .07% to 1.96% of explained phenotypic variance (Park et al., 2010). A recent analysis of published disease- and trait-associated SNP associations corroborates these estimates with an observed range of odds ratios (ORs) from 1.04 to 29.4 and a median of 1.33 (Hindorff et al., 2009). Such small effect sizes require inordinately large samples to be detected, especially for variants with low MAF: for example, over 10,000 cases and 10,000 controls would be required to detect an association between a dichotomous trait and a disease variant with an OR of 1.3 at MAF < 10% with 80% power at p < 10−6 (Wang, Barratt, Clayton, & Todd, 2005). While power increases with larger effect sizes and variant MAF for binary traits, GWAS studies still typically require samples on the order of 2,000 to 3,000 cases and controls to detect significant associations using currently available genotyping platforms (Spencer et al., 2009). It is worth noting that statistical power is slightly higher for detecting GWAS associations with quantitative rather than binary traits under some conditions (Yang, Wray, &

394

Molecular Genetics Methods for Developmental Scientists

Visscher, 2010). Other factors such as the heritability of the trait, disease prevalence, the number of causal variants, the presence of epistatic (gene by gene or GxG) interactions, the degree of LD between the typed marker and the causal variant, the case/control proportion in the study sample, and the presence of covariates also all influence statistical power, and extreme care must be exercised at the stage of GWAS study design to address these issues. Over 2,000 GWAS studies have been published since 2005, and curated catalogues have been constructed to facilitate dissemination of GWAS findings (Welter et al., 2014). Although GWAS studies have produced an array of results for a variety of traits, several issues have become apparent once these vast amounts of data have been accumulated. First, although multiple susceptibility loci have been identified for many complex traits, many were likely missed due to power limitations mentioned previously and characteristics of individual GWAS studies. Second, nonreplication and false association issues of genetic association studies, when extended to large-scale GWASes, are severe (Ioannidis, 2007; Lazzeroni, Lu, & Belitskaya-Levy, 2014), thus requiring extensive and careful validation in carefully designed replication studies, transforming a GWAS study into a two-stage design where association signals detected in the discovery cohort are validated in the replication cohort. This strategy can be augmented by GWAS association results prioritization strategies (Cantor, Lange, & Sinsheimer, 2010; Hou & Zhao, 2013) that optimize further gene, SNP, and sample selection. Third, when pooled together, disease- and trait-associated SNPs typically explain well under 10% of the phenotypic variance of complex traits (Visscher, Brown, McCarthy, & Yang, 2012). When compared with heritability estimates from behavioral genetic studies, these results suggest that trait-associated SNPs explain only a minor fraction of the heritability estimated in behavioral genetic studies. The remaining, unexplained portion has been correspondingly called missing heritability (Manolio et al., 2009), and this issue has received a lot of attention in the literature in the recent years. Several factors have been proposed to account for the missing heritability of complex traits, ranging from the acknowledgment of the important role of rare and structural variants (see following sections of this chapter) to the suggestions that missing heritability does not imply the presence of missing/undetected genetic variants but rather an overestimation of heritability in studies that ignore interactions among genes, their variants, and corresponding etiological pathways (Zuk, Hechter, Sunyaev, & Lander, 2012).

One interesting approach to increasing statistical power in genetic association studies, put forward in the psychiatric genetics literature, is to shift attention from direct behavioral phenotypes (e.g., disease status or performance levels) to intermediate phenotypes that are more proximal to the etiology of the trait than syndromic phenotypes. These so-called endophenotypes can be, for example, metabolic (e.g., peripheral hormone levels), neurophysiological (e.g., resting state EEG power), neuroanatomic (e.g., volumetric measures), or neurocognitive if the latter are proximal to the trait (e.g., phonological working memory measures in studies of developmental language disorders). Componential, quantitative endophenotypes reduce the complexity of end-point behavioral phenotypes and provide an intermediate link between biological and behavioral-syndromic levels of analysis (Cannon & Keller, 2006). The evident increase of interest in endophenotypes of complex disorders and traits was accompanied by both theoretical and methodological changes in the approaches to these traits that emphasized multivariate, dimensional and continuous investigations of multiple elements or facets of complex traits (Insel et al., 2010), in particular neurodevelopmental traits and disorders (Grigorenko et al., 1997; Smith et al., 2010). Until now we have been describing ideal situations where, aside from design, power and interpretation issues, genetic association studies (including GWAS studies) have been otherwise properly conducted. Yet, there is an additional layer of complexity to such studies related to the technical aspects of DNA sample type, collection and handling, genotyping errors and quality control procedures. Although behavioral scientists undoubtedly recognize the complexity of the application of genetic methods, we would like to bring the attention of the reader to these issues to further facilitate their recognition and appreciation of the multi-layered uncertainty that is fundamental to modern molecular genetics methods. High-throughput genotyping platforms have specific requirements for the volume, concentration, and quality of DNA samples submitted for genotyping. Correspondingly, in many studies, genomic DNA isolated from whole (venous) blood samples has been used as the preferred sample type. Collecting blood samples from development populations, however, might be challenging due to the complexities and potential complications of phlebotomy procedures. Inexpensive and easy to use saliva collection kits such as Oragene DNA Saliva kits (DNA Genotek, Inc) have been proposed as a viable alternative to standard procedures that typically yield ∼8 ml of

Molecular Genetic Study Designs and Genotyping

anticoagulant-preserved whole blood per sample. DNA yields from saliva samples are typically lower than those from blood samples, and protein contamination rates are higher (Abraham et al., 2012). However, several studies have demonstrated that despite a slight drop in genotyping quality for saliva samples, concordance rates between blood and saliva samples are very high, and saliva samples perform well both in small-scale and large-scale genotyping projects (Abraham et al., 2012; Bahlo et al., 2010; Philibert, Zadorozhnyaya, Beach, & Brody, 2008). Once the sample type has been selected, sample collection for large-scale projects typically takes place at multiple data collection sites. Multiple confounding issues can arise at this stage if cases and controls are overrepresented at different sites, genotyping for them is performed using different genotyping platforms and/or on separate sets of sample plates. Analyses of existing data sets demonstrated the presence of significant batch effects during genotyping that led to spurious associations due to confounding case–control status with such variables as experimental order, sample type, and confounding of family member status in family trios with plate sets that masked true genetic differences between families (Lambert & Black, 2012). Such effects are far from trivial and sometimes lead to the retraction of potentially high-impact publications, as has been the case with the New England Centenarian Study of human longevity GWAS published in and subsequently retracted from Science (Sebastiani et al., 2010) due to spurious associations attributable to combining data from multiple genotyping platforms. These issues are directly related to the principles of experimental design, and corresponding precautions can be taken to prevent confounding by using such procedures familiar to most scientists as blocked randomization. Most current GWAS studies, however, rely not only on careful study design and procedures for quality assurance but also on the application of a complex set of procedures for quality control (Weale, 2010). Although the rationale behind each of these procedures is beyond the scope of this chapter, it is worth mentioning that it is common to assess the quality of genotyping per sample and per genetic marker (typically filtering out markers and samples with a genotype call rate below 95%), evaluate concordance between reported gender and gender inferred from genotypic data (i.e., based on the X chromosome heterozygosity) and between technical replicates, remove markers with low MAF (i.e., below 5%) and markers that show evidence for Mendelian segregation errors (one of the benefits of adopting a family-based design), assess population

395

stratification using methods such as principal components analysis (PCA), and evaluate genetic markers for their deviations from Hardy-Weinberg equilibrium.17 Given the complexity of genetic association study (including GWAS) design, execution, analysis, and interpretation, a special set of reporting guidelines called STREGA (Strengthening the Reporting of Genetic Association Studies) has been developed to facilitate reporting transparency and ease of interpretation (Little et al., 2009). In our view, these guidelines can also be used to facilitate optimal study design by bringing all of the above issues (and more) to the attention of researchers, funding agencies, and reviewers. GWAS studies of developmental psychopathology traits are relatively scarce (Addington & Rapoport, 2012) and findings from these studies have been mixed. For example, although eight GWAS studies of ADHD have been performed, none have identified variants that would survive corrections for multiple testing. However, a recently published study used a two-stage GWAS approach with 607 adult ADHD cases and 584 controls in the discovery and 2104 ADHD cases and 1901 controls in replication samples, obtained suggestive evidence for the involvement of FBX033 in ADHD and frontal gray matter volume (Sanchez-Mora et al., 2015). Augmenting Genetic Association Studies Using Recent Statistical Developments: Mixed Models, Chip Heritabilities, Gene-Based Associations, and Meta-analytic Approaches Some of the quality control procedures and adjustments can be performed in a GWAS setting simultaneously with statistical analysis. A particularly efficient method for adjusting for population structure confounders such as population stratification and familial relatedness (including so-called cryptic relatedness which arises when related samples are included in a population-based genetic association study) is the application of a set of analytical procedures known as linear mixed models (Kang et al., 2010; Listgarten et al., 2012; Zhou & Stephens, 2012). These models generally treat a genetic effect (e.g., number of minor alleles for a SNP) as a fixed effect in a regression model that also includes an additional random effect that models genetic relatedness (genetic covariance) 17

Hardy-Weinberg equilibrium (HWE) describes the relationship between genotype and allele frequencies where there are no selection forces (such as mutation, selective survival or genetic drift) influencing a particular locus in a stable population. In the context of GWAS, departures from HWE can indicate genotyping errors.

396

Molecular Genetics Methods for Developmental Scientists

between individuals in the study sample. In the GWAS context, available SNPs are used to calculate measures of genome-wide similarity between pairs of individuals (e.g., a kinship matrix) that capture the probabilities that these pairs share causal alleles and estimate their contribution to the phenotype. Developed explicitly for the application to complex quantitative traits, these methods are generally computationally intensive but are extremely efficient at controlling for a variety of confounding factors, sometimes in an SNP-specific (rather than sample-specific) manner, thereby reducing both Type I and Type II error rates. Although these methods present a viable alternative to standard association testing approaches, they do not solve other GWAS issues such as missing heritability. Several recently developed methods that are gaining popularity in the complex trait genetics literature are aimed at addressing the missing heritability issue by estimating the proportion of phenotypic variance captured by all of the information provided by high-throughput genotyping technologies, rather than relying on single-SNP testing. This estimated proportion is called genetic heritability or chip heritability. For example, genome-wide complex trait analysis (GCTA) has been proposed as method for the polygenic modeling of complex traits (Yang, Lee, Goddard, & Visscher, 2011). Under the assumption that individuals who are more genetically similar should also be more phenotypically similar, GCTA utilizes genome-wide SNP data to estimate genetic relatedness between pairs of individuals in the sample. It estimates chip heritability by using the extension of linear mixed models to partition phenotypic variance into additive genetic variance explained by all SNPs, additive– environment interaction effects, and residual variance. However, this approach is not applicable when gene-gene interactions or dominance effects are at play, and effectively assumes that every genetic variant affects the trait and effect sizes are normally distributed. For some traits, however, it might be more realistic to assume that a proportion (but not all) of SNPs affect the trait of interest, and methods based on sparse regression models (e.g., Bayesian variable selection regression models) have been developed for such situations. Given that for the majority of traits the true genetic architecture is unknown, hybrid polygenic modeling approaches that combine the linear mixed modeling and Bayesian sparse regression modeling have been developed (Zhou, Carbonetto, & Stephens, 2013). Hybrid polygenic models allow for both a small number of large effects and a large number of small genetic effects to be estimated simultaneously, with the balance between the two inferred directly from the data. They are not only efficient at estimating chip heritabilities when combinations of larger and smaller effect size SNPs are

at play, but also can be used to predict disease risk and quantitative trait values. Of note is that these methods are new, and few empirical reports have been published to date. Two other classes of analytical procedures have been proposed to increase the yield of GWAS and candidate gene association studies. While polygenic modeling approaches are aimed at estimating the contribution of all of the (genome-wide distributed) genotyped SNPs to the trait, gene-based tests are aimed at evaluating the association between a trait and all markers within a particular gene (Li, Gui, Kwan, & Sham, 2011; Liu et al., 2010). This approach is especially powerful when a gene contains more than one causal variant, each of which is only marginally associated with the trait. Gene-based tests effectively combine the effects of all of typed SNPs in a gene and estimating a test statistics that is adjusted for non-independence between these SNPs. While genetic variants can exhibit different allele frequencies, LD structure, and heterogeneity in different populations, gene-based analyses might produce more consistent and replicable findings across populations than single-marker analyses. Gene-based tests are not only efficient in candidate gene studies, but also in the GWAS context, particularly by reducing the multiple comparisons burden as tens of thousands (number of evaluated genes) rather than hundreds of thousands or millions of statistical tests are performed. Extensions of gene-based testing procedures include the incorporation of functional information on SNPs that are assigned a prior weight based on their involvement in specific pathways and protein-protein interaction networks. Another approach to evaluating the robustness and increasing the power for detecting genetic associations is to perform a genetic association meta-analysis. Even with relatively large samples, for the majority of complex traits the yield of reported and replicated genetic associations has been relatively modest. While GWAS studies typically employ a two-staged design where each of the stages is analyzed individually, there is a growing appreciation of the utility of directly combining statistical evidence for associations from multiple samples both within the same GWAS study and across published GWAS data sets (Begum, Ghosh, Tseng, & Feingold, 2012). Meta-analytic procedures, familiar to most behavioral scientists, offer an important toolset for detecting small effects that are consistent across multiple studies and cohorts, and revealing heterogeneity between studies and samples. They are especially relevant for so-called mega-analyses of association studies that are based on multitudes of individual studies performed within research consortia formed specifically to

Structural Variation in the Human Genome and Copy Number Variant (CNV) Association Studies

increase statistical power. Genetic meta-analyses typically use straightforward statistical methods for combining (weighted or unweighted; e.g., using Fisher’s method or a weighted Z-score method) p-values or effect sizes within either fixed or random effects models (Nakaoka & Inoue, 2009). Although powerful, meta-analytic techniques require special care with respect to ensuring the validity and robustness of the application of study exclusion and inclusion criteria, similarity of phenotype definition, interchangeability of SNP coding, consistent application of data cleaning procedures across data sets, and genotype imputation when studies differ with respect to the genotyping platform, as well as meta-analysis model selection. The development of the approaches covered in this section of the chapter is reflective of the maturation of genetic association research in general and GWAS research in particular as a field that has moved beyond simple association testing, and has already contributed to disease- and trait-related gene and variant discovery. For example, Kornilov et al. (2015) performed a GWAS study of developmental language disorder in a small sample (n = 359) from an isolated population with an increased prevalence of the disorder using a mixed linear modeling approach. Although no individual SNP reached genome-wide significance, a gene-based analysis identified a gene statistically significantly (after correction for the number of genes for which the tests have been performed) associated with a part of the quantitative multivariate language phenotype in the population, SETBP1. Although the functional significance of this gene is currently unknown, several independent case study reports also indicated that SETBP1 haploinsufficiency (i.e., when only one copy of the gene is functional) is associated with expressive language deficits in children. Yet the methods described in this section have only rarely been applied to developmental traits, and although for other complex traits (mostly a variety of somatic disorders) the accumulation of data led to a certain level of awareness of both benefits and limitations of common variant genetic association testing, psychological, psychiatric, and developmental genetics still lives in the pre-GWAS era or, optimistically, the GWAS era, compared with the post-GWAS era status for other traits. STRUCTURAL VARIATION IN THE HUMAN GENOME AND COPY NUMBER VARIANT (CNV) ASSOCIATION STUDIES Although SNPs represent most of the interindividual variation in the human genome, other types of variants have been receiving increasing attention in the last decade.

397

These include structural genetic variants such as insertions, deletions, duplications, inversions, and translocations of genomic segments and segmental duplications. The field of human genetics has been extensively characterizing genetic variants at the nucleotide (e.g., SNPs) and karyotype levels (e.g., by studying aneuploidies), but structural variants that are smaller than ∼3 Mb have generally been poorly characterized with respect to their number, distribution, size, allele frequencies, and functional impact. The human genome contains several types of structural genetic variants that are frequently (albeit relatively arbitrarily) classified according to their size. Small insertions and deletions called indels represent the smallest class of structural genetic variants and generally range in size from 1 bp to 1 kb. Although initial estimates of the number of indels in the human genome were around 400,000 indels (Mills et al., 2006), with an average density of one indel per ∼7 kb of genome, more recent data from the 1000 Genomes Project indicated that the human genome might contain over 1,000,000 indels combined over different populations (1000 Genomes Project Consortium, 2010). Indel density in the human genome is correlated with SNP density, and approximately a third of indels are found in the coding regions of the genome, thus potentially having important functional effects on the affected genes. Structural variants that involve duplication or deletion of DNA segments over 1 kb long are called copy number variants (CNVs), and when structural variants of this size do not involve changes in the amount of genetic material (e.g., as in the case of inversions where the DNA sequence is flipped around with no actual gain or loss of genetic material), they are called structural variants (SVs). The distribution of CNVs and SVs across the human genome is not uniform: for example, as deletions are frequently associated with loss of protein function, they tend to be evolutionarily biased away from genes, and CNVs in general are biased away from ultraconserved genomic regions that are known to be under strong selection (Cooper, Nickerson, & Eichler, 2007). Structural variant frequency declines with variant length, with most (Shen et al., 2013; Wong et al., 2007; Zogopoulos et al., 2007) of the structural variants being classified as low-frequency (.50% to 1–5% MAF) or rare (< .50% MAF), and variants found at higher frequencies are called copy number polymorphisms (CNP); correspondingly, the human genome is estimated to have fewer CNVs and CNPs (e.g., ∼20,000 estimated by the 1000 Genomes Project) than indels or SNPs. Note that, although SNPs vastly outnumber structural variants, given the size of the latter, the cumulative sum of base pairs that can be affected by them can actually be larger than the sum of base pairs affected by SNPs, thus

398

Molecular Genetics Methods for Developmental Scientists

representing the second major source of genetic variation among individuals. Like SNPs, structural variants can be benign or damaging with respect to their impact, and functional effects of this type of genetic variation are diverse. Gene expression association analyses revealed that CNVs account for 17.7% of the total genetic variation in gene expression levels (Stranger et al., 2007), and CNVs that directly lead to the increase or decrease of number of copies of a particular gene can affect that gene’s transcript abundance (i.e., the amount of gene transcription product) and gene dosage. SVs, on the other hand, can disrupt transcription by creating breakpoints within transcripts, and large SVs can result in the alteration of the spatial distribution of chromosomes in the nucleus, thereby affecting important regulatory genetic interactions. CNVs, indels, and SVs located in regulatory regions might also disrupt gene expression, and these effects have been observed for loci that are sometimes millions of bp away from the target gene. The specifics of these effects partly depend on the location of the variant and the type of the regulatory element affected by it (i.e., activator or repressor). Likewise, structural variants located in noncoding regions can modulate transcription and even produce more severe effects like chromosomal fragility. Our understanding of the landscape of structural variation in the human genome and its role in human traits has been limited in part due to the technological challenges specific to methods used for structural variant discovery (i.e., identification of the variant in a population, its validation, and characterization at the sequence level) and genotyping. CNV discovery and genotyping has relied greatly on the development of microarray approaches, primarily array comparative genomic hybridization (array CGH) and SNP microarrays, and in both cases the evaluation of the copy number state is done by comparing probe signal intensities to those of the reference sample. Array CGH relies on comparative and competitive hybridization of two samples labeled with different fluorophores to a set of targets, typically represented by long oligos or bacterial artificial chromosome (BAC) clones (Oostlander, Meijer, & Yistra, 2004). The fluorescence intensity signal ratio (Cy3:Cy5) between the two samples is normalized, converted to a log2 ratio, and the latter serves as the main source of information for copy number inference: increase in this ratio is indicative of the increase (or gain) in a copy number relative to the reference, while a decrease is indicative of a copy number decrease (or loss). The initial resolution of CGH studies was low and allowed for the evaluation of copy number states for variants larger

than 1 kb. Commercially available whole-genome CGH arrays manufactured by Roche NimbleGen and Agilent Technologies rely on evaluating signals from multiple (i.e., 3 and more) consecutive probes to infer copy number state for each region and can genotype structural variants that are ∼500 bp long. Although the CGH method is not sensitive to structural variants that do not alter the amount of genomic materials (i.e., SVs such as transversions), it is currently the leading method for CNV detection in clinical settings that previously relied on karyotyping, especially in the contexts of searching for the potential causes of unexplained neurodevelopmental disorders (Miller et al., 2010). SNP microarray panels discussed in the previous sections have also been suggested as a method for structural variant genotyping (Cooper, Zerr, Kidd, Eichler, & Nickerson, 2008). Unlike for CGH, in this case hybridization is performed using one sample rather than two, and total normalized intensity as well as allelic intensity ratios are used for CNV inference to compare the copy number state of the tested sample with the aggregated intensity estimates of another individual or, more typically, a group of individuals (e.g., all of the genotyped individuals in a study sample). Although recent instantiations of these microarray panels include (nonpolymorphic) probes that specifically tag common CNVs and indels (especially in regions with low SNP density or coverage), in general the signal to noise ratio for CNV detection is lower for SNP panels compared to that of CGH panels (Alkan, Coe, & Eichler, 2011; Greshock et al., 2007), frequently leading to false positives (i.e., a discovery of the CNV that is not further validated using methods such as quantitative PCR). Microarrays have important limitations with respect to structural variant detection. For instance, copy number differences are evaluated with respect to the reference used to design the probes, and no information is available with respect to the actual location of duplicated or deleted regions or the position of structural breakpoints; in addition, single-copy gains are harder to detect than losses, and, as mentioned above, SVs that do not affect copy numbers are undetected by microarray panels. Furthermore, analysis of copy number states in and around repeat-rich and duplicated regions of the human genome is challenging given that microarray platforms rely on the assumption that each location is diploid in the reference, which might or might not be true, and is important given that CNVs actually correlate with segmental duplications in the human genome (Itsara et al., 2009). In addition, microarray CNV genotyping and discovery rely on the application of computational copy number detection

Structural Variation in the Human Genome and Copy Number Variant (CNV) Association Studies

algorithms (Colella et al., 2007; Wang et al., 2007), which, while performing well with large CNVs, are not as accurate in detecting small structural variants, and multiple algorithms have been shown to differ with respect to identified variants with concordance rates between different algorithms sometimes falling below 50% and validation rates below 70% (Pinto et al., 2011). Alternative methods for structural variant genotyping that overcome some of the aforementioned limitations of microarrays are single-molecule methods and nextgeneration sequencing. The former include cytogenetic methods like fluorescence in situ hybridization (FISH), fiber-FISH, and spectral karyotyping. These methods provide visualization of alterations of genomic structure at the single molecule level and have been used for structural variant discovery in the past. However, the fact that these methods are low throughput and have low resolution limits their application to several individuals and large structural aberrations (generally 500 kb to 5 Mb). Recently developed visualization techniques such as optical mapping and nanochannel visualization provide a more fine-grained analysis of genome structure and detect multiple types of structural variants (including complex rearrangements) and their locations but their application is also highly limited by a low throughput; these methods are reviewed elsewhere (Alkan et al., 2011; Das et al., 2010; Levy-Sakin & Ebenstein, 2013; Yamada et al., 2011). Structural variant association study designs and analytical approaches are generally very similar to those described in this chapter for SNP-based studies. They rely on the recruitment of samples of cases and controls (less frequently—population-based samples to study associations with quantitative traits) and testing for association between the structural variant (typically a CNV) alleles (e.g., the number of copies) and the trait, either in a candidate–gene or whole-genome fashion. Given their high potential for being functionally damaging, CNVs are important vehicles for both studying candidate genes and identifying new trait-related genes. Recall, however, that although they are collectively common and account for a significant proportion of human genetic variation, most structural variants are individually rare. In fact, structural variants in disease association studies have become synonymous with rare variants. In the previous sections we called GWAS studies hypothesis-free. This is not entirely correct given that GWAS study design, technologies used for genotyping, and statistical considerations guiding the analysis of GWAS studies are rooted in the common disease/common variant (CDCV) hypothesis. CDCV assumes that complex

399

traits are primarily influenced by small to moderate effect size and low penetrance variants that are common in the population (Reich & Lander, 2001). Although initially promising, this approach has faced serious criticism in light of the landscape of findings from GWAS studies published in the last decade, most importantly the failure of common variants identified as disease susceptibility loci or QTLs to account for the major portion of genetic variance in the corresponding trait or provide evidence for their functional involvement in that trait’s etiology. Correspondingly, the common disease/rare variant (CDRV) predicts that multiple rare variations in the DNA with high penetrance are the major (or at least complementary to common variants) contributors to complex trait variance and manifestation (Schork, Murray, Frazer, & Topol, 2009). Several arguments have been put forward to support CDRV. For example, it has been argued that population-level processes such as purifying selection and random genetic drift regulate the frequency of deleterious genetic variants, and rare genetic variants with recent (i.e., past 2 centuries) mutational history are more likely to contribute to common disease etiology than older common variants (Pritchard, 2001). Much like the way recent GWAS polygenic modeling studies provide some support for the CDCV hypothesis, structural variant and DNA sequencing studies published in recent years have been pivotal in providing evidence for CDRV. The debate between the two has not been resolved, and the authors of this chapter would like to emphasize that recent attempts to reconcile the two are more likely to result in advances in our understanding of the genetic architecture of complex traits (Gibson, 2012; Schork et al., 2009). One approach is to evaluate the relative contributions of the two types of variants to the trait. A complementary approach would also seek to provide a theoretical and mechanistic account of how common and rare variants operate together, for example, by extending the well-known threshold-dependent response model to incorporate both types of variants. In this case, complex traits are regulated by a large number of polymorphisms that modulate biochemical traits contributing to the overall phenotypic trait, and individuals at the extremes of the continuous liability conferred by these polymorphisms exhibit disease states when additional risk is conferred by rare variants, the variable effects of which are conditional on the background liability (Gibson, 2012). The digression in the previous two paragraphs and the distinction between the two classes of genetic variants is important for understanding the most commonly used structural variant association designs and problems

400

Molecular Genetics Methods for Developmental Scientists

specific to this class of genetic association studies. CNPs with relatively high MAF can be analyzed in a relatively straightforward fashion in case–control association studies by comparing CNP frequencies in cases and controls, and the usual association study validity (e.g., as related to population stratification) considerations also apply in this case. If the analysis is performed using SNP microarray-derived data, quality control procedures developed for SNP data (e.g., call rate thresholding) are supplemented by strict quality control procedures that attempt to account for the high signal-to noise ratio of microarray-based intensity data and imprecisions in CNV detection (e.g., by adjusting probe data for batch effects and so-called genomic waves that affect probe signal intensities, and evaluating consensus between CNV calling algorithms; Diskin et al., 2008). Another concern in CNV association studies is related to the difficulties in locating structural breakpoints specific to a particular structural variant: both true heterogeneity (i.e., presence of overlapping but not fully identical CNV segments associated with the trait) and genotyping/inference errors (i.e., when the same CNV is identified as having somewhat different start–end points for several samples) confound CNV association studies. Several strategies have been proposed to account for this issue by integrating CNVs in a study sample in a way that evaluates the boundaries between different CNVs and, for example, treats all common and unique CNV segments in the study sample as separate copy number variant regions (CNVRs). Since CNV association studies are relatively new in the field, comprehensive tools that perform CNVR detection and association analyses in an integrated fashion have only appeared recently (Forer et al., 2010; Kim et al., 2012). These tools also incorporate CNV association testing strategies for quantitative traits, most of which are based on familiar GLM techniques such as linear regression. CNV association studies can also be performed in a family-based setting using designs similar to the ones used by SNP association studies, e.g., by using family trios or extended pedigrees. These data can be analyzed by recoding CNV data as pseudo-SNPs (and assigning minor allele counts to each sample based on the deviation from reference copy number) and using corresponding analytic techniques such as TDT. However, this two-step procedure (copy-number assignment and association testing), common to case–control CNV GWAS studies, has been criticized for artificial discretization of distributed intensity values and loss of power. Correspondingly, the FBAT approach that has been extended to CNV data in

the GWAS context uses normalized probe signal intensities instead of assigned copy numbers to test for association between the trait (binary, quantitative, time-to-onset, and multivariate) and each marker (Ionita-Laza et al., 2008). Other association testing techniques for family-based designs based on variance component analysis for quantitative traits (Eleftherohorinou et al., 2011) or the further refining of the FBAT approach (Murphy et al., 2010) have also been proposed. Family-based designs offer an important advantage over case–control CNV association designs as they permit evaluating the de novo (noninherited) status of CNVs when parental CNV genotypes are available. Determining the de novo status of a structural variant can be performed both after and during CNV calling (Chu et al., 2013). The popularity of the de novo or spontaneous/sporadic framework for CNV analysis is illustrative of the recent attention to the high per-generation genetic mutation rate in humans. For example, on average a newborn is expected to have ∼70 new single nucleotide mutations in their genome with most of these mutations seemingly being of paternal origin (Keightley, 2012; Lynch, 2010). Although structural variants have not been precisely evaluated for their mutational properties yet, some estimates suggest that their mutational rate is lower than that of single nucleotide variants, and decreases with size. For example, new large CNVs over 100 kb long are estimated to occur once every 42 births (Itsara et al., 2010). However, de novo CNVs affect a significantly higher number of base pairs (16–50 kb) overall than single nucleotide variants (61 bp) per birth (Campbell & Eichler, 2013), and their deleterious potential is very high. As we mentioned above, CNVs became nearly synonymous with rare variants in genetic association studies, and in some cases the researchers’ attention is exclusively focused on the role of rare de novo structural variants in the etiology of complex traits and disorders. Although structural variant association studies are subject to the same statistical concerns and limitations as SNP studies, some of the issues are extremely pronounced in GWAS CNV investigations. Given the significant overlap in analytical approaches used for SNP and CNV association testing in the case/control setting, statistical power for CNV association tests can be evaluated using general power calculation tools. Recall that power is in part related to the frequency of the minor allele. Given their overall low frequency and genotyping issues that decrease power to detect CNVs described earlier, structural variants in general and CNVs in particular become difficult to detect in smaller-n studies (Wineinger & Tiwari, 2012). For example,

DNA Sequencing: From Sanger Sequencing to High-Throughput Next-Generation Sequencing Technologies

for single-marker tests, sample sizes above n = 4,000 might be required to detect an association even with higher MAF CNPs (Barnes et al., 2008). Evaluating the impact of CNVs on a trait can be boosted by employing different collapsing methods, for example, by using CNV burden testing that evaluates the differences in estimated overall CNV frequencies or cumulative CNV length (burden) across different length bands between cases and controls. This powerful strategy can be augmented in a candidate–gene fashion by evaluating burden estimates for specific gene sets, such as for genes implicated in the candidate (or related) pathways and networks. In recent years, CNV burden has been linked to a host of developmental outcomes. For example, in one study individuals with intellectual disability and combined intellectual disability and multiple congenital anomalies have been found to harbor more large (>1 Mb) CNVs than control individuals or individuals with autism or dyslexia (Girirajan et al., 2011), and similar findings have been recently reported for ADHD (Yang et al., 2013).

DNA SEQUENCING: FROM SANGER SEQUENCING TO HIGH-THROUGHPUT NEXT-GENERATION SEQUENCING TECHNOLOGIES So far we have been discussing genotyping methods and study designs that largely rely on the availability of genetic markers with known genomic locations and allele frequencies estimated for populations of different ancestries. In the case of linkage studies and in most cases of GWAS investigations, the statistical (linkage/association) signal is indicative of the location of the potentially causal variant which by itself is neither known precisely nor genotyped. Under these conditions, the causal mutation can be identified if the precise order of nucleotides in the DNA sequence is established for the candidate region. The Human Genome Project, the largest public scientific collaboration project in biology that took 13 years and $3 billion to complete, resulted in the publication of the working draft of the full sequence (also called the reference sequence or human genome build) of human genome (International Human Genome Sequencing Consortium, 2001, 2004), which revolutionized genetic science in terms of our knowledge about evolution, the number, location, and structure of human genes, the proteins they code for, and noncoding elements, led to the emergence of new fields such as proteomics (concerned with the identification and quantification of protein content in different organs and even cellular organelles) and directly or indirectly led to

401

the development of technological, computational, and analytical tools that form the basis of the modern genetic science (Hood & Rowen, 2013). The Human Genome Project relied on a combination of DNA sequencing technologies, most importantly the golden standard method of DNA sequencing based on Sanger’s chemistry, developed in the 1970s (Sanger & Coulson, 1975; Sanger, Nicklen, & Coulson, 1977). Modern implementations of Sanger or first-generation DNA sequencing methods (e.g., methods that combine Sanger sequencing with capillary electrophoresis), albeit being faster, cheaper, and more automated these days, rely on the same sequencing approach. Sanger sequencing begins with the amplification of the target region of an unknown DNA sequence (called a template). A synthetic primer that is complementary to a region of the template is then hybridized to the single-stranded template, forming a double-stranded region that serves as a starting point for DNA chain extension by a DNA polymerase. The latter extends the sequence beyond the primer using complementary base pairing. A key element of Sanger sequencing is the addition of a small amount of dideoxynucleotide triphosphates (ddNTPs) to the reaction mix that also includes normal deoxynucleoside triphosphates (dNTPs). dNTPs (nucleotides) are standard building blocks of the DNA molecule, while ddNTPs lack a hydroxil group at the 3’ position of the sugar that forms a phosphodiester bond between two nucleotides. The incorporation of a ddNTP into the growing DNA chain effectively leads to the termination of chain extension. In the solution where the same DNA chain is continuously synthesized, chain termination due to ddNTPs occurs randomly at all positions where the ddNTP nucleotide can be added to the chain. Thus, one reaction produces a set of synthetic products (sequencing ladders) of varying lengths, which can be separated according to their molecular weight. Four separate reactions are performed—one for each type of the nucleotide (A, T, C, and G). Originally, determining the actual genotype at each locus using the Sanger method relied on radioactive labeling of ddNTPs read via autoradiography of the sequencing gel that contained four adjacent lanes (one per nucleotide type). Radioactive labels were replaced with fluorescent labels and automated laser-induced fluorescence detection methods in the 1980s and 1990s. Current Sanger capillary sequencing methods use four different fluorophores in a single reaction; the sorting of DNA fragments by their molecular weight is performed using capillary electrophoresis, and less DNA is required for sequencing reactions due to multiple rounds of primer extensions coupled with the development of more sensitive detection

402

Molecular Genetics Methods for Developmental Scientists

systems. Although Sanger sequencing remains the gold standard for targeted DNA sequencing, the application of this method in large-scale human projects (i.e., with a large number of individuals or multiple long DNA fragments) is limited by its relatively high cost and low throughput. Modern sequencing platforms outperform Sanger sequencing by a factor of 100–1,000 in terms of daily bp throughput while reducing the cost of sequencing to .10–4% of Sanger sequencing costs when sequencing long (over 1 Mb) regions (Kircher & Kelso, 2010). These methods are typically referred to as next-generation DNA sequencing to reflect the paradigmatic change associated with the advent of commercially available high-throughput sequencing solutions in mid 2000s. For example, the 454 sequencing platform (Roche, Base) was the first non-Sanger technology to sequence an individual human genome (Wheeler et al., 2008). Unlike Sanger sequencing, this approach relies on pyrosequencing (Rothberg & Leamon, 2008) and can be used to sequence the whole genome or target regions of the genome. During DNA library preparation for pyrosequencing, the DNA that has been cut into short (1–1.2 kb) fragments is separated into single strands, and ligated to universal adapter primers. These adapter-ligated fragments are joined with microscopic beads (about 1 micron in diameter, with ∼1 fragment per bead), which are then captured within droplets of oil–water emulsion, amplified using a PCR reaction (so that each droplet contains millions of copies of the original DNA fragment immobilized on the bead surface), and deposited into the picotiter plate (i.e., a plate that contains multiple wells that act as small test tubes) that contains fixed-location wells that can hold up to 1 bead each. Pyrosequencing reactions are then performed by sequentially washing single-nucleotide reagents and reaction enzymes over the plate (separately for each nucleotide). The actual sequencing relies on the detection of primer extension DNA polymerase activity using the so-called enzymatic cascade that generates light through the release of inorganic pyrophosphate when a specific (to the reaction) nucleotide is incorporated in a growing DNA strand. The pyrophosphate is converted to ATP by the enzyme sulfurylase, and a chemiluminescent enzyme, luciferase, then converts luciferin to oxyluciferin using ATP as a substrate. Light emission is monitored using cameras with fiber-optic bundles glued to a charge-coupled device (CCD). The emission is recorded at each step of the reaction and processed offline to infer the genotype at each nucleotide position. Illumina’s Genome Analyzer (also refer to as Solexa sequencing technology) is based on the four-color

sequencing-by-synthesis approach that uses reversible terminator-based chemistry similar to that of Sanger sequencing (Bentley et al., 2008). For this platform, the DNA library also consists of adaptor-flanked fragments that are nevertheless shorter than those used with the 454 platform (generally up to several hundred bp). The DNA template is amplified using a bridge PCR method. In bridge PCR, both forward and reverse primers complementary to the adaptors’ sequences are hybridized to a solid surface (flow cell) that contains eight independent channels. The immobilized DNA is then amplified to form spatially condensed clusters that contain multiple (∼1,000) copies of the original template fragment in the form of arched single-stranded molecules. The resulting molecules are used as templates for DNA polymerase-based extension reactions that incorporate fluorophore-dyed terminator dNTPs into the synthesized DNA strand. Once a terminator dNTP has been incorporated, a CCD camera captures the fluorescent signal produced by laser excitation to identify the terminator nucleotide. The terminator group and the fluorophore are then removed for the next synthesis cycle (i.e., to determine the identity of the next base pair). Both 454 sequencing and Genome Analyzer platforms can be used to sequence the whole genome by cutting it into smaller fragments and sequencing each of the fragments in parallel in the same reaction (massively parallel sequencing). This approach, called next-generation whole-genome sequencing (WGS), can be used effectively in three different applications: mutation discovery for Mendelian diseases that involve studying multiple affected individuals, discovery of the genetic architecture of complex traits and disorders, and clinical diagnostics of childhood disorders (Goldstein et al., 2013). The first two cases generally utilize designs familiar to the reader from the linkage and association sections of this chapter. Thus, DNA sequencing can be performed using samples of unrelated individuals that represent cases–controls or extremes of the quantitative phenotype variation continuum; allele frequencies can then be compared between the two cohorts at all genotyped loci that exhibit variation. In family-based designs, samples of related individuals are used to both detect mutations within the shared regions of the genome in extended pedigrees and reveal de novo mutations in family trios. Given the high cost of DNA sequencing, some studies only sequence affected individuals. Although causal mutations (i.e., for certain Mendelian disorders) can be identified based on their functional annotation and sharing among only the affected individuals, such studies sometimes utilize borrow controls from public repositories of sequencing data that have become available recently.

DNA Sequencing: From Sanger Sequencing to High-Throughput Next-Generation Sequencing Technologies

In this case, causal variant identification relies on multistep variant annotation (e.g., by obtaining information on the functional class and predicted effects of the variant) and filtration (e.g., by excluding common variants and mutations found in populations reportedly unaffected by the disorder in question). Note that DNA sequencing methods not only detect single nucleotide variants but also are capable of discovering small structural variants (indels), although the latter represent a certain bioinformatic challenge. While Sanger sequencing is primarily used for targeted region sequencing (i.e., to sequence genes and regions under the linkage peaks or around significant association signals), WGS studies are similar to GWAS studies in the sense that they adopt a hypothesis-free approach by interrogating the whole genome. Given the high cost of next-generation sequencing in the context of association studies, however, targeted rather than whole-genome applications of next-generation sequencing have become popular in recent years. These applications and designs use DNA sequence capture chemistry to isolate regions of interest that are then subjected to sequencing. One of the most popular applications of targeted next-generation sequencing is whole-exome sequencing (WES) that is aimed at sequencing only the coding regions of the DNA by performing exome capture reactions prior to sequencing. Given the cumulative abundance of rare coding variants of potentially high functional impact in the human genome and the coding nature of most mutations underlying Mendelian disorders, WES has recently become the state-of-the-art tool for Mendelian disorder gene discovery (Bamshad et al., 2011). Next-generation sequencing studies rely not only on an impressive set of recent technological but also on computational developments that permit analysis of extremely large data sets. Recall that massively parallel sequencing means that a multitude of genomic fragments are sequenced in parallel. For example, Illumina’s HiSeq 2000 systems that are frequently used for WGS/WES applications are based on the Solexa method described above and generate 600 Gb of data in 12 days; these data include 3 billion ∼100 bp reads per run. The generation of raw sequencing reads is only the first step in what became one of the most computationally intensive bionformatic paradigms in biological science. After extensive data preprocessing aimed at removal of low-quality reads, read trimming and removal of adapter sequences, the reads needs to be aligned to the reference genome. Since the reference sequence of the human genome is available, millions of short reads of unknown location generated by DNA sequencing platforms can be aligned (or mapped) to the

403

reference genome. A variety of tools have been developed (Li & Homer, 2010), and their use typically requires both a certain expertise in bioinformatics and the availability of computational resources (e.g., a computing cluster). Most alignment algorithms construct reference genome indexes, read sequence indexes, or both. These indexes are then used in conjunction with statistical inference algorithms that find the part of the genome that corresponds to each read. Additional quality control procedures are typically employed after the reads have been aligned and sorted according their location, and are aimed, for example, at increasing the quality of indel detection by performing local realignment (or de novo assembly) for regions that contain indels. The third essential step is called variant calling and is performed on multiple overlapping aligned reads (e.g., typically with a goal of obtaining ∼20x coverage at each genomic location). At the time of the writing of this chapter, a comprehensive bioinformatics pipeline developed by the Broad Institute at MIT has become the gold standard in WGS/WES variant calling. This pipeline uses a freely available Genome Analysis Toolkit (McKenna et al., 2010) to perform variant calling using HaplotypeCaller, an algorithm that uses a combination of de novo assembly and a Hidden Markov Model to identify possible haplotypes in the data and assign sample genotypes. GATK has been shown to outperform other variant calling tools in the majority of current human applications of next-generation sequencing methodologies (Ghoneim, Myers, Tuttle, & Paciorkowski, 2014; Liu, Han, Wang, Gelernter, & Yang, 2013). Once the genotypes have been assigned to samples, the data produced by DNA sequencing can be analyzed using statistics similar to those developed for single nucleotide and structural variants described in the section on association studies. Given the large amount of data produced by WGS/WES studies (e.g., an individual genome is expected to have ∼20,000 departures from the reference sequence), next-generation sequencing studies heavily rely on variant filtration, annotation, and prioritization to reduce the search space for the causal variant. Some of these methods use algorithms we referred to before when describing the evaluation of the functional impact of SNPs in association studies; these and other algorithms are frequently used in combination with each other to integrate data on the predicted effects of the genetic variant (e.g., using amino-acid substitution and phylogenetic conservation analysis; Hu et al., 2013) with other data such as the similarity between the phenotype in question with that of genetically modified model organisms (e.g., mouse; Robinson et al., 2014), or the presence of the variant in public databases or linkage

404

Molecular Genetics Methods for Developmental Scientists

data (Zhang et al., 2013). Over 200 tools have been developed in the past decade for the analysis of WGS/WES data. A comprehensive overview of these tools is available elsewhere (Pabinger et al., 2014). The statistical power of classic single-marker association tests is low for rare genetic variants, and while large samples (similar and exceeding those for GWAS studies) are required for studies of complex traits and common disorders, they are currently unattainable given the high cost18 of next-generation sequencing. A new generation of statistical tests for rare-variant association testing has been developed to address the needs of next-generation sequencing studies. These methods include region- and gene-based multimarker tests that collapse rare variants into genetic scores (burden tests) and are useful when many rare variants are causal and produce effects in the same direction, and variance-component or random-effects methods that evaluate distributions of (weighted) aggregated score statistics of individual variants instead of collapsing them. These are more useful in case the region contains multiple non-causal variants and variants have different directions of association (Lee, Abecasis, Boehnke, & Lin, 2014; Moutsianas & Morris, 2014). Specific association methods have been developed for family-based sequencing studies, and these methods have been shown to outperform population-based association methods in case of dichotomous traits, while population-based tests offer higher power for continuous traits (Ionita-Laza, Lee, Makarov, Buxbaum, & Lin, 2013). While low statistical power remains the main limitation of next-generation sequencing studies of human traits and disorders, the reader should in general be cautious when interpreting results from WGS/WES studies. First, technical errors can dramatically affect the false positive rate of next-generation sequencing studies. For example, one of the biggest technical challenges of next-generation sequencing is the presence of multiple repetitive DNA segments in the human genome that lead to ambiguities in read mapping and variant detection (Treangen & Salzberg, 2012). Second, variant filtration typically requires some 18 The arrival of Illumina’s HiSeq X Ten ultra-high-throughput sequencer in 2014 marked a milestone in DNA sequencing prices, which have been dropping since the advent of next-generation sequencing platforms: i.e., the cost decreased to ∼$1,000 per genome (compare with $100,000,000 in 2001 and ∼$50,000 in 2010; van Dijk, Auger, Jaszczyszyn, & Thermes, 2014). It is important to note, however, that this estimate is based on projected sequencing of 18,000 genomes per year over a four-year operational period, and only several HiSeq X Ten sequencers have been installed worldwide at the time of the writing of this chapter.

sort of decision criteria to be imposed on the data set, and the researcher risks excluding the causal variant from the search set. Third, certain types of genetic variants of particular interest to researchers such as loss of function variants or variants located in regulatory regions are actually enriched for artifacts (MacArthur et al., 2012) or have low sequence coverage (Wang, Wei, Lam, & Wang, 2011), and, correspondingly, higher coverage (e.g., 60x coverage) and cost are required to accurately infer the genotype. Fourth, the same population ancestry and stratification concerns that apply to GWAS association studies also apply to WGS/WES studies. Fifth, current implementations of next-generation sequencing are limited with respect to their ability to accurately detect structural variants larger than indels (Zhao, Wang, Wang, Peilin, & Zhao, 2013). Finally, unlike the field of GWAS studies that has seen the emergence of study design, genotyping, analysis, replication, and reporting standards in recent years, the field of next-generation sequencing at the moment lacks these community-approved standards (Pulit, Leusink, Menelaou, & de Bakker, 2014), which significantly complicates the interpretation of next-generation sequencing studies that have already produced an extraordinary amount of data. Next-generation sequencing revolutionized the field of genetics, and this technology offers the field of developmental psychopathology a new tool for identifying the genetic architecture of complex traits in a comprehensive fashion. For example, Sanders and colleagues (2012) performed a WES study of de novo single nucleotide mutations in 238 families from the Simons Simplex Collection, a cohort of pedigrees with at least one child affected with autism spectrum disorder (ASD), two unaffected parents, and (in 200 families) an unaffected sibling. This study estimated that disruptive de novo mutations are found in 14% of probands with ASD and showed that rare non-synonymous disruptive mutations in brain-expressed genes (in particular, SCN2A, KATNAL2, and CHD8) are associated with ASD with effect sizes similar to those for previously identified large CNVs. WGS/WES studies are rising, slowly replacing (or at least augmenting) GWAS studies of complex traits, and while the challenges outlined in the previous paragraph limit the utility of WGS/WES-generated data, next-generation sequencing is likely to become the new standard in molecular genetic studies of complex traits, signifying the transition to the post-GWAS era or next-generation sequencing era. In this section, we reviewed only two out of several commercially available next-generation sequencing platforms, and the interested reader is welcome to further familiarize

Gene Expression Profiling

themselves with these platforms by referring to specialized sources (Harismendy et al., 2009; Mardis, 2013). The next-generation sequencing era will likely be replaced by the next-next (or third-) generation sequencing era grounded in recently developed technologies that do not rely on the amplification of DNA fragments that leads to polymerase errors during library construction, preferential amplification of certain regions over others, and dilution of DNA modifications (e.g., DNA methylation). Single-molecule next-next-generation sequencing approaches such as nanopore sequencing (detection of DNA bases based on their effect on an electrical current or optical signal during the passing of the molecule or the bases through small natural or synthetic nanopores) can potentially overcome some limitations of next-generation sequencing platforms by offering affordable, quicker, and accurate (e.g., due to the absence of the amplification step and longer resulting reads) solutions for DNA sequencing (Schadt, Turner, & Kasarskis, 2010). These methods will likely bring a host of new methodological and analytical challenges with them. Until that time has come, however, current WGS/WES studies rely on next-generation sequencing platforms and approaches described above.

GENE EXPRESSION PROFILING As we mentioned in the beginning of this chapter, variation in gene expression underlies such key developmental processes as cellular proliferation, differentiation, and plasticity in response to both external and internal factors. Recall that the first stage in protein synthesis, transcription, results in the synthesis of an mRNA molecule that is used as a template at the translation stage. Thus, the concentration of a particular RNA product is indicative of the level of the expression of the corresponding gene in the studied tissue. The full set of gene transcripts (transcribed from genomic DNA) in a cell at a particular stage of development and their quantity is called a transcriptome, and gene expression analysis (also called transcriptome profiling) has become one of the most widely applied methods in developmental biology. The main objective of gene expression analyses is to establish patterns of differential gene expression between different samples that might be obtained from the same group of individuals (e.g., different areas of the brain in a within-subject design) or from several groups of individuals (e.g., comparing peripheral blood gene expression levels in individuals who were or were not exposed to toxic stress in a between-subject design). Correspondingly, genes that are found to be

405

differentially expressed between the two samples are likely involved in one or more etiological pathways for the trait. The majority of gene expression studies rely on the microarray technology that we covered in the SNP genotyping section (in fact, we indirectly mentioned this method when touching upon eQTLs—expression microarray data remain the main source for endophenotypic data on gene expression). Library preparation for such studies requires using RNA to synthesize a complementary DNA (cDNA) molecule that is more stable than the RNA molecule during a process known as reverse transcription. Oligonucleotide probes that are located on the microarray are partially complementary to the most unique regions of each studied gene, and the evaluation of gene expression is based on the analysis of laser-induced fluorescent signals emitted by the fluorophore-labeled amplified cDNA hybridized to the microarray surface. For example, Affymetrix GeneChip Tiling Arrays, one of the most commonly used platforms, uses multiple probes (a probe set containing 11–20 pairs of probes—each pair includes a perfect match probe and a mismatch probe that is created by changing the middle base of the perfect match probe; it is used to measure nonspecific binding and correct for technical artifacts) per gene to arrive at the aggregate signal indicative of transcript abundance in a particular sample. Note that probe intensity confounds the true hybridization signal with other signals (e.g., from cross-hybridization, nonspecific binding, or system noise), and preprocessing methods that adjust for background noise (e.g., by subtracting the mismatch probe intensities from perfect match intensities) and batch effects are necessarily applied to the signal before further analyses. Gene expression microarray studies rely on the comparison of normalized expression patterns between different samples. For a long time, fold change (the A/B ratio where A and B are intensity levels for a particular gene transcript for two samples) has been used as the main statistic to identify differentially expressed genes with relatively arbitrary fold change value (e.g. >2) cutoffs. However, this approach can produce biased estimates of gene expression and is insensitive to differential expression of genes that have extremely high as well as extremely low expression values (intensities), and inferential statistical methods such as the two-sample t-test (as well as its robust modifications that are less sensitive to outliers) have been proposed to evaluate differences in average expression levels between samples. Like GWAS studies, gene expression studies face the familiar problems of multiple comparisons and low statistical power (e.g., requiring ∼32 samples per group to detect a twofold change in the 75% least variable genes

406

Molecular Genetics Methods for Developmental Scientists

at 90% power and p = .001; Wei, Li, & Bumgamer, 2004; note that the current cost for a whole-genome microarray expression profiling is ∼$500–750 per sample), and similar approaches have been taken to account (correct) for multiple comparisons or reduce the number of tests by analyzing differential expressions of sets rather than individual genes. For example, the Gene Set Enrichment Analysis (GSEA) focuses on sets of genes that share a common biological function, genomic location, or regulation (Subramanian et al., 2005) instead of identifying individual genes that are differentially expressed in two or more samples. In addition to partially ameliorating the multiple comparisons problem, GSEA has a number of other advantages over single-gene analyses such improved signal-to-noise ratio and its ability to identify and formally evaluate changes in pathways and processes relevant to the studied phenomenon. In addition to microarray methods, sequence-based approaches have been developed as a method for gene expression profiling. These include early methods that used Sanger sequencing of cDNA and, most recently, RNA sequencing (RNA-Seq). RNA-Seq methods rely on the application of existing high-throughput sequencing technologies (e.g., 454 sequencing reviewed above) to cDNA libraries. RNA-Seq approaches, compared with microarrays, have low background signal given successful mapping of cDNA to the reference transcript sequence and, correspondingly, detect expression levels in the greater dynamic fold range; they are also more accurate at quantifying expression levels, leading to a higher degree of reproducibility (Marioni, Mason, Mane, Stephens, & Gilad, 2008; Wang, Gerstein, & Snyder, 2009). However, sample library preparation and RNA/cDNA fragmentation procedures involved in it can produce specific biases with respect to expression quantification, and special care should be exercised during sample library construction and data analysis to account for these biases. Transcription levels are typically inferred using the total number of reads mapped onto gene exons normalized by the length of exons that can be uniquely mapped (Morin et al., 2008). Like next-generation DNA sequencing studies, RNA-Seq studies face a number of bioinformatic challenges such as the necessity to perform extensive quality control procedures (e.g., to correct for sequencing platform biases), dependence of statistical power on read depth which is related to both gene expression level and its length (i.e., longer transcripts typically have larger numbers of reads at the same expression level), with appreciable effects on the results of expression profile comparisons

(Bullard, Purdom, Hansen, & Dudoit, 2010; Tarazona, García-Alcalde, Dopazo, Ferrer, & Conesa, 2011). Like DNA sequencing studies, RNA-Seq studies in human also greatly depend on the availability of the reference genome builds and the accumulating data regarding the spectrum of transcripts found in human cells. The presence of alternative transcripts (i.e., when the same gene gives rise to multiple transcripts/proteins due to a process known as alternative splicing) and problems aligning reads spanning exon junctions significantly complicate the mapping stage of the RNA-Seq studies; correspondingly, instead of the reference genomic DNA sequence, these studies sometimes use de novo assembly of transcripts or specific exon junction-modified sequence libraries to map the raw reads. Read mapping, normalization, expression level summarization and analysis of differential expression represent the main stages of bioinformatic analyses of RNA-Seq data, and multiple alternative computational approaches and tools have been proposed to tackle issues specific to RNA-Seq studies. These tools are comprehensively reviewed elsewhere (Bullard et al., 2010; Oshlack, Robinson, & Young, 2010; Soneson & Delorenzi, 2013). Gene expression studies investigate expression levels in specific tissues (e.g., whole blood or brain gray matter). This specificity is one of the key features of expression studies, but it calls for extra care during study design, sample preparation, data analysis, and interpretation of results. Human expression studies that focus on complex neuropsychiatric and developmental phenotypes depend on the availability of relevant tissue samples. A natural strategy in this type of study is to examine differential gene expression in tissues relevant for the phenotype—that is, the brain. However, that requires justification of the selection of particular brain regions for the analysis that itself is based on data obtained from post mortem brain samples. In addition to ethical and practical concerns associated with obtaining these samples, such studies face the issue of post mortem RNA stability, and ensuring RNA integrity in these progressively degrading samples is of pivotal importance for precise expression quantification (Birdsill, Walker, Lue, Sue, & Beach, 2011). In the absence of post-mortem samples, it would be tempting to evaluate global differential gene expression patterns between two samples using whole blood samples, as peripheral blood is easier to acquire. However, whole blood gene expression levels are not perfectly correlated with gene expression levels in central nervous system tissues (Sullivan, Fan, & Perou, 2006). Correspondingly, researchers should be extra careful when interpreting results obtained from peripheral

Conclusion

blood samples, especially with respect to genes that are not expressed in either brain or blood, or, in general, show differential expression between blood and brain. Sample type selection represents just one of multiple issues in studying gene expression in developmental populations. The human transcriptome is dynamic and its temporal dynamics reflect regulatory (re)organization of gene expression throughout the lifespan, which is especially pronounced in the human brain prenatally and neonatally (Colantuoni et al., 2011; Kang et al., 2011). The dynamic nature of the transcriptome, coupled with its tissue specificity (e.g., high variability between different brain regions and even different cortical layers) effectively requires researchers to carefully examine the hypothesis space generated by the combination of these factors and their research question. However, as the field matures, refines its methodology and produces enough publicly available data (e.g., as is realized in the BrainAtlas, http://www .brainspan.org, a public resource for spatiotemporal expression patterns across anatomical brain structures and different developmental stages), it will likely significantly advance our understanding of the biological machinery underlying both typical and atypical development.

FUTURE DIRECTIONS Within the next few years (or at least for the duration prior to the publication of the next volume of this Handbook), the applications of molecular-genetic findings in the developmental science in general and developmental psychopathology in particular will likely only grow in popularity. Therefore, it is particularly important that the field introduces high standards for the design and methods employed in such studies. Combining different methodologies in the same study, although challenging and costly, will likely generate a more comprehensive view of the molecular genetic architecture of developmental psychopathology traits. This statement applies to the combination of different phenotyping methods (e.g., by augmenting binary disease phenotypes with quantitative subphenotypes or endophenotypes such as neuroimaging data), as well as molecular genetic methods (e.g., by combining GWAS and WGS/WES methodologies to increase resolution and power). It is also important, as it is in studies of somatic disorders, to encourage different groups interested in the same phenotype, either typical or atypical, to homogenize the methods used for generating both the phenotypic

407

and molecular data used in the developmental sciences. Finally, the field should exercise high standards for the procedures followed when both original and secondary (partially or completely) analyzed data are available for independent meta- and mega-studies. Homogenization of phenotyping, genotyping, and analyses enables big science projects like the formation of large consortia that are aimed at investigating the genetic etiology of traits based on the pooled data from multiple populations and research groups—a crucial advantage given that small sample sizes remain one of the main obstacles in trait-associated genetic variant discovery.

CONCLUSION The main goal of this chapter was to provide the readers with a succinct yet comprehensive sampling of the molecular genetics methods used to study human traits, in particular developmental and developmental psychopathology traits. These methods range from the discovery of the regions of the genome shared by individuals affected with a particular disorder to the quantification of the degree to which a particular gene (or genes) is expressed in a particular tissue. This variety of methods can be applied in a targeted (candidate) or a whole-genome fashion, and the choice ultimately depends on researchers’ agendas and prior knowledge about the phenotype and its genetic and molecular etiology. We attempted to build our narrative to familiarize developmental researchers with this complex methodological toolbox in a non-technical way. Given the scope of the field, however, interested readers are welcome to consult with a variety of original research and review sources referenced in this chapter. In this section of the chapter, we would like to offer several brief concluding remarks. Although the title of the chapter (correctly) assumes that all of the reviewed molecular genetics methods can be used to study human development, the extent to which they can be applied to developmental phenotypes in general, and developmental psychopathology in particular, has been limited for multiple reasons. Recruiting developmental populations for interdisciplinary, biologically oriented studies and characterizing them with respect to complex phenotypes is inherently a difficult task, undermined by uncertainties in and costs of phenotyping, issues that deserve a methodological treatment in their own separate chapter. Perhaps even more difficult is the integration of multiple levels of analysis that such studies demand.

408

Molecular Genetics Methods for Developmental Scientists

As the fields of developmental science and developmental psychopathology are just crystallizing, it is not surprising that comprehensive frameworks for such integration are yet to be developed. The research methods and studies reviewed in this chapter point at their complexity but also at their heterogeneity with respect to the establishment of gold standards in the field. For example, GWAS studies are probably familiar to most researchers interested in the genetic architecture of complex traits, and although the accumulation of evidence produced by these studies has been rapid, the field has had enough time to reflect upon its methodological but also theoretical limitations and challenges. On the other hand, DNA and RNA sequencing studies that use next-generation sequencing approaches have had a much shorter history, and the authors of this chapter can attest to the significant amount of methodological inconsistencies across these studies. These inconsistencies are in part related to the field’s responses to the technological and computational/analytical challenges it is only now becoming familiar with. They are likely to be at least partially resolved in the near future, and public data sharing, although plagued with its own problems (e.g., ethical concerns), will likely play a pivotal role in this process. Crucially, however, most of the progress in complex trait genetics has been tied to studies of somatic disease or adult psychopathology rather than psychological phenotypes and developmental psychopathology. In that respect, the field is open to the new generation of developmental scientists and the interdisciplinary studies that will necessarily capitalize on this progress but that will also face their own challenges. For better or for worse, molecular genetic methods for the research of complex phenotypes have been embedded into the developmental sciences just as strongly as any method in the neighboring sciences. This, of course, assumes that whatever rigor is characteristic of the utilization of these methods in their native disciplines, such as genetics or genomics, the same should be exercised in the developmental sciences. This, of course, does not assume that there should not be a certain amount of reasonable adaptation of these methods with regard to how and why these methods should be utilized in developmental sciences, but rather that their adaptation should be carried out with caution and with direct and indirect references to the lessons learned from the utilization of these methods in their native disciplines, genetic epidemiology and genetics and genomics.

One of the most important steps in our progress toward understanding the molecular underpinnings of human development via molecular genetics methods is the facilitation of cross-talk between specialists trained in complementary albeit compartmentalized subfields that serve as the foundation for the emergent fields mentioned above. To this end, it is our hope that we have succeeded in providing an overview that can be used as a starting point for the new generation of developmental scientists, but also for seasoned researchers coordinating (or hoping to coordinate) such interdisciplinary research.

REFERENCES 1000 Genomes Project Consortium. (2010). A map of human genome variation from population-scale sequencing. Nature, 467(7319), 1061–1073. doi: 10.1038/nature09534 Abraham, J. E., Maranian, M. J., Spiteri, I., Russell, R., Ingle, S., Luccarini, C., . . . Caldas, C. (2012). Saliva samples are a viable alternative to blood samples as a source of DNA for high throughput genotyping. BCM Medical Genomics, 5, 19. doi: 10.1186/1755–8794–5–19 Addington, A. M., & Rapoport, J. L. (2012). Annual research review: Impact of advances in genetics in understanding developmental psychopathology. Journal of Child Psychology and Psychiatry, 53(5), 510–518. doi: 10.1111/j.1469–7610.2011.02478.x Alberts, B., Johnson, A., Lewis, J., Raff, M., Roberts, K., & Walter, P. (2002). Molecular biology of the cell (4th ed.). New York, NY: Garland Science. Alkan, C., Coe, B. P., & Eichler, E. E. (2011). Genome structural variation discovery and genotyping. Nature Reviews Genetics, 12(5), 363–376. doi: 10.1038/nrg2958 Almasy, L., & Blangero, J. (1998). Multipoint quantitative-trait linkage analysis in general pedigrees. American Journal of Human Genetics, 62(5), 1198–1211. doi: 10.1086/301844 Ameres, S. L., & Zamore, P. D. (2013). Diversifying microRNA sequence and function. Nature Reviews Molecular Cell Biology, 14, 475–488. doi: 10.1038/nrm3611 Armitage, P. (1955). Tests for linear trends in proportions and frequencies. Biometrics, 11, 375–386. Asher, J. H., Morell, R., & Friedman, T. B. (1991). Waardenburg syndrome (WS): The analysis of a single family with a WS1 mutation showing linkage to RFLP markers on human chromosome 2q. American Journal of Human Genetics, 48(1), 43–52. Bahlo, M., Stankovich, J., Danoy, P., Hickey, P. F., Taylor, B. V., Browning, S. R., . . . Rubio, J. P. (2010). Saliva-derived DNA performs well in large-scale, high-density single-nucleotide polymorphism microarray studies. Cancer Epidemiology Biomarkers & Prevention, 19(3), 794–798. doi: 10.1158/1055–9965.EPI-09–0812 Balding, D. J. (2006). A tutorial on statistical methods for population association studies. Nature Reviews Genetics, 7, 781–791. doi: 10.1038/nrg1916 Bamshad, M. J., Ng, S. B., Bigham, A. W., Tabor, H. K., Emond, M. J., Nickerson, D. A., & Shendure, J. (2011). Exome sequencing as a tool for Mendelian disease gene discovery. Nature Review Genetics, 12(11), 745–755. doi: 10.1038/nrg3031

References Barnes, C., Plagnol, V., Fitzgerald, T., Redon, R., Marchini, J., Clayton, D., & Hurles, M. E. (2008). A robust statistical method for case–control association testing with copy number variation. Nature Genetics, 40(10), 1245–1252. doi: 10.1038/ng.206 Begum, F., Ghosh, D., Tseng, G. C., & Feingold, E. (2012). Comprehensive literature review and statistical considerations for GWAS meta-analysis. Nucleic Acids Research, 40(9), 3777–3784. doi: 10.1093/nar/gkr1255 Bentley, D. R., Balasubramanian, S., Swerdlow, H. P., Smith, G. P., Milton, J., Brown, C. G., . . . Smith, A. J. (2008). Accurate whole human genome sequencing using reversible terminator chemistry. Nature, 456(7218), 53–59. doi: 10.1038/nature07517 Birdsill, A., Walker, D., Lue, L., Sue, L., & Beach, T. (2011). Postmortem interval effect on RNA and gene expression in human brain tissue. Cell and Tissue Banking, 12(4), 311–318. doi: 10.1007/s10561–010–9210–8 Botstein, D., & Risch, N. (2003). Discovering genotypes underlying human phenotypes: Past successes for mendelian disease, future approaches for complex disease. Nature Genetics, 33(supplement), 228–237. doi: 10.1038/ng1090 Boycott, K. M., Vanstone, M. R., Bulman, D. E., & MacKenzie, A. E. (2013). Rare-disease genetics in the era of next-generation sequencing: Discovery to translation. Nature Reviews Genetics, 14(10), 681–691. doi: 10.1038/nrg3555 Brunner, H. G., & Oti, M. (2007). The modular ntaure of genetic diseases. Clinical Genetics, 71, 1–11. doi: 10.1111/j.1399–0004.2006.00708.x Bullard, J. H., Purdom, E., Hansen, K. D., & Dudoit, S. (2010). Evaluation of statistical methods for normalization and differential expression in mRNA-Seq experiments. BMC Bioinformatics, 11(94).doi: 10.1186/1471–2105–11–94 Calafell, F., Shuster, A., Speed, W. C., Kidd, J. R., & Kidd, K. K. (1998). Short tandem repeat polymorphism evolution in humans. European Journal of Human Genetics, 6, 38–49. Campbell, C. D., & Eichler, E. E. (2013). Properties and rates of germline mutations in humans. Trends in Genetics, 29(10), 575–584. doi: 10.1016/j.tig.2013.04.005 Cannon, T. D., & Keller, M. C. (2006). Endophenotypes in the genetic analyses of mental disorders. Annual Review of Clinical Psychology, 2, 267–290. doi: 10.1146/annurev.clinpsy.2.022305.095232 Cantor, R. M., Lange, K., & Sinsheimer, J. S. (2010). Prioritizing GWAS results: A review of statistical methods and recommendations for their application. The American Journal of Human Genetics, 86(1), 6–22. doi: 10.1016/j.ajhg.2009.11.017 Cardon, L. R., & Palmer, L. J. (2003). Population stratification and spurious allelic association. Lancet, 361(9357), 598–604. doi: 10.1016/ S0140–6736(03)12520–2 Chen, W.-M., Manichaikul, A., & Rich, S. S. (2009). A generalized family-based association test for dichotomous traits. American Journal of Human Genetics, 85(3), 364–376. doi: 10.1016/j.ajhg.2009.08.003 Chu, J.-H., Rogers, A., Ionita-Laza, I., Darvishi, K., Mills, D. L., Lee, C., & Raby, B. A. (2013). Copy number variation genotyping using family information. BMC Bioinformatics, 14, 157. doi: 10.1186/ 1471–2105–14–157 Colantuoni, C., Lipska, B. K., Ye, T., Hyde, T. M., Tao, R., Leek, J. T., . . . Kleinman, J. E. (2011). Temporal dynamics and genetic control of transcription in the human prefrontal cortex. Nature, 478(7370), 519–523. doi: 10.1038/nature10524 Colella, S., Yau, C., Taylor, J. M., Mirza, G., Butler, H., Clouston, P., . . . Ragoussis, J. (2007). QuantiSNP: An objective Bayes hidden-Markov model to detect and accurately map copy number

409

variation using SNP genotyping data. Nucleic Acids Research, 35(6), 2013–2025. doi: 10.1093/nar/gkm076 Cooper, G. M., Nickerson, D. A., & Eichler, E. E. (2007). Mutational and selective effects on copy-number variants in the human genome. Nature Genetics, 39(S), 22–29. doi: 10.1038/ng2054 Cooper, G. M., Zerr, T., Kidd, J. M., Eichler, E. E., & Nickerson, D. A. (2008). Systematic assessment of copy number variant detection via genome-wide SNP genotyping. Nature Genetics, 40(10), 1199–1203. doi: 10.1038/ng.236 Cordell, H. J., Barratt, B. J., & Clayton, D. G. (2004). Case/pseudocontrol analysis in genetic association studies: A unified framework for detection of genotype and haplotype associations, gene–gene and gene–environment interactions, and parent-of-origin effects. Genetic Epidemiology, 26(3), 167–185. doi: 10.1002/gepi.10307 Dadd, T., Weale, M. E., & Lewis, C. M. (2009). A critical evaluation of genomic control methods for genetic association studies. Genetic Epidemiology, 33(4), 290–298. doi: 10.1002/gepi.20379 Das, S. K., Austin, M. D., Akana, M. C., Deshpande, P., Cao, H., & Xiao, M. (2010). Single molecule linear analysis of DNA in nano-channel labeled with sequence specific fluorescent probes. Nucleic Acids Research, 38(18), e177. doi: 10.1093/nar/gkq673 de Bakker, P. I. W., Yelensky, R., Pe’er, I., Gabriel, S. B., Daly, M. J., & Altshuler, D. (2005). Efficiency and power in genetic association studies. Nature Genetics, 37(11), 1217–1223. doi: 10.1038/ng1669 Devlin, B., Roeder, K., & Wasserman, L. (2001). Genomic control, a new approach to genetic-based association studies. Theoretical Population Biology, 60(3), 155–166. doi: 10.1006/tpbi.2001.1542 Dimas, A. S., Deutsch, S., Stranger, B. E., Montgomery, S. B., Borel, C., Attar-Cohen, H., . . . Antonarakis, S. E. (2009). Common regulatory variation impacts gene expression in a cell type–dependent manner. Science, 325(5945), 1246–1250. doi: 10.1126/science.1174148 Diskin, S. J., Li, M., Hou, C., Yang, S., Glessner, J., Hakonarson, H., . . . Wang, K. (2008). Adjustment of genomic waves in signal intensities from whole-genome SNP genotyping platforms. Nucleic Acids Research, 36(19), e126. doi: 10.1093/nar/gkn556 Dixon, A. L., Liang, L., Moffatt, M. F., Chen, W., Heath, S., Wong, K. C. C., . . . Cookson, W. O. C. (2007). A genome-wide association study of global gene expression. Nature Genetics, 39(10), 1202–1207. doi: 10.1038/ng2109 Drysdale, C. M. (2000). Complex promoter and coding region b2adrenergic receptor haplotypes alter receptor expression and predict in vivo responsiveness. Proceedings of National Academy of Sciences, 97, 10483–10488. doi: 10.1073/pnas.97.19.10483 Duggal, P., Gillanders, E. M., Holmes, T. N., & Bailey-Wilson, J. E. (2008). Establishing an adjusted p-value threshold to control the family-wide Type 1 error in genome wide association studies. BMC Genomics, 9, 516. doi: 10.1186/1471–2164–9–516 Eleftherohorinou, H., Andersson-Assarsson, J. C., Walters, R. G., El-Sayed Moustafa, J. S., Coin, L., Jacobson, P., . . . Falchi, M. (2011). famCNV: copy number variant association for quantitative traits in families. Bioinformatics, 27(13), 1873–1875. doi: 10.1093/bioinformatics/btr264 ENCODE Project Consortium. (2012). An integrated encyclopedia of DNA elements in the human genome. Nature, 489(7414), 57–74. doi: 10.1038/nature11247 Fan, J. B., Gunderson, K. L., Bibikova, M., Yeakley, J. M., Chen, J., Wickham Garcia, E., . . . Barker, D. (2006). Illumina universal bead arrays. In K. Alan & O. Brian (Eds.), Methods in enzymology (Vol. 410, pp. 57–73). New York, NY: Academic Press.

410

Molecular Genetics Methods for Developmental Scientists

Fisher, S. E., Vargha-Khadem, F., Watkins, K. E., Monaco, A. P., & Pembrey, M. E. (1998). Localisation of a gene implicated in a severe speech and language disorder. Nature Genetics, 18(2), 168–170. doi: 10.1038/ng0298–168 Forer, L., Schonherr, S., Weissensteiner, H., Haider, F., Kluckner, T., Gieger, C., . . . Kloss-Brandstatter, A. (2010). CONAN: Copy number variation analysis software for genome-wide association studies. BMC Bioinformatics, 11, 318. doi: 10.1186/1471–2105–11–318 Gao, X., Becker, L. C., Becker, D. M., Starmer, J. D., & Province, M. A. (2010). Avoiding the high Bonferroni penalty in genome-wide association studies. Genetic Epidemiology, 34(1), 100–105. doi: 10.1002/gepi.20430 Gerstein, M. B., Bruce, C., Rozowsky, J. S., Zheng, D., Du, J., Korbel, J. O., . . . Snyder, M. (2007). What is a gene, post-ENCODE? History and updated definition. Genome Research, 17(6), 669–681. doi: 10.1101/gr.6339607 Ghoneim, D. A., Myers, J. R., Tuttle, E., & Paciorkowski, A. R. (2014). Comparison of insertion/deletion calling algorithms on human next-generation sequencing data. BMC Research Notes, 7(1), 864. doi: 10.1186/1756–0500–7–864 Gibson, G. (2012). Rare and common variants: Twenty arguments. Nature Reviews Genetics, 13(2), 135–145. doi: 10.1038/nrg3118 Girirajan, S., Brkanac, Z., Coe, B. P., Baker, C., Vives, L., Vu, T. H., . . . Eichler, E. E. (2011). Relative burden of large CNVs on a range of neurodevelopmental phenotypes. PLoS Genetics, 7(11), e1002334. doi: 10.1371/journal.pgen.1002334 Goldstein, D. B., Allen, A., Keebler, J., Margulies, E. H., Petrou, S., Petrovski, S., & Sunyaev, S. (2013). Sequencing studies in human genetics: design and interpretation. Nature Review Genetics, 14(7), 460–470. doi: 10.1038/nrg3455 Goring, H. H. H., Curran, J. E., Johnson, M. P., Dyer, T. D., Charlesworth, J., Cole, S. A., . . . Blangero, J. (2007). Discovery of expression QTLs using large-scale transcriptional profiling in human lymphocytes. Nature Genetics, 39(10), 1208–1216. doi: 10.1038/ng2119 Greshock, J., Feng, B., Nogueira, C., Ivanova, E., Perna, I., Nathanson, K., . . . Chin, L. (2007). A comparison of DNA copy number profiling platforms. Cancer Research, 67(21), 10173–10180. doi: 10.1158/0008–5472.CAN-07–2102 Grigorenko, E. L., Ngorosho, D., Jukes, M., & Bundy, D. (2006). Reading in able and disabled readers from around the world: Same or different? An illustration from a study of reading-related processes in a Swahili sample of siblings. Journal of Research in Reading, 29(1), 104–123. doi: 10.1111/j.1467–9817.2006.00295.x Grigorenko, E. L., Wood, F. B., Meyer, M. S., Hart, L., Speed, W. C., Shuster, A., & Pauls, D. L. (1997). Susceptibility loci for distinct components of developmental dyslexia on chromosomes 6 and 15. American Journal of Human Genetics, 60, 27–39. Gudbjartsson, D. F., Jonasson, K., Frigge, M. L., & Kong, A. (2000). Allegro, a new computer program for multipoint linkage analysis. Nature Genetics, 25(1), 12–13. doi: 10.1038/75514 Guichoux, E., Lagache, L., Wagner, S., Chaumeil, P., LÉGer, P., Lepais, O., . . . Petit, R. J. (2011). Current trends in microsatellite genotyping. Molecular Ecology Resources, 11(4), 591–611. doi: 10.1111/j.1755–0998.2011.03014.x Ha, N.-T., Freytag, S., & Bickeboeller, H. (2014). Coverage and efficiency in current SNP chips. European Journal of Human Genetics, 22(9), 1124–1130. doi: 10.1038/ejhg.2013.304 Han, B., Kang, H. M., & Eskin, E. (2009). Rapid and accurate multiple testing correction and power estimation for millions of correlated markers. PLoS Genetics, 5(4), e1000456. doi: 10.1371/journal.pgen.1000456

Harismendy, O., Ng, P. C., Strausberg, R. L., Wang, X., Stockwell, T. B., Beeson, K. Y., . . . Frazer, K. A. (2009). Evaluation of next generation sequencing platforms for population targeted sequencing studies. Genome Biology, 10(3), 1–13. doi: 10.1186/gb-2009–10–3-r32 Haseman, J. K., & Elston, R. C. (1972). The investigation of linkage between a quantitative trait and a marker locus. Behavior Genetics, 2, 3–19. doi: 10.1007/BF01066731 Hattersley, A. T., & McCarthy, M. I. (2005). What makes a good genetic association study? Lancet, 366(9493), 1315–1323. doi: 10.1016/S0140–6736(05)67531–9 Hindorff, L. A., Sethupathy, P., Junkins, H. A., Ramos, E. M., Mehta, J. P., Collins, F. S., & Manolio, T. A. (2009). Potential etiologic and functional implications of genome-wide association loci for human diseases and traits. Proceedings of the National Academy of Sciences, 106(23), 9362–9367. doi: 10.1073/pnas.0903103106 Hood, L., & Rowen, L. (2013). The human genome project: Big science transforms biology and medicine. Genome Medicine, 5(79), 1–9. doi: 10.1186/gm483 Horvath, S., Xu, X., & Laird, N. M. (2001). The family based association test method: strategies for studying general genotype–phenotype associations. European Journal of Human Genetics, 9(4), 301–306. Hou, L., & Zhao, H. (2013). A review of post-GWAS prioritization approaches. Frontiers in Genetics, 4, e280. doi: 10.3389/fgene .2013.00280 Hu, H., Huff, C. D., Moore, B., Flygare, S., Reese, M. G., & Yandell, M. (2013). VAAST 2.0: Improved variant classification and disease-gene identification using a conservation-controlled amino acid substitution matrix. Genetic Epidemiology, 37(6), 622–634. doi: 10.1002/gepi.21743 Hublin, C., Partinen, M., Koskenvuo, M., & Kaprio, J. (2011). Heritability and mortality risk of insomnia-related symptoms: A genetic epidemiologic study in a population-based twin cohort. Sleep, 34(7), 957–964. doi: 10.5665/SLEEP.1136 Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., . . . Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167(7), 748–751. doi: 10.1176/appi .ajp. 2010.09091379 International HapMap Consortium. (2005). A haplotype map of the human genome. Nature, 437(7063), 1299–1320. doi: 10.1038/ nature04226 International HapMap Consortium. (2007). A second generation human haplotype map of over 3.1 million SNPs. Nature, 449(7164), 851–861. doi: 10.1038/nature06258 International Human Genome Sequencing Consortium. (2001). Initial sequencing and analysis of the human genome. Nature, 409(6822), 860–921. doi: 10.1038/35057062 International Human Genome Sequencing Consortium. (2004). Finishing the euchromatic sequence of the human genome. Nature, 431(7011), 931–945. doi: 10.1038/nature03001 Ioannidis, J. P. A. (2007). Non-replication and inconsistency in the genome-wide association setting. Human Heredity, 64, 203–213. doi: 10.1159/000103512 Ioannidis, J. P. A., Ntzani, E. E., Trikalinos, T. A., & ContopoulosIoannidis, D. G. (2001). Replication validity of genetic association studies. Nature Genetics, 29(3), 306–309. doi: 10.1038/ng749 Ionita-Laza, I., Lee, S., Makarov, V., Buxbaum, J. D., & Lin, X. (2013). Family-based association tests for sequence data, and comparisons with population-based association tests. European Journal of Human Genetics, 21(10), 1158–1162. doi: 10.1038/ejhg.2012.308 Ionita-Laza, I., Perry, G. H., Raby, B. A., Klanderman, B., Lee, C., Laird, N. M., . . . Lange, C. (2008). On the analysis of copy-number variations in genome-wide association studies: a translation of the

References family-based association test. Genetic Epidemiology, 32(3), 273–284. doi: 10.1002/gepi.20302 Itsara, A., Cooper, G. M., Baker, C., Girirajan, S., Li, J., Absher, D., . . . Eichler, E. E. (2009). Population analysis of large copy number variants and hotspots of human genetic disease. American Journal of Human Genetics, 84(2), 148–161. doi: 10.1016/j.ajhg.2008.12.014 Itsara, A., Wu, H., Smith, J. D., Nickerson, D. A., Romieu, I., London, S. J., & Eichler, E. E. (2010). De novo rates and selection of large copy number variation. Genome Research, 20(11), 1469–1481. doi: 10.1101/gr.107680.110 Johnson, R., Nelson, G. W., Troyer, J. L., Lautenberger, J. A., Kessing, B. D., Winkler, C. A., & O’Brien, S. J. (2010). Accounting for multiple comparisons in a genome-wide association study (GWAS). BMC Genomics, 11, 724. doi: 10.1186/1471–2164–11–724 Kang, H. J., Kawasawa, Y. I., Cheng, F., Zhu, Y., Xu, X., Li, M., . . . Sestan, N. (2011). Spatio-temporal transcriptome of the human brain. Nature, 478(7370), 483–489. doi: 10.1038/nature10523 Kang, H. M., Sul, J. H., Service, S. K., Zaitlen, N. A., Kong, S.-y., Freimer, N. B., . . . Eskin, E. (2010). Variance component model to account for sample structure in genome-wide association studies. Nature Genetics, 42(4), 348–354. doi: 10.1038/ng.548 Keightley, P. D. (2012). Rates and fitness consequences of new mutations in humans. Genetics, 190(2), 295–304. doi: 10.1534/genetics.111.134668 Kim, J.-H., Hu, H.-J., Yim, S.-H., Bae, J. S., Kim, S.-Y., & Chung, Y.-J. (2012). CNVRuler: A copy number variation-based case– control association analysis tool. Bioinformatics, 28(13), 1790–1792. doi: 10.1093/bioinformatics/bts239 Kircher, M., & Kelso, J. (2010). High-throughput DNA sequencing— Concepts and limitations. Bioessays, 32, 524–536. doi: 10.1002/ bies.200900181 Klein, R. J., Zeiss, C., Chew, E. Y., Tsai, J.-Y., Sackler, R. S., Haynes, C., . . . Hoh, J. (2005). Complement factor H polymorphism in age-related macular degeneration. Science, 308(5720), 385–389. doi: 10.1126/science.1109557 Kornilov, S. A., Rakhlin, N., Koposov, R. A., Lee, M., Yrigollen, C., Chang, J. T., . . . Grigorenko, E. L. (2015). A genome-wide association and whole-exome sequencing study of developmental language disorder in an isolated Russian population. Manuscript submitted for publication. Kovas, Y., Hayiou-Thomas, M. E., Oliver, B., Dale, P. S., Bishop, D. V. M., & Plomin, R. (2005). Genetic influences in different aspects of language development: The etiology of language skills in 4.5-year-old twins. Child Development, 76(3), 632–651. doi: 10.1111/j.1467–8624 .2005.00868.x Kruglyak, L., Daly, M. J., Reeve-Daly, M. P., & Lander, E. S. (1996). Parametric and nonparametric linkage analysis: A unified multipoint approach. American Journal of Human Genetics, 58(6), 1347–1363. Kruglyak, L., & Lander, E. S. (1995). Complete multipoint sib-pair analysis of qualitative and quantitative traits. American Journal of Human Genetics, 57, 439–454. Lai, C. S. L., Fisher, S. E., Hurst, J. A., Vargha-Khadem, F., & Monaco, A. P. (2001). A forkhead-domain gene is mutated in a severe speech and language disorder. Nature, 413(6855), 519–523. doi: 10.1038/35097076 Laird, N. M., & Lange, C. (2006). Family-based designs in the age of large-scale gene-association studies. Nature Reviews Genetics, 7, 385–394. doi: 10.1038/nrg1839 Laird, N. M., & Lange, C. (2011). The fundamentals of modern statistical genetics. New York, NY: Springer. Lambert, C. G., & Black, L. J. (2012). Learning from our GWAS mistakes: From experimental design to scientific method. Biostatistics, 13(2), 195–203. doi: 10.1093/biostatistics/kxr055

411

Lange, C., DeMeo, D., Silverman, E. K., Weiss, S. T., & Laird, N. M. (2004). PBAT: Tools for family-based association studies. American Journal of Human Genetics, 74(2), 367–369. doi: 10.1086/381563 Lange, C., DeMeo, D. L., & Laird, N. M. (2002). Power and design considerations for a general class of family-based association tests: Quantitative traits. American Journal of Human Genetics, 71(6), 1330–1341. doi: 10.1086/344696 Lange, C., Silverman, E. K., Xu, X., Weiss, S. T., & Laird, N. M. (2003). A multivariate family-based association test using generalized estimating equations: FBAT-GEE. Biostatistics, 4(2), 195–206. doi: 10.1093/biostatistics/4.2.195 Lazzeroni, L. C., Lu, Y., & Belitskaya-Levy, I. (2014). P-values in genomics: Apparent precision masks high uncertainty. Molecular Psychiatry, 19(12), 1336–1340. doi: 10.1038/mp. 2013.184 Lee, S., Abecasis, G. a. R., Boehnke, M., & Lin, X. (2014). Rare-variant association analysis: Study designs and statistical tests. American Journal of Human Genetics, 95(1), 5–23. doi: 10.1016/j.ajhg.2014 .06.009 Levy-Sakin, M., & Ebenstein, Y. (2013). Beyond sequencing: optical mapping of DNA in the age of nanotechnology and nanoscopy. Current Opinion in Biotechnology, 24(4), 690–698. doi: 10.1016/j.copbio .2013.01.009 Li, H., & Homer, N. (2010). A survey of sequence alignment algorithms for next-generation sequencing. Briefings in Bioinformatics, 11(5), 473–483. doi: 10.1093/bib/bbq015 Li, M.-X., Gui, H.-S., Kwan, Johnny S. H., & Sham, Pak C. (2011). GATES: A rapid and powerful gene-based association test using extended Simes procedure. American Journal of Human Genetics, 88(3), 283–293. doi: 10.1016/j.ajhg.2011.01.019 Listgarten, J., Lippert, C., Kadie, C. M., Davidson, R. I., Eskin, E., & Heckerman, D. (2012). Improved linear mixed models for genome-wide association studies. Nature Methods, 9(6), 525–526. doi: 10.1038/nmeth.2037 Little, J., Higgins, J. P. T., Ioannidis, J. P. A., Moher, D., Gagnon, F., von Elm, E., . . . Birkett, N. (2009). Strengthening the reporting of genetic association studies (STREGA)—An extension of the STROBE statement. PLoS Medicine, 6(2), e1000022. doi: 10.1371/ journal.pmed.1000022 Liu, G., Wang, Y., & Wong, L. (2010). FastTagger: an efficient algorithm for genome-wide tag SNP selection using multi-marker linkage disequilibrium. BMC Bioinformatics, 11, 66. doi: 10.1186/1471 –2105–11–66 Liu, J. Z., McRae, A. F., Nyholt, D. R., Medland, S. E., Wray, N. R., Brown, K. M., . . . Macgregor, S. (2010). A versatile gene-based test for genome-wide association studies. American Journal of Human Genetics, 87(1), 139–145. doi: 10.1016/j.ajhg.2010.06.009 Liu, X., Han, S., Wang, Z., Gelernter, J., & Yang, B.-Z. (2013). Variant callers for next-generation sequencing data: A comparison study. PLoS ONE, 8(9), e75619. doi: 10.1371/journal.pone.0075619 Lynch, M. (2010). Rate, molecular spectrum, and consequences of human mutation. Proceedings of the National Academy of Sciences, 107(3), 961–968. doi: 10.1073/pnas.0912629107 MacArthur, D. G., Balasubramanian, S., Frankish, A., Huang, N., Morris, J., Walter, K., . . . Tyler-Smith, C. (2012). A systematic survey of loss-of-function variants in human protein-coding genes. Science, 335(6070), 823–828. doi: 10.1126/science.1215040 Mah, J. T. L., Low, E. S. H., & Lee, E. (2011). In silico SNP analysis and bioinformatics tools: A review of the state of the art to aid drug discovery. Drug Discovery Today, 16(17–18), 800–809. doi: 10.1016/j.drudis.2011.07.005 Manolio, T. A., Collins, F. S., Cox, N. J., Goldstein, D. B., Hindorff, L. A., Hunter, D. J., . . . Visscher, P. M. (2009). Finding the missing

412

Molecular Genetics Methods for Developmental Scientists

heritability of complex diseases. Nature, 461(7265), 747–753. doi: 10.1038/nature08494 Mardis, E. (2013). Next-generation sequencing platforms. Annual Review of Analytical Chemistry, (6), 287–303. doi: 10.1146/annurev-anchem -062012–092628 Marioni, J. C., Mason, C. E., Mane, S. M., Stephens, M., & Gilad, Y. (2008). RNA-seq: An assessment of technical reproducibility and comparison with gene expression arrays. Genome Research, 18(9), 1509–1517. doi: 10.1101/gr.079558.108 Marlow, A. J., Fisher, S. E., Francks, C., MacPhie, I. L., Cherny, S. S., Richardson, A. J., . . . Cardon, L. R. (2003). Use of multivariate linkage analysis for dissection of a complex cognitive trait. American Journal of Human Genetics, 72(3), 561–570. doi: 10.1086/368201 Mattick, J. S. (2004). RNA regulation: A new genetics? Nature Reviews Genetics, 5(4), 316–323. doi: 10.1038/nrg1321 McGinnis, R., Shifman, S., & Darvasi, A. (2002). Power and efficiency of the TDT and case–control design for association scans. Behavior Genetics, 32(2), 135–144. doi: 10.1023/A:1015205924326 McKenna, A., Hanna, M., Banks, E., Sivachenko, A., Cibulskis, K., Kernytsky, A., . . . DePristo, M. A. (2010). The genome analysis toolkit: A MapReduce framework for analyzing next-generation DNA sequencing data. Genome Research, 20(9), 1297–1303. doi: 10.1101/gr.107524.110 Metzgar, D., Bytof, J., & Wills, C. (2000). Selection against frameshift mutations limits microsatellite expansion in coding DNA. Genome Research, 10, 72–80. Miller, D. T., Adam, M. P., Aradhya, S., Biesecker, L. G., Brothman, A. R., Carter, N. P., . . . Ledbetter, D. H. (2010). Consensus statement: Chromosomal microarray is a first-tier clinical diagnostic test for individuals with developmental disabilities or congenital anomalies. American Journal of Human Genetics, 86(5), 749–764. doi: 10.1016/j.ajhg.2010.04.006 Mills, R. E., Luttig, C. T., Larkins, C. E., Beauchamp, A., Tsui, C., Pittard, W. S., & Devine, S. E. (2006). An initial map of insertion and deletion (INDEL) variation in the human genome. Genome Research, 16(9), 1182–1190. doi: 10.1101/gr.4565806 Moreau, Y., & Tranchevent, L. (2012). Computational tools for prioritizing candidate genes: boosting disease gene discovery. Nature Reviews Genetics, 13, 523–536. doi: 10.1038/nrg3253 Morin, R. D., Bainbridge, M., Fejes, A., Hirst, M., Krzywinski, M., Pugh, T. J., . . . Marra, M. A. (2008). Profiling the HeLa S3 transcriptome using randomly primed cDNA and massively parallel short-read sequencing. BioTechniques, 45(1), 81–94. doi: 10.2144/000112900 Moutsianas, L., & Morris, A. P. (2014). Methodology for the analysis of rare genetic variation in genome-wide association and re-sequencing studies of complex human traits. Briefings in Functional Genomics, 13(5), 362–370. doi: 10.1093/bfgp/elu012 Murphy, A., Won, S., Rogers, A., Chu, J.-H., Raby, B. A., & Lange, C. (2010). On the genome-wide analysis of copy number variants in family-based designs: methods for combining family-based and population-based information for testing dichotomous or quantitative traits, or completely ascertained samples. Genetic Epidemiology, 34(6), 582–590. doi: 10.1002/gepi.20515 Nakaoka, H., & Inoue, I. (2009). Meta-analysis of genetic association studies: methodologies, between-study heterogeneity and winner’s curse. Journal of Human Genetics, 54(11), 615–623. doi: 10.1038/jhg.2009.95 Nelson, S. C., Doheny, K. F., Pugh, E. W., Romm, J. M., Ling, H., Laurie, C. A., . . . Laurie, C. C. (2013). Imputation-based genomic coverage assessments of current human genotyping arrays. G3: Genes| Genomes|Genetics, 3(10), 1795–1807. doi: 10.1534/g3.113.007161

Nicolae, D. L., Gamazon, E., Zhang, W., Duan, S., Dolan, M. E., & Cox, N. J. (2010). Trait-Associated SNPs are more likely to be eQTLs: Annotation to enhance discovery from GWAS. PLoS Genetics, 6(4), e1000888. doi: 10.1371/journal.pgen.1000888 Oostlander, A. E., Meijer, G. A., & Yistra, B. (2004). Microarray-based comparative genomic hybridization and its applications in human genetics. Clinical Genetics, 66, 488–495. doi: 10.1111/j.1399–0004 .2004.00322.x Oshlack, A., Robinson, M. D., & Young, M. D. (2010). From RNA-seq reads to differential expression results. Genome Biology, 11(220), 1–10. doi: 10.1186/gb-2010–11–12–220 Pabinger, S., Dander, A., Fischer, M., Snajder, R., Sperk, M., Efremova, M., . . . Trajanoski, Z. (2014). A survey of tools for variant analysis of next-generation genome sequencing data. Briefings in Bioinformatics, 15(2), 256–278. doi: 10.1093/bib/bbs086 Park, J.-H., Wacholder, S., Gail, M. H., Peters, U., Jacobs, K. B., Chanock, S. J., & Chatterjee, N. (2010). Estimation of effect size distribution from genome-wide association studies and implications for future discoveries. Nature Genetics, 42(7), 570–575. doi: 10.1038/ng.610 Patnala, R., Clements, J., & Batra, J. (2013). Candidate gene association studies: a comprehensive guide to useful in silico tools. BMC Genetics, 14(39), 1–11. doi: 10.1186/1471–2156–14–39 Pearson, H. (2006). Genetics: What is a gene? Nature, 441, 398–401. doi: 10.1038/441398a Philibert, R. A., Zadorozhnyaya, O., Beach, S. R. H., & Brody, G. H. (2008). A comparison of the genotyping results using DNA obtained from blood and saliva. Psychiatric genetics, 18(6), 275–281. doi: 10.1097/YPG.0b013e3283060f81 Pinto, D., Darvishi, K., Shi, X., Rajan, D., Rigler, D., Fitzgerald, T., . . . Feuk, L. (2011). Comprehensive assessment of array-based platforms and calling algorithms for detection of copy number variants. Nature Biotechnology, 29(6), 512–520. doi: 10.1038/nbt.1852 Pritchard, J. K. (2001). Are rare variants responsible for susceptibility to complex diseases? American Journal of Human Genetics, 69, 124–137. doi: 10.1086/321272 Pritchard, J. K., & Rosenberg, N. A. (1999). Use of unlinked genetic markers to detect population stratification in association studies. American Journal of Human Genetics, 65(1), 220–228. doi: 10.1086/302449 Pritchard, J. K., Stephens, M., & Donnelly, P. (2000). Inference of population structure using multilocus genotype data. Genetics, 155(2), 945–959. Pulit, S. L., Leusink, M., Menelaou, A., & de Bakker, P. I. W. (2014). Association claims in the sequencing era. Genes, 5(1), 196–213. doi: 10.3390/genes5010196 Reich, D. E., & Lander, E. S. (2001). On the allelic spectrum of human disease. Trends in Genetics, 17(9), 502–510. doi: 10.1016/S0168 –9525(01)02410–6 Riethoven, J. J. (2010). Regulatory regions in DNA: Promoters, enhancers, silencers, and insulators. Methods in Molecular Biology, 674, 33–42. doi: 10.1007/978–1–60761–854–6_3 Risch, N. (2000). Searching for genetic determinants in the new millennium. Nature, 405, 847–856. doi: 10.1038/35015718 Robinson, P. N., Köhler, S., Oellrich, A., Project, S. M. G., Wang, K., Mungall, C. J., . . . Smedley, D. (2014). Improved exome prioritization of disease genes through cross-species phenotype comparison. Genome Research, 24(2), 340–348. doi: 10.1101/gr.160325.113 Rothberg, J. M., & Leamon, J. H. (2008). The development and impact of 454 sequencing. Nature Biotechnology, 26(10), 1117–1124. doi: 10.1038/nbt1485

References Russell, P. J. (2009). iGenetics: A molecular approach. NY: BenjaminCummings Publishing Company. Sanchez-Mora, C., Ramos-Quiroga, J. A., Bosch, R., Corrales, M., Garcia-Martinez, I., Nogueira, M., . . . Ribases, M. (2015). Case– control genome-wide association study of persistent attention-deficit hyperactivity disorder identifies FBXO33 as a novel susceptibility gene for the disorder. Neuropsychopharmacology. 40(4), 915–926. doi: 10.1038/npp. 2014.267 Sanders, S. J., Murtha, M. T., Gupta, A. R., Murdoch, J. D., Raubeson, M. J., Willsey, A. J., . . . State, M. W. (2012). De novo mutations revealed by whole-exome sequencing are strongly associated with autism. Nature, 485(7397), 237–241. doi: 10.1038/nature10945 Sandkuijl, L. A. (1989). Analysis of affected sib-pairs using information from extended families. In R. C. Elston, M. A. Spence, S. E. Hodge, & J. W. MacCluer (Eds.), Multipoint mapping and linkage based upon affected pedigree members: Genetic analysis workshop 6 (p. 117–122). New York: Alan R. Liss. Sanger, F., & Coulson, A. R. (1975). A rapid method for determining sequences in DNA by primed synthesis with DNA polymerase. Journal of Molecular Biology, 94(3), 441–448. doi: 10.1016/0022 –2836(75)90213–2 Sanger, F., Nicklen, S., & Coulson, A. R. (1977). DNA sequencing with chain-terminating inhibitors. Proceedings of the National Academy of Sciences, 74(12), 5463–5467. Schadt, E. E., Turner, S., & Kasarskis, A. (2010). A window into third-generation sequencing. Human Molecular Genetics, 19, 227–240. doi: 10.1093/hmg/ddq416 Schaid, D. J., Rowland, C. M., Tines, D. E., Jacobson, R. M., & Poland, G. A. (2002). Score tests for association between traits and haplotypes when linkage phase is ambiguous. American Journal of Human Genetics, 70(2), 425–434. doi: 10.1086/338688 Schork, N. J., Murray, S. S., Frazer, K. A., & Topol, E. J. (2009). Common vs. rare allele hypotheses for complex diseases. Current Opinion in Genetics & Development, 19(3), 212–219. doi: 10.1016/ j.gde.2009.04.010 Schutte, N. M., Hansell, N. K., de Geus, E. J., Martin, B., Wright, M. J., & Smit, D. J. A. (2013). Heritability of resting state EEG functional connectivity patterns. Twin Research and Human Genetics, 16(5), 962–969. doi: 10.1017/thg.2013.55 Sebastiani, P., Solovieff, N., Puca, A., Hartley, S. W., Melista, E., Andersen, S., . . . Perls, T. T. (2010). Genetic signatures of exceptional longevity in humans. Science (retracted). doi: 10.1126/science .1190532 Sham, P. C., Purcell, S., Cherny, S. S., & Abecasis, G. R. (2002). Powerful regression-based quantitative-trait linkage analysis of general pedigrees. American Journal of Human Genetics, 71(2), 238–253. doi: 10.1086/341560 Shen, G.-Q., Abdullah, K., & Wang, Q. (2009). The TaqMan method for SNP genotyping. In A. A. Komar (Ed.), Single nucleotide polymorphisms (Vol. 578, pp. 293–306). New York, NY: Humana Press. Shen, H., Li, J., Zhang, J., Xu, C., Jiang, Y., Wu, Z., . . . Deng, H.-W. (2013). Comprehensive characterization of human genome variation by high coverage whole-genome sequencing of forty four Caucasians. PLoS ONE, 8(4), e59494. doi: 10.1371/journal.pone.0059494 Shen, R., Fan, J.-B., Campbell, D., Chang, W., Chen, J., Doucet, D., . . . Oliphant, A. (2005). High-throughput SNP genotyping on universal bead arrays. Mutation Research/Fundamental and Molecular Mechanisms of Mutagenesis, 573(1–2), 70–82. doi: 10.1016/j.mrfmmm .2004.07.022 Sicotte, H., Rider, D. N., Poland, G. A., Dhiman, N., & Kocher, J. A. (2011). SNPPicker: High quality tag SNP selection across multiple

413

populations. BMC Bioinformatics, 12(129), 1–19. doi: 10.1186/1471 –2105–12–129 Silventoinen, K., Sammalisto, S., Perola, M., Boomsma, D. I., Cornes, B. K., Davis, C., . . . Kaprio, J. (2003). Heritability of adult body height: A comparative study of twin cohorts in eight countries. Twin Research and Human Genetics, 6(05), 399–408. doi: 10.1375/ 136905203770326402 Smith, S. D., Grigorenko, E. L., Willcutt, E., Pennington, B. F., Olson, R. K., & DeFries, J. C. (2010). Etiologies and molecular mechanisms of communication disorders. Journal of Developmental & Behavioral Pediatrics, 31(7), 555–563. doi: 10.1097/DBP.0b013e3181ee3d9e Sobreira, N. L. M., Cirulli, E. T., Avramopoulos, D., Wohler, E., Oswald, G. L., Stevens, E. L., . . . Goldstein, D. B. (2010). Whole-genome sequencing of a single proband together with linkage analysis identifies a Mendelian disease gene. PLoS Genetics, 6(6), e1000991. doi: 10.1371/journal.pgen.1000991 Soneson, C., & Delorenzi, M. (2013). A comparison of methods for differential expression analysis of RNA-seq data. BCM Bioinformatics, 14, 91. doi: 10.1186/1471–2105–14–91 Spencer, C. C. A., Su, Z., Donnelly, P., & Marchini, J. (2009). Designing genome-wide association studies: Sample size, power, imputation, and the choice of genotyping chip. PLoS Genetics, 5(5), e1000477. doi: 10.1371/journal.pgen.1000477 Spielman, R. S., McGinnis, R. E., & Ewens, W. J. (1993). Transmission test for linkage disequilibrium: the insulin gene region and insulindependent diabetes mellitus (IDDM). American Journal of Human Genetics, 52, 506–516. Stephens, M., & Donnelly, P. (2003). A comparison of Bayesian methods for haplotype reconstruction from population genotype data. American Journal of Human Genetics, 73(5), 1162–1169. doi: 10.1086/379378 Stranger, B. E., Forrest, M. S., Dunning, M., Ingle, C. E., Beazley, C., Thorne, N., . . . Dermitzakis, E. T. (2007). Relative impact of nucleotide and copy number variation on gene expression phenotypes. Science, 315(5813), 848–853. doi: 10.1126/science.1136678 Subramanian, A., Tamayo, P., Mootha, V. K., Mukherjee, S., Ebert, B. L., Gillette, M. A., . . . Mesirov, J. P. (2005). Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences, 102(43), 15545–15550. doi: 10.1073/pnas.0506580102 Sullivan, P. F., Fan, C., & Perou, C. M. (2006). Evaluating the comparability of gene expression in blood and brain. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 141B(3), 261–268. doi: 10.1002/ajmg.b.30272 Sun, S., Greenwood, C. M. T., & Neal, R. M. (2007). Haplotype inference using a Bayesian hidden Markov model. Genetic Epidemiology, 31(8), 937–948. doi: 10.1002/gepi.20253 Tallal, P., Hirsch, L. S., Realpe-Bonilla, T., Miller, S., Brzustowicz, L. M., Bartlett, C., & Flax, J. F. (2001). Familial aggregation in specific language impairment. Journal of Speech, Language & Hearing Research, 44, 1172–1182. doi: 10.1044/1092–4388(2001/091) Tarazona, S., García-Alcalde, F., Dopazo, J., Ferrer, A., & Conesa, A. (2011). Differential expression in RNA-seq: A matter of depth. Genome Research, 21(12), 2213–2223. doi: 10.1101/gr.124321.111 Tavtigian, S. V., Simard, J., Teng, D. H. F., Abtin, V., Baumgard, M., Beck, A., . . . Cannon-Albright, L. A. (2001). A candidate prostate cancer susceptibility gene at chromosome 17p. Nature Genetics, 27(2), 172–180. doi: 10.1038/84808 Teare, M. W., & Barrett, J. H. (2005). Genetic linkage studies. Lancet, 366, 1036–1044. doi: 10.1016/S0140–6736(05)67382–5 Tenesa, A., & Haley, C. S. (2013). The heritability of human disease: Estimation, uses and abuses. Nature Reviews Genetics, 14(2), 139–149. doi: 10.1038/nrg3377

414

Molecular Genetics Methods for Developmental Scientists

Thomson, G. (1995). Mapping disease genes: family-based association studies. American Journal of Human Genetics, 57, 487–498. Tomblin, B. J. (1989). Familial concentration of developmental language impairment. Journal of Speech and Hearing Disorders, 54, 287–295. doi: 10.1044/jshd.5402.287 Tranchevent, L.-C., Capdevila, F. B., Nitsch, D., De Moor, B., De Causmaecker, P., & Moreau, Y. (2011). A guide to Web tools to prioritize candidate genes. Briefings in Bioinformatics, 12(1), 22–32. doi: 10.1093/bib/bbq007 Treangen, T. J., & Salzberg, S. L. (2012). Repetitive DNA and next-generation sequencing: computational challenges and solutions. Nature Reviews Genetics, 13(1), 36–46. doi: 10.1038/nrg3117 Trucco, E. M., Villafuerte, S., Heitzeg, M. M., Burmeister, M., & Zucker, R. A. (2014). Rule breaking mediates the developmental association between GABRA2 and adolescent substance abuse. Journal of Child Psychology and Psychiatry, 55(12), 1372–1379. doi: 10.1111/jcpp. 12244 Tuvblad, C., Bezdjian, S., Raine, A., & Baker, L. A. (2014). The heritability of psychopathic personality in 14- to 15-year-old twins: A multirater, multimeasure approach. Psychological Assessment, 26(3), 704–716. doi: 10.1037/a0036711 van Dijk, E. L., Auger, H., Jaszczyszyn, Y., & Thermes, C. (2014). Ten years of next-generation sequencing technology. Trends in Genetics, 30(9), 418–426. doi: 10.1016/j.tig.2014.07.001 Visscher, P. M., Brown, M. A., McCarthy, M. I., & Yang, J. (2012). Five years of GWAS discovery. American Journal of Human Genetics, 90, 7–24. doi: 10.1016/j.ajhg.2011.11.029 Wang, K., Li, M., Hadley, D., Liu, R., Glessner, J., Grant, S. F. A., . . . Bucan, M. (2007). PennCNV: An integrated hidden Markov model designed for high-resolution copy number variation detection in whole-genome SNP genotyping data. Genome Research, 17(11), 1665–1674. doi: 10.1101/gr.6861907 Wang, W., Wei, Z., Lam, T.-W., & Wang, J. (2011). Next generation sequencing has lower sequence coverage and poorer SNP-detection capability in the regulatory regions. Science Reports, 1(55). doi: 10.1038/srep00055 Wang, W. Y. S., Barratt, B. J., Clayton, D. G., & Todd, J. A. (2005). Genome-wide association studies: theoretical and practical concerns. Nature Reviews Genetics, 6(2), 109–118. doi: 10.1038/nrg1522 Wang, Z., Gerstein, M., & Snyder, M. (2009). RNA-Seq: A revolutionary tool for transcriptomics. Nature Reviews Genetics, 10(1), 57–63. doi: 10.1038/nrg2484 Weale, M. E. (2010). Quality control for genome-wide association studies. In M. R. Barnes & G. Breen (Eds.), Genetic variation (Vol. 628, pp. 341–372). New York, NY: Humana Press. Wei, C., Li, J., & Bumgamer, R. E. (2004). Sample size for detecting differentially expressed genes in microarray experiments. BMC Genomics, 5, 87. doi: 10.1186/1471–2164–5–87 Weinberg, C. R. (1999). Methods for detection of parent-of-origin effects in genetic studies of case-parents triads. American Journal of Human Genetics, 65(1), 229–235. doi: 10.1086/302466 Wellcome Trust Case Control Consortium. (2007). Genome-wide association study of 14,000 cases of seven common diseases and 3,000 shared controls. Nature, 447(7145), 661–678. doi: 10.1038/nature05911 Welter, D., MacArthur, J., Morales, J., Burdett, T., Hall, P., Junkins, H., . . . Parkinson, H. (2014). The NHGRI GWAS catalog, a curated resource of SNP-trait associations. Nucleic Acids Research, 42(Database issue), D1001–1006. doi: 10.1093/nar/gkt1229 Wheeler, D. A., Srinivasan, M., Egholm, M., Shen, Y., Chen, L., McGuire, A., . . . Rothberg, J. M. (2008). The complete genome of an individual

by massively parallel DNA sequencing. Nature, 452(7189), 872–876. doi: 10.1038/nature06884 Williams, J. T., & Blangero, J. (1999). Power of variance component linkage analysis to detect quantitative trait loci. Annals of Human Genetics, 63(6), 545–563. doi: 10.1046/j.1469–1809.1999.6360545.x Wineinger, N. E., & Tiwari, H. K. (2012). The impact of errors in copy number variation detection algorithms on association results. PLoS ONE, 7(4), e32396. doi: 10.1371/journal.pone.0032396 Wong, K. K., deLeeuw, R. J., Dosanjh, N. S., Kimm, L. R., Cheng, Z., Horsman, D. E., . . . Lam, W. L. (2007). A comprehensive analysis of common copy-number variations in the human genome. American Journal of Human Genetics, 80(1), 91–104. doi: 10.1086/510560 Wu, R., & Lin, M. (2006). Functional mapping—How to map and study the genetic architecture of dynamic complex traits. Nature Reviews Genetics, 7, 229–237. doi: 10.1038/nrg1804 Wu, R., Ma, C., & Casella, G. C. (2007). Statistical genetics of quantitative traits: Linkage, maps and QTL. New York, NY: Springer-Verlag. Xu, X., Weiss, S., Xu, X., & Wei, L. J. (2000). A unified Haseman-Elston method for testing linkage with quantitative traits. American Journal of Human Genetics, 67(4), 1025–1028. doi: 10.1086/303081 Xu, Z., Kaplan, N. L., & Taylor, J. A. (2007). Tag SNP selection for candidate gene association studies using HapMap and gene resequencing data. European Journal of Human Genetics, 15(10), 1063–1070. doi: 10.1038/sj.ejhg.5201875 Yamada, N. A., Rector, L. S., Tsang, P., Carr, E., Scheffer, A., Sederberg, M. C., . . . Brothman, A. R. (2011). Visualization of fine-scale genomic structure by oligonucleotide-based high-resolution FISH. Cytogenetic and Genome Research, 132(4), 248–254. doi: 10.1159/000322717 Yang, J., Lee, S. H., Goddard, M. E., & Visscher, P. M. (2011). GCTA: A tool for genome-wide complex trait analysis. American Journal of Human Genetics, 88(1), 76–82. doi: 10.1016/j.ajhg.2010.11.011 Yang, J., Wray, N. R., & Visscher, P. M. (2010). Comparing apples and oranges: Equating the power of case–control and quantitative trait association studies. Genetic Epidemiology, 34, 254–257. doi: 10.1002/gepi.20456 Yang, L., Neale, B. M., Liu, L., Lee, S. H., Wray, N. R., Ji, N., . . . Psychiatric, G. C. A. S. (2013). Polygenic transmission and complex neuro developmental network for attention deficit hyperactivity disorder: Genome-wide association study of both common and rare variants. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 162(5), 419–430. doi: 10.1002/ajmg.b.32169 Zeggini, E., Rayner, W., Morris, A. P., Hattersley, A. T., Walker, M., Hitman, G. A., . . . McCarthy, M. I. (2005). An evaluation of HapMap sample size and tagging SNP performance in large-scale empirical and simulated data sets. Nature Genetics, 37(12), 1320–1322. doi: 10.1038/ng1670 Zhang, L., Zhang, J., Yang, J., Ying, D., Lau, Y. l., & Yang, W. (2013). PriVar: a toolkit for prioritizing SNVs and indels from next-generation sequencing data. Bioinformatics, 29(1), 124–125. doi: 10.1093/bioinformatics/bts627 Zhao, M., Wang, Q., Wang, Q., Peilin, J., & Zhao, Z. (2013). Computational tools for copy number variation (CNV) detection using next-generation sequencing data: Features and perspectives. BCM Bioinformatics, 14, S1. doi: 10.1186/1471–2105–14-S11-S1 Zheng, Y., Heagerty, P. J., Hsu, L., & Newcomb, P. A. (2010). On combining family-based and population-based case–control data in association studies. Biometrics, 66(4), 1024–1033. doi: 10.1111/ j.1541–0420.2010.01393.x Zhou, X., Carbonetto, P., & Stephens, M. (2013). Polygenic modeling with Bayesian sparse linear mixed models. PLoS Genetics, 9(2), e1003264. doi: 10.1371/journal.pgen.1003264

References Zhou, X., & Stephens, M. (2012). Genome-wide efficient mixed-model analysis for association studies. Nature Genetics, 44(7), 821–824. doi: 10.1038/ng.2310 Zhu, M., & Zhao, S. (2007). Candidate gene identification approach: progress and challenges. International Journal of Biological Sciences, 3(7), 420–427. Ziegler, A., König, I. R., & Pahlke, F. (2010). A statistical approach to genetic epidemiology: Concepts and applications. New York, NY: Wiley-Blackwell.

415

Zogopoulos, G., Ha, K. C., Naqib, F., Moore, S., Kim, H., Montpetit, A., . . . Gallinger, S. (2007). Germ-line DNA copy number variation frequencies in a large North American population. Human Genetics, 122(3–4), 345–353. doi: 10.1007/s00439–007–0404–5 Zuk, O., Hechter, E., Sunyaev, S. R., & Lander, E. S. (2012). The mystery of missing heritability: Genetic interactions create phantom heritability. Proceedings of the National Academy of Sciences, 109(4), 1193–1198. doi: 10.1073/pnas.1119675109

Trim Size: 8.5in x 11in

Cicchetti

c11.tex printer - Volume II - 01/12/2016

CHAPTER 11

Epigenetic Mechanisms in the Development of Behavior KATHRYN HILL and TANIA L. ROTH

INTRODUCTION 416 EPIGENETICS: THE BASICS 416 EPIGENETIC FACTORS AND OUTCOMES ASSOCIATED WITH PRENATAL ENVIRONMENTAL EVENTS 418 EPIGENETIC FACTORS AND OUTCOMES ASSOCIATED WITH EARLY-LIFE POSTNATAL ENVIRONMENTAL EVENTS 421 EPIGENETIC FACTORS AND OUTCOMES ASSOCIATED WITH EVENTS LATER IN LIFE 425 EPIGENETIC FACTORS ASSOCIATED WITH BEHAVIORAL DISORDERS 427 Rett Syndrome 428

Fragile X 428 Rubinstein-Taybi 428 Schizophrenia 429 Posttraumatic Stress Disorder 429 Suicide and Mood Disorders 430 Alzheimer’s Disease 430 EVIDENCE OF EPIGENETIC TRANSMISSION OF PHENOTYPES 430 TRANSLATIONAL IMPLICATIONS 432 FUTURE DIRECTIONS 433 CONCLUDING REMARKS 434 REFERENCES 434

INTRODUCTION

and a route through which behavior could arise. The birth of epigenetics research thus has major implications for the field of developmental psychopathology and its tenets (Cicchetti, 1993, 2006; Sroufe & Rutter, 1984). In this chapter, we discuss key studies that have helped define our current understanding of epigenetics in the development of behavior. We draw on work using animal models with relevance to developmental psychopathology and studies in which the translation of these findings has been made to humans. As the development of behavior is a process that extends throughout the life span, a central theme that is also explored in this review is that epigenetic regulation of genes is not a process exclusive to developing organisms, but is an active process that continues throughout life into senescence. Throughout our chapter and within the conclusion, we also raise several questions that need to be addressed in order to advance our understanding of the link between epigenetics and behavioral outcomes.

Both genes and environmental events have long been recognized for their significant contribution to the development of behavior, and in turn, phenotypic variation among individuals. Recent work with animal models of early-life stress, depression, and addiction along with human studies focused on infant behavior and stress have revealed that genes and the environment are inextricably linked by epigenetic factors, whereby the environment can create biochemical changes that affect gene activity (Branchi, 2009; Champagne, 2010; Dudley, Li, Kobor, Kippin, & Bredy, 2011; Wong, Mill, & Fernandes, 2011). This revelation has afforded a new framework for our understanding of how psychological and social-contextual factors can interact with our biology. These factors have given us too a valuable measure informative of norms and aberrations present at the molecular level. Furthermore, these factors give us a mechanism to understand how risk and protective factors can leave their mark on our genome This work was supported by grant number 1P20GM103653– 01A1 from the National Institute of General Medical Sciences and a grant from the University of Delaware Research Foundation. 1 Color versions of Figures 11.1, 11.2 and 11.3 are available at http://onlinelibrary.wiley.com/book/10.1002/9781119125556

EPIGENETICS: THE BASICS Before discussing studies of relevance to the field of developmental psychopathology, it is necessary to understand some basics of epigenetics. In a cell’s nucleus, DNA is 416

8:24pm

Page 416

Epigenetics: The Basics

417

Figure 11.1 In the nucleus, DNA is wrapped around histone proteins, forming a DNA-protein complex called chromatin. DNA-histone interactions occur at the N-terminal tail of histones. On these tails reside many sites for epigenetic marking, including acetylation (Ac-blue), methylation (Me-red), and phosphorylation (P-gray). Acetylation occurs when there is the addition of an acetyl group via an enzyme called histone acetyltransferase (HAT). Deacetylation occurs when the acetyl group is removed by enzymes called histone deacetylases (HDACs). In a similar fashion, methyl groups can be added via histone methyltransferases (HMTs) or removed by histone demethylases (HDMs). Within the bottom portion of the figure is a schematic of DNA methylation, in which methyl groups (Me-tan) can be added to cytosine-guanine dinucleotides by DNA methyltransferase (DNMTs) enzymes. Gadd45b and TETs can facilitate active DNA demethylation. See footnote 1.

wrapped around histone proteins, forming a complex called chromatin (Figure 11.1). An easy way to consider histones is to think of them as spools and of the DNA as the thread that wraps around them. The backbone of DNA is negatively charged so it wraps around the positively charged histones. Without the histones (spools) to wrap around, DNA could not fit inside a cell nucleus. To help determine whether DNA is accessible for gene transcription, chemical modifications to histones and the DNA strands themselves govern whether chromatin structure is in a state permissive for gene transcription (open and less tightly packaged, a structure referred to as euchromatin) or not (closed and tightly packaged, or heterochromatin). The histone proteins have amino acid tails that protrude beyond the DNA, and these amino acid residues are prone to chemical modifications that include acetylation and methylation (Berger, 2007). Histone acetylation, or the addition of acetyl groups to lysine residues on a histone tail, is a process catalyzed by enzymes called histone acetyltransferases (HATs). CREB binding protein (CBP) is one known HAT regulating chromatin structure.

The addition of an acetyl group neutralizes the positive charge on histones, thereby decreasing the interaction with the negatively charged phosphates of DNA. This helps remodel (loosen) chromatin to a state permissive for gene transcription. Histone acetylation is rapid and reversible in an experience-dependent manner, but it too can be long-lived. Histone deacetylases (HDACs) are enzymes that remove the acetyl groups and reverse the effects of HATs. These help drive the state of chromatin back to a more closed form permissive to silencing gene activity. As an aside here, HDACs have much structural diversity and because of this are receiving research attention as possible targets of therapeutic interventions to increase gene activity. Though far less studied in the behaviors we will review in this chapter, we are learning that histone methylation likewise is a crucial regulator of behavioral change and this modification can either suppress or activate gene transcription depending upon which amino acid residue of the histone is targeted and the degree of methylation that occurs. For example, up to three methyl groups can

418

Epigenetics and the Development of Behavior

be added to lysines, producing mono-, di-, and trimethylation patterns (Akbarian & Huang, 2009). H3K4 diand trimethylation and H3K9 monomethylation are associated with gene transcription whereas H3K27 di- and trimethylation and H3K9 di- and trimethylation are linked with gene suppression. Histone methylation is catalyzed by enzymes called histone methyltransferases (HMTs), whereas enzymes called histone demethylases (HDMs) catalyze demethylation. Together, histone modifications help regulate gene activity by integrating numerous responses to signal biochemical cascades and repelling/recruiting chromatin remodeling and transcription factors making gene loci either more or less available to transcriptional modulation (Berger, 2007; Kouzarides, 2007). DNA methylation is another epigenetic process that affects gene activity, and as will become clear in later sections, is an epigenetic factor that has really emerged as a leading candidate biological pathway linking gene–environment interactions to long-term and even multigenerational trajectories in behavioral development. DNA methylation often occurs at cytosine residues of cytosine-guanine (CG) dinucleotides, sequences that are highly underrepresented in the genome but often clustered within so-called CpG islands located within gene regulatory regions. DNA methyltransferases (DNMTs) are enzymes that catalyze the transfer of a methyl group, and de novo DNMTs (DNMT3a and DNMT3b) methylate previously unmethylated CG sites in DNA (Moore, Le, & Fan, 2013). The maintenance DNMT isoform (DNMT1, though it appears to regulate de novo methylation in some cases) perpetuates methylation marks after cell division, regenerating the methyl-cytosine marks on the newly synthesized complementary DNA strand that arises from DNA replication (Moore et al., 2013). The predominant view in the literature is that methylation of DNA is associated with suppression of gene transcription. The precise molecular processes through which this occurs are complex, but in essence methylated cytosines in turn can bind repressor proteins, including the methyl-binding domain protein MeCP2 and HDACs (Moore et al., 2013). DNA methylation (and deacetylation of histones) promotes a higher affinity interaction between DNA and the histone core, ultimately blocking access for transcriptional machinery and thus suppressing gene transcription. It is important to note that while most evidence indicates that DNA methylation is associated with reduced gene activity, a handful of studies have indicated that DNA methylation can also be associated with transcriptional activation (Chahrour et al., 2008; Uchida et al., 2011). Regardless of suppressing or activating gene activity, DNA

methylation is recognized as one of the most stable epigenetic processes affecting gene expression. As for a way to remove the methyl groups, historically in the epigenetics field, demethylation has been viewed largely as a passive process in differentiated cells, whereby multiple rounds of cell division without DNMT-mediated re-methylation is necessary to erase epigenetic marks. The idea of whether there is active DNA demethylation in post-mitotic cells, such as neurons, has been controversial. Our current understanding of active DNA demethylation is via a DNA repair-like mechanism catalyzed by Gadd45b (Ma, Guo, Ming, & Song, 2009) or oxidation by ten-eleven translocation (TET) proteins (Guo, Su, Zhong, Ming, & Song, 2011; Williams, Christensen, & Helin, 2012). A demethylase, MBD2, may also be responsible for active DNA demethylation in some cases (Detich, Theberge, & Szyf, 2002). Though histone modifications and DNA methylation have been the most-studied epigenetic mechanisms in regards to behavior to date, increasing evidence is showing that genes (and behavior) can also be epigenetically regulated by noncoding RNAs, or RNA transcripts that have no apparent protein product. For example, microRNAs (miRNAs) are small, single-stranded RNAs with around 22 base pairs that can silence gene expression through mRNA degradation, inhibition of translation, and destabilization (Bartel, 2009). While at present we do not have a complete picture of the role of noncoding RNAs in gene regulation and behavior, it is becoming increasingly clearer that dysregulation of noncoding RNA-epigenetic pathways has relevance to understanding behavior and disease etiology (Bredy, Lin, Wei, Baker-Andresen, & Mattick, 2011; Sato, Tsuchiya, Meltzer, & Shimizu, 2011).

EPIGENETIC FACTORS AND OUTCOMES ASSOCIATED WITH PRENATAL ENVIRONMENTAL EVENTS The intrauterine environment has long been recognized for its major contributory role in optimizing growth and development of an individual, and environmental events within this context are well recognized for their ability to affect behavioral development. For example, exposure of pregnant women to war-related stressors (including dietary restrictions) in the 1940s has been associated with an increase in prevalence of schizophrenia (SZ) and anxiety disorders in their offspring (Susser, Hoek, & Brown, 1998; van Os & Selten, 1998). Maternal emotion and psychological distress during pregnancy have been shown to predict

Epigenetic Factors and Outcomes Associated With Prenatal Environmental Events

children’s later emotional behavior (O’Connor, Heron, Golding, Beveridge, & Glover, 2002). Though such data (and numerous others) provide compelling evidence that early-life factors can predispose individuals to particular behavioral outcomes over the life course, it has not been until the past decade or so that empirical evidence has yielded a plausible mechanism for this notion. Beginning in the early to mid-2000s, work using rodents where the prenatal environment could be experimentally altered began to provide direct evidence that one way prenatal events could alter developmental trajectories was indeed through bona fide epigenetic changes to genes. By altering nutritional cues within the prenatal environment while holding all other environmental factors and animals’ experiences constant, researchers began to establish causality between prenatal nutritional exposure and phenotypic outcome (Waterland, 2006; Waterland & Jirtle, 2003). In these classic studies, pregnant Agouti mouse dams were provided dietary supplements comprised of methyl donors (e.g., methionine and folate). Offspring born to mothers who had not been fed the special methyl-donor diet had developmental outcomes characterized with yellow coat colors and overeating tendencies. In contrast, offspring born to mothers who had been fed the special diet had brown coat colors and did not overeat. The difference in developmental outcomes between the cohorts of mice was shown to be attributable to DNA methylation differences at genes controlling coat color and feeding behavior. Since this seminal work, studies focusing on the developmental origin of health and disease have continued to identify links between early nutrition, epigenetic processes, and phenotypic outcomes. Additional work in rodents provides further empirical evidence of epigenetic factors linking maternal over-nutrition to adverse health outcomes in her offspring. For example, consumption of a high-fat diet during pregnancy has been shown to produce offspring that prefer sucrose and fat in adulthood, a phenotype that is accompanied by DNA hypomethylation (mean less methylation) and increased expression of several reward-related genes (including those encoding for the mu-opioid receptor and dopamine reuptake transporter) within the nucleus accumbens or prefrontal cortex, brain epicenters of appetite and reward (Vucetic, Kimmel, Totoki, Hollenbeck, & Reyes, 2010). Individuals who were prenatally exposed to famine during the Dutch hunger winter of 1944–45 have been shown, 6 decades later, to have less methylation of DNA associated with an important imprinted gene regulating body growth (insulin-like growth factor 2 [Igf2]) compared with unexposed, same-sex siblings (Heijmans et al., 2008). A recent

419

study in Gambia has suggested DNA methylation patterns in first generation offspring differ depending on season of conception, in a context where diet varies greatly according to season (Waterland et al., 2010). Finally, mothers who reportedly consumed an unbalanced diet during pregnancy (one that has been linked to elevated blood pressure and cortisol in offspring) have offspring that display altered methylation of genes involved in cortisol regulation and body growth (11𝛽-hydroxysteroid dehydrogenase type 2 [HSD2], glucocorticoid receptor [GR], and Igf2 [Drake et al., 2012]). A growing body of work has also established epigenetic factors in pathways responsible for the negative impact of environmental toxins on current and subsequent generational offspring. During embryogenesis, exposure to the endocrine disruptor Vinclozolin (achieved by exposing gestating females) causes epigenetic programming of the germ-line that has transgenerational consequences for behavior and disease for future generations. For example, behavioral studies performed on Vinclozolin-exposed F3 generation animals show increased anxiety-like behaviors, hyperactivity, and impulsivity that coincided with alterations in gene expression networks within the amygdala (Skinner, Anway, Savenkova, Gore, & Crews, 2008). In more recent work from this line of research, investigators have shown that a single exposure to Vinclozolin likewise alters male descendants’ (three generations removed) response to chronic restraint stress experienced during adolescence (Crews et al., 2012). Specifically, males with only the experience of adolescent stress (no prenatal fungicide exposure) show lower anxiety levels in an open-field test, whereas males with the cumulative experience of Vinclozolin in their lineage (i.e., their ancestors were exposed) and adolescent stress show higher levels of anxiety in the same test. Observations based on their study design are consistent with a two-hit model where exposure to chronic stress during adolescence (second hit) clearly influences subsequent brain development and behavior but they themselves are altered by ancestral exposures and epigenetic programming (first hit). In 2008, a seminal investigation highlighted the role of epigenetic mechanisms in the maladaptive effects of prenatal stress on adult hypothalamic-pituitary-adrenal (HPA) responsivity and behavior (Mueller & Bale, 2008). Researchers found that adult males born to mothers who had been subjected to gestational stress exhibited marked changes in expression of the corticotropin-releasing factor (Crf) and GR genes, increased HPA-axis responsivity, and a depressive-like phenotype (anhedonia for example). CRF is secreted by the hypothalamus in response to stress, with

420

Epigenetics and the Development of Behavior

consequent secretion of adrenocorticotropin from the anterior pituitary and glucocorticoids from the adrenal gland. Examination of the Crf gene in these adults indicated that the prenatal stress regimen had significantly reduced methylation of specific cytosines with the regulatory region of the Crf gene in both the hypothalamus and amygdala. Reduced methylation parallels their observation of increased Crf gene expression, which together contributed to the depressive-like phenotype present in the offspring. Additional work in rodent models has been consistent in showing the ability of psychological stress experienced directly by the mother to render epigenetic consequences on offspring. For example, restraint of the mother produces high levels of Dnmt1 and 3a (genes encoding enzymes that are responsible for adding methyl groups to DNA) mRNA in the frontal cortex and hippocampus of her male offspring (Matrisciano et al., 2012). This same study showed that prenatally stressed mice also displayed increased methylation of DNA associated with synaptic plasticity genes (including reelin) and several schizophrenic-like behaviors, including hyperactivity, deficits in social interaction, altered prepulse inhibition, and memory capacity. Restraint stress of pregnant rats has also been associated with changes in methylation of a placental and fetal gene (HSD2) whose product helps protect the fetus against the damaging effects of stress hormones (Jensen Peña, Monk, & Champagne, 2012). Further, variations in the magnitude of the same stressor can produce dramatically different epigenetic consequences. One group has shown that offspring of dams who had been briefly subjected to the stress of being on an elevated platform showed higher levels of global DNA methylation in their frontal cortex and hippocampus (Mychasiuk, Ilnytskyy, Kovalchuk, Kolb, & Gibb, 2011). In contrast, offspring of mothers that had spent much longer time on the platform showed a dramatic reduction in global DNA methylation levels. As discussed in an earlier section, increasing evidence shows that genes (and in turn behavior) can also be epigenetically regulated by miRNAs. A recent study has provided a link between stress during gestation and long-term alterations in miRNA regulation in offspring (Zucchi et al., 2013). In this study, pregnant rats were exposed to a stressful environment twice daily from gestational day 12 to 18. The two stressors consisted of exposure to restraint for 20 minutes and forced swimming for 5 minutes. Investigators then assessed whether their stress regimen produced any changes in frontal cortex miRNA expression in dams or changes in whole brain miRNA expression of male offspring. They found alterations in miRNA expression in the frontal cortex of dams

that correlated with stress-induced behavioral alterations (a decrease in total time spent tail-chasing antepartum compared with nonstressed dams). Specifically, gestational stress caused a downregulation of many miRNAs that are related to stress responses, metabolism, hormonal regulation, and neurological diseases, including miR-329, miR-20a, miR-500, miR-181, and miR-186. Prenatal stress also upregulated several miRNAs, including miR-24–1, which play a role in destabilization and translational inhibition of target mRNA. Significant changes were also observed in the brains (whole) of male offspring. Downregulation occurred in miRNAs that target genes that are related to neurodevelopment, stress responsivity, and brain pathologies. Upregulation also occurred in miRNAs that play a role in cell signaling and stress response, miR-98 (alters neural inflammatory responses), and miR-219 (a putative marker of SZ and bipolar disorder [BD]). While these observations cannot be directly linked to the changes in antepartum behavior observed, they nonetheless demonstrate the capacity of stress experienced by the dam to alter both her and her offspring’s miRNA levels. Fascinating evidence is beginning to emerge demonstrating that simply watching another individual experience stress can also render epigenetic consequences for offspring. Specifically, one report has shown global DNA methylation (frontal cortex and hippocampus), gene expression, and behavioral alterations in offspring of pregnant rats that were housed with another female who was repeatedly subjected to stress outside of the homecage (Mychasiuk et al., 2011). In this study, pregnant Long-Evans rats (during gestational days 12–16) were housed with another female rat that underwent elevated platform and bright light stress (30 minutes twice daily). Female offspring (10–13 days old) prenatally exposed to stress entered significantly less squares in an open field test, indicative of either locomotor deficits or their hesitation to explore novel environments. Investigation of global DNA methylation patterns in bystander-stressed offspring showed an increase in methylation within the frontal cortex and the hippocampus. A total of 558 genes were differentially expressed between the stressed and nonstressed control offspring, including genes involved in biological processes controlling neural development and function. These data effectively illustrate the ability of bystander stress to impact methylation levels within multiple brain regions (with distinct differences between sexes that we did not discuss here). Our first hint that methylation status of the human genome is equally sensitive to maternal stress and emotion came in 2008. In this study, it was shown that infants

Epigenetic Factors and Outcomes Associated With Early-Life Postnatal Environmental Events

born to mothers that reported high levels of depression and anxiety during their third trimester of pregnancy exhibited increased methylation of the human GR gene (Nr3c1) promoter in cord blood cells (Oberlander et al., 2008). Increased methylation was also associated with increased HPA stress reactivity (change in cortisol levels from baseline) in response to novel visual stimuli at the 3-month visit. These data provided seminal evidence consistent with the notion of fetal programming of the infant’s HPA axis by maternal mood. Since that study, this same group has also shown the programming capacity of maternal depressed mood on behavior and development that involves epigenetic changes within the serotonergic system. Specifically, increased maternal depressed mood scores have been associated with decreased maternal (peripheral leukocytes) and infant (umbilical cord) serotonin transporter (SLC6A4) promoter methylation (Devlin, Brain, Austin, & Oberlander, 2010). More recent work has shown an association between extreme psychosocial stress experienced while pregnant and methylation of the GR gene in newborns (Mulligan, D’Errico, Stees, & Hughes, 2012). In this study, war stress (vs. material deprivation or mundane stress) experienced by mothers in the Democratic Republic of Congo had the strongest correlation with birth weight, able to account for 35% of variance in birth weight. Birth weight and stress were also strongly correlated with GR methylation (the higher stress, the greater methylation and lower birth weight). Together, the studies reviewed in this section indicate that nutritional factors, environmental insults, and life stress during pregnancy are risk factors for a wide range of behavioral outcomes in offspring. More importantly, these studies provide compelling evidence that DNA methylation (or demethylation) is a mechanism whereby prenatal stressors and environmental events could give rise to behavior, including altered stress responsivity and depression. As we discuss in the next section, we have also learned that epigenetic factors are at play during early-life development outside of the womb, and are thus recognized as a mechanism whereby early-life environmental events can create biochemical changes that dictate gene activity and the development of behavior.

EPIGENETIC FACTORS AND OUTCOMES ASSOCIATED WITH EARLY-LIFE POSTNATAL ENVIRONMENTAL EVENTS The programming capacity of early environmental events outside the womb has become equally clear over decades of

421

research. For example, children and adults with a history of early abuse or neglect show high rates of depression, anxiety, substance abuse and have impaired social functioning and deficits in cognitive performance (Cicchetti & Toth, 2005; De Bellis, 2005; Heim, Shugart, Craighead, & Nemeroff, 2010). One neural impact long presumed responsible for these outcomes is experience-induced changes in the function of the HPA axis. Indeed many clinical studies have demonstrated that early-life stress canalizes HPA axis function, as indicated by hypo- or hypercortisolism and the dysregulation of the circadian rhythm of cortisol production typical of children and adults with a history of maltreatment (Carpenter et al., 2009; Gunnar, Quevedo, & De Kloet, 2007; Rogosch, Dackis, & Cicchetti, 2011). HPA axis dysregulation is also a prominent symptom in adults with a history of child maltreatment and the diagnosis of major depressive disorder or posttraumatic stress disorder (PTSD) (Bauer, Wieck, Lopes, Teixeira, & Grassi-Oliveira, 2010; De Bellis & Thomas, 2003). Nonhuman primate models of early-life adversity likewise demonstrate the negative impact of early-life stress and maltreatment on the HPA axis, showing for example that separation of infants from their attachment figure can provoke increased activity of the HPA axis (Gunnar et al., 2007; Howell & Sanchez, 2011). Variable foraging paradigms that disrupt mother–infant interactions (approximating a neglectful caregiving environment for the infant) yield offspring that are more fearful and have high levels of CRF as well as alterations in noradrenergic and serotonergic function (Gunnar et al., 2007). Together, these observations suggest that epigenetic programming of HPA axis regulation is one way in which early-life environmental events could render an individual with lifelong altered stress responses and increased susceptibility to later-emergent stress- and mood-related psychiatric disorders. In 2004, investigators brought forth the first direct evidence in support of this notion. At the same time, these data challenged the previously held view that epigenetic mechanisms are static and unresponsive to the environment outside of periods of embryonic development and cellular differentiation. Data they presented indicated that methylation of DNA associated with the GR, a gene underlying stress-responsivity through its regulation of HPA activity, was directly associated with the type of caregiving experienced during the first postnatal week (Weaver et al., 2004). Specifically, they showed that adult male rats that had been reared by nurturing mothers that exhibited high levels of pup licking and grooming (LG) had low levels of methylation of DNA associated with the

422

Epigenetics and the Development of Behavior

GR gene within their hippocampus, while adults who had been raised by less nurturing (low LG) mothers exhibited hypermethylation of GR DNA. These observations were consistent with GR gene expression patterns and anxiety-related behavior of the animals. Animals with low methylation had higher expression of the GR gene and exhibited less stress-responsivity and anxiety-like behavior (less fearful response to stimuli, decreased defensive response, and a more modest HPA response to a stress challenge), while animals with higher methylation had lower gene expression and increased anxiety-like behavior. It is important to point out here that DNA methylation changes often occur in conjunction with other chromatin changes, including histone acetylation. Histone acetylation at the ninth lysine (K9) residue of histone 3 is a marker of active chromatin, and adult offspring of high-LG mothers when compared with offspring of low-LG mothers had higher acetylated K9 levels. Then through a series of cross-fostering studies, they were able to demonstrate that the levels of GR promoter methylation were a direct product of the mother’s behavior during the postnatal period. These data were key in providing an association between the levels of caregiving behavior and DNA methylation of the GR gene promoter. Finally, in an effort to help establish a causal link between the observed epigenetic modifications, gene expression patterns, and adult behavior, they demonstrated that pharmacologically manipulating methylation patterns, via treatment of adult animals with a HDAC inhibitor (trichostatin A), removed group differences in DNA methylation, histone acetylation, and behavior. Another way to examine the ability of incant-caregiver interactions to program the HPA axis is to separate infant rodents from their mother for a specified period of time. Since rodent pups are dependent on their mother for survival, separation from the mother produces an array of responses, including alterations in HPA activity (Faturi et al., 2010). We have learned that behavioral outcomes in separated infants are also in part due to epigenetic programming of the HPA axis (Murgatroyd et al., 2009). In this study, mice as far as 1 year out from the last infant separation experience were shown to have hypomethylation of arginine vasopressin (AVP) DNA in their paraventricular nucleus (PVN). The product of the AVP gene is a hypothalamic component that induces the synthesis and release of adrenocorticotropin from the pituitary. The lower levels of DNA methylation inversely correlated with AVP expression, as well as reduced binding of MeCP2, a DNA methyl-binding protein typically

associated with gene suppression. They found that these molecular effects occurred in parallel with increases in corticosterone secretion (in response to an acute stress challenge and under basal conditions) and memory deficits in an inhibitory avoidance task. AVP antagonism was shown to partially reverse the phenotype. Overall, these results indicate that an early-life stressor activates a key group of stress regulating neurons that are part of the HPA axis, leading to stables changes in MeCP2 function and programing of a key gene involved in the stress regulatory pathway. Since this study, infant separation has been shown to produce less methylation of CG sites associated with the Crf promoter within the PVN, an effect that coincides with heightened plasma corticosterone response to an acute stress challenge (Chen et al., 2012). Akin to the rodent literature, the human GR gene appears equally responsive to early-life environmental events. For example, increased methylation of GR and ribosomal RNA genes (and a corresponding decrease in transcript levels) have been found in hippocampal samples derived from adult males who committed suicide and had a history of childhood abuse (Labonté, Suderman, et al., 2012; McGowan et al., 2009; McGowan et al., 2008). In another study linking the effects of early-life adversity with epigenetic modification of the human GR gene, parental loss, childhood maltreatment, and disruptions in parental care were found to be associated with increased GR promoter DNA methylation (as determined from leukocyte DNA) and attenuated cortisol responses to a stress challenge (Tyrka, Price, Marsit, Walters, & Carpenter, 2012). Additional work has shown that childhood sexual abuse and its severity show a positive correlation with GR methylation (Perroud et al., 2011). In this study, GR methylation was explored in individuals with Borderline Personality Disorder (BPD), Major Depressive Disorder (MDD) or individuals with a co-morbidity of MDD and PTSD. While the severity of childhood sexual abuse and the number of maltreatment types correlated positively with GR methylation in all of these individuals, those with BPD who experienced repeated sexual abuse with penetration and other abuses showed the highest methylation percentage. This is yet another example linking childhood maltreatment, epigenetic changes, and adult psychiatric disorders. If we return to rodent models, the broad array of phenotypes affected by maternal care in the rodent model of high vs. low LG, including hippocampal-dependent memory, sexual behavior, fear-like behavior, maternal care, and alterations in drug-seeking behavior (Kaffman & Meaney, 2007) would suggest the ability of early-life events

Epigenetic Factors and Outcomes Associated With Early-Life Postnatal Environmental Events

to epigenetically program a multitude of genes and brain systems. Subsequent research has shown that this is indeed the case. Maternal care has been shown to influence levels of cytosine methylation of the estrogen-receptor alpha (ER-α) gene promoter in the medial preoptic area, an area intimately associated with maternal behavior (Champagne, Weaver, Diorio, Dymov, Szyf, & Meaney, 2006). Activation of ER-α supports maternal behavior expression, and expression of this gene is regulated by a signaling cascade known as the Janus kinase-Stab5b pathway. Lower levels of ER-α gene DNA methylation, increased binding of Stab5b, and increased expression of ER-α were found in adult female offspring of high-LG mothers. These animals showed high levels of licking and grooming towards their own offspring, as well as displaying an arched back nursing position toward their offspring. The opposite was found for adult offspring of low LG mothers. Reminiscent of the outcome regarding fear- and anxiety-like behaviors, variations in maternal care also confer differences in future mothering behavior via epigenetic mechanisms. LG maternal behavior has since been shown to affect 𝛾-aminobutyric acid (GABA) inhibitory circuits, as males that are raised by low-LG mothers have reduced hippocampal levels of the rate-limiting enzyme in GABA synthesis called glutamic acid decarboxylase (GAD1) which is associated with increased methylation of GAD1 promoter DNA (Zhang et al., 2010). Epigenetic changes in this model are known to occur on a much broader genome-wide scale within the hippocampus (McGowan et al., 2011). Work in other models likewise demonstrates the ability of early-life experiences to epigenetically alter multiple genes and brain regions. For example, infant male rats experiencing repeated separation from their mother and nest environment show altered hippocampal methylation and expression of estrogen receptor (ER𝛽) DNA (a gene that encodes a receptor responsive to estrogen) (Wang, Meyer, & Korz, 2012) and increased methylation (and therefore reduced expression) of the hippocampal synaptic plasticity gene reelin (Qin et al., 2011). With some of our own work, we have provided a link between caregiver maltreatment, altered DNA methylation patterns of the Brain-derived neurotrophic factor (Bdnf) gene, and aberrant behavioral outcomes (Roth, Lubin, Funk, & Sweatt, 2009). A known stressor of rat mothers is resource deprivation (nesting material), and we and others have used this to elicit aberrant caregiving behaviors (Ivy, Brunson, Sandman, & Baram, 2008; Raineki, Moriceau, & Sullivan, 2010). In our laboratory, we use this stressor in conjunction with a within litter design to examine the ability of

423

adverse caregiving to alter DNA methylation. We employ three experimental conditions: normal maternal care, cross-foster care, and caregiver maltreatment. The normal maternal care condition consists of leaving infants in their home cage with their biological mother in the room in which they were bred (infants are however briefly handled for marking and weighing purposes). The cross-foster care condition consists of placing infants in a different room and in a different apparatus with a lactating dam. This female is given an hour to habituate to the chamber before receiving infants. She is also given adequate amounts of bedding so that she may nest and nurture infants. The caregiver maltreatment condition consists of placing infants in a different room and apparatus with a lactating dam. This female is not given time to habituate to her new surroundings and is not provided adequate bedding to construct a nest. Infant rats are repeatedly exposed to their designated conditions for 30 minutes daily from postnatal day one through seven. Nurturing caregiving behaviors that we see in the normal care and cross-foster conditions include frequent observations of infant grooming, anogenital licking (to void their bladder), and nursing. Adverse caregiving behaviors displayed in the maltreatment condition include infant dragging, dropping, roughly handling, and active avoidance of the pups (Figure 11.2a). Dams within this condition also spend less time crouching over or nursing the infants and decreased amounts of licking and grooming. In some of our earliest work with this model (Roth et al., 2009), we first explored the capacity of our maltreatment conditions to alter Bdnf gene expression. We found that this was the case, as adult animals that were exposed to the maltreatment condition during infancy had significantly less Bdnf mRNA in their prefrontal cortex (as a whole) than controls (either normal or cross-foster care). We then explored whether DNA methylation could be involved. We showed that normal adults (i.e., adults with a history of normal infancy) had either no or very little CG dinucleotide methylation within an important regulatory region of the Bdnf gene (DNA associated with exon IV). This was in sharp contrast to the adults who had experienced maltreatment during infancy, where sequencing of that same regulatory region revealed that the same CG dinucleotides were all highly methylated (Figure 11.2b). Next, to establish that the methylation changes were contributing to the gene deficits, we showed that chronic treatment with a DNA methylation inhibitor called Zebularine was able to rescue the adult maltreatment-caused deficits in Bdnf gene expression. Zebularine treatment also lowered methylation levels in maltreated animals.

424

Epigenetics and the Development of Behavior

A

Nurse - 21%

Step on - 22%

Lick - 9%

Drop - 8% Drag 5%

Roughly handle - 22%

Actively avoid - 13%

Bdnf (IV) DNA Methylation (%)

B 25 20

pA). Among adolescent girls, Furman, Chen, and Gotlib (2011) found that homozygosity for the G allele was associated with decreased amygdala volume. Furman et al. posited that this relationship may be due to decreased regulation of corticosteroids by oxytocin, which may then be associated with atrophy of the amygdala. Given that the G allele is found in higher frequencies in patients with depression, this study provides another example of a link between a gene and a neural structure (the amygdala) frequently implicated in depression, helping to delineate the developmental pathways through which risk genes may have their effects. Although each of these candidate genes outlined above may explain a relatively small amount of variance in brain structure and function, cumulatively and in interaction with environmental stressors, they are likely to have greater predictive power in helping to determine which individuals

are at risk for disrupted development of corticolimbic circuitry. For example, individuals with multiple genotypes associated with altered amygdala structure and function (the low-expressing allele of 5-HTTLPR, the Met allele of the Val66Met BDNF polymorphism, and homozygous G alleles in the rs2254298 polymorphism of the oxytocin receptor gene) may be at particularly high risk for the development of anxiety or depressive disorders, whereas individuals with only one of these risk alleles may be relatively buffered from this effect. Future research examining multiple genes and their associations with multiple regions (prefrontal cortex, amygdala, striatum) will be necessary to examine potential additive and interactive effects (e.g., Nikolova et al., 2011). Finally, whereas studies described in this chapter have almost exclusively relied on a candidate gene approach to select functional polymorphisms with known effects on a neurotransmitter system or protein of interest, an alternative approach uses GWAS to search for genes that are associated with the functioning of a candidate neural region. Using this approach, a novel candidate polymorphism in the DOK5 gene, which is associated with neuronal development, was found to be associated with amygdala activation in adolescents with bipolar disorder and among typically developing controls (Liu et al., 2010). Studies using GWAS and other related exploratory techniques could be used to identify other potential candidate genes that may be linked to corticolimbic function and the development of internalizing disorders. The downside to GWAS studies is that they require very large sample sizes, which are difficult and costly to collect when examining neuroimaging phenotypes. Biological Mechanisms Underlying Neurogenetics and GxE Interaction Findings As this chapter has highlighted, there is now evidence that the low-expressing allele of 5-HTTLPR interacts with childhood maltreatment or stressful life events to predict the development of depression in adolescents and adults (Caspi et al., 2003; Kumsta et al., 2010; Petersen et al., 2012; although see Laucht et al. (2009) for evidence of the high-expressing allele predicting greater internalizing problems in adolescents). As we have noted in previous descriptions of neurogenetics and IGxE, emerging research suggests that a major molecular pathway through which GxE interactions affect behavior is via brain chemistry (Hyde, Bogdan, et al., 2011). In particular, emerging research in epigenetics suggests a potential pathway through which environmental stress

Neurogenetics and Youth Internalizing Disorders

interacts with 5-HTTLPR genotype to influence the development of internalizing disorders. For example, preliminary evidence suggests that DNA methylation of cytosine-phosphate-guanosine (CpG) sites in the promoter region of the 5-HTT gene may play a role in the previously observed GxE interactions in internalizing disorders (Beach, Brody, Gunter, et al., 2010; Kinnally et al., 2010; Kinnally et al., 2011; Philibert et al., 2007; Philibert et al., 2008). In particular, methylation of CpG islands in the promoter region of genes is thought to make the promoter region less accessible to transcription factors, decreasing the expression of the gene (for more extended reviews on this topic see Bagot & Meaney, 2010; Meaney, 2010; van IJzendoorn, Bakermans-Kranenburg, & Ebstein, 2011). DNA methylation of the 5-HTT promoter region is one potential mechanism through which GxE interaction effects may occur. Research from the Iowa Adoption Studies has suggested that methylation of this region may be sensitive to environmental stressors, providing evidence that a history of child abuse is associated with increased methylation in the promoter region of the 5-HTT gene (Beach, Brody, Todorov, Gunter, & Philibert, 2010). Greater methylation at certain CpG sites in the serotonin transporter gene region has also been shown to be associated with decreased mRNA transcription, and a lifetime history of major depression (Philibert et al., 2008). Maltreatment thus appears to be associated with a molecular mechanism that decreases the transcription of the serotonin transporter, and there is a trend for this mechanism to be associated with a greater risk for having a lifetime history of depression. Taken together, these studies provide preliminary evidence that early environmental stressors, such as child abuse, alter methylation levels of CpG sites in the serotonin transporter gene region. This altered methylation appears to result in decreased expression of the serotonin transporter, which could potentially affect corticolimbic development and function. Importantly from a neurogenetics perspective, these mechanisms may differ further by genotype. For example, in nonhuman primate work, Kinnally and colleagues (2010) found that the low-expressing allele of the rh5-HTTLPR was associated with increased methylation of CpG sites on the 5-HTT gene in macaques, which in turn was associated with decreased levels of serotonin transporter mRNA. This pattern of associations suggests that, similar to findings in humans, higher methylation leads to reduced serotonin transporter expression in nonhuman primates. Additionally, increased methylation was found to interact with early life stress (separation from mother or unpredictable food availability)

467

to predict higher scores on a behavioral measure of stress reactivity (Kinnally et al., 2010; Kinnally et al., 2011). This emerging body of research, when considered alongside pioneering research on rodents by Meaney and colleagues (e.g., McGowan et al., 2009; Meaney, 2010; Weaver et al., 2004), demonstrates potential neurogenetics mechanisms by which well-replicated GxE interactions appear to affect brain neurochemistry and function, and subsequent risk for psychopathology. Taken together, research from humans, nonhuman primates, and rodents suggests a potential pathway for epigenetic influences on the development of psychopathology. In particular, increased methylation of the 5-HTT gene leads to reduced transcription of the serotonin transporter (Philibert et al., 2008) and heightened stress reactivity in nonhuman primates (Kinnally et al., 2011). Importantly from a neurogenetics perspective, early environmental stress (e.g., child abuse) leads to increased methylation of the serotonin transporter promoter region, and methylation levels may be moderated by 5-HTTLPR genotype (Beach, Brody, Todorov, et al., 2010; Kinnally et al., 2010), thus providing a mechanism for a GxE interaction in which individuals with the low-expressing allele who are exposed to stressful life events are at increased risk for developing psychopathology. These preliminary findings signal an exciting potential direction for future research. However, incorporating epigenetics into IGxE designs will require careful consideration of a number of theoretical and methodological concerns. First, although we can assume that DNA sequences assessed in adulthood have not changed from early development, the same may not necessarily be true for epigenetic modifications, given that they are subject to environmental influence (Houston et al., 2013). Thus, as noted throughout, although much of the research to date has examined experiences that occur in childhood (e.g., maltreatment), studies carried out in humans have typically not examined outcomes in youth, nor focused on childhood trajectories of these behaviors. Thus longitudinal work, especially with a developmental psychopathology focus on mechanisms and trajectories of behaviors, is needed to chart these complex molecular pathways across development and complex trajectories of behavior in youth. Second, a major limitation of research to date is that methylation levels have been assessed peripherally in blood cells, which may not match methylation levels in the brain, especially given that methylation levels vary by cell type (Houston et al., 2013). As yet, it appears there is no straightforward association between peripherally measured methylation levels and methylation in the brain.

468

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

In particular, it is possible that the association between peripheral and central nervous system methylation levels could vary depending on the gene of interest, or the brain structure in which methylation levels are examined (Ursini et al., 2011). Third, the mechanisms linking increased methylation of the 5-HTT gene to increased risk for psychopathology need to be further delineated. Although studies have shown that methylation of the 5-HTT gene alters mRNA transcription of the serotonin transporter gene, the developmental consequences of methylation for brain structure and function remain to be examined. Given the aforementioned limitations, progress in this area will require research that spans methodologies to triangulate a developmental pathway, while accounting for the strengths and limitations of each method. Animal research will clearly be important in this regard because it is possible to assess methylation levels in brain regions of interest, and test whether these match methylation levels assessed peripherally. Animal models, such as those used by Kinally and colleagues, also allow for experimental manipulation of environmental stress and rearing conditions (Kinnally et al., 2010; Kinnally et al., 2011). However, animal models necessarily limit the conclusions that can be drawn regarding complex human phenomena, such as social anxiety or depression. Postmortem studies in humans have also examined methylation levels directly in brain regions of interest and linked these to potential environmental stressors (e.g., McGowan et al., 2009). There are, however, a number of limitations to these studies, including the inability to assess the consequences of observed methylation levels for brain function. An example of a research design incorporating a number of these proposed research directions examined a common functional variant in the gene regulating Catechol-O-methyltransferase (COMT), an enzyme that breaks down dopamine in the prefrontal cortex (Ursini et al., 2011). The COMT gene has a common functional polymorphism: Val158Met. The Val allele is associated with greater COMT activity, lower levels of prefrontal dopamine, and reduced prefrontal efficiency. In contrast, the Met allele is associated with less COMT activity via decreased thermostability, greater prefrontal dopamine, and, subsequently, increased prefrontal efficiency. This polymorphism may be an important candidate for GxE interactions because methylation at the region investigated in this study is possible on the Val allele, where there is a CpG site, but not on the Met allele (Wiers, 2012). Because COMT breaks down dopamine in the prefrontal cortex, greater methylation of the Val allele leads to reduced

expression of COMT, leading to greater availability of dopamine in the prefrontal cortex and greater prefrontal cortex efficiency. Incorporating research from animal models into their study, Ursini and colleagues demonstrated that COMT methylation levels measured peripherally in blood cells were significantly associated with COMT methylation levels in the prefrontal cortex in rats, suggesting that peripherally measured methylation levels correspond to those in the brain region of interest. Then, incorporating a human IGxE design, Ursini and colleagues examined the association between peripherally-measured COMT methylation levels, environmental stress, and prefrontal cortex function during a working memory task performed during fMRI scanning. In this portion of the study, they found that environmental stress predicted reduced methylation in Val/Val participants. Moreover, reduced methylation was associated with reduced working memory performance. The imaging component provided evidence that greater life stress and reduced COMT methylation were associated with decreased prefrontal cortex efficiency during the working memory task. When applying these findings to what we know about the genetic and molecular pathways affected by COMT variation, this study could be interpreted to show that methylation actually biased individuals toward the sequence-based variation conferring greater flexibility and adaptability via increased phasic dopamine signaling; stress-induced changes in methylation may then alter this bias and confer risk for psychopathology. This study demonstrates how research can be integrated across a number of domains to better elucidate the processes through which genes and the environment interact to influence neural function, thus exemplifying much of the promise of neurogenetics. Using an animal model, Ursini and colleagues provided preliminary evidence that peripherally measured COMT methylation levels matched those in the region of interest (the prefrontal cortex). In humans, they demonstrated that environmental stress reduces COMT methylation (measured peripherally) and that reduced methylation coupled with greater stress is associated with prefrontal cortex function. Because methylation can only occur on the Val allele, this research has begun to indicate a potential pathway through which a GxE interaction may occur where the presence or absence of environmental stress has a stronger influence on Val allele carriers. Overall, these types of studies signal an important need for greater cross-talk and collaboration across a range of disciplines, as it is clear that no one method alone will be sufficient for identifying the mechanisms through

Future Directions

which GxE interactions influence risk and resilience in the development of psychopathology. Summary of Research Examining the Neurogenetics of Internalizing Disorders To summarize, neurogenetics studies have begun to elucidate a number of developmental pathways through which GxE interaction effects may contribute to the development of internalizing disorders (Figure 12.3). Studies have highlighted a number of genetic polymorphisms that affect neural development and neurotransmission, including polymorphisms in genes that influence the reception or secretion of serotonin, BDNF, dopamine, and oxytocin. These candidate genes can in turn be linked to the development and function of neural regions that are affected by the proteins these genes code for. For instance, the 5-HTTLPR genotype, which modulates serotonin reception, can be plausibly linked to altered function of corticolimbic circuitry, including reduced functional connectivity between the prefrontal cortex and amygdala. Administration of oxytocin has been shown to dampen amygdala activity (Kirsch et al., 2005), creating a potential link between OXTR genotype and amygdala function. BDNF is expressed in the amygdala, and thus BDNF Val66Met genotype may contribute to individual differences in plasticity and fear learning in the amygdala (Casey et al., 2009). Amygdala function, in turn, is linked to processes, such as fear conditioning and attention bias to threat, whereas prefrontal cortex-amygdala connectivity is associated with fear extinction and emotion regulation, all intermediate behavioral phenotypes associated with internalizing disorders. These neural phenotypes serve as key bridges between genetic variation and behavioral and psychological outcomes. Additionally, environmental stress has been shown to interact with genotype to predict amygdala function (Canli et al., 2006), and environmental support has been shown to moderate the association between amygdala function and anxiety (Hyde, Gorka, et al., 2011). Thus, the environment interacts with genotype and neural function at multiple levels in this model, further elucidating the developmental pathways through which genes and environment interact to influence the development of psychopathology. However, despite the progress made to date, a developmental perspective has still been largely lacking from such explorations. As discussed above and illustrated by the relatively small number of studies described examining neurogenetics of internalizing in youth, the majority of imaging genetics

469

research has been conducted in adults, which may be problematic for informing our theories of the development of psychopathology. When studies are conducted in adults, they are either performed in healthy controls (who may be relatively resilient to the development of psychopathology) or in adult patients (whose neural function may reflect the consequence of disorder rather than the cause). When studies have been conducted in children or adolescents, they often include a wide cross-section of ages and have generally not considered moderation of GxE interaction effects by development or developmental pathways. Thus, an important direction for future research will be to incorporate a developmental perspective into neurogenetics designs, particularly as outlined above considering points from developmental psychopathology. FUTURE DIRECTIONS Ongoing Challenges and Future Directions for Neurogenetics Research of Internalizing Disorders To return to the questions posed at the beginning of this section, research in the nascent field of developmental neurogenetics as applied to internalizing disorders has demonstrated the importance of incorporating a developmental perspective into research. Initial studies, including those of Wiggins et al. (2012) and Casey et al. (2009), provide preliminary evidence for “age–genotype” interaction effects on behavioral and neural phenotypes, indicating that the association between a gene and a phenotype of interest will vary depending on when the association is measured. There are several reasons that we could see such developmental effects, each of which has important implications for future directions in this field. Many brain structures, including the prefrontal cortex and amygdala, undergo protracted development throughout adolescence and thus the association between genes and brain function may depend on age. For instance, several recent studies have shown that there are cross sectional linear declines with age in amygdala reactivity to emotional faces (Gee et al., 2013; Swartz, Carrasco et al., 2014). Thus, if amygdala reactivity is normatively higher in children, the association between genotype and phenotype may differ in children relative to other earlier or later developmental stages. If brain functioning can be measured and examined as a trajectory over time, we may be able to examine how individuals differ from the decreasing trajectory that occurs during adolescence or if variations from this trajectory predict maladaptive outcomes. Moreover, some

470

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

effects may be latent, or may be only activated after certain developmental stages, such as when puberty triggers changes in gene expression. Casey and colleagues (2009) have proposed another mechanism through which “age × –genotype” interactions may occur: because expressed levels of proteins can vary over development, the effect of a genetic variation that decreases or increases expression of that protein may vary by developmental stage because the absolute levels of protein available vary by age. Thus, a genotype that codes for an “ideal” amount of protein expression in one environment at one developmental stage may not be ideal at another developmental stage. Thus, although relatively few IGxE studies that include an additional assessment of moderation by age have been conducted to date, initial theoretical and empirical considerations suggest that there is great promise in taking a developmental perspective to neurogenetics approaches. A related question has been to pinpoint the biological mechanisms through which genes and the environment interact to influence development. Again, we have barely scratched the surface when it comes to addressing this question, but preliminary studies incorporating epigenetics have suggested a plausible pathway through which environmental stress can “get under the skin” and influence the development and functioning of the brain. This area of research is subject to some novel methodological difficulties given the lack of straightforward associations between peripherally measured methylation levels and methylation levels in brain regions of interest. However, these limitations also highlight the promise of a neurogenetics approach. In particular, studies of epigenetics will benefit greatly from the incorporation of animal models, imaging genetics, GxE approaches, and PET and multimodal imaging approaches to better link experience to the complex molecular pathways leading from gene to brain to behavior. Other questions to date remain largely unanswered. In particular, we know very little about whether patterns of neural development in childhood or adolescence could be used to predict onset of disorder. Moreover, we are yet to examine whether knowing about neural development patterns, could better identify genetic risk markers that could be used to relatively inexpensively and noninvasively assess an individual’s risk for the development of psychopathology. Given the current gaps in the literature, we outline several directions for future research that could lead to significant progress in this area: First, the reliance on cross-sectional research to date indicates a need for the mapping of neural development using longitudinal

approaches. There are several large-scale longitudinal studies that have done such mapping in structural brain development (e.g., Gogtay et al., 2004; Shaw et al., 2008), but to date, only a handful of longitudinal studies have examined changes in brain function (e.g., Koolschijn, Schel, de Rooij, Rombouts, & Crone, 2011; Ordaz, Foran, Velanova, & Luna, 2013; Pfeifer et al., 2013). Mapping the longitudinal development of brain function, including amygdala reactivity and prefrontal cortex-amygdala connectivity, is an important first step in forming hypotheses about how development might moderate GxE interaction effects on neural function, or how trajectories of neural development could mediate GxE interaction effects on disorder onset. Additionally, studies that compare trajectories of typical brain development to trajectories of brain development in individuals at risk for disorder (e.g., individuals with a high number of risk alleles of candidate genes or heightened familial risk due to a family history of disorder) can help to highlight early markers of the boundary between adaptive and maladaptive brain development. This step will be important in determining whether altered patterns of brain development could be used as early indicators of risk for disorder. These future directions point to a need both for large-scale, longitudinal functional neuroimaging studies and the training of neuroscientists in statistical approaches for the analysis of large data sets and longitudinal measures of behavior and brain functioning (e.g., Nagin & Tremblay, 2001). Second, a notable limitation of many imaging genetics studies conducted to date is that sample sizes have been too small to investigate moderation of effects by age, sex, or race/ethnicity. In general, these studies have analyzed all participants together, and when race/ethnicity is considered, the approach is usually to select for White/Caucasian participants or reanalyze results to ensure they hold in the subsample of participants that are White/Caucasian. Lacking the power to test for moderation effects of gender is problematic given the noted gender differences in internalizing disorders (and differential effects of pubertal hormones on brain development), with females tending to have higher rates of depression and anxiety from adolescence onward. Moreover, lack of power to test for differences across race/ethnicity categories is also problematic, given that allelic frequencies vary by race/ethnicity and may even have different or opposite effects in different groups. For instance, there is some evidence to suggest that the association between the 5-HTTLPR and brain function in Asian populations is opposite to that found in Caucasian populations, with the high-expressing allele predicting greater anxiety and reduced prefrontal

Future Directions

cortex-amygdala connectivity (Long et al., 2013). Thus, failure to take into account potential differences could mask significant effects or lead to conclusions that are not accurate for all racial/ethnic groups. Treatment Implications There is significant promise in using neurogenetics and developmental psychopathology to inform translational efforts (Bogdan, Hyde, et al., 2012; Hyde, 2014). For example, the promise of examining subgrouping and person-centered approaches to studies of psychopathology, particularly those examining neural and genetic correlates, is that if these studies identify a group of youth or adults with a distinct etiology (e.g., those with early-onset depression), then we may be better able to tailor interventions to these individuals (e.g., Hyde, Waller, & Burt, 2014). Moreover, if empirical studies identify factors (i.e., early anxiety symptoms, certain genetic polymorphisms) that predict a different course of a disorder, then these factors may be important in identifying those at highest risk early for preventative interventions (e.g., Dishion et al., 2008; Hankin, 2012). Genetic variation and brain function may also help to identify treatment response and thus be considered before interventions are even started (Bryant et al., 2008; Ising et al., 2007; Uhr et al., 2008). Thus, as medicine moves toward both a more tailored and personalized model of care at the individual level, and a more preventative model of care at the population level, identifying factors that delineate subgroups of individuals that need different treatments or that can be targeted earlier with preventative interventions will be increasingly important and may help to increase the effectiveness of both prevention and intervention models (Simon & Perlis, 2010; Willard & Ginsburg, 2009). More specifically, by deconstructing the molecular mechanisms underlying gene–brain–behavior pathways, especially in collaboration with animal models and in vitro research, neurogenetics has the potential to inform novel therapeutic targets. For example, combining evidence from studies linking HTR1A rs6295 with anxiety through amygdala reactivity with prior work demonstrating effects of the polymorphism on the capacity for negative feedback inhibition, suggests that targeting 5-HT1A autoreceptors, perhaps as an adjuvant to SSRI treatment, may produce greater clinical effect. In fact, a recent study in a transgenic mouse model of 5-HT1A autoreceptor function demonstrated that reducing autoreceptor levels prior to SSRI administration converted nonresponders into responders (Richardson-Jones et al., 2010). Thus, neurogenetics

471

research with HTR1A rs6295 has identified a novel therapeutic target (i.e., antagonism of 5-HT1A autoreceptors) and a marker that could be used to individually tailor treatment (i.e., for C allele homozygotes). Another example of how neurogenetics can inform treatment and prevention comes from research on TREK1, a background potassium channel. Inspired by a TREK1 knockout mouse study showing that deletion of TREK1 results in a depression resistant phenotype (Heurteaux et al., 2006), human studies have linked variation in the human TREK1 gene (KCNK2) to depression (Liou et al., 2009), blunted striatal response to reward (a neural profile associated with depression; Dillon et al., 2010), as well as antidepressant treatment response (Perlis et al., 2008). More recently these findings have led researchers to develop antidepressant medications that antagonize TREK1. One such compound designed to inhibit TREK1 has been associated with a positive antidepressant response, hippocampal neurogenesis, and increased serotonergic signaling in rodents (Mazella et al., 2010). Though the potential of this novel treatment mechanism has yet to be tested in humans, it does further document how neurogenetics research can spur therapeutic advancements by identifying novel targets. These examples highlight the great potential of neurogenetics to inform targets for novel interventions and ways to better personalize psychiatric treatment. However, they also underline the importance of additional research in this area that is translational in nature. Next Steps in Research Drawing across this chapter, there are a number of important future steps arising for the next generation of research. In particular, there are several key methodological steps and developments needed in the tools we use to examine gene–brain–behavior links. Further, there are theoretical advancements needed in the way that we model complex and conditional relationships, including the effects of development, further moderation of GxE interaction by additional genes or environments, and the importance of incorporating epigenetic mechanisms into models. First, there is need for imaging genetics studies to draw out indirect pathways from gene to brain to behavior, instead of focusing solely on gene to brain associations. In particular, we need research that examines the brain as a mediator of gene–behavior pathways longitudinally. As outlined, within a neurogenetics framework, this research will benefit from leveraging what we know from complementary biological techniques that probe the chemistry of the brain, including pharmacologic challenge studies and

472

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

animal models. We also need ways to increase sample sizes within imaging genetics research, which will support the testing of increasingly complex pathways, particularly if this can be done with repeated neuroimaging data collection across development. Looking ahead, neurogenetics will benefit from scanning larger and larger samples of individuals at a population level, using data sharing, consortium models, and neuroimaging meta-analyses. The use of larger sample sizes will also enable closer examination of the potential moderation of GxE interaction effects by sex or race/ethnicity, which are too often overlooked in studies that focus on homogenous subgroups, such as White/Caucasian participants. Failure to take into account potential differences according to gender or race/ethnicity may mask significant effects or lead to conclusions that are not accurate for all subgroups. As a field, we also need to be open for further collaboration with colleagues who are developing novel statistical modeling methods, including advanced computational and statistical techniques, as well as data mining approaches such as machine learning and graph theory. Finally, as we develop increasingly complex models that incorporate the influence of the environment in IGxE interaction studies, we need to be more precise in our operationalization of the environment. In particular, by incorporating an emphasis highlighted in developmental psychopathology on adopting an interdisciplinary approach, neurogenetics will benefit from the use of ecological and complex systems models that will help to inform changing views of the structure of psychopathology and maladaptive behaviors. Finally, by adopting more genetically informed designs we can begin to highlight which environmental effects are independent of genotype. Second, from both methodological and theoretical perspectives, we need to be better at modeling increasingly complex conditional relationships, including GxGxE and GxExE. In considering the potential for further moderation of GxE interaction effects by genes, we will be able to build on a significant body of literature examining cumulative genetic profiles (Nikolova et al., 2011; Purcell et al., 2009). Likewise, modeling the role of the environment is usefully informed by cumulative risk models of environmental exposures within developmental psychopathology, highlighting the fact that an accumulation of “risk” is often more important than any one ‘single’ risk factor. Combining cumulative genetics models with cumulative experiential models is a key next step in neurogenetics research. Third, a further source of complexity is the role of developmental stage (D) in influencing GxE and IGxE

interaction effects (i.e., GxExD). As outlined, we believe that neurogenetics has much to gain from incorporating key tenets of developmental psychopathology to create an integrated developmental neurogenetics framework (see Hyde, 2015). Through the integration of these approaches, future next steps in neurogenetics research will involve an examination of broader liability for psychopathology (i.e., within a general-specific modeling framework), deconstruction of diagnostic classifications within a person-centered framework to identify even greater homogeneity with subgroups of individuals, incorporation of the concepts of equifinality and multifinality within models, and a focus on the temperamental, personality, or endophenotypic building blocks of psychopathology that serve as more objective outcomes with which to assess gene–brain–behavior links. Importantly, models from developmental psychopathology can inform modeling of these complex, multilevel questions and will help to guide efforts to examine the role of development in these processes. In particular, future neurogenetics research needs to be more sensitive to potential developmental effects. First, studies are needed that include an examination of the effects of “age x gene” interactions within imaging genetics studies to determine whether the influence of genes on brain function varies at different ages or developmental stages. A second and related aim is for studies to examine how genes affect trajectories of brain development using prospective longitudinal designs. Longitudinal research on cortical thickness and IQ suggests that the trajectory of brain development can be more predictive of outcomes than the end points of development (Shaw et al., 2006). Thus, the association between genes and trajectories of brain development during childhood and adolescence may provide greater predictive power than that afforded by examining genes and brain function solely in adulthood (see also Casey et al., 2009). Indeed, although there is preliminary evidence from behavioral and health research of age x gene effects, we have barely scratched the surface on how these manifest as differences in the development of brain structure and function. Third, in relation to neurogenetics research examining the development of youth internalizing disorders more specifically, there are some overlapping directions for future research that could lead to significant progress in this area. Importantly, there is a need for prospective longitudinal functional neuroimaging studies that can prospectively examine changes in brain function and whether trajectories of neural development mediate GxE interaction effects on anxiety or depression. In relation to a developmental

References

psychopathology framework, studies that examine individuals with high cumulative experiential or genetic risk are also needed. In addition, we are only just beginning to understand the full extent of the implications associated with epigenetic mechanisms and in particular, how environments interact with genotype to affect an outcome. However, the transmission of epigenetic markers from parent to offspring is a further source of risk for individuals, and one that precipitates examining environments, not just in the lives of individuals, but also that of their parents. Related to this point is the need for research to move beyond a single candidate gene approach. As mentioned above, candidate genes only account for a small percentage of variation in neural phenotypes of interest. One approach, as we have suggested, is to incorporate cumulative genetic profiles (Nikolova et al., 2011) or to use statistical methodology to identify how individual genes fit with a neural or behavioral phenotype (e.g., Gruenewald et al., 2006). Likewise, using GWAS approaches with neural phenotypes (or developmental trajectories of neural phenotypes) as outcomes of interest has the potential to identify novel genetic polymorphisms that may provide additional predictive power. Ultimately, a more nuanced approach to characterizing genetic risk, environmental risk, GxE statistical effects, and neural development will hopefully lead to more precise characterization of the developmental processes through which genes and environment interact to influence risk or resilience for psychopathology, which will have important translational implications for the early detection of risk for disorder and early intervention and prevention efforts. Though these types of developmental neurogenetics models are relatively novel, they hold great promise and are an exciting new approach to understanding the development of psychopathology. REFERENCES Achenbach, T. M. (1966). The classification of children’s psychiatric symptoms: A factor-analytic study. Psychological Monographs: General and Applied, 80, 1–37. doi: 10.1037/h0093906 Ansorge, M. S., Morelli, E., & Gingrich, J. A. (2008). Inhibition of serotonin but not norepinephrine transport during development produces delayed, persistent perturbations of emotional behaviors in mice. Journal of Neuroscience, 28, 199–207. doi: 10.1523/JNEUROSCI. 3973–07.2008 Appleyard, K., Egeland, B., Dulmen, M. H. M., & Sroufe, A. L. (2005). When more is not better: The role of cumulative risk in child behavior outcomes. Journal of Child Psychology and Psychiatry, 46, 235–245. doi: 10.1111/j.1469–7610.2004.00351.x Astolfi, L., De Vico Fallani, F., Cincotti, F., Mattia, D., Marciani, M., Bufalari, S., . . . & Babiloni, F. (2007). Imaging functional brain connectivity patterns from high resolution EEG and fMRI via graph

473

theory. Psychophysiology, 44, 880–893. doi: 10.1111/j.1469–8986 .2007.00556.x Bagot, R. C., & Meaney, M. J. (2010). Epigenetics and the biological basis of gene × environment interactions. Journal of the American Academy of Child and Adolescent Psychiatry, 49, 752–771. doi: 10.1016/j.jaac.2010.06.001 Bar-Haim, Y., Lamy, D., Pergamin, L., Bakermans-Kranenburg, M. J., & van, I. M. H. (2007). Threat-related attentional bias in anxious and nonanxious individuals: a meta-analytic study. Psychological Bulletin, 133, 1–24. doi: 10.1037/0033–2909.133.1.1 Barr, C. S., Newman, T. K., Becker, M. L., Parker, C. C., Champoux, M., Lesch, K. . . . & Higley, J. D. (2003). The utility of the non-human primate model for studying gene by environment interactions in behavioral research. Genes, Brain and Behavior, 2, 336–340. doi: 10.1046/j.1601–1848.2003.00051.x Barr, C. S., Newman, T. K., Shannon, C., Parker, C., Dvoskin, R. L., Becker, M. L. . . . & Higley, J. D. (2004). Rearing condition and rh5-HTTLPR interact to influence limbic-hypothalamicpituitary-adrenal axis response to stress in infant macaques. Biological Psychiatry, 55, 733–738. doi: 10.1016/j.biopsych.2003.12.008 Battaglia, M., Zanoni, A., Taddei, M., Giorda, R., Bertoletti, E., Lampis, V. . . . & Tettamanti, M. (2012). Cerebral responses to emotional expressions and the development of social anxiety disorder: A preliminary longitudinal study. Depression and Anxiety, 29, 54–61. doi: 10.1002/da.20896 Beach, S. R., Brody, G. H., Todorov, A. A., Gunter, T. D., & Philibert, R. A. (2010). Methylation at SLC6A4 is linked to family history of child abuse: An examination of the Iowa Adoptee sample. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 153, 710–713. doi: 10.1002/ajmg.b.31028 Beach, S. R. H., Brody, G. H., Gunter, T. D., Packer, H., Wernett, P., & Philibert, R. A. (2010). Child maltreatment moderates the association of MAOA with symptoms of depression and antisocial personality disorder. Journal of Family Psychology, 24, 12–20. doi: 10.1037/a0018074 Beauchaine, T. P., McNulty, T. (2013). Comorbidities and continuities as ontogenic processes: toward a developmental spectrum model of externalizing psychopathology. Developmental Psychopathology, 25(4), 1505–1528. Beaudet, A. L. (2007). Autism: highly heritable but not inherited. Nature Medicine, 13, 534–536. doi: 10.1038/nm0507–534 Beesdo, K., Lau, J. Y., Guyer, A. E., McClure-Tone, E. B., Monk, C. S., Nelson, E. E. . . . & Pine, D. S. (2009). Common and distinct amygdala-function perturbations in depressed vs anxious adolescents. Archives of General Psychiatry, 66, 275–285. doi: 10.1001 /archgenpsychiatry.2008.545 Belsky, J., Jonassaint, C., Pluess, M., Stanton, M., Brummett, B., & Williams, R. (2009). Vulnerability genes or plasticity genes? Molecular Psychiatry, 14, 746–754. doi: 10.1038/mp. 2009.44 Belsky, J., & Pluess, M. (2009). Beyond diathesis stress: Differential susceptibility to environmental influences. Psychological Bulletin, 135, 885–908. doi: 10.1037/a0017376 Bennett, A., Lesch, K., Heils, A., Long, J., Lorenz, J., Shoaf, S. . . . & Higley, J. D. (2002). Early experience and serotonin transporter gene variation interact to influence primate CNS function. Molecular Psychiatry, 7, 118–122. doi: 10.1038/sj.mp. 4000949 Bigos, K. L., Mattay, V. S., Callicott, J. H., Straub, R. E., Vakkalanka, R., Kolachana, B. . . . & Weinberger, D. R. (2010). Genetic variation in CACNA1C affects brain circuitries related to mental illness. Archives of General Psychiatry, 67, 939–945. doi: 10.1001 /archgenpsychiatry.2010.96 Bigos, K. L., Pollock, B. G., Aizenstein, H. J., Fisher, P. M., Bies, R. R., & Hariri, A. R. (2008). Acute 5-HT reuptake blockade

474

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

potentiates human amygdala reactivity. Neuropsychopharmacology, 33, 3221–3225. Bigos, K. L., & Weinberger, D. R. (2010). Imaging genetics—Days of future past. Neuroimage, 53, 804–809. doi: 10.1038/npp. 2008.52 Blair, K. S., Geraci, M., Korelitz, K., Otero, M., Towbin, K., Ernst, M. . . . & Pine, D. S. (2011). The pathology of social phobia is independent of developmental changes in face processing. American Journal of Psychiatry, 168, 1202–1209. doi: 10.1176/appi.ajp. 2011.10121740 Blair, R. J. R. (2010). Neuroimaging of psychopathy and antisocial behavior: A targeted review. Current Psychiatry Reports, 12, 76–82. doi: 10.1007/s11920–009–0086-x Blakely, R., & Veenstra-Vanderweele, J. (2011). Genetic indeterminism, the 5-HTTLPR, and the paths forward in neuropsychiatric genetics. Archives of General Psychiatry, 68, 457–458. doi: 10.1001 /archgenpsychiatry.2011.34 Bogdan, R., Fitzgibbon, H., Woolverton, W., Bethea, C., Iyo, A., Stockmeier, C. . . . & Austin, M. C. (2011). 5-HTTLPR genotype and gender, but not chronic fluoxetine administration, are associated with cortical TREK1 protein expression in rhesus macaques. Neuroscience Letters, 503, 83–86. doi: 10.1016/j.neulet.2011.08.005 Bogdan, R., Hyde, L., & Hariri, A. (2012). A neurogenetics approach to understanding individual differences in brain, behavior, and risk for psychopathology. Molecular Psychiatry, 18, 288–299. doi: 10.1038 /mp. 2012.35 Bogdan, R., Williamson, D. E., & Hariri, A. R. (2012). Mineralocorticoid receptor iso/val (rs5522) genotype moderates the association between previous childhood emotional neglect and amygdala reactivity. American Journal of Psychiatry, 169, 515–522. doi: 10.1176 /appi.ajp. 2011.11060855 Bookheimer, S. Y., Strojwas, M. H., Cohen, M. S., Saunders, A. M., Pericak-Vance, M. A., Mazziotta, J. C., et al. (2000). Patterns of brain activation in people at risk for Alzheimer’s disease. New England Journal of Medicine, 343, 450–456. doi: 10.1056/NEJM200008173430701 Brody, G. H., Beach, S. R. H., Philibert, R. A., Chen, Y., & Murry, V. M. B. (2009). Prevention effects moderate the association of 5-HTTLPR and youth risk behavior initiation: Gene× Environment hypotheses tested via a randomized prevention design. Child Development, 80, 645–661. doi: 10.1111/j.1467–8624.2009.01288.x Bronfenbrenner, U., & Ceci, S. J. (1994). Nature–nuture reconceptualized in developmental perspective: A bioecological model. Psychological Review, 101, 568–586. doi: 10.1037/0033–295X.101.4.568 Brunner, H. G., Nelen, M., Breakefield, X., Ropers, H., & Van Oost, B. (1993). Abnormal behavior associated with a point mutation in the structural gene for monoamine oxidase A. Science, 262, 578–580. doi: 10.1126/science.8211186 Bryant, R., Felmingham, K., Kemp, A., Das, P., Hughes, G., Peduto, A. . . . & Williams, L. (2008). Amygdala and ventral anterior cingulate activation predicts treatment response to cognitive behaviour therapy for post-traumatic stress disorder. Psychological Medicine, 38, 555–562. doi: 10.1017/S0033291707002231 Buckholtz, J., Sust, S., Tan, H., Mattay, V., Straub, R., Meyer-Lindenberg, A. . . . & Callicott, J. H. (2007). fMRI evidence for functional epistasis between COMT and RGS4. Molecular Psychiatry, 12, 893–895. doi: 10.1038/sj.mp. 4002008

Buckholtz, J. W., Treadway, M. T., Cowan, R. L., Woodward, N. D., Li, R., Ansari, M. S. . . . & Zald, D. H. (2010). Dopaminergic network differences in human impulsivity. Science, 329, 532–532. doi: 10.1126 /science.1185778 Burt, S. A. (2008). Genes and popularity evidence of an evocative gene–environment correlation. Psychological Science, 19, 112–113. doi: 10.1111/j.1467–9280.2008.02055.x Butterworth, P., Cherbuin, N., Sachdev, P., & Anstey, K. J. (2011). The association between financial hardship and amygdala and hippocampal volumes: Results from the PATH through life project. Social Cognitive and Affective Neuroscience, 7, 548–556 doi: 10.1093/scan /nsr027. Canli, T., Omura, K., Haas, B. W., Fallgatter, A., Constable, R. T., & Lesch, K. P. (2005). Beyond affect: A role for genetic variation of the serotonin transporter in neural activation during a cognitive attention task. Proceedings of the National Academy of Sciences of the United States of America, 102, 12224–12229. doi: 10.1073/pnas.0503880102 Canli, T., Qiu, M., Omura, K., Congdon, E., Haas, B. W., Amin, Z. . . . & Lesch, K. P. (2006). Neural correlates of epigenesis. Proceedings of the National Academy of Sciences, 103, 16033–16038. doi: 10.1073 /pnas.0601674103 Cardon, L. R., & Palmer, L. J. (2003). Population stratification and spurious allelic association. Lancet, 361, 598–604. doi: 10.1016 /S0140–6736(03)12520–2 Casey, B. J., Glatt, C. E., Tottenham, N., Soliman, F., Bath, K., Amso, D. . . . & Lee, F. S. (2009). Brain-derived neurotrophic factor as a model system for examining gene by environment interactions across development. Neuroscience, 164, 108–120. doi: 10.1016 /j.neuroscience.2009.03.081 Casey, B. J., Tottenham, N., Liston, C., & Durston, S. (2005). Imaging the developing brain: what have we learned about cognitive development? Trends in Cognitive Science, 9, 104–110. doi: 10.1016/j.tics.2005.01.011 Caspi, A., Hariri, A. R., Holmes, A., Uher, R., & Moffitt, T. E. (2010). Genetic sensitivity to the environment: The case of the serotonin transporter gene and its implications for studying complex diseases and traits. American Journal of Psychiatry, 8, 509–527. doi: 10.1176 /appi.ajp. 2010.09101452 Caspi, A., Houts, R. M., Belsky, D. W., Goldman-Mellor, S. J., Harrington, H., Israel, S. . . . & Moffitt, T. E. (2013). The p factor one general psychopathology factor in the structure of psychiatric disorders.? Clinical Psychological Science, 2, 119–137. doi: 10.1177 /2167702613497473 Caspi, A., McClay, J., Moffitt, T. E., Mill, J., Martin, J., Craig, I. W. . . . & Poulton, R. (2002). Role of genotype in the cycle of violence in maltreated children. Science, 297, 851–854. doi: 10.1126/science.1072290 Caspi, A., & Moffitt, T. E. (2006). Gene–environment interactions in psychiatry: joining forces with neuroscience. Nature Reviews: Neuroscience, 7, 583–590. doi: 10.1038/nrn1925 Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H. . . . & Poulton, R. (2003). Influence of life stress on depression: moderation by a polymorphism in the 5-HTT gene. Science, 301, 386–389. doi: 10.1126/science.1083968

Buckholtz, J. W., & Meyer-Lindenberg, A. (2008). MAOA and the neurogenetic architecture of human aggression. Trends in Neurosciences, 31, 120–129. doi: 10.1016/j.tins.2007.12.006

Charney, D. S., Barlow, D. H., Botteron, K., Cohen, J. D., Goldman, D., Gur, R. E. . . . & Zalcman, S.J. (2002). Neuroscience research agenda to guide development of a pathophysiologically based classification system. In Kupfer, D. J. (Ed), First, M. B. (Ed), & Regier, D.A. (Ed), (2002). A research agenda for DSM—V (pp. 31–83). Arlington, VA, US: American Psychiatric Association.

Buckholtz, J. W., & Meyer-Lindenberg, A. (2012). Psychopathology and the human connectome: toward a transdiagnostic model of risk for mental illness. Neuron, 74, 990–1004. doi: 10.1016/j.neuron. 2012.06.002

Chen, Z.-Y., Jing, D., Bath, K. G., Ieraci, A., Khan, T., Siao, C.-J. . . . & Lee, .F. S. (2006). Genetic variant BDNF (Val66Met) polymorphism alters anxiety-related behavior. Science, 314, 140–143. doi: 10.1126 /science.1129663

References Chiao, J. Y., & Cheon, B. K. (2010). The weirdest brains in the world. Behavioral and Brain Sciences, 33, 88–90. doi: 10.1017 /S0140525X10000282 Cicchetti, D., & Rogosch, F. A. (1996). Equifinality and multifinality in developmental psychopathology. Development and Psychopathology, 8, 587–600. doi: 10.1017/S0954579400007318 Cicchetti, D., Rogosch, F.A., & Oshri, A. (2011). Interactive effects of cortiotropin releasing hormone receptor 1, serotonin transporter linked polymorphic region, and child maltreatment on diurnal cortisol regulation and internalizing symptomatology. Development and Psychopathology, 23, 1125–1138. doi: 10.1017/S0954579411000599 Cicchetti, D., Rogosch, F. A., & Thibodeau, E. L. (2012). The effects of child maltreatment on early signs of antisocial behavior: Genetic moderation by tryptophan hydroxylase, serotonin transporter, and monoamine oxidase A genes. Development and Psychopathology, 24, 907. doi: 10.1017/S0954579412000442 Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50, 16–25. doi: 10.1111/j.1469–7610.2008.01979.x Clark, L. A., Watson, D., & Reynolds, S. (1995). Diagnosis and classification of psychopathology: Challenges to the current system and future directions. Annual Review of Psychology, 46, 121–153. doi: 10.1146 /annurev.ps.46.020195.001005 Cole, S. W., Arevalo, J. M., Manu, K., Telzer, E. H., Kiang, L., Bower, J. E. . . . & Fuligni, A. J. (2011). Antagonistic pleiotropy at the human IL6 promoter confers genetic resilience to the pro-inflammatory effects of adverse social conditions in adolescence. Developmental Psychology, 47, 1173. doi: 10.1037/a0023871 Connolly, C. G., Wu, J., Ho, T. C., Hoeft, F., Wolkowitz, O., Eisendrath, S. . . . & Yang, T. T. (2013). Resting-state functional connectivity of subgenual anterior cingulate cortex in depressed adolescents. Biological Psychiatry, 74(12), 898–907. doi: 10.1016/j.biopsych.2013.05.036 Coolidge, F. L., Thede, L. L., & Jang, K. L. (2001). Heritability of personality disorders in childhood: A preliminary investigation. Journal of Personality Disorders, 15, 33–40. doi: 10.1521/pedi.15.1.33.18645 Costello, E. J., Compton, S. N., Keeler, G., & Angold, A. (2003). Relationships between poverty and psychopathology. Journal of the American Medical Association, 290, 2023–2029. doi: 10.1001/jama.290.15.2023 Cousijn, H., Rijpkema, M., Qin, S., van Marle, H. J., Franke, B., Hermans, E. J. . . . & Fernandez, G. (2010). Acute stress modulates genotype effects on amygdala processing in humans. Proceedings of the National Academy of Sciences, 107, 9867–9872. doi: 10.1073/pnas.1003514107 Cullen, K. R., Klimes-Dougan, B., Muetzel, R., Mueller, B. A., Camchong, J., Houri, A. . . . & Lim, K. O. (2010). Altered white matter microstructure in adolescents with major depression: A preliminary study. Journal of the American Academy of Child and Adolescent Psychiatry, 49, 173–183. Cuthbert, B. N., & Kozak, M. J. (2013). Constructing constructs for psychopathology: The NIMH research domain criteria. Journal of Abnormal Psychology, 122, 928–937. doi: 10.1037/a0034028 David, S. P., Murthy, N. V., Rabiner, E. A., Munafó, M. R., Johnstone, E. C., Jacob, R. . . . & Grasby, P. M. (2005). A functional genetic variation of the serotonin (5-HT) transporter affects 5-HT1A receptor binding in humans. The Journal of Neuroscience, 25, 2586–2590. doi: 10.1523/JNEUROSCI.3769–04.2005 De Raedt, R., Leyman, L., Baeken, C., Van Schuerbeek, P., Luypaert, R., Vanderhasselt, M.A. . . . & Dannlowski, U. (2010). Neurocognitive effects of HF-rTMS over the dorsolateral prefrontal cortex on the attentional processing of emotional information in healthy women: an event-related fMRI study. Biological Psychology, 85, 487–495. doi: 10.1016/j.biopsycho.2010.09.015

475

Dillon, D. G., Bogdan, R., Fagerness, J., Holmes, A. J., Perlis, R. H., & Pizzagalli, D. A. (2010). Variation in TREK1 gene linked to depression-resistant phenotype is associated with potentiated neural responses to rewards in humans. Human Brain Mapping, 31, 210–221. doi: 10.1002/hbm.20858. Dishion, T. J., Shaw, D. S., Connell, A., Gardner, F., Weaver, C., & Wilson, M. (2008). The Family check-up with high-risk indigent families: Preventing problem behavior by increasing parents’ positive behavior support in early childhood. Child Development, 79, 1395–1414. doi: 10.1111/j.1467–8624.2008.01195.x Drabant, E. M., Ramel, W., Edge, M. D., Hyde, L. W., Kuo, J. R., Goldin, P. R. . . . & Gross, J. J. (2012). Neural mechanisms underlying 5-HTTLPR-related sensitivity to acute stress. American Journal of Psychiatry, 169, 397–405. doi: 10.1176/appi.ajp. 2011.10111699 Egan, M. F., Goldberg, T. E., Kolachana, B. S., Callicott, J. H., Mazzanti, C. M., Straub, R. E. . . . & Weinberger, D. R. (2001). Effect of COMT Val108/158 Met genotype on frontal lobe function and risk for schizophrenia. Proceedings of the National Academy of Sciences, 98, 6917–6922. doi: 10.1073/pnas.111134598 Ellis, B. J., & Boyce, W. T. (2011). Differential susceptibility to the environment: Toward an understanding of sensitivity to developmental experiences and context. Development and Psychopathology, 23, 1–5. doi: 10.1017/S095457941000060X Erickson, K. I., Banducci, S. E., Weinstein, A. M., MacDonald, A. W., Ferrell, R. E., Halder, I. . . . & Manuck, S. B. (2013). The brain-derived neurotrophic factor Val66Met polymorphism moderates an effect of physical activity on working memory performance. Psychological Science, 24, 1770–1779. doi: 10.1177/0956797613480367 Etkin, A., Prater, K. E., Hoeft, F., Menon, V., & Schatzberg, A. F. (2010). Failure of anterior cingulate activation and connectivity with the amygdala during implicit regulation of emotional processing in generalized anxiety disorder. American Journal of Psychiatry, 167, 545–554. doi: 10.1176/appi.ajp. 2009.09070931 Etkin, A., & Wager, T. D. (2007). Functional neuroimaging of anxiety: A meta-analysis of emotional processing in PTSD, social anxiety disorder, and specific phobia. American Journal of Psychiatry, 164, 1476–1488. doi: 10.1176/appi.ajp. 2007.07030504 Fakra, E., Hyde, L. W., Gorka, A., Fisher, P. M., Munoz, K. E., Kimak, M. . . . & Hariri, A. R. (2009). Effects of HTR1A C(-1019)G on amygdala reactivity and trait anxiety. Archives of General Psychiatry, 66, 33–40. doi: 10.1001/archpsyc.66.1.33 Falk, E. B., Hyde, L. W., Mitchell, C., Faul, J., Gonzalez, R., Heitzeg, M. M. . . . & Schulenberg, J. (2013). What is a representative brain? Neuroscience meets population science. Proceedings of the National Academy of Sciences, 110, 17615–17622. doi: 10.1073/pnas.1310134110 Feder, A., Nestler, E.J., & Charney, D.S. (2009). Psychobiology and molecular genetics of resilience. Nature Reviews Neuroscience, 10, 446–457. doi: 10.1038/nrn2649 Fennema-Notestine, C., Gamst, A. C., Quinn, B. T., Pacheco, J., Jernigan, T. L., Thal, L. . . . & Gollub, R. L. (2007). Feasibility of multi-site clinical structural neuroimaging studies of aging using legacy data. Neuroinformatics, 5, 235–245. doi: 10.1007/s12021–007–9003–9 Fenno, L., Yizhar, O., & Deisseroth, K. (2011). The development and application of optogenetics. Annual Review of Neuroscience, 34, 389–412. doi: 10.1146/annurev-neuro-061010–113817 Fisher, P., & Hariri, A. (2012). Linking variability in brain chemistry and circuit function through multimodal human neuroimaging. Genes, Brain and Behavior, 11, 633–642. doi: 10.1111/j.1601–183X .2012.00786.x Fisher, P. M., & Hariri, A. R. (2013). Identifying serotonergic mechanisms underlying the corticolimbic response to threat in humans. Philosophical Transactions of the Royal Society B: Biological Sciences, 368, epub ahead of print; doi: 10.1098/rstb.2012.0192.

476

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

Fisher, P. M., Meltzer, C. C., Price, J. C., Coleman, F. L., Ziolko, S. K., Becker, C. . . . & Hariri, A. R. (2009). Medial prefrontal cortex 5-HT2A density is correlated with amygdala reactivity, response habituation, and functional coupling. Cerebral Cortex, 19, 2499–2507. doi: 10.1093/cercor/bhp022 Fisher, P. M., Meltzer, C. C., Ziolko, S. K., Price, J. C., & Hariri, A. R. (2006). Capacity for 5-HT1A–mediated autoregulation predicts amygdala reactivity. Nature Neuroscience, 9, 1362–1363. doi: 10.1038/nn1780 Fjell, A. M., Walhovd, K. B., Brown, T. T., Kuperman, J. M., Chung, Y., Hagler, D. J. . . . & Dale, A. M. (2012). Multimodal imaging of the self-regulating developing brain. Proceedings of the National Academy of Sciences, 109, 19620–19625. doi: 10.1073/pnas.1208243109 Foland-Ross, L. C., & Gotlib, I. H. (2012). Cognitive and neural aspects of information processing in major depressive disorder: an integrative perspective. Frontiers in psychology, 3. doi: 10.3389/fpsyg.2012.00489 Forbes, E. E., Brown, S. M., Kimak, M., Ferrell, R. E., Manuck, S. B., & Hariri, A. R. (2009). Genetic variation in components of dopamine neurotransmission impacts ventral striatal reactivity associated with impulsivity. Molecular Psychiatry, 14, 60–70. doi: 10.1038 /sj.mp. 4002086 Fortier, E., Noreau, A., Lepore, F., Boivin, M., Pérusse, D., Rouleau, G. A., & Rouleau, G. A. (2010). Early impact of 5-HTTLPR polymorphism on the neural correlates of sadness. Neuroscience Letters, 485, 261–265. doi: 10.1016/j.neulet.2010.09.026 Fowler, J. S., Alia-Klein, N., Kriplani, A., Logan, J., Williams, B., Zhu, W. . . . & Wang, G-J. (2007). Evidence that brain MAO A activity does not correspond to MAO A genotype in healthy male subjects. Biological Psychiatry, 62, 355–358. doi: 10.1016/j.biopsych .2006.08.038 Franklin, T. B., & Mansuy, I. M. (2010). Epigenetic inheritance in mammals: Evidence for the impact of adverse environmental effects. Neurobiology of Disease, 39, 61–65. doi: 10.1016/j.nbd.2009.11.012 Fujimura, J. H., & Rajagopalan, R. (2011). Different differences: The use of ‘genetic ancestry’ versus race in biomedical human genetic research. Social Studies of Science, 41, 5–30. doi: 10.1177/0306312710379170 Furman, D. J., Chen, M. C., & Gotlib, I. H. (2011). Variant in oxytocin receptor gene is associated with amygdala volume. Psychoneuroendocrinology, 36, 891–897. doi: 10.1016/j.psyneuen.2010.12.004

Gianaros, P. J., Jennings, J. R., Sheu, L. K., Greer, P. J., Kuller, L. H., & Matthews, K. A. (2007). Prospective reports of chronic life stress predict decreased grey matter volume in the hippocampus. Neuroimage, 35, 795–803. doi: 10.1016/j.neuroimage.2006.10.045 Gianaros, P. J., Manuck, S. B., Sheu, L. K., Kuan, D. C. H., Votruba-Drzal, E., Craig, A. E., & Hariri, A. R. (2011). Parental education predicts corticostriatal functionality in adulthood. Cerebral Cortex, 21, 896–910. doi: 10.1093/cercor/bhq160 Glatt, C. E., & Freimer, N. B. (2002). Association analysis of candidate genes for neuropsychiatric disease: the perpetual campaign. Trends in Genetics, 18, 307–312. doi: 10.1016/S0168–9525(02)02670–7 Glaser, Y. G., Zubieta, J.-K., Hsu, D. T., Villafruerte, S., Mickey, B. J., Trucco, E. M. . . . & Heitzeg, M. M. (2014). Indirect effect of corticotropin-releasing hormone receptor 1 gene variation on negative emotionality and alcohol use via right ventrolateral prefrontal cortex. Journal of Neuroscience, 34, 4009–4107. doi: 10.1523/JNEUROSCI .3672–13.2014 Glenn, A. L. (2011). The other allele: Exploring the long allele of the serotonin transporter gene as a potential risk factor for psychopathy: A review of the parallels in findings. Neuroscience & Biobehavioral Reviews, 35, 612–620. doi: 10.1016/j.neubiorev.2010.07.005 Gogtay, N., Giedd, J. N., Lusk, L., Hayashi, K. M., Greenstein, D., Vaituzis, A. C . . . & Thompson, P. M. (2004). Dynamic mapping of human cortical development during childhood through early adulthood. Proceedings of the National Academy of Sciences of the United States of America, 101, 8174–8179. doi: 10.1073/pnas.0402680101 Grishkevich, V., & Yanai, I. (2013). The genomic determinants of genotype× environment interactions in gene expression. Trends in Genetics, 29, 479–487. doi: 10.1016/j.tig.2013.05.006 Groenewold, N. A., Opmeer, E. M., de Jonge, P., Aleman, A., & Costafreda, S. G. (2013). Emotional valence modulates brain functional abnormalities in depression: evidence from a meta-analysis of fMRI studies. Neuroscience and Biobehavioral Reviews, 37, 152–163. doi: 10.1016/j.neubiorev.2012.11.015 Gross, C., Zhuang, X., Stark, K., Ramboz, S., Oosting, R., Kirby, L. . . . & Hen, R. (2002). Serotonin1A receptor acts during development to establish normal anxiety-like behaviour in the adult. Nature, 416, 396–400. doi: 10.1038/416396a

Furman, D. J., Hamilton, J. P., Joormann, J., & Gotlib, I. H. (2011). Altered timing of amygdala activation during sad mood elaboration as a function of 5-HTTLPR. Social Cognitive and Affective Neuroscience, 6, 270–276. doi: 10.1093/scan/nsq029

Gruenewald, T. L., Seeman, T. E., Ryff, C. D., Karlamangla, A. S., & Singer, B. H. (2006). Combinations of biomarkers predictive of later life mortality. Proceedings of the National Academy of Sciences of the United States of America, 103, 14158–14163. doi: 10.1073 /pnas.0606215103

Furmark, T., Appel, L., Henningsson, S., Åhs, F., Faria, V., Linnman, C. . . . & Fredrikson, M. (2008). A link between serotoninrelated gene polymorphisms, amygdala activity, and placebo-induced relief from social anxiety. Journal of Neuroscience, 28, 13066–13074. doi: 10.1523/JNEUROSCI.2534–08.2008

Gunduz-Cinar, O., Macpherson, K., Cinar, R., Gamble-George, J., Sugden, K., Williams, B. . . . & Holmes, A. (2013). Convergent translational evidence of a role for anandamide in amygdala-mediated fear extinction, threat processing and stress-reactivity. Molecular Psychiatry, 18, 813–823. doi: 10.1038/mp. 2012.72

Gee, D. G., Humphreys, K. L., Flannery, J., Goff, B., Telzer, E. H., Shapiro, M. . . . Tottenham, N. (2013). A developmental shift from positive to negative connectivity in human amygdala–prefrontal circuitry. Journal of Neuroscience, 33, 4584–4593. doi: 10.1523 /JNEUROSCI.3446–12.2013

Guyer, A. E., Lau, J. Y., McClure-Tone, E. B., Parrish, J., Shiffrin, N. D., Reynolds, R. C . . . & Nelson, E. E. (2008). Amygdala and ventrolateral prefrontal cortex function during anticipated peer evaluation in pediatric social anxiety. Archives of General Psychiatry, 65, 1303–1312. doi: 10.1001/archpsyc.65.11.1303

Gerritsen, L., Tendolkar, I., Franke, B., Vasquez, A., Kooijman, S., Buitelaar, J., . . . & Rijpkema, M. (2011). BDNF Val66Met genotype modulates the effect of childhood adversity on subgenual anterior cingulate cortex volume in healthy subjects. Molecular Psychiatry, 17, 597–603. doi: 10.1038/mp. 2011.51.

Hackman, D. A., Farah, M. J., & Meaney, M. J. (2010). Socioeconomic status and the brain: mechanistic insights from human and animal research. Nature Reviews Neuroscience, 11, 651–659. doi: 10.1038 /nrn2897

Gianaros, P. J., Horenstein, J. A., Hariri, A. R., Sheu, L. K., Manuck, S. B., Matthews, K. A. . . . & Cohen, S. (2008). Potential neural embedding of parental social standing. Social Cognitive and Affective Neuroscience, 3, 91–96. doi: 10.1093/scan/nsn003

Hahn, A., Stein, P., Windischberger, C., Weissenbacher, A., Spindelegger, C., Moser, E. . . . & Lanzenberger, R. (2011). Reduced resting-state functional connectivity between amygdala and orbitofrontal cortex in social anxiety disorder. Neuroimage, 56, 881–889. doi: 10.1016 /j.neuroimage.2011.02.064

References Hankin, B. L. (2012). Future directions in vulnerability to depression among youth: Integrating risk factors and processes across multiple levels of analysis. Journal of Clinical Child & Adolescent Psychology, 41, 695–718. doi: 10.1080/15374416.2012.711708 Hariri, A. R. (2009). The neurobiology of individual differences in complex behavioral traits. Annual Review of Neuroscience, 32, 225–247. doi: 10.1146/annurev.neuro.051508.135335 Hariri, A. R., Drabant, E. M., Munoz, K. E., Kolachana, B. S., Mattay, V. S., Egan, M. F. . . . & Weinberger, D. R. (2005). A susceptibility gene for affective disorders and the response of the human amygdala. Archives of Gen Psychiatry, 62, 146–152. doi: 10.1001/archpsyc.62 .2.146 Hariri, A. R., Drabant, E. M., & Weinberger, D. R. (2006). Imaging genetics: perspectives from studies of genetically driven variation in serotonin function and corticolimbic affective processing. Biological Psychiatry, 59, 888–897. doi: 10.1016/j.biopsych.2005.11.005 Hariri, A. R., Gorka, A., Hyde, L. W., Kimak, M., Halder, I., Ducci, F. . . . & Manuck, S. B. (2009). Divergent effects of genetic variation in endocannabinoid signaling on human threat- and reward-related brain function. Biological Psychiatry, 66, 9–16. doi: 10.1016/j.biopsych .2008.10.047 Hariri, A. R., & Holmes, A. (2006). Genetics of emotional regulation: the role of the serotonin transporter in neural function. Trends in Cognitive Sciences, 10, 182–191. doi: 10.1016/j.tics.2006.02.011 Hariri, A. R., Mattay, V. S., Tessitore, A., Kolachana, B., Fera, F., Goldman, D. . . . & Weinberger, D. R. (2002). Serotonin transporter genetic variation and the response of the human amygdala. Science, 297, 400. doi: 10.1126/science.1071829 Hariri, A. R., & Weinberger, D. R. (2003). Imaging genomics. British Medical Bulletin, 65, 259. doi: 10.1093/bmb/65.1.259 Hasler, G., & Northoff, G. (2011). Discovering imaging endophenotypes for major depression. Molecular Psychiatry, 1, 1–16. doi: 10.1038 /mp. 2011.23 Hattingh, C. J., Ipser, J., Tromp, S. A., Syal, S., Lochner, C., Brooks, S. J. . . . & Stein, D. J. (2012). Functional magnetic resonance imaging during emotion recognition in social anxiety disorder: an activation likelihood meta-analysis. Front Hum Neurosci, 6, 347. doi: 10.3389 /fnhum.2012.00347 Heinz, A., Goldman, D., Jones, D. W., Palmour, R., Hommer, D., Gorey, J. G. . . . & Weinberger, D. R. (2000). Genotype influences in vivo dopamine transporter availability in human striatum. Neuropsychopharmacology, 22, 133–139. doi: 10.1016/S0893–133X(99)00099–8 Heinz, A., Jones, D. W., Mazzanti, C., Goldman, D., Ragan, P., Hommer, D. . . . & Weinberger, D. R. (2000). A relationship between serotonin transporter genotype and in vivo protein expression and alcohol neurotoxicity. Biological Psychiatry, 47, 643–649. doi: 10.1016/S0006–3223(99)00171–7 Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world. Behavioral and Brain Sciences, 33, 61–83. doi: 10.1017 /S0140525X0999152X Heurteaux, C., Lucas, G., Guy, N., El Yacoubi, M., Thümmler, S., Peng, X.-D. . . . & Lazdunski, M. (2006). Deletion of the background potassium channel TREK-1 results in a depression-resistant phenotype. Nature Neuroscience, 9, 1134–1141. doi: 10.1038/nn1749 Hizer, S. E., Wright, T. M., & Garcia, D. K. (2004). Genetic markers applied in regression tree prediction models. Animal Genetics, 35, 50–52. doi:10.1046/j.1365–2052.2003.01068.x Holmes, A. (2001). Targeted gene mutation approaches to the study of anxiety-like behavior in mice. Neuroscience & Biobehavioral Reviews, 25, 261–273. doi: 10.1016/S0149–7634(01)00012–4

477

Holmes, A. (2008). Genetic variation in cortico-amygdala serotonin function and risk for stress-related disease. Neuroscience and Biobehavioral Reviews, 32, 1293–1314. doi: 10.1016/j.neubiorev.2008.03.006 Holtzheimer, P. E., & Mayberg, H. S. (2011). Deep brain stimulation for psychiatric disorders. Annual Review of Neuroscience, 34, 289–307. doi: 10.1146/annurev-neuro-061010–113638 Honey, G., & Bullmore, E. (2004). Human pharmacological MRI. Trends in Pharmacological Sciences, 25, 366–374. doi: 10.1016/j.tips .2004.05.009 Houston, I., Peter, C. J., Mitchell, A., Straubhaar, J., Rogaev, E., & Akbarian, S. (2013). Epigenetics in the human brain. Neuropsychopharmacology, 38, 183–197. doi: 10.1038/npp. 2012.78 Hu, X.-Z., Lipsky, R. H., Zhu, G., Akhtar, L. A., Taubman, J., Greenberg, B. D. . . . & Goldman, D. (2006). Serotonin transporter promoter gain-of-function genotypes are linked to obsessive-compulsive disorder. American Journal of Human Genetics, 78, 815–826. doi: 10.1086 /503850 Hyde, L. W. (2015). Developmental psychopathology in an era of molecular genetics and neuroimaging: A developmental neurogenetics approach. Development and Psychopathology, 27, 587–613. doi: 10.1017/S0954579415000188 Hyde, L. W., Bogdan, R., & Hariri, A. R. (2011). Understanding risk for psychopathology through imaging gene–environment interactions. Trends in Cognitive Sciences, 15, 417–427. doi: 10.1016 /j.tics.2011.07.001 Hyde, L. W., Byrd, A. L., Hariri, A. R., & Manuck, S. B. (2014). Amygdala reactivity and negative emotionality: Divergent correlates of antisocial personality and psychopathy traits in a community sample. Journal of Abnormal Psychology, 123, 214–224. doi: 10.1037/a0035467 Hyde, L. W., Gorka, A., Manuck, S. B., & Hariri, A. R. (2011). Social support moderates the link between amygdala reactivity and trait anxiety. Neuropsychologia, 49, 651–656. doi: 10.1016 /j.neuropsychologia.2010.08.025 Hyde, L. W., Shaw, D. S., & Hariri, A. R. (2013). Neuroscience, developmental psychopathology and youth antisocial behavior: Review, integration, and directions for research. Developmental Review, 33, 168–223. doi: 10.1016/j.dr.2013.06.001 Hyde, L. W., Waller, R., & Burt, S. A. (2014). Commentary on Dadds et al. (2013): Improving treatment for youth with callous-unemotional traits through the intersection of basic and applied science. Journal of Child Psychology and Psychiatry, in press. Insel, T. R., Cuthbert, B. N., Garvey, M. A., Heinssen, R. K., Pine, D. S., Quinn, K. J. . . . & Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167, 748–751. doi: 10.1176/appi.ajp. 2010.09091379 Ising, M., Horstmann, S., Kloiber, S., Lucae, S., Binder, E. B., Kern, N. . . . & Holsboer, F. (2007). Combined dexamethasone/corticotropin releasing hormone test predicts treatment response in major depression–A potential biomarker? Biological Psychiatry, 62, 47–54. doi: 10.1016/j.biopsych.2006.07.039 Jack, C. R., Bernstein, M. A., Fox, N. C., Thompson, P., Alexander, G., Harvey, D. . . . & Weiner, M. W. (2008). The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Resonance Imaging, 27, 685–691. doi: 10.1002/jmri.21049 Jaffee, S. R. (2011). Genotype–environment correlations: Definitions, methods of measurement, and implications for research on adolescent psychopathology. In K. S. Kendler, S. R. Jaffee, & D. Romer (Eds.), The dynamic genome and mental health (pp. 79–102). New York, NY: Oxford University Press.

478

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

Jaffee, S. R., Caspi, A., Moffitt, T. E., Dodge, K. A., Rutter, M., Taylor, A., & Tully, L. A. (2005). Nature X nurture: Genetic vulnerabilities interact with physical maltreatment to promote conduct problems. Development and Psychopathology, 17, 67–84. doi: 10.1017 /S0954579405050042

Kendler, K. S., Kuhn, J. W., Vittum, J., Prescott, C. A., & Riley, B. (2005). The interaction of stressful life events and a serotonin transporter polymorphism in the prediction of episodes of major depression: areplication. Archives of General Psychiatry, 62, 529. doi: 10.1001/archpsyc.62.5.529

Jahanshad, N., Kochunov, P., Sprooten, E., Mandl, R. C., Nichols, T. E., Almassy, L. . . . & McIntosh, A. M. (2014). Multi-site genetic analysis of diffusion images and voxelwise heritability analysis: A pilot project of the ENIGMA–DTI working group. Neuroimage, 81, 455–469. doi: 10.1016/j.neuroimage.2013.1004.1061.

Kendler, K. S., & Myers, J. (2009). A developmental twin study of church attendance and alcohol and nicotine consumption: amodel for analyzing the changing impact of genes and environment. American Journal of Psychiatry, 166, 1150–1155. doi: 10.1176/appi.ajp.2009 .09020182

Jahanshad, N., Kohannim, O., Hibar, D. P., Stein, J. L., McMahon, K. L., de Zubicaray, G. I. . . . & Thompson, P. M. (2012). Brain structure in healthy adults is related to serum transferrin and the H63D polymorphism in the HFE gene. Proceedings of the National Academy of Sciences, 109, E851–E859. doi: 10.1073/pnas.1105543109

Kilpatrick, D. G., Koenen, K. C., Ruggiero, K. J., Acierno, R., Galea, S., Resnick, H. S. . . . & Gelernter, J. (2007). The serotonin transporter genotype and social support and moderation of posttraumatic stress disorder and depression in hurricane-exposed adults. American Journal of Psychiatry, 164, 1693–1699. doi: 10.1176/appi.ajp.2007 .06122007

Jahanshad, N., Rajagopalan, P., Hua, X., Hibar, D. P., Nir, T. M., Toga, A. W. . . . & The Alzheimer’s Disease Neuroimaging Initiative (2013). Genome-wide scan of healthy human connectome discovers SPON1 gene variant influencing dementia severity. Proceedings of the National Academy of Sciences, 110, 4768–4773. doi: 10.1073 /pnas.1216206110 Jang, J. H., Jung, W. H., Choi, J.-S., Choi, C.-H., Kang, D.-H., Shin, N. Y. . . . & Kwon, J. S. (2011). Reduced prefrontal functional connectivity in the default mode network is related to greater psychopathology in subjects with high genetic loading for schizophrenia. Schizophrenia Research, 127, 58–65. doi: 10.1016/j.schres.2010.12.022 Jedema, H. P., Gianaros, P. J., Greer, P. J., Kerr, D. D., Liu, S., Higley, J. D. . . . & Bradberry, C. W. (2010). Cognitive impact of genetic variation of the serotonin transporter in primates is associated with differences in brain morphology rather than serotonin neurotransmission. Molecular Psychiatry, 15, 512–522. doi: 10.1038/mp. 2009.90 Joeyen-Waldorf, J., Nikolova, Y. S., Edgar, N., Walsh, C., Kota, R., Lewis, D. A. . . . & Sibille, E. (2012). Adenylate cyclase 7 is implicated in the biology of depression and modulation of affective neural circuitry. Biological Psychiatry, 71, 627–632. doi: 10.1016/j.biopsych.2011.11.029 Kalin, N., Shelton, S., Fox, A., Rogers, J., Oakes, T., & Davidson, R. (2008). The serotonin transporter genotype is associated with intermediate brain phenotypes that depend on the context of eliciting stressor. Molecular Psychiatry, 13, 1021–1027. doi: 10.1038/mp. 2008.37 Karg, K., Burmeister, M., Shedden, K., & Sen, S. (2011). The serotonin transporter promoter variant (5-HTTLPR), stress, and depression meta-analysis revisited: Evidence of genetic moderation. Archives of General Psychiatry, 68, 444–454. doi: 10.1001/archgenpsychiatry .2010.189 Kauffman, S. (1996). At home in the universe: The search for the laws of self-organization and complexity. New York, NY: Oxford University Press. Kaufman, J., Yang, B. Z., Douglas-Palumberi, H., Houshyar, S., Lipschitz, D., Krystal, J. H., & Gelernter, J. (2004). Social supports and serotonin transporter gene moderate depression in maltreated children. Proceedings of the National Academy of Sciences of the United States of America, 101, 17316–17321. doi: 10.1073/pnas.0404376101

Kim, D.-K., Tolliver, T. J., Huang, S.-J., Martin, B. J., Andrews, A. M., Wichems, C. . . . & Murphy, D. L. (2005). Altered serotonin synthesis, turnover and dynamic regulation in multiple brain regions of mice lacking the serotonin transporter. Neuropharmacology, 49, 798–810. doi: 10.1016/j.neuropharm.2005.08.010 King, A. P., & Liberzon, I. (2009). Assessing the neuroendocrine stress response in the functional neuroimaging context. Neuroimage, 47, 1116–1124. doi: 10.1016/j.neuroimage.2009.05.055 Kinnally, E. L., Capitanio, J. P., Leibel, R., Deng, L., LeDuc, C., Haghighi, F. . . . & Mann, J. (2010). Epigenetic regulation of serotonin transporter expression and behavior in infant rhesus macaques. Genes, Brain and Behavior, 9, 575–582. doi: 10.1111/j.1601–183X.2010 .00588.x Kinnally, E. L., Feinberg, C., Kim, D., Ferguson, K., Leibel, R., Coplan, J. D. . . . & Mann, J. (2011). DNA methylation as a risk factor in the effects of early life stress. Brain, Behavior, and Immunity, 25, 1548–1553. doi: 10.1016/j.bbi.2011.05.001 Kirsch, P., Esslinger, C., Chen, Q., Mier, D., Lis, S., Siddhanti, S. . . . & Meyer-Lindenberg, A. (2005). Oxytocin modulates neural circuitry for social cognition and fear in humans. Journal of Neuroscience, 25, 11489–11493. doi: 10.1523/JNEUROSCI.3984–05.2005 Koolschijn, P. C. M., Schel, M. A., de Rooij, M., Rombouts, S. A., & Crone, E. A. (2011). A three-year longitudinal functional magnetic resonance imaging study of performance monitoring and test–retest reliability from childhood to early adulthood. Journal of Neuroscience, 31, 4204–4212. doi: 10.1523/JNEUROSCI.6415–10.2011 Krueger, R. F. (1999). The structure of common mental disorders. Archives of General Psychiatry, 56, 921–926. doi: 10.1001/archpsyc .56.10.921 Krueger, R. F., & Markon, K. E. (2006). Reinterpreting comorbidity: A model-based approach to understanding and classifying psychopathology. Annual Review of Clinical Psychology, 2, 111–133. doi: 10.1146/annurev.clinpsy.2.022305.095213 Krueger, R. F., & Markon, K. E. (2011). A dimensional-spectrum model of psychopathology: progress and opportunities. Archives of General Psychiatry, 68, 10. doi: 10.1001/archgenpsychiatry.2010.188

Kempton, M. J., Salvador, Z., Munafò, M. R., Geddes, J. R., Simmons, A., Frangou, S. . . . & Williams, S. C. R. (2011). Structural neuroimaging studies in major depressive disorder: Meta-analysis and comparison with bipolar disorder. Archives of General Psychiatry, 68, 675–690. doi: 10.1001/archgenpsychiatry.2011.60

Krueger, R. F., Markon, K. E., Patrick, C. J., Benning, S. D., & Kramer, M. D. (2007). Linking antisocial behavior, substance use, and personality: an integrative quantitative model of the adult externalizing spectrum. Journal of Abnormal Psychology, 116, 645–666. doi: 10.1037 /0021–843X.116.4.645

Kendler, K. S. (2011). A conceptual overview of gene–environment interaction and correlation in a developmental context. In K. S. Kendler, S. R. Jaffee, & D. Romer (Eds.), The dynamic genome and mental health (pp. 5–28). New York, NY: Oxford University Press.

Kumsta, R., Stevens, S., Brookes, K., Schlotz, W., Castle, J., Beckett, C., et al. (2010). 5HTT genotype moderates the influence of early institutional deprivation on emotional problems in adolescence: evidence from the English and Romanian Adoptee (ERA) study. Journal of

References Child Psychology and Psychiatry and Allied Disciplines, 51, 755–762. doi: 10.1111/j.1469–7610.2010.02249.x Kupfer, D. J., & Regier, D. A. (2011). Neuroscience, clinical evidence, and the future of psychiatric classification in DSM-5. American Journal of Psychiatry, 168, 672–674. doi: 10.1176/appi.ajp. 2011.11020219 Lahey, B. B. (2009). Public health significance of neuroticism. American Psychologist, 64, 241–256. doi: 10.1037/a0015309 Lahey, B. B., Applegate, B., Hakes, J. K., Zald, D. H., Hariri, A. R., & Rathouz, P. J. (2012). Is there a general factor of prevalent psychopathology during adulthood? Journal of Abnormal Psychology, 121, 971–977. doi: 10.1037/a0028355 Lau, J. Y., Goldman, D., Buzas, B., Fromm, S. J., Guyer, A. E., Hodgkinson, C. . . . & Ernst, M. (2009). Amygdala function and 5-HTT gene variants in adolescent anxiety and major depressive disorder. Biological Psychiatry, 65, 349–355. doi: 10.1016/j.biopsych.2008.08.037 Lau, J. Y., Goldman, D., Buzas, B., Hodgkinson, C., Leibenluft, E., Nelson, E. . . . & Ernst, M. (2010). BDNF gene polymorphism (Val66Met) predicts amygdala and anterior hippocampus responses to emotional faces in anxious and depressed adolescents. Neuroimage, 53, 952–961. Laucht, M., Treutlein, J., Blomeyer, D., Buchmann, A. F., Schmid, B., Becker, K. . . . & Banaschewski, T. (2009). Interaction between the 5-HTTLPR serotonin transporter polymoprhism and environmental adversity for mood and anxiety psychopathology: Evidence from a high-risk community sample of young adults. International Journal of Neuropsychopharmacology, 12, 737–747. doi: 10.1017 /S1461145708009875 Lemonde, S., Turecki, G., Bakish, D., Du, L., Hrdina, P. D., Bown, C. D. . . . & Alberty, P. R. (2003). Impaired repression at a 5-hydroxytryptamine 1A receptor gene polymorphism associated with major depression and suicide. Journal of Neuroscience, 23, 8788. Lenroot, R. K., & Giedd, J. N. (2011). Annual research review: Developmental considerations of gene by environment interactions. Journal of Child Psychology and Psychiatry, 52, 429–441. doi: 10.1111 /j.1469–7610.2011.02381.x Lesch, K. P., Bengel, D., Heils, A., Sabol, S. Z., Greenberg, B. D., Petri, S. . . . & Murphy, D. L. (1996). Association of anxiety-related traits with a polymorphism in the serotonin transporter gene regulatory region. Science, 274, 1527–1531. doi: 10.1126/science.274.5292.1527 Leve, L. D., Kerr, D. C. R., Shaw, D., Ge, X., Neiderhiser, J. M., Scaramella, L. V. . . . & Reiss, D. (2010). Infant pathways to externalizing behavior: evidence of Genotype × Environment interaction. Child Development, 81, 340–356. doi: 10.1111/j.1467–8624.2009.01398.x Lichtenstein, P., & Annas, P. (2000). Heritability and prevalence of specific fears and phobias in childhood. Journal of Child Psychology and Psychiatry, 41, 927–937. doi: 10.1111/1469–7610.00680 Liou, Y. J., Chen, T. J., Tsai, S. J., Yu, Y. W., Cheng, C. Y., & Hong, C. J. (2009). Support for the involvement of the KCNK2 gene in major depressive disorder and response to antidepressant treatment. Pharmacogenet Genomics, 19, 735–741. doi: 10.1097/FPC .0b013e32832cbe61 Lissek, S., Kaczkurkin, A. N., Rabin, S., Geraci, M., Pine, D. S., & Grillon, C. (2013). Generalized anxiety disorder is associated with overgeneralization of classically conditioned fear. Biological Psychiatry, 75, 909–915. doi: 10.1016/j.biopsych.2013.07.025 Lissek, S., Powers, A. S., McClure, E. B., Phelps, E. A., Woldehawariat, G., Grillon, C. . . . & Pine, D. S. (2005). Classical fear conditioning in the anxiety disorders: a meta-analysis. Behaviour Research and Therapy, 43, 1391–1424. doi: 10.1016/j.brat.2004.10.007 Liu, X., Akula, N., Skup, M., Brotman, M. A., Leibenluft, E., & McMahon, F. J. (2010). A genome-wide association study of amygdala activation in youths with and without bipolar disorder. Journal of the

479

American Academy of Child and Adolescent Psychiatry, 49, 33–41. doi: 10.1016/j.jaac.2009.10.006 Long, H., Liu, B., Hou, B., Wang, C., Li, J., Qin, W. . . . & Jian, T. (2013). The long rather than the short allele of 5-HTTLPR predisposes Han Chinese to anxiety and reduced connectivity between prefrontal cortex and amygdala. Neuroscience bulletin, 29, 4–15. doi:10.1007/s12264–013–1299-x Lu, Q., Li, H., Luo, G., Wang, Y., Tang, H., Han, L., & Yao, Z. (2012). Impaired prefrontal-amygdala effective connectivity is responsible for the dysfunction of emotion process in major depressive disorder: A dynamic causal modeling study on MEG. Neuroscience Letters, 523, 125–130. doi: 10.1016/j.neulet.2012.06.058 Maher, B. (2008). Personal genomes: The case of the missing heritability. Nature, 456, 18–21. doi: 10.1038/456018a Malhotra, A. K., & Goldman, D. (1999). Benefits and pitfalls encountered in psychiatric genetic association studies. Biological Psychiatry, 45, 544–550. doi:10.1016/S0006–3223(98)00365–5 Manuck, S. B. (2009). The reaction norm in gene × environment interaction. Molecular Psychiatry, 15, 881–882. doi: 10.1038/mp. 2009.139 Manuck, S. B. (2013). Gene–Environment Interaction: Once and future prospect. Annual Review of Psychology, 65, 41–70. doi: 10.1146 /annurev-psych-010213–115100 Manuck, S. B., & McCaffery, J. M. (2010). Genetics of stress: gene–stress correlation and interaction. In A. Steptoe (Ed.), Handbook of behavioral medicine: Methods and applications (pp. 454–478). New York, NY: Springer. Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G., Morris, J. C., & Buckner, R. L. (2007). Open access series of imaging studies (OASIS): cross-sectional MRI data in young, middle aged, nondemented, and demented older adults. Journal of Cognitive Neuroscience, 19, 1498–1507. doi: 10.1162/jocn.2007.19.9.1498 Masten, A. S. (2001). Ordinary magic: Resilience processes in development. American Psychologist, 56, 227. doi: 10.1037/0003–066X .56.3.227 Masten, A. S., & Cicchetti, D. (2010). Developmental cascades. Development and Psychopathology, 22, 491–495. doi: 10.1017/S09545794 10000222 Mazella, J., Petrault, O., Lucas, G., Deval, E., Beraud-Dufour, S., Gandin, C. . . . & Borsotto, M. (2010). Spadin, a sortilin-derived peptide, targeting rodent TREK-1 channels: a new concept in the antidepressant drug design. PLoS Biol, 8, e1000355. doi: 10.1371 /journal.pbio.1000355 McClure, E. B., Monk, C. S., Nelson, E. E., Parrish, J. M., Adler, A., Blair, R. J. R. . . . & Pine, D. S. (2007). Abnormal attention modulation of fear circuit function in pediatric generalized anxiety disorder. Archives of General Psychiatry, 64, 97–106. doi: 10.1001/archpsyc.64.1.97 McCrory, E. J., De Brito, S. A., Kelly, P. A., Bird, G., Sebastian, C. L., Mechelli, A. . . . & Viding, E. (2013). Amygdala activation in maltreated children during pre-attentive emotional processing. British Journal of Psychiatry, 202. 269–276. doi: 10.1192/bjp.bp. 1112.116624. McGowan, P. O., Sasaki, A., D’Alessio, A. C., Dymov, S., Labonté, B., Szyf, M. . . . & Meaney, M. J. (2009). Epigenetic regulation of the glucocorticoid receptor in human brain associates with childhood abuse. Nature Neuroscience, 12, 342–348. doi: 10.1038/nn.2270 Meaney, M. J. (2010). Epigenetics and the biological definition of gene × environment interactions. Child Development, 81, 41–79. doi: 10.1111/j.1467–8624.2009.01381.x Meyer-Lindenberg, A. (2011). Neurogenetic mechanisms of gene–environment interactions. In K. A. Dodge & M. Rutter (Eds.), Gene–environment interactions in developmental psychopathology (pp. 71–87). New York, NY: Guilford.

480

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

Meyer-Lindenberg, A., Buckholtz, J. W., Kolachana, B., A, R. H., Pezawas, L., Blasi, G. . . . & Weinberger, D. R. (2006). Neural mechanisms of genetic risk for impulsivity and violence in humans. Proceedings of the National Academy of Sciences of the United States of America, 103, 6269–6274. doi: 10.1073/pnas.0511311103 Milad, M. R., & Quirk, G. J. (2012). Fear extinction as a model for translational neuroscience: ten years of progress. Annual Review of Psychology, 63, 129–151. doi: 10.1146/annurev.psych.121208.131631 Moffitt, T. E., Caspi, A., & Rutter, M. (2005). Strategy for investigating interactions between measured genes and measured environments. Archives of General Psychiatry, 62, 473–481. doi: 10.1001/archpsyc .62.5.473 Moffitt, T. E., Caspi, A., & Rutter, M. (2006). Measured gene–environment interactions in psychopathology. Perspectives on Psychological Science, 1, 5. doi: 10.1111/j.1745–6916.2006.00002.x Moldin, S. O., & Gottesman, I. I. (1997). Genes, experience, and chance in schizophrenia—Positioning for the 21st century. Schizophrenia Bulletin, 23, 547–561. doi: 10.1093/schbul/23.4.547 Monk, C. S., Nelson, E. E., McClure, E. B., Mogg, K., Bradley, B. P., Leibenluft, E. . . . & Pine, D. S. (2006). Ventrolateral prefrontal cortex activation and attentional bias in response to angry faces in adolescents with generalized anxiety disorder. American Journal of Psychiatry, 163, 1091–1097. doi: 10.1176/appi.ajp. 163.6.1091 Monk, C. S., Telzer, E. H., Mogg, K., Bradley, B. P., Mai, X., Louro, H. M. . . . & Pine, D. S. (2008). Amygdala and ventrolateral prefrontal cortex activation to masked angry faces in children and adolescents with generalized anxiety disorder. Archives of General Psychiatry, 65, 568–576. doi: 10.1001/archpsyc.65.5.568 Munafo, M. R., Brown, S. M., & Hariri, A. R. (2008). Serotonin transporter (5-HTTLPR) genotype and amygdala activation: a meta-analysis. Biological Psychiatry, 63, 852–857. doi: 10.1016 /j.biopsych.2007.08.016 Munafò, M. R., Brown, S. M., & Hariri, A. R. (2008). Serotonin transporter (5-HTTLPR) genotype and amygdala activation: a meta-analysis. Biological Psychiatry, 63, 852–857. doi: 10.1016 /j.biopsych.2007.08.016 Munoz, K. E., Hyde, L. W., & Hariri, A. R. (2009). Imaging genetics. Journal of the American Academy of Child and Adolescent Psychiatry, 48, 356–361. doi: 10.1097/CHI.0b013e31819aad07 Nagin, D. S., & Tremblay, R. E. (2001). Analyzing developmental trajectories of distinct but related behaviors: A group-based method. Psychological Methods, 6, 18–34. doi: 10.1037/1082–989X.6.1.18 Nikolova, Y. S., Ferrell, R. E., Manuck, S. B., & Hariri, A. R. (2011). Multilocus genetic profile for dopamine signaling predicts ventral striatum reactivity. Neuropsychopharmacology, Epub online; doi: 10.1038/npp. 2011.82. Nikolova, Y. S., Singhi, E. K., Drabant, E. M., & Hariri, A. R. (2013). Reward-related ventral striatum reactivity mediates gender-specific effects of a galanin remote enhancer haplotype on problem drinking. Genes, Brain and Behavior, 12, 516–524 doi: 10.1111/gbb.12035 O’Connor, T. G., Deater-Deckard, K., Fulker, D., Rutter, M., & Plomin, R. (1998). Genotype–environment correlations in late childhood and early adolescence: Antisocial behavioral problems and coercive parenting. Developmental Psychology, 34, 970–981. doi: 10.1037 /0012–1649.34.5.970 Ochsner, K. N., & Gross, J. J. (2005). The cognitive control of emotion. Trends in Cognitive Sciences, 9, 242–249. doi: 10.1016 /j.tics.2005.03.010 Oler, J. A., Fox, A. S., Shelton, S. E., Christian, B. T., Murali, D., Oakes, T. R. . . . & Kalin, N. H. (2009). Serotonin transporter availability in the amygdala and bed nucleus of the stria terminalis predicts anxious

temperament and brain glucose metabolic activity. Journal of Neuroscience, 29, 9961–9966. doi: 10.1523/JNEUROSCI.0795–09.2009 Ordaz, S. J., Foran, W., Velanova, K., & Luna, B. (2013). Longitudinal growth curves of brain function underlying inhibitory control through adolescence. Journal of Neuroscience, 33, 18109–18124. doi: 10.1523/JNEUROSCI.1741–13.2013 Pacheco, J., Beevers, C. G., Benavides, C., McGeary, J., Stice, E., & Schnyer, D. M. (2009). Frontal-limbic white matter pathway associations with the serotonin transporter gene promoter region (5-HTTLPR) polymorphism. Journal of Neuroscience, 29, 6229–6233. doi: 10.1523/JNEUROSCI.0896–09.2009 Papaleo, F., Crawley, J. N., Song, J., Lipska, B. K., Pickel, J., Weinberger, D. R. . . . & Chen, J. (2008). Genetic dissection of the role of catechol-O-methyltransferase in cognition and stress reactivity in mice. The Journal of Neuroscience, 28, 8709–8723. doi: 10.1523 /JNEUROSCI.2077–08.2008 Pardini, D., & Frick, P. J. (2013). Multiple developmental pathways to conduct disorder: Current conceptualizations and clinical implications. Journal of the Canadian Academy of Child and Adolescent Psychiatry, 22, 20–25. Paus, T. (2010). Population neuroscience: Why and how. Human Brain Mapping, 31, 891–903. doi: 10.1002/hbm.21069 Pereira, F., Mitchell, T., & Botvinick, M. (2009). Machine learning classifiers and fMRI: A tutorial overview. Neuroimage, 45, S199–S209. doi: 10.1016/j.neuroimage.2008.11.007 Perez-Edgar, K., Hardee, J. E., Guyer, A. E., Benson, B., Nelson, E. E., Gorodetsky, E. . . . & Ernst, M. (2014). DRD4 and striatal modulation of the link between childhood behavioral inhibition and adolescent anxiety. Social Cognitive and Affective Neuroscience, 9, 445–453. doi: 10.1093/scan/nst001 Perlis, R. H., Moorjani, P., Fagerness, J., Purcell, S., Trivedi, M. H., Fava, M. . . . & Smoller, J. W. (2008). Pharmacogenetic analysis of genes implicated in rodent models of antidepressant response: Association of TREK1 and treatment resistance in the STAR*D study. Neuropsychopharmacology, 33, 2810–2819. doi: 10.1038/npp. 2008.6 Perlman, G., Simmons, A. N., Wu, J., Hahn, K. S., Tapert, S. F., Max, J. E. . . . & Yang, T. T. (2012). Amygdala response and functional connectivity during emotion regulation: A study of 14 depressed adolescents. Journal of Affective Disorders, 139, 75–84. doi: 10.1016/j.jad .2012.01.044 Pessoa, L., & Ungerleider, L. G. (2004). Neuroimaging studies of attention and the processing of emotion-laden stimuli. Progress in Brain Research, 144, 171–182. doi: 10.1016/S0079–6123(03)14412–3 Petersen, I. T., Bates, J. E., Goodnight, J. A., Dodge, K. A., Lansford, J. E., Pettit, G. S. . . . & Dick, D. M. (2012). Interaction between serotonin transporter polymorphism (5-HTTLPR) and stressful life events in adolescents’ trajectories of anxious/depressed symptoms. Developmental Psychology, 48, 1463. doi: 10.1037/a0027471 Pezawas, L., Meyer-Lindenberg, A., Drabant, E. M., Verchinski, B. A., Munoz, K. E., Kolachana, B. S. . . . & Weinberger, D. R. (2005). 5-HTTLPR polymorphism impacts human cingulate-amygdala interactions: a genetic susceptibility mechanism for depression. Nature Neuroscience, 8, 828–834. doi: 10.1038/nn1463 Pfeifer, J. H., Kahn, L. E., Merchant, J. S., Peake, S. J., Veroude, K., Masten, C. L. . . . & Dapretto, M. (2013). Longitudinal change in the neural bases of adolescent social self-evaluations: Effects of age and pubertal development. Journal of Neuroscience, 33, 7415–7419. doi:10.1523/JNEUROSCI.4074–12.2013 Phan, K. L., Fitzgerald, D. A., Nathan, P. J., & Tancer, M. E. (2006). Association between amygdala hyperactivity to harsh faces and severity of social anxiety in generalized social phobia. Biological Psychiatry, 59, 424–429. doi: 10.1016/j.biopsych.2005.08.012

References Phan, K. L., Orlichenko, A., Boyd, E., Angstadt, M., Coccaro, E. F., Liberzon, I. . . . & Arfanakis, K. (2009). Preliminary evidence of white matter abnormality in the uncinate fasciculus in generalized social anxiety disorder. Biological Psychiatry, 66, 691–694. doi: 10.1016/j.biopsych.2009.02.028 Philibert, R., Madan, A., Andersen, A., Cadoret, R., Packer, H., & Sandhu, H. (2007). Serotonin transporter mRNA levels are associated with the methylation of an upstream CpG island. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 144, 101–105. doi: 10.1002/ajmg.b.30414 Philibert, R. A., Sandhu, H., Hollenbeck, N., Gunter, T., Adams, W., & Madan, A. (2008). The relationship of 5HTT (SLC6A4) methylation and genotype on mRNA expression and liability to major depression and alcohol dependence in subjects from the Iowa Adoption Studies. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 147, 543–549. doi: 10.1002/ajmg.b.30657 Pizzagalli, D. A. (2010). Frontocingulate dysfunction in depression: toward biomarkers of treatment response. Neuropsychopharmacology, 36, 183–206. doi: 10.1038/npp. 2010.166 Plomin, R. (2005). Finding genes in child psychology and psychiatry: When are we going to be there? Journal of Child Psychology and Psychiatry and Allied Disciplines, 46, 1030–1038. doi:10.1111 /j.1469–7610.2005.01524.x Plomin, R., Haworth, C., & Davis, O. (2009). Common disorders are quantitative traits. Nature Reviews Genetics, 10, 872–878. doi: 10.1038/nrg2670 Plomin, R., Owen, M. J., & McGuffin, P. (1994). The genetic basis of complex human behaviors. Science, 264, 1733–1739. doi: 10.1126/science .8209254 Pluess, M., & Belsky, J. (2013). Vantage sensitivity: Individual differences in response to positive experiences. Psychological Bulletin, 139, 901–916. doi: 10.1037/a0030196 Potkin, S. G., & Ford, J. M. (2009). Widespread cortical dysfunction in schizophrenia: the FBIRN imaging consortium. Schizophrenia Bulletin, 35, 15–18. doi: 10.1093/schbul/sbn159 Prater, K. E., Hosanagar, A., Klumpp, H., Angstadt, M., & Phan, K. L. (2013). Aberrant amygdala-frontal cortex connectivity during perception of fearful faces and at rest in generalized social anxiety disorder. Depression and Anxiety, 30, 234–241. doi: 10.1002/da.22014 Preacher, K. J., Rucker, D. D., & Hayes, A. F. (2007). Addressing moderated mediation hypotheses: Theory, methods, and prescriptions. Multivariate Behavioral Research, 42, 185–227. doi: 10.1080/00273170701341316 Price, J. L., & Drevets, W. C. (2010). Neurocircuitry of mood disorders. Neuropsychopharmacology, 35, 192–216. doi: 10.1038/npp. 2009.104 Purcell, S. M., Wray, N. R., Stone, J. L., Visscher, P. M., O’Donovan, M. C., Sullivan, P. F. . . . & The International Schizophrenia Consortium. (2009). Common polygenic variation contributes to risk of schizophrenia and bipolar disorder. Nature, 460, 748–752. doi: 10.1038 /nature08185. Reiss, D., & Leve, L. D. (2007). Genetic expression outside the skin: Clues to mechanisms of genotype × environment interaction. Development and Psychopathology, 19, 1005–1027. doi: 10.1017/S0954579407000508 Ressler, K. J., Mercer, K. B., Bradley, B., Jovanovic, T., Mahan, A., Kerley, K. . . . & May, V. (2011). Post-traumatic stress disorder is associated with PACAP and the PAC1 receptor. Nature, 470, 492–497. doi: 10.1038/nature09856 Rhee, S., & Waldman, I. (2002). Genetic and environmental influences on antisocial behavior: A meta-analysis of twin and adoption studies. Psychological Bulletin, 128, 490–527. doi:10.1037 /0033–2909.128.3.490

481

Richardson-Jones, J. W., Craige, C. P., Guiard, B. P., Stephen, A., Metzger, K. L., Kung, H. F. . . . & Leonardo, E. D. (2010). 5-HT1A autoreceptor levels determine vulnerability to stress and response to antidepressants. Neuron, 65, 40–52. doi: 10.1016/j.neuron.2009.12.003 Roth, T. L., & Sweatt, J. D. (2011). Annual Research Review: Epigenetic mechanisms and environmental shaping of the brain during sensitive periods of development. Journal of Child Psychology and Psychiatry and Allied Disciplines, 52, 398–408. doi:10.1111 /j.1469–7610.2010.02282.x Rutter, M. (2006). Implications of resilience concepts for scientific understanding. Annals of the New York Academy of Sciences, 1094, 1–12. doi: 10.1196/annals.1376.002 Rutter, M., & Dodge, K. A. (2011). Gene–environment interactions: the state of science. In K. A. Dodge & M. Rutter (Eds.), Gene–environment interaction in developmental psychopathology (pp. 87–101). New York, NY: Guilford. Rutter, M., Moffitt, T. E., & Caspi, A. (2006). Gene–environment interplay and psychopathology: multiple varieties but real effects. Journal of Child Psychology and Psychiatry, 47, 226–261. doi:10.1111 /j.1469–7610.2005.01557.x Sadeh, N., Javdani, S., Jackson, J. J., Reynolds, E. K., Potenza, M. N., Gelernter, J. . . . & Verona, E. (2010). Serotonin transporter gene associations with psychopathic traits in youth vary as a function of socioeconomic resources. Journal of Abnormal Psychology, 119, 604–609. doi: 10.1037/a0019709 Sameroff, A. J. (1995). General systems theories and developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental psychopathology, Vol 1: Theory and methods (Vol. 1, pp. 659–695). Oxford, England: John Wiley and Sons. Sameroff, A. J. (2000). Developmental systems and psychopathology. Development and Psychopathology, 12, 297–312. doi: 10.1017 /S0954579400003035 Sameroff, A. J. (2010). A unified theory of development: A dialectic integration of nature and nurture. Child Development, 81, 6–22. doi:10.1111/j.1467–8624.2009.01378.x Sameroff, A. J., Seifer, R., Zax, M., & Barocas, R. (1987). Early indicators of developmental risk: Rochester longitudinal study. Schizophrenia Bulletin, 13, 383–394. doi: 10.1093/schbul/13.3.383 Satterthwaite, T. D., Ruparel, K., Loughead, J., Elliott, M. A., Gerraty, R. T., Calkins, M. E. . . . & Wolf, D. H. (2012). Being right is its own reward: Load and performance related ventral striatum activation to correct responses during a working memory task in youth. Neuroimage, 61, 723–729. doi: 10.1016/j.neuroimage.2012.03.060 Schumann, G., Loth, E., Banaschewski, T., Barbot, A., Barker, G., Büchel, C. . . . & The IMAGEN consortium (2010). The IMAGEN study: reinforcement-related behaviour in normal brain function and psychopathology. Molecular Psychiatry, 15, 1128–1139. doi: 10.1038 /mp. 2010.4 Schwarz, A. J., Gozzi, A., Reese, T., & Bifone, A. (2007). In vivo mapping of functional connectivity in neurotransmitter systems using pharmacological MRI. Neuroimage, 34, 1627–1636. doi: 10.1016 /j.neuroimage.2006.11.010 Seligman, M. E., & Csikszentmihalyi, M. (2000). Positive psychology: an introduction. American Psychologist, 55, 5–14. doi:10.1037 /0003–066X.55.1.5 Sen, S., Burmeister, M., & Ghosh, D. (2004). Meta-analysis of the association between a serotonin transporter promoter polymorphism (5-HTTLPR) and anxiety-related personality traits. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 127, 85–89. doi: 10.1002/ajmg.b.20158

482

Neurogenetics Approaches to Mapping Pathways in Developmental Psychopathology

Shaw, P., Greenstein, D., Lerch, J., Clasen, L., Lenroot, R., Gogtay, N. . . . & Giedd, J. (2006). Intellectual ability and cortical development in children and adolescents. Nature, 440, 676–679. doi: 10.1038/nature04513 Shaw, P., Kabani, N. J., Lerch, J. P., Eckstrand, K., Lenroot, R., Gogtay, N. . . . & Wise, S. P. (2008). Neurodevelopmental trajectories of the human cerebral cortex. Journal of Neuroscience, 28, 3586–3594. doi:10.1523/JNEUROSCI.5309–07.2008 Sibille, E., & Lewis, D. A. (2006). SERT-ainly involved in depression, but when? American Journal of Psychiatry, 163, 8–11. doi: 10.1176 /appi.ajp.163.1.8 Simon, G. E., & Perlis, R. H. (2010). Personalized medicine for depression: can we match patients with treatments? American Journal of Psychiatry, 167, 1445–1455. doi: 10.1176/appi.ajp.2010.09111680 Sroufe, L. A., & Rutter, M. (1984). The domain of developmental psychopathology. Child Development, 55, 17–29. doi: 10.2307/1129832 Swartz, J. R., Carrasco, M., Wiggins, J. L., Thomason, M. E., & Monk, C. S. (2014). Age-related changes in the structure and function of prefrontal cortex–amygdala circuitry in children and adolescents: A multi-modal imaging approach. Neuroimage, 86, 212–220. doi: 10.1016/j.neuroimage.2013.08.018 Swartz, J. R., & Monk, C. S. (in press). Functional magnetic resonance imaging in developmental psychopathology: Using neural function as a window into the development and treatment of psychopathology. In M. D. Lewis & K. Rudolph (Eds.), Handbook of Developmental Psychopathology (3rd ed.). Swartz, J. R., Phan, K. L., Angstadt, M., Fitzgerald, K. D., & Monk, C. S. (2014). Dynamic changes in amygdala activation and functional connectivity in children and adolescents with anxiety disorders. Development and Psychopathology, 26, 1305–1319. doi: 10.1017/s0954579414001047 Talkowski, M. E., Kirov, G., Bamne, M., Georgieva, L., Torres, G., Mansour, H. . . . & Nimgaonkar, V. L. (2008). A network of dopaminergic gene variations implicated as risk factors for schizophrenia. Human Molecular Genetics, 17, 747–758. doi: 10.1093/hmg/ddm347 Thomas, K. M., Drevets, W. C., Dahl, R. E., Ryan, N. D., Birmaher, B., Eccard, C. H. . . . & Casey, B. J. (2001). Amygdala response to fearful faces in anxious and depressed children. Archives of General Psychiatry, 58, 1057–1063. doi: 10.1001/archpsyc.58.11.1057 Thomason, M. E., Henry, M. L., Paul Hamilton, J., Joormann, J., Pine, D. S., Ernst, M. . . . & Gotlib, I. H. (2010). Neural and behavioral responses to threatening emotion faces in children as a function of the short allele of the serotonin transporter gene. Biological Psychology, 85, 38–44. doi: 10.1016/j.biopsycho.2010.04.009 Thomason, M. E., & Thompson, P. M. (2011). Diffusion imaging, white matter, and psychopathology. Annual Review of Clinical Psychology, 7, 63–85. doi: 10.1146/annurev-clinpsy-032210–104507 Thyreau, B., Schwartz, Y., Thirion, B., Frouin, V., Loth, E., Vollstädt-Klein, S. . . . & The IMAGEN Consortium. (2012). Very large fMRI study using the IMAGEN database: Sensitivity–specificity and population effect modeling in relation to the underlying anatomy. Neuroimage, 61, 295–303. doi: 10.1016/j.neuroimage.2012.02.083 Toga, A. W., Clark, K. A., Thompson, P. M., Shattuck, D. W., & Van Horn, J. D. (2012). Mapping the human connectome. Neurosurgery, 71, 1–5. doi: 10.1227/NEU.0b013e318258e9ff Toga, A. W., Thompson, P. M., & Sowell, E. R. (2006). Mapping brain maturation. Trends in Neurosciences, 29, 148–159. doi: 10.1016 /j.tins.2006.01.007 Tottenham, N., Hare, T., Millner, A., Gilhooly, T., Zevin, J., & Casey, B. (2011). Elevated amygdala response to faces following early deprivation. Developmental Science, 14, 190–204. doi: 10.1111 /j.1467–7687.2010.00971.x

Tottenham, N., Hare, T. A., Quinn, B. T., McCarry, T. W., Nurse, M., Gilhooly, T. . . . & Casey, B. J. (2010). Prolonged institutional rearing is associated with atypically large amygdala volume and difficulties in emotion regulation. Developmental Science, 13, 46–61. doi: 10.1111/j.1467–7687.2009.00852.x Tromp, D. P., Grupe, D. W., Oathes, D. J., McFarlin, D. R., Hernandez, P. J., Kral, T. R. . . . & Nitschke, J. B. (2012). Reduced structural connectivity of a major frontolimbic pathway in generalized anxiety disorder. Archives of General Psychiatry, 69, 925–934. doi: 10.1001 /archgenpsychiatry.2011.2178 Tsuang, M. T., Lyons, M. J., & Faraone, S. V. (1990). Heterogeneity of schizophrenia. Conceptual models and analytic strategies. British Journal of Psychiatry, 156, 17–26. doi: 10.1192/bjp.156.1.17 Uhr, M., Tontsch, A., Namendorf, C., Ripke, S., Lucae, S., Ising, M. . . . & Holsboer, F. (2008). Polymorphisms in the drug transporter geneABCB1 predict antidepressant treatment response in depression. Neuron, 57, 203–209. doi: 10.1016/j.neuron.2007.11.017 Ursini, G., Bollati, V., Fazio, L., Porcelli, A., Iacovelli, L., Catalani, A. . . . & Bertolino, A. (2011). Stress-Related methylation of the catechol-o-methyltransferase Val158 allele predicts human prefrontal cognition and activity. Journal of Neuroscience, 31, 6692–6698. doi:10.1523/JNEUROSCI.6631–10.2011 van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., & Ebstein, R. P. (2011). Methylation matters in child development: Toward developmental behavioral epigenetics. Child Development Perspectives, 5, 305–310. doi: 10.1111/j.1750–8606.2011.00202.x Viding, E., Williamson, D. E., & Hariri, A. R. (2006). Developmental imaging genetics: challenges and promises for translational research. Development and Psychopathology, 18, 877–892. doi: 10.1017/S0954579406060433 Vrieze, S. I., Iacono, W. G., & McGue, M. (2012). Confluence of genes, environment, development, and behavior in a post genome-wide association study world. Development and Psychopathology, 24, 1195–1214. doi: 10.1017/S0954579412000648 Walsh, N. D., Dalgleish, T., Dunn, V. J., Abbott, R., St Clair, M. C., Owens, M. . . . & Goodyer, I. M. (2012). 5-HTTLPR-environment interplay and its effects on neural reactivity in adolescents. Neuroimage, 63, 1670–1680. doi: 10.1016/j.neuroimage.2012.07.067 Watson, K. K., Ghodasra, J. H., & Platt, M. L. (2009). Serotonin transporter genotype modulates social reward and punishment in rhesus macaques. PloS one, 4, e4156. doi: 10.1371/journal.pone .0004156 Weaver, I. C. G., Cervoni, N., Champagne, F. A., D’Alessio, A. C., Sharma, S., Seckl, J. R. . . . & Meaney, M. J. (2004). Epigenetic programming by maternal behavior. Nature Neuroscience, 7, 847–854. doi: 10.1038/nn1276 Weissman, M. M., Leckman, J. F., Merikangas, K. R., Gammon, G. D., & Prusoff, B. A. (1984). Depression and anxiety disorders in parents and children: results from the Yale family study. Archives of General Psychiatry, 41, 845. doi: 10.1001/archpsyc.1984.01790200 027004 Wellman, C. L., Camp, M., Jones, V. M., MacPherson, K. P., Ihne, J., Fitzgerald, P. . . . & Holmes, A. (2013). Convergent effects of mouse Pet-1 deletion and human PET-1 variation on amygdala fear and threat processing. Experimental Neurology, 250, 260–269. doi: 10.1016/j.expneurol.2013.09.025 Wenten, M., Gauderman, W. J., Berhane, K., Lin, P. C., Peters, J., & Gilliland, F. D. (2009). Functional variants in the catalase and myeloperoxidase genes, ambient air pollution, and respiratory-related school absences: An example of epistasis in gene–environment interactions. American Journal of Epidemiology, 170, 1494–1501. doi: 10.1093/aje/kwp310

References White, M., Bogdan, R., Fisher, P., Munoz, K., Williamson, D., & Hariri, A. (2012). FKBP5 and emotional neglect interact to predict individual differences in amygdala reactivity. Genes, Brain and Behavior, 11, 869–878. doi: 10.1111/j.1601–183X.2012.00837.x Whitfield-Gabrieli, S., Thermenos, H. W., Milanovic, S., Tsuang, M. T., Faraone, S. V., McCarley, R. W. . . . & Seidman, L. J. (2009). Hyperactivity and hyperconnectivity of the default network in schizophrenia and in first-degree relatives of persons with schizophrenia. Proceedings of the National Academy of Sciences, 106, 1279–1284. doi: 10.1073 /pnas.0809141106 Wiers, C. E. (2012). Methylation and the human brain: Towards a new discipline of imaging epigenetics. European Archives of Psychiatry and Clinical Neuroscience, 262, 271–273. doi: 10.1007/s00406–011–0261-z Wiggins, J. L., Bedoyan, J. K., Carrasco, M., Swartz, J. R., Martin, D. M., & Monk, C. S. (2012). Age-related effect of serotonin transporter genotype on amygdala and prefrontal cortex function in adolescence. Human Brain Mapping, epub ahead of print; doi: 10.1002/hbm.22208. Willard, H. F., & Ginsburg, G. S. (2009). Essentials of genomic and personalized medicine. San Diego, CA: Academic Press. Willeit, M., & Praschak-Rieder, N. (2010). Imaging the effects of genetic polymorphisms on radioligand binding in the living human brain: a review on genetic neuroreceptor imaging of monoaminergic systems in psychiatry. Neuroimage, 53, 878–892. doi: 10.1016/j.neuroimage .2010.04.030

483

Williams, L. M., Gatt, J. M., Schofield, P. R., Olivieri, G., Peduto, A., & Gordon, E. (2009). “Negativity bias” in risk for depression and anxiety: brain–body fear circuitry correlates, 5-HTT-LPR and early life stress. Neuroimage, 47, 804–814. doi: 10.1016/j.neuroimage .2009.05.009 Williamson, D. E., Ryan, N. D., Birmaher, B., Dahl, R. E., Kaufman, J., Rao, U. . . . & Puig-Antich, J. (1995). A case–control family history study of depression in adolescents. Journal of the American Academy of Child and Adolescent Psychiatry, 34, 1596–1607. doi: 10.1097 /00004583–199512000–00010 Yan, C.-G., Craddock, R. C., Zuo, X.-N., Zang, Y.-F., & Milham, M. P. (2013). Standardizing the intrinsic brain: Towards robust measurement of inter-individual variation in 1000 functional connectomes. Neuroimage, 80, 246–262. doi: 10.1016/j.neuroimage .2013.04.081 Yang, T. T., Simmons, A. N., Matthews, S. C., Tapert, S. F., Frank, G. K., Max, J. E. . . . & Paulus, M. P. (2010). Adolescents with major depression demonstrate increased amygdala activation. Journal of the American Academy of Child and Adolescent Psychiatry, 49, 42–51. Yu, Q., Texieria, C.M., Mahadevia, D., Huang, Y., Balsam, D., Mann, J.J., et al. (2014). Dopamine and serotonin signaling during two sensitive developmental periods differentially impact adult aggressive and affective behaviors in mice. Molecular Psychiatry, 19, 688–698. doi: 10.1038/mp.2014.10

CHAPTER 13

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior CLANCY BLAIR, C. CYBELE RAVER, and ERIC D. FINEGOOD

EXPERIENTIAL CANALIZATION 485 A MODEL OF SELF-REGULATION DEVELOPMENT 486 EXPERIENTIAL CANALIZATION OF SELF-REGULATION DEVELOPMENT 490 Experiential Canalization as Life-Course Strategy 491 Sensitivity to Context 492 Gene Expression 492 EXPERIENTIAL CANALIZATION OF SELF-REGULATION DEVELOPMENT AND RISK FOR PSYCHOPATHOLOGY 493 Typical Development of Executive Functions 493 Measurement of Executive Functions 494 Executive Functions and Developmental Psychopathology 496 Application to the Study of Anxiety and Mood Disorders 497 THE EXPERIENTIAL CANALIZATION MODEL OF SELF-REGULATION IN THE STUDY OF EMOTIONAL DEVELOPMENT AND RISK FOR THE DEVELOPMENT OF PSYCHOPATHOLOGY 498 Attention to Emotional Information 499

Higher Order Cognitive Processing of Emotional Information 500 The Body as a Source of Emotional Information: Subjective Feelings of Negative Affect 501 Theoretical and Methodological Challenges and Innovations THE ROLE OF EARLY CAREGIVING IN THE EXPERIENTIAL CANALIZATION MODEL OF SELF-REGULATION: IMPLICATIONS FOR RISK AND RESILIENCE IN DEVELOPMENTAL PSYCHOPATHOLOGY 504 The Neuroendocrinology of Maternal Behavior 504 Development of the Caregiving System 505 The Prenatal Period 506 Motivational Model of Maternal Behavior 507 LONGITUDINAL EVIDENCE IN SUPPORT OF THE EXPERIENTIAL CANALIZATION OF SELF-REGULATION AND RISK FOR THE DEVELOPMENT OF PSYCHOPATHOLOGY 509 CONCLUSIONS AND IMPLICATIONS FOR FUTURE RESEARCH AND INTERVENTION 511 REFERENCES 515

A central problem in development concerns origins; understanding how the multiple sources of information that direct the development of the organism originate and interrelate over time (Oyama, 2000). From its initial definition as “the study of the origins and course of individual patterns of behavioral maladaptation . . . however complex the developmental pattern may be” (Sroufe & Rutter, 1984, p. 18) the discipline of developmental psychopathology has directly addressed the need to understand the origins of mature forms of behavior in terms of developmental processes. Developmental psychopathology has served as a driving theoretical force in shaping the vast array of scientific disciplines that are directly concerned with development, with “the dynamic interplay of processes across time frames, levels of analysis, and contexts” (Carolina Consortium on Human Development, 1996). Recognition of the common focus on development across disciplines

has given rise to what is known as developmental science and to a set of principles by which developmental science is defined. As stated by Magnusson and Cairns (1996) these principles include (1) an understanding of development as holistic and requiring an examination of reciprocity among systems across levels; (2) the contextually bound nature of development; (3) the role of novelty in development; (4) the centrality of time and timing in development, on multiple scales and metrics; (5) the problem of reductionism and the tendency to study one aspect of development in isolation; and (6) recognition of what is known as correlated constraint, or the organized and conservative nature of development. In combination, these principles describe an understanding of the person situated within time frames, levels of analysis, and contexts and offer insight into a developmental understanding of psychopathology. 484

503

Experiential Canalization

EXPERIENTIAL CANALIZATION In the application of general principles of developmental science to the study of developmental psychopathology, we focus on the holistic and contextually bound nature of development and the principle of correlated constraint. We are particularly interested in these principles as they relate to the psychobiological process referred to as experiential canalization (Gottlieb, 1991). The concept of canalization represents an attempt to formally describe the phenomenon of differentiation in development; the emergence of specific and well-ordered forms and behaviors from initially undifferentiated prior states. Canalization, as a descriptive term, contains no prior assumptions about the sources of information that direct the canalizing or shaping process in development. What is implied in the term, however, is the idea of a progressive narrowing of possibilities—of paths taken and foreclosed. As such, it is an apt term for describing the principle of correlated constraint, the constraining of development by the process of development itself. In research on psychological development, this constraining of possibilities is a central feature of cognitive and perceptual development. A canonical example is found in research on language development in infancy. With repeated exposure to a given language in infancy, the child gains experience discriminating phonetic contrasts common to that language while losing the ability to discriminate phonetic contrasts not encountered in that language (Kuhl et al., 1992; Werker & Tees, 1986). Simultaneously with such gain through loss in language development, however, infants acquire an increased ability to integrate sensory experience from multiple sources—visual, auditory, and affective—leading to the further tuning and elaboration of perceptual networks in response to experience (Pons et al., 2009). As it is generally and historically understood, the concept of canalization describes inevitability or determinism in development; the assumption that information is present in the organism or in the environment that directs development to specific endpoints. As reviewed by Gottlieb (1991), canalization was first invoked by Holt (1931) to describe motor development from the prenatal to the early postnatal period. This initial use of the term was framed within learning theory and the shaping of behavior by conditioned responses. In keeping with the general tenor of psychological research in the first half of the twentieth century, this early use of the term was framed within the assumption that experience is responsible for the shaping of behavior, and that the process of development is shaped by general but ubiquitous experiences acting on initially diffuse and undifferentiated patterns of motor activity.

485

In contrast to the initial use of canalization from the perspective of learning theory and implied assumptions about the experiential source of the information in directing the development of the organism, a very different and more commonly understood view of the concept was subsequently proposed by Conrad Waddington. Waddington’s conceptual model proposes genes as the primary source of information shaping the development of the organism. In contrast to a conception of development as a progressive narrowing or shaping of abilities by experience, the use of the term as associated with Waddington describes a developmental process that is resistant to the influence of experience and that is characterized by a self-righting tendency in response to variation in experience. Distinct from both the learning theory and genetic interpretations of canalization, the meaning of canalization from the perspective of developmental psychobiology differs from the views proposed by Holt and Waddington (Gottlieb, 1991). It recognizes both biological and experiential influences across multiple levels of analysis and situates them within the individual’s developmental history. The framing of canalization proposed by Gottlieb and Kuo (1965) integrates biological and experiential information to address the question of how development occurs. As such, it is concerned not so much with the ultimate origin or the endpoint of development as it is with the pathway, with the developmental trajectory and consequences for later development conditioned on earlier occurring developmental processes. Without positing a primacy to any particular source of information, biological or experiential, it proposes an understanding of psychological development as essentially probabilistic as opposed to predetermined, but one that can appear to be highly stable and resistant to change. Identifying the ways the different constituent components of development combine to create stability in physical and psychological development is the objective of the experiential canalization approach; this then allows for the identification of potential points of malleability and specific time points and variables that can be manipulated to bring about change in the developing system. In this way, the concept addresses the paradoxical nature of development as a process of change that is characterized by stability and resistance to change. Perhaps the most direct illustration of the experiential canalization approach as the combined action of biological and experiential influences on development is seen in Gottlieb’s research on the recognition of the maternal call in wood and mallard duck hatchlings. Gottlieb’s embryological experiments indicated that the seemingly innate and instinctual tendency of hatchlings to identify and maintain proximity to conspecifics (duck mothers) is

486

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

dependent on prenatal experience. Specifically, in order to exhibit the prototypical imprinting behavior exhibited by hatchlings, it is necessary that the embryo experience its own hypersonic vocalizations. In the absence of this experience, imprinting can occur to a variety of species (Gottlieb, 1997). That is, the developmental system—including genetic, physiological, and environmental input, pre-, peri-, and postnatally—is plastic and open to canalization. Here, behavior can be considered a leading edge of development (Cairns, 1991), as the embryonic ducklings vocalization is the primary mechanism of development. That is, behavior organizes the components of the developmental system into a new and stable configuration and acts to maintain that configuration. A related example in humans, referred to already, is the development of intersensory perception and perceptual narrowing in infancy. The integration of multimodal information occurring very early in development—prenatally and postnatally—provides the basis for the development of increasingly precise perceptual abilities, reducing the amount of information that must be processed in order to bring about a particular behavioral response to stimulation (Lewkowicz, 2000). Perceptual integration can provide for the facilitation or inhibition of sensory processing in one modality by another, for example enhanced auditory perception by the visual modality. From an experiential canalization perspective, the timing and type of modality specific stimulation is central to the process of perceptual development. Experience, primarily through temporal rate and frequency, drives a process whereby the organism acquires increasingly integrated and differentiated information about the world. The act of perceiving integrated information provides the basis for further and more differentiated integration of perceptual processes. Similar processes are also described in dynamic systems theory approaches to motor development, and the idea that as perception shapes action, so action shapes perception (Thelen, 1995).

In this broad definition, self-regulation includes influences at multiple levels of analysis ranging from the genetic to the neural and physiological to the behavioral and social-cultural. Here, self-regulation can be understood to be composed of a set of interacting processes with well-defined biological substrates that are hierarchically organized and integrated by development into particular forms and behaviors. A model of these components and their hierarchical organization is presented in Figure 13.1. This figure presents an architecture or architectonics of self-regulation that is open to the canalizing process of experience at each level. In this, it can be understood both as a model of influences on behavior at a given point in time as well as developmentally over time, as processes at one level influences the development of processes at multiple other levels, as in the general developmental psychobiological model in Figure 13.2. At the top of the model in Figure 13.1, the fifth level, are the executive functions, defined as cognitive abilities associated with holding information in mind in working memory, the inhibition of highly automatic or prepotent responses to stimulation, and the volitional shifting of

Executive Functions Working Memory, Inhibitory Control, Mental Flexibility

Attention Alerting, Orienting, Executive Attention

Emotion Positive and Negative Reactivity and Regulation

A MODEL OF SELF-REGULATION DEVELOPMENT To apply the experiential canalization model to selfregulation development, we look to prior thinking on self-regulation (Luu & Tucker, 2004; Posner & Rothbart, 2000) and define it as the volitional and non-volitional (conscious and non-conscious) management of emotional, attentional, and physiological responses to stimulation, both internally and externally generated, for the purposes of enacting intentional goal-directed, motivated behavior.

Physiology Sympathetic, Parasympathetic

Genes Sensitivity to Monoamines and Glucocorticoids

Figure 13.1

The architecture of self-regulation.

A Model of Self-Regulation Development

487

Environmental, contextual Behavioral, psychological Neural, physiological

Genetic Individual Development

Figure 13.2 The developmental psychobiological model.

attention among distinct but related dimensions of a given stimulus. Executive functions are widely studied in typical and atypical development (Diamond, 2013). The construct originated in neuropsychological research to describe an array of cognitive and behavioral deficits in individuals associated with planning and decision making, primarily in the instance of damage to specific brain areas in prefrontal cortex (PFC; Fuster, 2008; Luria, 1973). Deficits in this set of cognitive skills in various psychopathologies have led to increased interest in executive functions and to their typical and atypical development. Executive functions are one aspect of the volitional control of arousal and in some ways are synonymous with the motivated and intentional control of behavior. Importantly, for the model in Figure 13.1 and for a developmental model of psychopathology, however, executive functions are only one, and perhaps the least stable or most vulnerable aspect of the self-regulation system. Executive functions are, relatively speaking, chronometrically slow, volitionally effortful, and resource intensive aspects of cognition. Neurobiologically, executive functions are dependent on the integrity of prefrontal cortex (PFC) and the neural circuitry linking PFC with multiple brain regions, including limbic, temporal, and parietal areas. These brain areas that are interconnected with PFC and that both influence and are influenced by PFC, are those represented by the third and fourth levels in the model in Figure 13.1, namely emotional and attentional reactivity and regulation. The brain areas associated with attention include three networks that underlie three functionally, anatomically, and neurochemically distinct aspects of attention, namely altering, orienting, and executive attention, or the volitional control of the focus of attention. The alerting system is activated by norepinephrine originating in midbrain and involves specific areas of frontal and parietal cortex. The orienting system includes the temporal parietal junction, superior parietal cortex, and

frontal eye fields and is activated by the cholinergic system. The executive system includes anterior cingulate cortex, insula, and basal ganglia and its primary neurotransmitter is dopamine (Petersen & Posner, 2012; Posner & Rothbart, 2007; Posner, Rothbart, Sheese, & Voelker, 2012). The brain areas associated with emotional reactivity and regulation include structures of the limbic system, notably the amygdala, anterior cingulate, and ventral, lateral, and medial areas of PFC (Davidson, 2000; Dolan, 2002; Ochsner, et al., 2004). All of these brain areas are activated by the above neurotransmitters as well as serotonin. Activity in the circuitry that underlies reactivity in attention and emotion systems is chronometrically fast relative to executive functioning, and consequently stimulation can drive emotional and attentional responses in ways that shape brain development and support or undercut executive function abilities. The neural basis for this relation occurs in part through the activity of the neurotransmitters described above. The action of these neuromodulators is represented in the second level of the model in Figure 13.1 as the stress response. The stress response system includes the distinct but interrelated sympathetic-adrenal system and resulting levels of catecholamines, primarily dopamine and norepinephrine, and the hypothalamic-pituitary-adrenal (HPA) axis and its end product, the glucocorticoid hormone cortisol. Cortisol results from a cascade of activity in the hypothalamic-pituitary-adrenal (HPA) axis in which stimulation initiates the release of corticotropin releasing hormone (CRH) from the paraventricular nucleus of the hypothalamus leading to the secretion of adrenocorticotropic hormone from the pituitary and resulting release of cortisol from the adrenals. Circulating cortisol then feeds back on the HPA system to inhibit CRH and the resulting production of cortisol (Gunnar & Quevedo, 2007). CRH also potentiates the sympathetic-adrenal system leading to complex interactions between the HPA axis and

488

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

the autonomic nervous system (ANS). Although, the time course of cortisol is slow, taking minutes to reach peak levels, cortisol has more rapid effects at the cell membrane occurring in a time frame of seconds rather than minutes. Studies focusing on physiologically faster actions of glucocorticoids have indicated ways the HPA axis interacts with the ANS (Groeneweg, Karst, de Kleot, & Joels, 2011). CRH facilitates catecholamine release, leading to what is termed a permissive effect of glucocorticoids on indicators of ANS activity such as heart rate (Sapolsky, Romero, & Munck, 2000). Here, interactions among the components of stress physiology are intricate and more research is needed on how they are related developmentally. One investigation of the combined effects of glucocorticoids and catecholamines on behavior in an adult rat model indicated that stress-induced elevations in dopamine are blocked by a glucocorticoid receptor antagonist and that this manipulation was associated with higher levels of working memory performance on a complex maze learning task following stress induction (Butts, Weinberg, Young, & Phillips, 2011). Also in a rat model, high levels of glucocorticoids acting at receptors in PFC both impair working memory and enhance memory consolidation for emotionally arousing events. Importantly, these effects are mediated by a membrane bound steroid receptor, not by activity within the cell nucleus, and dependent on levels of norepinephrine (Barsegyan, MacKenzie, Kurose, McGaugh, & Roozendaal, 2010). More generally, the feed forward-feedback nature of the stress response system is central to higher and lower aspects of the self-regulation model. As levels of NE, DA, and cortisol rise, they potentiate activity in the emotional and attentional networks that influence executive function abilities. This activation is understood as signaling from limbic and brainstem structures to PFC that levels of stimulation are occurring that require effortful attention and control. That is, as neurotransmitters rise to moderate levels, they increase neural activity in PFC. At very high levels, however, hormone increase is associated with decreased activity in PFC and increased activity in subcortical brain areas associated with reactive responses to stimulation, the capture of attention, and also with robust formation of declarative memories of highly emotional events, the so-called flashbulb memory phenomenon (Diamond, Campbell, Park, Halonen, & Zoladz, 2007). As demonstrated primarily by the research of Goldman-Rakic (1995) and Arnsten (2009), the functional form of the relation of hormone levels to executive function abilities such as working memory is the classic inverted U-shaped curve

first identified by Yerkes and Dodson (1908). The inverted U-shaped curve describes the well-known colloquial relation between performance and anxiety in which moderate levels of stress are understood to facilitate performance. Notably, however, in this model and in the original paper by Yerkes and Dodson describing it, moderate levels of stimulation in rats, operationalized as mild electrical shock, was associated with increased performance only on complex learning tasks. In contrast, performance on simple reactivity tasks, such as fear conditioning, was linearly and positively associated with stimulation; with simple tasks, the higher the increase in stress hormones, the better the learning, as in the emotional memory phenomenon. The mechanism linking moderate stress with complex learning has to do with the fact that neural activity in PFC is sensitive to levels of catecholamines. At moderate levels of catecholamine increase, synaptic activity in PFC is increased and synaptic long-term potentiation, the process through which neurons communicate and form networks, takes place. At very high levels of catecholamine increase, however, the opposite occurs. Instead of increasing synaptic activity and LTP, synaptic activity in PFC is decreased, as neurons in this brain area work less well when exposed to high levels of catecholamines. (The loose analogy that is sometimes used is that of an internal combustion engine, in which an optimal mixture of fuel and oxygen are needed in order for combustion to take place. When the amount of fuel or oxygen is too great or too small, the mixture is difficult to ignite and combustion does not occur.) In other brain areas, however, notably those associated with emotional, attentional, and motoric reactivity to stimulation, such as a vigilant and hyper attentive state, the relation between catecholamine levels and synaptic activity is linear. That is, as catecholamines reach levels that lead to decreased activity in PFC, shutting down brain areas associated with complex cognition and reflective thinking, they are turning on brain areas associated with reactive and less reflective responses, namely the amygdala, a key structure in orchestrating automatic and nonconscious responses to stimulation associated with fear, anxiety, and vigilance to threat. The Inverted U The model of self-regulation presented here suggests that the concept of the inverted U can be applied to selfregulation generally. From a bottom-up perspective, as levels of emotion and attentional reactivity and associated neuromodulators rise they support effortful self-regulation through executive functions that are important for

A Model of Self-Regulation Development

performance and learning on complex tasks. And vice versa, from a top-down perspective, executive functions allow the organism to reflect on information and to direct attention and emotional arousal with cascading consequences through the levels of the model in Figure 13.1. In the model, when stimulation becomes overwhelming or underwhelming, or when top-down control is compromised in some way, bottom-up processes take over and lead to reactive behavior. That is, behavior is more directly under the control of highly learned responses and outside stimulation and less under the control of internally generated information and control—in a word, less reflective. High levels of stimulation leading to a sustained stress response and rising catecholamine and glucocorticoid levels flip the brain from a top-down mode of control associated with PFC and reflective thinking to a bottom-up mode of control associated with the amygdala and reactive responses to stimulation (Arnsten, 2009). Such an aspect of brain function has clear benefits to survival and evolutionarily speaking is highly conserved across species. In times of threat, it is important for bottom-up processes to take control of attention, emotion, and motor responses to stimulation. The bottom-up form of control assists in the process of survival because it is faster from the perspective of neuronal activity than top-down control (see LeDoux, 1996, 2012). The mechanism by which neural activity is flipped from a top-down to a bottom-up mode of control is related to the types of neural receptors for catecholamines and glucocortioids that predominate in different brain areas and variation in their sensitivity to levels of these hormones. For example, a receptor for norepinephrine (NE), the alpha2 receptor, proliferates in PFC and has a high affinity for NE, meaning that it is active at low to moderate levels of NE and potentiates synaptic activity. Similarly, the dopamine D1 receptor is also present in PFC and sensitive to moderate increases in dopamine. The NE alpha2 and dopamine D1 receptors have complimentary roles in supporting working memory and are both sensitive to increase in NE and DA, respectively. At increases beyond a moderate level, however, they become inactive (Arnsten, 2009). In contrast, different types of receptors for NE are present in the region of the amygdala and hippocampus, the alpha1 and beta receptors, and they are less sensitive to increases in NE and become increasingly active as levels rise. Similar relations are observed for types of glucocorticoid receptors, and along with receptors for catecholamines, provide the neurobiological basis for the inverted U-shaped curve between stress and performance outlined previously.

489

Genetic Variation Notably for understanding the hierarchical relation among levels of analysis in the model of self-regulation in Figure 13.1 at the base of the model are variants in genes, both single nucleotide polymorphisms and variable number tandem repeats that are associated with sensitivity to and variation in levels of the relevant hormones that act as neuromodulators in the neural circuitry that underlies self-regulation. For example, a variant of a single nucleotide polymorphism (SNP), the catecholo-methyltransferase gene (COMT) has been widely studied, particularly as a potential contributing cause of schizophrenia. This gene is associated with the break down or catabolism of catecholamines, dopamine and norepinephrine, in PFC. Individuals who have a G to A substitution, resulting in a valine to methionine substitution at position 158, catabolize dopamine less efficiently than do individuals with the valine version of the gene. Individuals who are heterozygous for the gene display intermediate levels of dopamine catabolism relative to those who are homozygous for the valine or methionine version of the gene. The functional activity of the gene in relation to behavior is seen in the role it plays in regulating catecholamines, particularly dopamine in PFC. Given the relatively low density of synaptic dopamine transporters in PFC, dopamine removal from the synapse results more from the activity of enzymes that break down dopamine and norepinephrine into inert components than from the removal of dopamine from the synapse through cellular transporter mechanisms. Given the inverted U shape relation of catecholamines to synaptic activity in PFC, a number of studies have shown relations of COMT val158met genotype to executive functions. Generally speaking, the form of this relation has been one in which individuals homozygous for the met version of the gene, meaning that they break down catecholamines less quickly, exhibit higher levels of working memory and executive cognitive task performance due to the increased presence of catecholamines in PFC. That is, consistent with the inverted U-shape relation between stress and performance described previously, moderately higher levels of catecholamines, in this case due to the benefit of a genetically conferred characteristic, should be associated with a higher level of executive function ability. Studies of the COMT gene and similar genes associated with catecholamine activity for the most part, however, have been carried out in laboratory contexts and have not taken into account where on the inverted U shape curve an individual may be situated, experientially speaking. Several studies,

490

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

however, note that expectations concerning the relation of the COMT gene to behavior must take into account the context and temperament of the individual, or what might be considered the initial starting point or background state of the brain prior to completing an executive cognitive task (see Tunbridge et al., 2007, for review). For example, pharmacological manipulation of catecholamine levels through amphetamine administration results in a higher level of performance among val carriers relative to met carriers (Mattay et al., 2003), an effect opposite from what would be expected under typically occurring conditions.

EXPERIENTIAL CANALIZATION OF SELF-REGULATION DEVELOPMENT The experiential canalization model of self-regulation is one in which the developmental shaping of the self-regulation system over time is open to the combined influences of individual biology and experience at all levels, from the genetic to the social and cultural. In the model, it is therefore essential to take into account the history of the organism when examining one aspect of the individual, such as COMT genotype, in relation to any other. Clear examples of this point are seen in demonstrations of the relation of genetic background and early experience to risk for psychopathology in widely cited analyses from a longitudinal sample, a birth cohort, in Dunedin NZ (Caspi et al., 2002, 2003). Here, the relation of genes that regulate monoamine (MAOA) and serotonin (5-HTT) metabolism to later psychopathology were dependent upon the quality of early experience. Specifically, in the Dunedin sample, individuals carrying the version of the MAOA gene that confers less efficient metabolism of monoamines were increasingly likely to exhibit antisocial behavior as adults if they had experienced maltreatment in childhood (Caspi et al., 2002). Similarly, for individuals carrying the version of the serotonin transporter gene that confers less efficient metabolism of serotonin (lower serotonin transporter availability), the effect of stressful life events on depression was increased (Caspi et al., 2003). To be clear, in both of these studies, adverse early experience was a risk factor for later psychopathology for all individuals in the study. The risk associated with that experience, however, was even greater for individuals carrying the gene associated with less efficient metabolism of the relevant neurotransmitters (dopamine, norepinephrine, serotonin). Although impressive in many ways, what is most notable about these findings is the idea that what is beneficial under advantageous circumstances is detrimental

in adverse circumstances. That is, while the maintenance of catecholamines in the brain confers benefits to self-regulation and reflective thinking under low stress conditions, it is disadvantageous under high stress conditions. The neurobiological basis for the way in which acute stress affects psychological function and behavior by increasing catecholamines and flipping the brain from a more reflective to more reactive mode of control in the short run is well established. What is less clear, however, is the extent to which this neurobiology has long-term implications for psychological development and risk for the development of psychopathology, such as that in the examples above. Given the role of experience in brain development (Greenough, Black, & Wallace, 1987) and the formation of neural networks by processes of LTP occurring in response to experience (Hebb, 1949), it is conceivable that tendencies toward more reactive or more reflective modes of self-regulation could become increasing well established, potentially early in development that would be maintained behaviorally in a way that would have long term implications for psychological outcomes and risk for psychopathology. Much more longitudinal research is needed to address this central point in the development of self-regulation, for which an experiential canalization approach focusing on stress and stress hormones can be helpful. The basis for such an approach is rooted in concepts of allostasis and cybernetics (Pribram, 1960). Decades of research on stress physiology have demonstrated that stress physiology adaptively adjusts as a cybernetic system to current and expected challenges. Unlike homeostatic systems that must maintain functioning within a relatively narrow band around a given set point to support the optimal functioning of the organism (e.g., body temperature around 98.6∘ Fahrenheit), stress response systems are allostatic, able to adaptively adjust set points and ranges in response to experience as needed. In this process of allostatic adjustment, the brain plays the key mediating role as it is shaped by experience to adjust physiologic systems to meet an expected future (McEwen & Gianaros, 2011). From the joint perspectives of allostasis and cybernetics, the self-regulation of behavior is a manifestation of physiological set points and accompanying cognitive representations of these set points (Luu & Tucker, 2004). Experience in concert with biological systems acts to establish increasing stability in patterns of behavior through its influence on stability in neural networks (Derryberry & Rothbart, 1997; Posner & Rothbart, 2007). Numerous studies have confirmed that PFC is central to the regulation of stress response physiology and that levels

Experiential Canalization of Self-Regulation Development

of catecholamines and glucocorticoids in part shape the development of this area of the brain (Holmes & Wellman, 2009). Consistent with the experiential canalization model of self-regulation, stress early in life is understood to shape the strength of connections between prefrontal and limbic structures that underlie self-regulation (Cerqueira et al., 2007). For example, a series of studies has shown that induced stress in adult rats leads to changes in the structure of neurons in PFC with consequent effects on attention and set-shifting executive cognitive types of abilities (Liston et al, 2006; Radley et al., 2006). These studies indicate that early stress can lead to the types of functional deficits that are observed in various forms of psychopathology and strongly indicate early stress as one potential contributor to the pathogenesis of psychiatric disorders through its influence on the development of self-regulation. Available human data are consistent with the findings of stress manipulation studies in animal models and indicate the effect of stress hormones on PFC and executive functions. Pharmacological manipulations affecting glucocorticoids as well as catecholamine levels have demonstrated the role of these neurochemicals in executive function abilities (Alexander et al., 2007; Lupien et al., 2001). Of pressing interest are the various pathways through which naturally occurring chronic stress in individuals’ lives may lead to self-regulation difficulties and to increased risk for psychopathology. Early life stress, primarily operationalized in terms of the adverse events usually associated with the experience of growing up in poverty, has been shown to lead to alterations in resting levels of stress hormones such as cortisol in children (Blair et al., 2011; Evans, 2003; Lupien et al., 2005). As well, early poverty has been associated at the molecular level with differential expression of genes associated with adrenergic and glucocorticoid function. In one study, genome wide transcriptional profiling found that low SES in childhood was associated with up regulation of genes associated with adrenergic neural receptor function and down regulation of genes associated with glucocorticoid receptor function in adulthood; a pattern consistent with an expected higher level of reactivity and less prototypically effective regulation of physiological reactivity in individuals experiencing poverty in early childhood (Miller et al., 2009). Experiential Canalization as Life-Course Strategy Implicit in the idea of the allostatic shaping of stress response systems is the notion of a tradeoff between short-term benefit and long-term disadvantage. Central to the experiential canalization model and also the related

491

adaptive calibration model of stress response physiology (Del Giudice, Ellis, & Shirtcliff, 2012) is the idea that alterations to physiology occurring in response to stress are an adaptive and probabilistically normative response to conditions of disadvantage. Alterations to stress physiology and to the neurobiological self-regulation system in response to an expected environment, particularly early in development, are central to a developmental and evolutionary understanding of psychological and physical functioning. Evolutionary biological models across a variety of species indicate that, in conditions of high disadvantage and in which resources and support are low, effective strategies for garnering resources and for surviving to maturity and engaging in opportunities for procreation are optimized through the tuning of stress physiology to potentiate high reactivity and low reflective behavior. Although such reactivity may have some long term cost, the potential for surviving to realize this longer term cost is reduced in conditions of disadvantage. In contrast, in highly resourced and supportive environments, alteration of stress physiology in ways that support more reflective behavior with positive consequences for longer term benefit is advantageous. A well-resourced future is one that can best be utilized and optimized through planning and reflection on multiple possibilities. The idea that development is shaped by environmental context in ways that are designed to maximize fitness within that context is observed across a variety of species from plants, to fish, to mammals. A notable example of the shaping of such a defensive response in development, persisting across generations, is seen in the modest example of the radish. High levels of damage by herbivores induce phenotypic changes in radish leaf production and in seed production in ways that make the leaves less appealing as a food source and as a result promote fitness in environments in which herbivory is high (Agrawal, 1999). A similarly modest example, in this instance in an insect, the water flea, exposure of the mother to the chemical signal of a predator induces morphological changes in offspring that render them less vulnerable to that predator over subsequent generations, even in the absence of any exposure during development of the offspring to that signal. In these examples, environmental programming of individual development is mediated through the mother, through what is referred to as the maternal provision. In a review of the literature on environmental programming, Cameron et al. (2005) provided several examples across a variety of species from plants to insects to mammals in which the experience of the parent induces a phenotypic change in the offspring, often for successive generations.

492

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Sensitivity to Context A key point implicit in the foregoing is the realization that the relation between one level of influence in the selfregulation model, such as genetic background associated with catecholamine metabolism, and experience extends beyond merely increasing risk for later psychopathology. The gene variant or stress response that confers risk in adverse circumstances may be the one that confers benefit in more advantageous circumstances. The recognition of such increased susceptibility to experience (Belsky & Pluess, 2009) or biological sensitivity to context (Boyce & Ellis, 2005), as this phenomenon is referred to is perhaps central to understanding how individuals who look very similar in certain respects may end up with very different outcomes. Such multifinality in development, the emergence of divergent developmental pathways and endpoints from similar starting points, is expected and may hold some promise for understanding processes of risk and resilience in developmental psychopathology. Fulfilling this promise will involve some understanding of the short and long term implications of the inverted U-shape relation between stress and self-regulation for psychological and behavioral development (Blair, 2010). It may be that any longer term effects on brain development and synaptic functioning are malleable and may offer clues as to how effects of early experience can be overcome by later experience. As well, it may be for some, that later experiences can exert disadvantageous effects despite early advantage. An important point here concerns the fact that in samples demonstrating an association between a specific gene variant, be it MAOA or any other gene, and disadvantageous experience, the number of individuals with the relevant version of the gene and high adverse life experience are usually a small percentage of the total population (e.g., Caspi et al., 2002; Kim-Cohen et al., 2007). This indicates that the association between the specific gene variant and the adversity being examined can account for only a portion of the pathologic outcome being examined. That is, complementary to multifinality, there is also equifinality, or multiple ways individuals come to the same developmental outcome. Here, reported interactions between gene variants and environments in the prediction of outcomes are perhaps best seen as exemplary of the relation between biology and experience in the study of developmental psychopathology. Multiple genes are related to the activity of catecholamines and glucocorticoids and multiple types of experience can combine with those genes to lead to psychopathology or to resilience

despite risk. Discerning common pathways among these heterogeneous influences is a complex undertaking but one in which psychobiological theory and understanding of self-regulation as a recursive feed forward and feedback driven system can provide a useful organizing framework. Gene Expression An overarching point essential to issues of multifinality and equifinality in development is that the recursive nature of the model in Figure 13.1 extends to the genetic level. That is, just as executive function processes can both control and be controlled by levels of arousal in emotion and attention systems, so do stress physiology and genes interact through processes of gene expression that shape and are in turn shaped by levels of stress hormones. The HPA axis cascade and action of glucocorticoids on relevant neural targets is generally physiologically somewhat slow, occurring over minutes or hours. Here, glucocorticoid effects on neural activity and structure are largely gene mediated, meaning that they affect processes in the cell nucleus and influence DNA transcription (de Kloet, Karst, & Joels, 2008; Joels & Baram, 2009). As such, the self-regulation model posits that experience shapes development in two directions: As experience acts on stress physiology to influence levels of circulating hormones, it is working “upstream” to affect emotion, attention, and executive function skills that work to regulate stress physiology. It is also working downstream, through cortisol to affect gene expression in ways that also affect levels of stress hormones and, consequently, reactivity and regulation of emotion and attention and executive functions. In theory, the feedforward–feedback loop of self-regulation is understood to be developing over time in response to experience, perhaps rapidly early in the life course, to establish set points or sensitivity for reactive as opposed to regulated responses to stimulation (Luu & Tucker, 2004). Experience and biology are understood to combine to determine the sensitivity of the self-regulation system. The defining psychobiological model of a canalizing process in self-regulation development, although not necessarily referred to as such, is found in research on the effect of maternal behavior on development in a rat model. As shown in the rat, expression of a gene associated with glucocorticoid receptor density in the hippocampus is determined by specific types of maternal behaviors occurring within the first 8 postnatal days. Expression of this gene is associated with the ability to regulate hypothalamic-pituitary-adrenal (HPA) axis responses to

Experiential Canalization of Self-Regulation Development and Risk for Psychopathology

stress and, in turn, with self-regulation behaviors influenced by stress hormone activity. Specifically, offspring of rat mothers that lick and groom their pups frequently and nurse in an arched-back manner (LG-ABN) manifest reduced reactivity and enhanced regulation not only of glucocorticoid levels but also greater learning and memory and reduced avoidance in the face of novel and/or threatening stimuli (Liu et al., 1997; Liu et al., 2000). Notably from the perspective of experiential canalization and adaptive phenotypic plasticity, the high licking and grooming maternal behavior associated with glucocorticoid receptor gene expression in the hippocampus and increased cognitive and behavioral regulation in offspring is facilitated in resource rich environments but reduced in resource poor environments (Meaney, 2001). That is, to some extent the maternal behavior that offspring receive is influenced by the characteristics of the environment in which the mother is situated when conceiving and rearing her offspring. The idea here is that in high-resource environments, the mother rat would be foraging away from the nest for relatively short periods of time and therefore engaging in more of the LG-ABN behavior while in low resource environments the opposite would be true. Such an environmentally driven change in maternal behavior leading to changes in stress physiology and hence behavioral development of offspring is a canonical example of a canalizing process in development.

EXPERIENTIAL CANALIZATION OF SELF-REGULATION DEVELOPMENT AND RISK FOR PSYCHOPATHOLOGY Understanding the distribution of genes and processes of gene expression in populations and the ways genetic background and gene expression are related to one another developmentally is a key scientific goal for research in developmental psychopathology and human development generally; one for which the experiential canalization approach can provide a useful theoretical framework. Important in this goal is an understanding of the relation of biology and experience to developmental outcomes within the model of self-regulation. Here, the model of self-regulation development describes what can be considered an endophenotype, or intermediate phenotype (Gottesman & Gould, 2003). The endophenotype describes combinations of physiological, anatomical, cognitive, and neural processes that are implicated in but by no means unique to psychopathology. The endophenotype is

493

understood to encompass a well-defined and coherent set of processes and behaviors that in part characterize a given disorder. As such, it is well aligned with new directions in the identification and classification of mental disorders on the basis of genetics and brain structure and function as well as behavior referred to as Research Diagnostic Criteria (RDoC; Insel et al., 2010). The self-regulation system provides an elaborated model of the intermediate phenotype that provides insight into interaction among processes at multiple levels of analysis and various pathways to the occurrence of a given disorder and means to treat it. In the clinical and developmental literature on psychopathology broadly, self-regulation is widely studied in the form of the regulation of emotion and attention and also executive functions. Many aspects of the components to the self-regulation model outlined above are the topic of extensive research in schizophrenia, attention-deficit/hyperactivity disorder, anxiety disorders, and depression. In keeping with the general idea that the study of typical and atypical development are mutually informative, the processes of the typical development of executive function, attention, and emotion from the experiential canalization approach, particularly in the context of high risk, whether from poverty, prior family history or both, can inform understanding of self-regulation in research in developmental psychopathology. To this end, we review research on the typical development of self-regulation, focusing on our own and others work with the aim of outlining ways developmental processes may increase risk for psychopathology or alternatively promote resilience in the context of risk. Typical Development of Executive Functions Definitions of executive functions have ranged somewhat widely within a circumscribed set of abilities from the more to less inclusive. Given the role executive skills in planning and problem solving and effortfully regulating aspects of cognition and motivation, some have argued for relatively inclusive definitions of executive functions that involve general planning and problem solving skills (Welch, Pennington, & Grossier, 1991). Others, in contrast, have focused more narrowly (Aron, 2008). Increasingly, however, the field has settled on the domain general cognitive skills of working memory, inhibitory control, and attention shifting as the operational and conceptual definition of executive control (Diamond, 2013). That is, cognitive control generally involves the integration of these skills in task performance. This is seen in widely used measures

494

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

of executive function (Davidson, Amso, Anderson, & Diamond, 2006; Zelazo, 2006) and has tended to be validated in studies examining the relation of executive function components to performance on comprehensive executive function tasks (Miyake et al., 2000). A complete review of research on executive functions is not appropriate here; however, from the perspective of the experiential canalization model of self-regulation development and its application to developmental psychopathology, the protracted developmental course of executive function abilities and their association with prefrontal cortex, a slow maturing brain region (Gogtay, Giedd, et al., 2004), recommends these cognitive skills as potentially valuable indicators of experiential influences on development and on risk for and prevention of psychopathology. The genesis of the executive functions construct in the neuropsychological literature (Luria, 1963) and intense interest in executive functions in the instance of brain damage or dysfunction (Duncan, Burgess, & Emslie, 1995; Waltz et al., 1999) further illustrates the applicability of these cognitive abilities to a developmental understanding of psychopathology. In typically and atypically developing samples, studies of executive function have of necessity employed a variety of measures, primarily due to the restricted age range in which various measures can be used. This makes cross-study and cross-age comparisons difficult. As well, research on the development of executive functions has tended to be cross sectional and has focused on age-related change as well as issues relating to measurement and to construct definition and validity, including the differentiation of executive functions from other aspects of cognitive ability, particularly general intelligence (Blair, 2006). Longitudinal research on the development of executive functions in childhood or throughout the lifespan is relatively rare. Nonetheless, useful information about executive function development can be gleaned from several analyses (e.g., Anderson, 2002; Welsh, Pennington, & Grossier, 1991). Studies with older children and adults have examined typical tasks from the neuropsychological literature and found that performance on these batteries follows a trajectory in which ability increases steadily throughout adolescence into young adulthood. For example, Luciana and colleagues found in a cross sectional sample of 9–20-year-olds that recall-guided action of single units of spatial information develops until 11 to 12 years while the maintenance and manipulation of multiple spatial units develops until 13 to 15 years. More complex processing of information in which the individual must organize and structure information develops until ages 16 to 17 years (Luciana et al., 2005).

Complex processing of information in working memory has also been found to develop into early adulthood (20–29 years) with age related declines becoming apparent in the 50–64 years range (De Luca et al., 2003). Measurement of Executive Functions Information about the typical trajectory for executive function abilities in younger children has been limited by the absence of measures appropriate for longitudinal use. Although there are a number of effective measures of executive functions for preschool children, variation in performance on these measures is usually restricted to a relatively narrow age range with pronounced floor and ceiling effects outside of that range. Carlson (2005) reviewed several widely used measures in early childhood research and found that most are adequate within a relatively narrow age range. Similarly, Espy and collaborators have shown that an attention shifting and inhibitory control task known as the Shape School tracks age related changes in abilities in a cross sectional sample of 3 to 6 year olds (Espy et al., 2006). Noting the need for measures that can be used longitudinally to examine within person change in executive function and that address the tripartite division of executive function into working memory, inhibitory control, and cognitive flexibility domains, Willoughby and Blair (2010) developed a battery of six executive function tasks appropriate for 3–5-year-olds. This battery, similar to analysis of other executive function batteries (Wiebe, Espy, & Charak, 2008), was found to indicate a single underlying factor of executive function. It also demonstrated longitudinal measurement invariance across the age range studied and good test–retest reliability (Willoughby et al., 2011). Notably, findings in a large longitudinal sample (N = 1,292) indicated rapid and substantial development in early childhood of 2.8 SD on average in executive function between ages 3 and 5 years (Willoughby et al., 2012). The aforementioned studies provide some indication of the typical course of the development of executive function and indicate that the ability to process complex information continues to develop into young adulthood. There currently is not a complete normative picture of executive function development, however, against which atypical trajectories can be identified. Obtaining this information is a priority for research in psychopathology and for research on human development in the context of risk, generally. Deficits in executive function are one mechanism through which early adversity affects lifetime risk for mental as well as physical health disorders (Blair & Raver, 2012).

Experiential Canalization of Self-Regulation Development and Risk for Psychopathology

Atypical trajectories for executive function development in early childhood may provide some early indication of risk. An important step in obtaining normative information on executive function development has been the creation of what is known as the NIH Toolbox, which includes computerized versions of two widely used measures of executive function, the Dimensional Change Card Sort task and the Eriksen Flanker Task (Zelazo et al., 2013). Presumably normative information on these measures will be forthcoming and will provide a valuable resource for examining executive function deficits as potential indicators of risk for developing psychopathology. In developing longitudinal measures of executive functions, however, it is important to note that executive function abilities may be highly vulnerable to a variety of insults and therefore less informative as a definitive indicator of risk for psychopathology. Executive functions can be affected by a variety of things, from the mundane (lack of sleep and lack of exercise; Chadock-Heyman et al., 2013; Davis et al., 2011) to the catastrophic (structural and functional damage to PFC and related circuitry; Duncan, Burgess, & Emslie, 1995). Unlike, for example, semantic memory or speed of information processing, aspects of cognition that are more stable, executive functions can be relatively variable within person. The openness of executive functions to insult small and large means that within person change and the stability of between person differences may be great. To a considerable extent, this aspect of executive functioning is a manifestation of the relation of these cognitive abilities to the lower level processes in the model in Figure 13.1. As noted by Luu and Tucker (2004, p. 125), “Disembodied concepts of executive functions may be particularly inappropriate when dysfunctions of executive control are invoked to explain pathologies of mood and motivation. Rather than a disorder of an external supervisory control, these pathologies suggest fundamental alterations in the internal goals and set-points guiding self-regulation.” Executive functions are understood to be predictably variable due to influences on attention and emotion and stress physiology as a manifestation of self-regulation development. Notably, the inherent variability in executive functions creates problems for traditional psychometric analyses by which the validity of psychological assessments is typically established. Test–retest reliability may tend to be low and correlations among assessments of putatively the same construct administered concurrently may correlate weakly or even not at all (Rabbitt, 1997; Willoughby & Blair, 2011). These measurement concerns and considerations in executive function research are seen most clearly in

495

latent variable model analyses of executive function test batteries (Willoughby, Pek, Blair, & FLP Investigators, 2013). Latent variable modeling is an appropriate and widely used strategy for estimating and analyzing the common variance among a set of measures of ostensibly the same construct. Analyses of general mental ability and academic achievement are well-known examples of the utility of latent variable modeling to derive a single underlying construct. Task batteries of these constructs produce robust latent factors that account for upward of 70% of the variance in each of the individual measures (observed indicator variables) used to create the latent variable. Unlike test batteries for these constructs, however, executive function assessments in most latent variable analyses (e.g., Miyake et al., 2000; Willoughby et al., 2012) present low to moderate correlations among the observed indicators and as a result executive function latent variables usually account for only small amounts of variance in each of the indicator variables. This aspect of the task batteries places an upward limit on the reliability of the latent variables, referred to as maximal reliability, despite typically excellent global model fit (Willoughby, Pek, Blair, & the FLP Investigators, 2013). Excellent model fit in this instance cannot be taken as an indicator of the quality of the measurement of the underlying construct, raising questions about conclusions of the literature on executive function abilities using latent variable modeling. The foregoing characteristics of executive function measures highlight the inherent weaknesses as well as strengths of the construct. On one hand, traditional psychometric standards for establishing reliability and validity indicate that the construct is somewhat problematic. On the other hand, part of the value of executive functions lies in their inherent instability. Here, theory and empirical evidence indicate that the construct may be most valuable because of its potential malleability and suggest novel methodologies to study it. To this end, executive function research could benefit from the application of innovative but underutilized methods focusing on within person change (Molenaar, 2008; Nesselroade & Jones, 1990). The application of what is referred to as a multivariate, replicated, single-subject, repeated measures design focusing on intraindividual variability as opposed to interindividual variability could provide the basis for a more accurate estimate of the typical course of the development of executive functions and the relation of this development to risk for psychopathology. Such idiographic research methods were developed to address the fact that traditional analytical approaches that address inter-individual variation cannot provide accurate information about within person change.

496

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

To appropriately model the complexity of development, for current purposes relations among various aspects of the self-regulation system as they relate to within person change, it is necessary to measure individuals, as many as possible, as repeatedly as possible, and to analyze the data to examine how variables combine within persons (Ram, Brose, & Molenaar, 2013). This approach allows for conclusions about trajectories and processes of change within an individual as well as identification of patterns of similarity among groups of individuals, empirically capturing the reciprocal, dynamic nature of development so forcefully outlined by Magnusson and Cairns (1996) and by an experiential canalization framework. Executive Functions and Developmental Psychopathology Issues relating to the measurement of executive function abilities and appropriate methods for the study of change are heightened by the fact that developmental studies have indicated deficits in executive functions across a variety of disorders including attention-deficit/hyperactivity disorder (ADHD; Biederman et al., 2007; Willcutt et al., 2010) and autism (Ozonoff & McEvoy, 1994). Studies have also established that executive function deficits are characteristic of major depression and schizophrenia (Tan, Callicott, & Weinberger, 2009). Although executive function deficits are neither necessary nor sufficient for the diagnosis of any psychiatric disorder, knowledge of executive function development and its biological and experiential determinants can likely provide valuable information to further understand psychiatric disorders and their pathogenesis. Consistent with the framework of the experiential canalization model in Figure 13.1, however, executive function deficits across various disorders or within a single diagnostic category could originate from multiple different sources of influence. Executive function deficits in and of themselves are perhaps too general to be of use on their own in identifying risk for and potential course of a given disorder. Knowledge of the processes that underlie the development of typical executive function abilities, however, will likely prove to be useful in diagnosing and treating subsets of individuals with a given disorder. The specific etiology and course of psychopathology generally might be more readily discerned by considering the various levels in the model in Figure 13.1. Two examples are useful here: schizophrenia and ADHD. Both are characterized by executive function deficits and executive functions have been extensively studied in individuals with these disorders. More specifically, for present purposes, the COMT gene has figured largely in research on

schizophrenia. Given the role of this gene in the activity of dopamine in PFC, it serves as a logical candidate for understanding the etiology of schizophrenia. The association of COMT with schizophrenia is, however, weak at best (Fan et al., 2005). This is not surprising given the complexity of processes affecting executive function as well as risk for schizophrenia, among which dopamine availability associated with the activity of a single gene is one small influence. Research on the COMT polymorphism, which has been voluminous, has revealed, however, the centrality of taking into account the background physiological state of the individual. Given the inverted U shape curve defining the relation between neurotransmitter levels and neural activity in PFC, it may be advantageous under conditions of stress or threat, or even moderately elevated stress, to break dopamine down more efficiently, as do individuals with one or two copies of the val allele of COMT. Therefore, research on one aspect of the individual, such as COMT genotype, might prove to be a risk factor or a protective factor in the case of schizophrenia or other disorders affecting executive functions, depending on a host of other characteristics of the individual. Although relations among this host of factors is undoubtedly complex, the experiential canalization approach provides a framework for incorporating this complexity and for discerning relations among levels of influence that lead to the disorder or to its absence despite substantial risk. Here, the principles of developmental science articulated by Magnusson and Cairns (1996) listed at the beginning of the chapter suggest the importance of focusing on the timing of individual experience and variation in rates of development of the various components of the self-regulation system. Application of single-subject, repeated measures designs to the well-established relations among COMT, catecholamine levels, and executive function abilities can address complex relations among influences on an intermediate phenotype associated with risk for psychopathology. ADHD is also a complex disorder for which the experiential canalization approach is useful in discerning multiple etiologies and subtypes of the disorder. Prior models of ADHD have focused primarily on executive function deficits as its defining feature (Barkley, 1997; Nigg, 2001). Given the inherent complexity of executive functions and multiple pathways to problems with executive functions, however, identification of ADHD as a disorder of executive functions is too general and has required a focus on heterogeneity of subtypes and on bottom-up as well as top-down influences on cognitive ability and behavior. As outlined in the model of self-regulation, the bottom-up influences can affect top-down executive

Experiential Canalization of Self-Regulation Development and Risk for Psychopathology

control abilities. In ADHD research, this has been noted in a focus on sensitivity to reward and approach behavior (Nigg & Casey, 2005) as processes at the emotional and physiological levels that can override or subvert executive function abilities. Furthermore, the recognition of the multiplicity of influences and pathways in ADHD has resulted in a focus on the identification of heterogeneity in the disorder. That is, rather than taking a core deficit such as executive dysfunction as the sole focus of research on etiology, diagnosis, and treatment, investigators have turned to heterogeneity of psychological functioning in typically developing populations as a backdrop against which to understand heterogeneity in ADHD. This effort has pursued what is known as community detection, an approach to define clusters or communities of similarity among relevant measures of cognitive ability among a sample of children meeting criteria for ADHD and typically developing controls. Notably, the approach has yielded the insight that heterogeneity in ADHD is embedded in heterogeneity in typically developing children (Fair, Bathula, et al., 2012). As well, the heterogeneity in subtypes identified by community detection was observed in variation in functional connectivity in the brain for two of the most common ADHD subtypes (Fair, Nigg, et al., 2012). These results confirm to some extent the value of a focus in developmental psychopathology on both multi and equifinality and on the mutually informative relation between typical and atypical development. As well, these findings suggest the value of a developmental analysis as outlined in the seven principles enumerated in the introduction. It may be that distinct pathways to the disorder can be discerned early in development and that the timing and type of experience associated with each yield insights into its etiology and potential preventive intervention. Application to the Study of Anxiety and Mood Disorders A further illustrative example of the way in which an understanding of the hierarchical organization of influences on self-regulation is related to the study of psychopathology is found in its application to anxiety and mood disorders. Anxiety is an essential aspect of human functioning that has been conserved throughout evolution for its value in detecting and responding to threat. Executive functions, which enable the higher order processing of potentially threatening information, expands the link between detection and response, projecting it into the future and building upon past experience (Luu, Tucker, & Derryberry, 1998). For present purposes in the experiential canalization framework of self-regulation, this understanding

497

of executive control, as one that is intimately connected with threat detection (emotional, reactive) highlights the reciprocal relation between bottom-up and top-down processes for understanding anxiety and mood disorders. Specifically, bottom-up pathways linking amygdala and anterior cingulate cortex (ACC) have been identified as critical components of individuals’ regulation of fear and anxiety. Individuals prone to high levels of anxiety demonstrate clear differences in amygdala function, processing threatening versus nonthreatening stimuli differently than non-anxious controls in a bottom-up model of attention and emotion processing (Debiec, Diax-Mataiz, Bush, Doyere, & LeDoux, 2010; Dolan & Vuilleumier, 2003; Mogg & Bradley, 1998). Processing biases in depressed individuals’ perception of emotionally negative stimuli have also been found to be associated with heightened activation throughout additional ventral components of the network linking amygdala, insula, ventral striatum, and ventral ACC (Mayberg, 2003; Phillips et al., 2003). Findings extend to pediatric samples of nonmedicated children and adolescents experiencing high levels of depressive symptoms (Roberson-Nay, et al., 2006; Yang et al., 2010). These bottom-up processes of emotion regulation (ER) involved in responses to fear-evoking or threatening stimuli are, of course, tied to top-down processes involving PFC function (see Dolcos, Iordan & Dolcos 2011 for review). Specifically, executive functions can aid in modulating negative arousal as well as in interpreting and responding to emotionally negative stimuli (Ochsner, Bunge, Gross & Gabrieli, 2006; Ochsner & Gross, 2005; Posner & Peterson, 1990; Wang et al., 2008). Highly anxious individuals, however, have been found to show reduced executive functions, and consequently, difficulty in managing competing or conflicting cognitive demands relative to nonanxious controls (Bishop, 2009; Bishop, Duncan, & Lawrence, 2004: Rueda, et al. 2004). These differences have been hypothesized to play critical roles in anxious individuals’ difficulty in accurately appraising both the emotional states and intentions of others, in socially positive as well as socially negative situations. While past work has largely focused on elucidating the links between greater proneness to reactive versus reflective profiles of self-regulation and profiles of behavioral disinhibition (e.g., higher risk taking behavior, higher delinquency and higher risk of substance use problems), newer research considers the ways that top-down processes are implicated as symptom (as much or more so than cause) in the development of anxiety, depressive disorders and broader-band internalizing difficulty (Buckner, Mezzacappa & Beardsley, 2003; Zucker, Heitzeg & Nigg, 2012).

498

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Recent reviews examining the neural circuitry implicated in processing of emotional stimuli for individuals with depressive disorders suggest support for a dual systems model emphasizing bottom-up and top-down processes. Automatic responding to emotional stimuli primarily draws upon amygdala and ventral and medial prefrontal cortex including subgenual ACC. In contrast, voluntary behavioral and attentional processing of emotional stimuli has been found to involve dorsal regions, including dorsal ACC and dorsal lateral PFC (Rive et al., 2013). Neuroimaging studies of adults with depression diagnoses suggest that effective connectivity from the subgenual ACC and amygdala is increased, with depressed individuals needing to deploy greater medial PFC activation to maintain automatic attentional control relative to their non-depressed counterparts. This hypothesis has been supported by findings indicating increased activity in dorsal medial PFC and dorsal ACC in depressed participants relative to healthy controls during tasks requiring automatic emotion regulation such as facial matching and Stroop-like tasks where emotional stimuli are embedded. In short, activation of neural circuitry underlying executive functions appears to be increased in a compensatory fashion in individuals with depression to redirect attention away from emotionally negative and distracting stimuli (Etkin & Schatzberg, 2011; Rive et al., 2013). This added demand on executive cognitive abilities may indicate that deficits in executive cognitive ability are secondary to increased demand on attention to emotional processing. Patterns of hypoactive function in lateral prefrontal areas across a number of studies examining depressed participants’ performance on tasks involving voluntary cognitive control of negative emotions may indicate typical depletion of executive abilities in the face of repeated demands. Such findings are aligned with results suggesting the individuals with major depressive disorder have difficulty voluntarily stopping or inhibiting the processing of negative information (Gotlib & Joorman, 2010). Treatment studies using fMRI to examine the impact of pharmacological treatment using SSRIs such as paroxetine, sertraline and fluoxetine highlight the joint roles of amygdala, rostral ACC and the lateral PFC in emotion processing. In one recent study, decreases in amygdala activation as well as increased dorsal cortical activation (including increased rostral cingulate and DLPFC) were found as a function of antidepressant treatment, with evidence argued to support indirect pathways of influence (e.g., SSRI treatment leading to improved dorsal prefrontal control over dysregulated limbic activity; Ruhe et al., 2013).

Parallel to these findings, investigators focusing on HPA axis function among depressed individuals have highlighted corresponding patterns of disrupted neuroendocrine function that align well with our model of self-regulation. Specifically, depression has been found to be associated with higher cortisol and increased CRF activity. Importantly, melancholic depression characterized by hyperarousal of negative emotion, higher anxiety, and greater focus on the self is associated with a pattern of hyperfunction of the HPA axis (Struber, Struber & Roth, 2014)

THE EXPERIENTIAL CANALIZATION MODEL OF SELF-REGULATION IN THE STUDY OF EMOTIONAL DEVELOPMENT AND RISK FOR THE DEVELOPMENT OF PSYCHOPATHOLOGY In previous sections, we have outlined a case for the multiple ways that chronic exposure to adversity disrupts self-regulation through both top-down and bottom-up processes. In the sections that follow, we focus more specifically on the bottom-up processes of emotion regulation, through which stressors impact higher order cognitive functions and increase risk for poor longer term mental health outcomes. In the next section, we build on the idea that this model of emotion regulation can be productively understood through the lens of experiential canalization, and can be characterized as the emergence of specific and well ordered forms and behaviors that represent a progressive narrowing of developmental paths taken and foreclosed. Three domains of emotion regulation (ER) may be particularly relevant to the experiential canalization of psychopathology including (1) the individual’s attention to emotional information, (2) the capacity to comprehend and categorize that emotional information, and (3) the individual’s capacity to make sense of emotional information within the self, including subjectively experienced anxiety, fear, anger, and frustration. Following the basic principle that understanding of normal development must serve as the foundation for our understanding of psychopathology, we first outline a schematic or thumbnail representation of some of the basic neurobiological and developmental processes that underlie each of these three domains of ER (see Cicchetti, 1984, 2002; Werner, 1948). We then go on to review ways that these domains are shaped by adversity, drawing from a theoretical model of experiential

The Experiential Canalization Model of Self-Regulation in the Study of Emotional Development

canalization to highlight the ways that children’s emotion regulation is shaped by chronic exposure to environmental stressors. Attention to Emotional Information Attention represents the gate to engagement, the key, early step to detecting, processing, and optimally responding to emotionally salient stimuli (Petersen & Posner, 2012; Posner, 1994). As noted earlier, attention processes (to both emotionally negative and neutral stimuli) are understood to be composed of three key components (see Petersen & Posner, 2012; Posner & Petersen, 1990). These include a first alerting phase (i.e., the brain’s effort to handle the when of the presentation of a stimulus through vigilance), and involve key cholinergic neurotransmitters ascending through the basal forebrain (Stormer et al., 2012). The second component involves attentional orienting, where the brain focuses on location (or the where of the stimulus), involving the deployment of the parietal cortex. Finally, the third component of attention is composed of the executive network, where the DLPF and dorsal ACC are recruited in distinguishing the target stimulus from multiple competing stimuli with the involvement of multiple neuromodulators including serotonin and dopamine (Desimone & Duncan, 1995; Stormer, et al, 2012). Humans are clearly neurobiologically and behaviorally wired to attend to stimuli that have an emotional valence, with extensive laboratory studies demonstrating that adults and children orient to emotional stimuli more quickly than to non-emotional stimuli within a given visual array (LeDoux, Mogg, & Bradley, 1997; LoBue & DeLoache, 2008; Ohman et al., 2001). These findings have been offered as evidence for ways that human attention is evolutionarily tuned to the rapid detection of threatening stimuli. Similarly, emotional information captures attention making it harder for individuals to disengage from emotionally valenced stimuli than nonemotional stimuli (Steinmetz & Kensinger, 2013). Research demonstrates these forms of attention bias through the use of several laboratory paradigms including visual search, where the individual has to locate targets with emotionally salient vs. emotionally neutral distractors, and the dot probe paradigm, where individual has to locate targets after presentation of paired neutral-neutral versus neutral-emotional stimuli. Using these paradigms, researchers highlight the ways that emotion and attention are bidirectionally linked: For example, individuals with mood disorders (e.g., high anxiety) consistently

499

demonstrate increased attentional bias for emotionally negative stimuli (Bar-Haim et al 2007; Waters et al., 2011). These processes emerge in early infancy, are complex and extend over a relatively long time course, with cognitive theories of anxiety suggesting initially faster orientation to threat followed by later avoidance, or suppression of attention to threat in an effort to modulate anxious mood (Nakagawa & Sukigara, 2012). Faster initial orientation to threat suggests that individuals prone to high levels of anxiety are demonstrating differences in amygdala function, processing percepts differently than non-anxious controls at the preattentive stage, that is, early in the bottom-up model of attention and emotion processing (Debiec, Diax-Mataiz, Bush, Doyere, & LeDoux, 2010; Dolan & Vuilleumier, 2003; Mogg & Bradley, 1998). This body of research on individual differences in attention bias has also highlighted the roles of rostral ACC and the lateral PFC in postattentive phase of emotion processing, where highly anxious individuals show reduced functioning of the DLPFC and consequently, in managing and shifting attention across multiple competing, or conflicting cognitive demands relative to nonanxious controls (Bishop, 2009; Bishop, Duncan, & Lawrence, 2004: Pesoa, 2009; Rueda et al., 2004). This level of complexity across multiple components of attention (and respective activation of multiple corresponding brain regions) is important to keep in mind when mapping the development of individual differences in attention to threat. As outlined earlier, emerging research suggests that individual differences in both early facilitation of attention and later difficulty managing and shifting attention may be powerfully shaped by chronic exposure to a range of environmental stressors. Given the key role of the multiple neuromodulators listed earlier, molecular genetic processes (involving multiple polymorphisms of several candidate genes such as COMT and CHRNA4) are also likely to be involved in driving individual differences in both the early and late phases of attention bias (Posner, Rothbart, Sheese & Voelker, 2012; Vuilleumier & Pourtois, 2007). In the following section, we review ways that environmental and genetic processes may work additively and interactively to support or constrain individual differences in attention to threat. Past research on the link between trauma exposure and later symptomatology (across diverse conditions of war, community violence, and parental maltreatment) highlight the role of the environment in altering children’s development of attention, with well-replicated evidence of higher levels of difficulty shifting and maintaining

500

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

attention on clinical measures (Husain, Allwood, & Bell, 2008). Functional neuroimaging research with adults with PTSD yields consistent evidence of heightened amygdala response to rapidly presented emotional stimuli, and correspondingly, behavioral hypervigilance and exaggerated startle response (Bruce et al, 2013; Shinn et al, 2007). These findings are indicative of alterations in the early preattentive period where the “rapid and automatic alerting mechanism for responding to nonconscious signals of fear” are altered by exposure to trauma (Bryant et al., 2008, p. 518). Importantly, several studies have also yielded evidence for increased connectivity between the amygdala and other brain regions associated with the assessment of emotional stimuli and difficulties with the redirection of attention, including left and ventral anterior cingulate and medial PFC, and for hyporesponsiveness in the ventral medial PFC (Bryant et al, 2008; Liddell, et al, 2005). In short, trauma may have negative sequalae in both heightening the system responsible for initial alerting to threat and in lowering capacity for disengagement (including shifting and executive processing) occurring in the late or phase of attention deployment. These findings regarding fear circuitry from the PTSD literature are helpful in demonstrating that it may be that disruptions in both early and late phases of attention processes are at work when trauma-exposed individuals manifest tremendous difficulty in downregulating fear, once it has been triggered (Bryant et al, 2008). For example, some previous studies using behavioral and attentional paradigms (including dot-probe tasks) demonstrate that chronic exposure to both parents’ and peers’ anger and aggression tunes children’s attention for heightened vigilance to emotionally negative stimuli (Pollak, Messner, Kistler & Cohen, 2009; Pollak, Vardi, Bechner, & Curtin, 2005). Functional neuroimaging research with adults who experienced childhood maltreatment has also found evidence for reduced medial PFC volume, relative to nonmaltreated controls (Van Harmelen et al, 2013). Emerging work on attention bias among young children who have experienced maltreatment relative to controls has replicated some of these findings. Higher Order Cognitive Processing of Emotional Information A second key component of emotional self-regulation is the extent to which children appraise stimuli as angering, upsetting, or fear-inducing, that is, children’s capacity to cognitively control emotionally salient conditions through top-down cognitive processing (McRae, Ochsner, & Gross,

2011). The capacity to read others’ emotions is thought to have species-wide benefit, allowing individual members of a group to communicate signals of fear, anger, threat, nurturance, and appetitive desire with conspecifics in ways that promote survival and reduce the risk of predation (Damasio, 2004). Children’s appraisal of others’ signals (including facial expressions, vocal tone and posture) as capable of communicating emotional information has been a key area of past research: Past work has focused both on children’s discrete ability to label others’ facial expressions (termed declarative emotion knowledge) and to individual differences in their tendency to attribute particular emotions (such as anger) to others’ intent, termed emotion processing patterns (Schultz, Izard & Abe, 2005). Not surprisingly, children’s ability to identify others’ emotions relies heavily, but not exclusively on the attention processes described earlier. In addition, exposure to facial stimuli produces activation in the face-selective fusiform face area (FFA) of the fusiform gyrus (see Kawasaki et al., 2012) with concomitant orienting and attending to this specific type of emotional cue. Amygdala activation also plays a central role in cognitive processing of expressions of anger and fear, with studies of lesion to the amygdala demonstrating striking deficit in the ability to interpret emotion from others’ facial expression and tone of voice (Adolphs, Tranel, Damasio & Damasio, 1995; LeDoux, 1996; Vuiellemier et al, 2004). As children transition from infancy through middle childhood, they develop increasing skill in identifying and labeling others’ emotions using appropriate emotion language. Children’s also demonstrate increasingly strong tendencies to read sadness, anger, or fear into emotionally ambiguous social situations and into others’ intentions as they transition through early and middle childhood (Ackerman, Izard & Schultz, 2000). Children’s accurate perception of emotional information has consistently been associated with adaptive social responses to peers, for example, while inaccurate perceptions of others’ emotions has been predictive of children’s negative attributional biases regarding others as well as feeling of failure in social situations, social isolation and loneliness (Denham et al, 1990; Fine et al, 2003; Garner, Jones & Miner, 1994: Schultz et al., 2010). Difficulty or deficits in accurately labeling emotions in others as well as heightened tendencies to attribute anger in others has consistently been found to serve as a major stumbling block for some children as they struggle to make friends and attempt to navigate emotionally intense interactions in classrooms and on playgrounds (Arsenio, Cooper & Lover, 2000; Blair & Coles, 2000; Schultz, Izard, Ackerman & Youngstrom, 2001).

The Experiential Canalization Model of Self-Regulation in the Study of Emotional Development

Individual differences in children’s cognitive processing of emotional cues (both in terms of their declarative knowledge and their proneness to attributional bias) are clearly shaped by environment. Past correlational studies of family socialization and randomized control studies of emotions curricula have implicated adults’ use of emotion language, the quality of adults’ care, and broader emotional climate as key contributors in shaping children’s ability to accurately perceive, encode, and label others’ emotions (Durlak, Weissberg, Dymnicki, Taylor, & Schellinger, 2011). Findings also clearly highlight the extent to which chronic exposure to violence within and outside the home is associated with deficits in children’s cognitive processing of emotional information. Children exposed to higher levels of interparental conflict demonstrate compromised, less fully formed cognitive schema about emotions and relationships as a central feature of profiles of maladjustment (McCoy, Cummings, & Davies, 2009). Importantly, though children from high-conflict households show greater physiological arousal as well as greater behavioral distress, they perceive lower levels of angry affect when overhearing a simulated argument between adults than do children who live in low-conflict households (Cummings, Pellegrini, Notarius, & Cummings, 1989; El-Sheikh, 1994). Consistent with the theoretical model of self-regulation in Figure 13.1 and using the lens of experiential canalization, we and others have argued that this disruption in children’s perception of and response to others’ emotion is undergirded by environmentally shaped neuroendocrine and neurocognitive processes: Higher exposure to the acute and chronic dimensions of threat associated with parental fighting, aggression, and violence may lead to alterations in the adrenocortical and neurocognitive response among conflict-exposed children (Blair & Raver, 2012; Davies, Sturge-Apple, & Cicchetti, 2012). Prior research suggests that, for children exposed to high levels of parental harshness and aggression, prolonged exposure to threat increases children’s arousal to such a great extent that they are less able to make accurate attributions about their own and others’ emotions (Kim & Cicchetti, 2010; Sullivan, Carmody & Lewis, 2009). These effects may be exacerbated by exposure to neighborhood stressors: Recent research suggests that low-income children’s higher exposure to neighborhood violence is clearly associated with compromised ability to marshal attention at chronometrically slower, molar levels (as assessed by observer report) and with maladaptive patterns of attention to negative stimuli at the chronometrically faster, micro level (using neuropsychological assessments such as the dot probe) (McCoy, Raver, & Sharkey, in press; Sharkey et al., 2012).

501

Importantly, violent crime exposure was found to be differentially related to the risk of manifesting hypervigilant versus avoidant patterns of attention depending on children’s profiles of high versus low trait-based negative affectivity, illustrating the ways that these regulatory subsystems are complex, integrated, and recursive rather than independent (McCoy et al., in press; see also Solomon, O’Toole, Hong, & Dennis, 2014). The Body as a Source of Emotional Information: Subjective Feelings of Negative Affect Finally, children vary greatly in their struggles and successes with felt experiences of negative emotion, as indicated through large research literatures on self-reported negative affectivity (Ronan, Kendall & Rowe, 1994), negative affect regulation, self-reported state and trait anxiety, and subjective distress. The neurocognitive and affective substrata for children’s emerging understanding of their own feeling states is remarkably complex, including somatosensory changes in striated musculature, and first- and second-order changes that range across several major brain regions including the amygdala, thalamic regions, and left prefrontal, as well as medial PFC and dACC regions (McKrae et al., 2011). Additional research focusing the somatic marker hypothesis underscores the role of the vmPFC in linking somatosensory information and feeling states, with individuals with lesion to that area unable to access emotionally relevant internal cues and also less able to predict the emotional consequences (whether positive or negative) of their actions (Bechara, 2004; Damasio et al., 1996). In other studies, higher levels of vmPFC have been associated with self-awareness, higher levels of processing of self-relevant emotion words, and rumination, supporting its role in individuals’ ability to identify and think about their own internal states (termed interoceptive awareness; Struber et al., 2014). Across all these literatures, the individual’s conscious appraisal of his or her own affective state yields evidence that individuals can either perceive those emotions are well-managed versus overwhelming and out of control, with a sense of themselves as reasonably good at regulating emotion, or less so. Early in development, children may be less aware of their own emotional states, but by middle childhood, children can reliably report both on dispositional negative affectivity and on transient emotional states. Recent experimental evidence from studies with adults highlights ways emotional states can be internally generated through thoughts or internally evoked images, and that those emotional states are accompanied by

502

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

matching physiological responses (such as changes in eye blink startle) that support the veridicality of subjective report (Ray, McRae, Ochner, & Gross, 2010). High levels of dispositional negative affectivity have consistently been found to relate to higher levels of attention bias to threat, particularly when children have lower levels of behavioral self-control (Lonigan & Vasey, 2004). Subjective experiences of distress (including feelings of sadness, fear, and worry) that are chronic, overwhelming, or debilitating represent hallmark symptoms of mood disorder, particularly when accompanied by the attentional and cognitive biases previously described. Bidirectional relationships between these three components of ER have been established at neuroanatomical and behavioral levels, with evidence that fear circuitry is self-reinforcing, becoming increasingly canalized over time. For children diagnosed with social phobia, for example, reported subjective experience of symptoms of physiological arousal such as blushing and elevated heart rate are not only more easily triggered and more acute than subjectively experienced by controls, but also accompanied by faulty attributional biases (e.g., that others can see their nervousness) and by heightened or excessive self-focused attention (Kramer et al., 2011). Those cognitive biases are accompanied by heightened autonomic and neuroendocrine reactivity, including disrupted patterns of sAA, HR and cortisol levels across both basal and stressor conditions (e.g., social evaluative threat tasks such as the Trier Social Stress Test) and across multiple pediatric samples (van Veen et al., 2008). Our own and others’ work highlights the ways that bottom-up processes such as dysregulated feelings of self-reported trait anxiety can interfere with EF performance on standardized assessments such as Stroop and Hearts and Flowers tasks (Ursache & Raver, 2014). Additional research on the role of experienced anxiety, lower EF performance, and heightened attention to error suggests that as children get older, they not only perform less well on cognitively demanding tasks when experiencing high levels of anxiety, but their subjectively experienced difficulties with working memory and heightened attention to failure may magnify their experiences of negative emotions as overwhelming and spiraling out of control (Ramirez, Gunderson, Levine, & Beilock, 2013). Exposure to chronic adversity has consistently been found to be predictive of earlier onset, steeper trajectories, and more acute symptomatology for a wide range of internalizing mood disorders including anxiety, depression, across multiple developmental periods from early toddlerhood through adolescence and early adulthood (Grant et al., 2003). Certainly, mood disorders such as depression

and anxiety are understood to have multiple etiologies: There is a wealth of empirical emphasis for the direct and indirect pathways by which parents with genetic vulnerability to mood disorders may transmit that risk to their offspring. However, our model of experiential canalization provides a complementary neurobiologically and environmentally mediated mechanism by which conditions of high versus low adversity may substantially increase individuals’ risk. That model is well aligned with developmental cascades models suggesting that cumulative exposure to high levels of poverty-related adversity is jointly predictive of profiles of externalizing behavior problems and high risk of comorbid internalizing problems from childhood through early adulthood (Herrenkohl et al., 20011; Jaffee et al., 2003; McCarty et al., 2009). As outlined earlier, models of mood disorder and neurocognitive function have highlighted the extent to which disrupted or compromised limbic function and attention bias represent hallmarks of anxiety and depression and that those changes in neurocognitive function have been argued to reinforce individuals’ dysphoria over time. Importantly, cognitive processing in the context of mood disorder has been understood as a two-way street. That is, a large number of observational and experimental studies in developmental and clinical areas of research highlight the importance of higher order cognitive processing, considered through the lenses of cognitive appraisal and coping, as a mechanism through which some individuals are able to ward off depression and anxiety and as a powerful approach to ameliorating elevated symptomatology (Lazarus & Folkman, 1984; Moos & Holahan, 2003). Past work on rumination versus cognitive avoidance provides good examples of styles of cognitive appraisal (or coping) that have been identified as placing some individuals at greater risk for long-term trajectories of mental health difficulty relative to others. Individuals who ruminate extensively about situations and events as negative, unmanageable and uncontrollable and about themselves as less capable in handling those situations are at substantially greater risk of higher levels of depressive symptoms and more frequent bouts of depression over time, even after controlling for symptomatology at baseline (Lyubomirsky, Tucker, Caldwell, & Berg, 1999; Nolen-Hoeksema, 2000; 1999) Alternately, individuals struggling with high levels of distress and fear may resort to cognitive styles of coping that involve cognitive avoidance, characterized by high mental effort not to think about the problems, events, or situations that are understood as sources of worry or sadness. In contrast, individuals able to maintain shift and persist involving realistic appraisal of obstacles and

The Experiential Canalization Model of Self-Regulation in the Study of Emotional Development

challenge with long-term positive outlook or confidence in long-term persistence or success in overcoming of those obstacles strategies of cognitive appraisal have been found to weather major stressors with lower vulnerability to health and mental health problems, than other individuals facing similar psychosocial challenges (Chen & Miller, 2012). Importantly, treatment involving shifts in cognitive appraisal serve as the foundation of much cognitive behavioral therapy (CBT) and have been found consistently in meta-analyses of randomized control trials to be effective in lowering the odds of mild and moderate depressive symptomatology in childhood adolescence, and adolescence (Harrington, Whittaker, Shoebridge, & Campbell, 1998; Jakobsen et al., 2012) as well as in reducing the risk of relapse after pharmacological treatment of depression (Paykel, 2007). Recent work integrating interpersonal and psychobiological perspectives highlights the ways that multiple systems can interact to alter individuals’ HPA axis reactivity and chances of long-term mental health problems in the face of chronic exposure to stress (Cacciopo et al., 2002; Gianaros & Hackman, 2013). Specifically, emerging work on the stress-buffering role of social support in both animal and human models suggests that social interactions suppress stress-responsive HPA reactivity through direct and indirect mechanisms. For example, the presence of socially supportive conspecific or romantic partner during stressful conditions such as the TSST has been associated with increased release of the neuropeptide oxytocin, which in turn has been found in both rat and human randomized controlled trials to significantly lower HPA axis activity in the face of experimentally induced stress (Heinrichs et al., 2009; Struber at al., 2014). Research in this area has been complicated by difficulty of measuring oxytocin peripherally: in several recent studies, positive tactile contact from a socially supportive partner has been found to substantially reduce cortisol levels and heart rate for adults during stressful lab procedures but not to reduce plasma levels of oxytocin (Ditzen et al., 2007). Additional work on the salutary effects of parents’ provision of positive caregiving on lowering HPA reactivity through mediating mechanisms of oxytocin in children is promising and is reviewed now. Across this range of social interactions, theoretical and empirical perspectives have emphasized the need to consider both transitory and longer term consequences of positive provisioning of care: results yielded from longer term studies of individuals exposed to early trauma and prolonged periods of deprivation suggest the likelihood of inhibition of oxytocin receptor expression. In contrast, supportive caregiving during early

503

periods in the context of adversity may facilitate long-term patterns of adaptive HPA axis response mediated through longer term changes in serotonin and oxytocin expression (Struber et al., 2014). Theoretical and Methodological Challenges and Innovations One strength of the model that we have outlined is that it underscores developmental plasticity and malleability, for good or for ill, rather than trait-like fixity in considering the dynamic attentional, cognitive, and neuroendocrine processes involved in making sense of one’s emotional world. On the other hand, one grave implication of this model consistent with the developmental science principle of correlated constraint is that individuals undoubtedly contribute to the reinforcement or canalization of their own emotion regulatory trajectory by selecting, structuring and reappraising emotional information from social interaction and by subsequent patterns of approach versus withdrawal from those social interactions (Cicchetti, 2002). The implications of this theoretical shift are immense: Through this lens, for example, feedforward models of fear circuitry would need to be reexamined across both neurobiological and biobehavioral strata. How does the responsiveness of those early components of the alerting and orienting mechanism develop, over time, increasingly constrained by prior development and experience, as an anxious child, for example, increasingly opts out of new and potentially frightening social contexts? At the neuroendocrine level, we also know less about ways that individuals who experience higher levels of anxiety in socially ambiguous situations may effectively reduce opportunities to learn to entrain HPA axis regulation of high arousal by avoiding social engagement. In short, models of canalization underscore the extent to which processes of self-regulation are self-organizing. Yet we have only just begun to consider ways that those systems are interrelated, along multiple neurobiological, representational, and behavioral strata. A major problem in research on self-regulation and psychological development in the area of emotions research is the failure to replicate across samples facing varying levels of environmental risk, varying types of stimuli, and varying age groups. For example, studies that have explored attention bias have often used stimuli that are of human faces, with depressed and anxious individuals more prone to attend to emotionally threatening stimuli (see for example, Waters et al, 2011 where children ages 9 to 12 and high though not clinical levels of anxiety showed clear attentional bias for angry versus neutral faces on

504

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

dot-probe, with stimulus duration ranging from 500 to 1250 ms). Yet other studies find that children with high levels of anxiety may be more prone to demonstrate avoidance of threat, rather than bias toward (or hypervigilance for) threatening stimuli. This may be because higher acuity in detecting negative stimuli may arise from multiple sources that are themselves nonlinear—exposure to moderate levels of conflict or threat, for example, may tune the attentional system to be more alert, more vigilant, and the PFC to be more actively mobilized to resolve conflicting system in order to promote the capacity to act. On the other hand, exposure to high levels of threat may lead the system to a pattern of blunted, avoidant, or hyporesponsive profile. Second, the process of attention, interpretation, and somatosensory response to threat occurs over a time course that is rarely accounted for in extant emotion regulation research. Temporally early bottom-up attention processes in handling threatening, emotionally arousing information may be facilitated (as reflected in behavioral measures) for individuals chronically exposed to threat, while top-down, temporally late-breaking processes (such as memory, retrieval, and labeling of emotion terms) may be disrupted by high levels of stress and among individuals experiencing high threat. Those individuals exposed to very high levels of anxiety may show correspondingly lower performance on behavioral measures that tap complex processes involved in recognizing, identifying, and correctly naming facial emotions (Henckens, van Wingen, Joels, & Fernandez, 2012). Research in this area is still nascent: we are not yet able to carefully parse neurobiological, neuroendocrine, and behavioral responses across a microanalytic time course in ways that yield consistency in our findings. Thirdly, with respect to developmental time course, the model of self-regulation indicates the need to identify and track processes that are occurring very early in development, at conception and over the first year, and to consider self-regulation intergenerationally. Emotion and attention in infancy have been studied extensively in their own right, but few studies, with the exception of the research on temperament of Rothbart, Posner, and colleagues (Posner & Rothbart, 2000) have studied relations between the constructs developmentally. As well, although the role of caregiving in early development has been the primary stock in trade of child development researchers, few programs of research have focused on the psychobiology of caregiving and the way processes of self-regulation in caregivers influences both the biological and experiential environment in which child development takes place.

THE ROLE OF EARLY CAREGIVING IN THE EXPERIENTIAL CANALIZATION MODEL OF SELF-REGULATION: IMPLICATIONS FOR RISK AND RESILIENCE IN DEVELOPMENTAL PSYCHOPATHOLOGY In the foregoing sections we have outlined processes of self-regulation development as a function of multiple levels of influence, focusing primarily on the internal processes of the experiential canalization model of self-regulation. In the following sections we turn to external influences that are presumed to shape development through experiential input. As noted in the earlier sections, a primary shaper and canalizer early in development is thought to be stress hormones and a primary environmental influence on hormone levels is thought to be experience with caregivers. Information about the environment is transmitted to developing offspring, prenatally and postnatally through maternal experience and behavior. As such, in the experiential canalization model, parenting behavior is presumed to be a central if not the central mechanism through which information about the quality of the environment is communicated to the developing child. Notably, the psychobiological system is multigenerational and recursive. The primary caregiver is itself a self-regulating psychobiological system and caregiving a product of the caregiver’s psychobiological makeup and developmental history. The model in Figure 13.1 applies as readily to the parent as to the offspring. Here, inheritance must be understood broadly—that what is passed on from parents to offspring is not only a biological inheritance or an environmental inheritance but an interactive psychobiological system, an interacting manifold of biology and experience (Gottlieb, 1997). The Neuroendocrinology of Maternal Behavior Given the psychobiological nature of parental behavior and influence (Corter & Fleming, 2002; Rosenblatt, 1994), the experiential canalization model provides one framework for understanding the intergenerational transmission of risk for the development of psychopathology. In the parent as in the offspring, stress hormones are a primary influence on maternal behavior with implications for child physiology and behavior. A number of key hormones are involved in the maternal response to infant stimuli including cortisol, oxytocin, dopamine, and endogenous opioids, which facilitate and regulate typical maternal behaviors (for review see Swain et al., 2011). In their review of the literature, Swain and colleagues (2011) used the example of

The Role of Early Caregiving in the Experiential Canalization Model of Self-Regulation

the maternal behavioral response to a baby cry to illustrate how multiple hormonal influences come online to facilitate maternal behavior in the course of a typical interaction between mother and child. In this example, dopamine and oxytocin are the first to engage in response to baby cry, and promote maternal action by focusing attention and decision making on the part of the mother. In response to infant suckling, oxytocin is involved in milk ejection during lactation. Subsequent to the initial response of dopamine and oxytocin, the HPA axis regulates maternal behavior in ways that vary as a function of life stressors, postpartum time, age, and parity status. Once resolution is reached, endogenous opioids promote relaxation, and in addition, reinforce maternal behaviors via reward circuitry. Importantly, each step in this behavioral and psychobiological cascade must be situated in context; each step is a point of inquiry and intervention. The relation between the HPA axis and maternal thoughts and actions is complex, however, given that it varies as a function of the personal history of the mother, her parity status, and time since parturition (Barrett & Fleming, 2011). For instance, resting levels of cortisol measured within the first four postpartum days are positively associated with increases in affectionate behaviors in mothers (Fleming, Steiner, & Anderson, 1987) as well as with a mother’s increased attraction to her infant’s odors (Fleming, Steiner, & Corter, 1997). These findings suggest that maternal HPA axis activity during the first postpartum days may enhance the salience of infant cues and in turn increases sensitive caregiving by mothers (Fleming et al., 1997). The direction of association changes between cortisol levels and maternal behavior in the months after parturition, however. Specifically, heightened cortisol levels at 3 months postpartum are associated with negative mood and fatigue in middle-aged mothers (Krpan, Coombs, Zinga, Steiner, & Fleming, 2005), and cortisol reactivity at this time has been related to less synchronous or sensitive behaviors with infants (Thompson & Trevathan, 2008). Increases in basal cortisol levels at 6 months postpartum are associated with increases in negative and intrusive behaviors with infants (Mills-Koonce et al., 2009). It is also critical to keep in mind the environmental and social origins of maternal physiological dysregulation and parenting behaviors. As an example, interparental avoidance in marital relationships is positively related to cortisol reactivity in mothers, which, in turn, is related to increases in psychologically controlling parenting behaviors and inconsistent discipline of children (Sturge-Apple, Davies, Cicchetti, & Cummings (2009). In addition, experiencing adversity in early life, including experiencing inconsistent

505

care or maltreatment by past caregivers, is related to higher levels of diurnal cortisol in mothers as well as with decreases in maternal executive functions (Gonzalez, Jenkins, Steiner, & Fleming, 2012), both of which mediated the relation between early life adversity and maternal sensitivity in adult mothers. Thus, the context of physiological stress is an important component of the psychobiological relation between maternal physiology and behavior. It is clear that biological as well as experiential influences critically shape maternal behaviors that have powerful downstream implications for child development. Development of the Caregiving System Typical and atypical development of the caregiving system is mutually informative. The early parent–child relationship provides a foundation upon which child development proceeds, from early influences on emotion and attention and the stress response, with implications for gene expression, brain growth and development, and emerging higher order cognitive abilities (Blair & Raver, 2012). As described in a previous section, the canonical example of the experiential canalization model of development is found in the effect of maternal behavior in the rat on stress physiology and behavior in offspring (Francis, Diorio, Liu, & Meaney, 1999; Liu et al., 1997; Weaver et al., 2004). Cross-fostering studies highlight the psychobiological factors accounting for the intergenerational transmission of the stress response in rats. Specifically, the adult biological offspring of low licking and grooming (LG) and low arched back nursing (ABN) mothers who were fostered by high LG-ABN mothers show significantly reduced fear response to novelty and additionally, exhibit increased LG-ABN behaviors themselves compared with adult offspring from high LG-ABN mothers who were fostered by low LG-ABN mothers (Francis, Diorio, Liu, & Meaney, 1999). Thus, rearing experiences with caregivers in very early life contribute to the intergenerational transmission of both stress response and maternal behaviors later in adulthood. Indeed, this multilevel epigenetic process from maternal behavior down to the genetic level and back again to the behavioral, presents both a top-down and bottom-up process of the shaping of biology by experience—the central tenet of the experiential canalization model. The example of cross fostering is very interesting in that it highlights the potential plasticity of the developing self-regulation system early in life. Alterations to experience can shape the developing system in ways that are likely to lead to a prototypically well-regulated profile or to a more reactive and less well regulated behavioral

506

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

profile that can predispose to psychopathology. An important point here, however, is the rapidity with which this developmental process unfolds. The influence of maternal behavior on the outcomes of adult offspring in the rat model unfolds over the first 8 postnatal days. This is a remarkably short period that is presumably specific to rats. The historical emphasis on early experience in child development, however, perhaps extending over the first year or more, is consistent with such a rapid effect. The presence of such an early occurring effect, however, does not preclude the possibility of later malleability. Indeed, in the developmental psychobiological model, earlier occurring processes shape later processes, propagating an essentially conservative nature of development, referred to in the principle of correlated constraint. Reorganization of the system is possible, however, and has been demonstrated in the LG-ABN rat model through environmental manipulations occurring in the immediate post-weaning period (Meaney, 2001). In the cross-fostering example, pups of low LG-ABN mothers reared by high LG-ABN mothers exhibited behavior similar to the biological offspring of the high LG-ABN mothers (Francis et al., 1999a). In another example, low LG-ABN mothers increased LG and ABN when their pups were briefly handled, an expected result, and these animals exhibited high LG-ABN themselves as mothers, providing evidence of the causal role of experience in shaping later behavior (Francis et al., 1999b). While the effects of maternal behavior on development in the rat model have been shown to persist into adulthood, these effects have been shown to be functionally reversible. Specifically, Francis and colleagues (Francis, Diorio, Plotsky, & Meaney 2002) have shown that the effects of maternal separation on HPA-axis reactivity and behavioral responses to stress are attenuated in maternally separated rat offspring reared in enriched environments during the peripubertal period. The authors highlight, however, that this reversal in stress response physiology and behavioral response associated with enriched environments is not mediated by the reversal of underlying neural mechanisms (i.e., gene expression) but rather, that activity of brain areas including the PFC and hippocampus may play a role in compensating for earlier life stress (i.e., maternal separation). Similarly, Bredy, Humpartzoomian, Cain, and Meaney (2003) demonstrated that the effects of low maternal caregiving competence on hippocampally dependent learning and memory could be reversed at the behavioral level through environmental enrichment, but that the physiological function of the hippocampus was not altered in expected ways. Thus, the effects of environmental enrichment in functionally reversing the negative effects of

maternal separation and low maternal caregiving competence occur on some levels of analysis but not others. This is consistent with the psychobiological perspective on the development of self-regulation, which not only highlights the considerable plasticity of the developing self-regulation systems at the neural and behavioral levels, but also the idea that these systems, in and of themselves, should not be considered at isolated levels of analysis, but rather, at multiple levels of analysis ranging from the genetic, to the neural and physiological, to the behavioral. The Prenatal Period Given the emphasis on malleability but also concerning the essentially conservative nature of development, it is important to consider how influences on the maternal psychobiological state may shape the developing psychobiological state of the offspring from conception onward. The prenatal environment influences both offspring embryological development as well as postnatal development through the influence of the HPA axis (Barbazanges, Piazza, Le Moal, & Maccari, 1996a; O’Donnell, O’Connor, & Glover, 2009). The fetal programming hypothesis suggests that during sensitive periods in utero, the placenta affects fetal development in ways that are adaptive for future environments, as it is a proximal indicator of the external environment. In the case of low-resource, stressful environments, the evolutionary significance of the fetus’ interaction with proximal (placental) and distal (external) environments is that insults to the fetus may adaptively prepare offspring for potential harsh external environmental conditions. Research on maternal stress during the prenatal period in humans and non-human animals indicates effects on behavioral and physiological aspects of the development of the stress response in offspring with substantial implications for the development of psychopathology. One pivotal mechanism by which prenatal stress is passed on to offspring is via secretion of corticosterone from mother to fetus through the placenta. Indeed, blocking maternal secretion of corticosterone prenatally resulted in no difference between offspring of prenatally stressed and nonstressed mothers, suggesting a mechanism of stress transmission via corticosterone placental transfer (Barbazanges, Piazza, Le Moal, & Maccari, 1996b). Neurophysiologically,, high prenatal maternal stress is associated with decreased glucocorticoid receptor (GR) density in the hippocampus (Maccari et al., 2003), which may compromise negative feedback regulation of glucocorticoids (Sapolsky, Meaney, & McEwen, 1985). Behaviorally, prenatally stressed rats display abnormalities

The Role of Early Caregiving in the Experiential Canalization Model of Self-Regulation

in social behavior and dysregulation of the HPA axis and these outcomes are mediated by heightened levels of corticotrophin-releasing hormone (CRH) in the amygdala, reduced hippocampal GR density, and attenuated endogenous opioid GABA/BZ inhibition (Weinstock, 1997). Maternal stress in both the prenatal and early postnatal environment is associated with physiological and behavioral outcomes in rat offspring, and in both scenarios, the influences of biology and environment are intertwined in an irreducible way. Research on the effects of maternal stress in the prenatal period in humans suggests processes similar to those seen in animal models. Buss and colleagues have shown that prenatal maternal anxiety specifically regarding pregnancy is related to both decreased gray matter density in the PFC (Buss, Davis, Muftuler, Head, & Sandman, 2010) as well as to diminished executive function abilities (Buss, Davis, Hobel, & Sandman, 2011) in child offspring at 6- to 9-year longitudinal follow-up. Maternal prenatal psychological distress and anxiety have been associated with infant temperament (Huizink, Robles de Medina, Pascale, Mulder, Visser, & Buitelaar, 2002) cognitive impairment (Mennes, Stiers, Lagae, & Van den Bergh, 2006), ADHD, anxiety, and externalizing problems in children (Van den Bergh & Marcoen, 2004) and with flattened diurnal cortisol patterns associated with depression in females (Van den Bergh, Van Calster, Smits, Van Huffel, & Lagae, 2007). Motivational Model of Maternal Behavior In addition to the mechanism of transmission associated with stress physiology and effects of glucocorticoids on the developing brain, the shaping of self-regulation development in offspring postnatally is of course predominantly shaped by caregiving behavior. In addition to passing on elevated stress levels and an HPA axis that may be less effective in regulating stress hormones within an optimal zone, caregivers under stress shape offspring behavior through caregiving behaviors that may provide limited support for the development of prototypically optimal self-regulation. Caregiving behavior is itself shaped by an underlying neurobiology that influences offspring development for good or for ill. As reviewed by Numan (2006), animal models support a motivational model of maternal behavior, whereby the medial preoptic area (MPOA) of the hypothalamus and the neighboring ventral bed nucleus of the stria terminalis (vBST) regulate maternal behavior in the presence of infant stimuli by inhibiting competitive avoidance motivations and increasing approach–response motivations, via

507

downstream projections inhibiting a competitive medial amygdala (MeA) to periaqueductal gray (PAG) circuit and activating dopaminergic mesolimbic circuitry involving the ventral tegmental area (VTA) and nucleus accumbens (NA), respectively (Numan, 2006). Here, as with the experiential canalization model of self-regulation in children, hormones are the essential mechanisms through which the development of the maternal response is understood to be shaped. Maternal behavior in rats is exhibited directly after parturition (Numan, 2006) and in contrast virgin rats do not immediately display maternal behavior in the presence of pups (Fleming & Luebke, 1981; Fleming, Vaccarino, & Luebke, 1980) due to their lacking the necessary hormonal shift occurring in the late prepartum period. Thus, the experience of giving birth to offspring critically primes the mother for childrearing. The hormonal mix during the peripartum period includes a dramatic decrease in the hormone progesterone and simultaneous increase in estradiol and lactogen hormones, responsible for the regulation of maternal behavior (Numan, 2006; Numan & Insel, 2003). As evidence for this, injection of the hormones at levels similar to those found during the peripartum period induces maternal behavior in virgin rats, suggesting the integral role that neurohormones play in the transition to parenthood (Numan, 2006). Similar to the above mechanisms of motivation and attachment in mothers, attachment and odor preferences on the part of offspring can be formed even in the instance of very low quality care and abuse. Specifically, Sullivan and colleagues have shown that infant rats learn preferences for maternal odor that are accompanied by a neurobiological mechanism in which the locus coeruleus is hyperfunctioning, sending large amounts of norepinephrine (NE) to the olfactory cortex to ensure that a preference for the odor of the mother is formed (Landers & Sullivan, 2012). For rats, this is a primary means by which attachment to the mother is formed. As a consequence of this neurobiological cascade, infant rats have a reduced ability to acquire learned aversions during this time and do not regularly learn aversions until post weaning age. At that age, corticosterone increases and begins regulating amygdala function, which is useful at an age that is characterized by life in and out of a nest. Notably, attachment is formed in infants regardless of the quality of the care received; even maltreated infants successfully form attachments. These findings are significant within the lens of experiential canalization, where early experiences shape expectations of future environments. That the process of attachment is robust enough to persist

508

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

even in contexts of pain or adversity—where instead of avoiding the mother, the infant remains in close proximity to, and learns from the mother—is consistent with what is known as the maternal provision model (Cameron et al., 2005). This neurobiological mechanism of attachment is therefore essential for learning in early life. Animal and human models illustrate a critical multisystem psychobiological shift during the peripartum period in neuroendocrine and neural architecture that set the stage for the initiation of maternal behavior with consequences for offspring development. This maternal behavior is guided by the motivation to perceive infant stimuli and to provide necessary care. However, parenthood in humans is qualitatively more complex than in animal models. Research on the neural substrates of parenting in humans is beginning to expand upon what is already known about the transition to parenthood in animal models (Kim, Leckman, Mayes, Feldman, et al., 2010; Lambert & Kinsley, 2012; Swain, 2011). In humans, structural maternal brain changes occur over the postpartum period. Specifically, Kim and colleagues (2010) compared magnetic resonance images of brain volumes during the early postpartum (2–4 postpartum weeks) and in the later postpartum (3–4 postpartum months) and found increases in gray matter in prefrontal cortex (PFC), parietal lobes, and midbrain. Moreover, the authors show that increases in gray matter volumes in hypothalamus, substantia nigra, and amygdala were related to positive perceptions of infants (Kim et al., 2010). These neural and behavioral developments taking place over the first postpartum year in mothers are thus the result of a combination of neurobiological and experiential input, consistent with the psychobiological model. Research highlights specific neural circuits involved in typically occurring maternal behavior in response to experience with implications for ways this circuitry might result in atypical maternal behavior (Champagne & Meaney, 2006; Jans & Woodside, 1990; Nephew & Bridges, 2011; Terkel & Rosenblatt, 1971). Gestational stress is associated with decreases in maternal licking and grooming behaviors in high LG-ABN dams, but not in low LG-ABN dams (Champagne & Meaney, 2006). Additionally, after a gestational stress condition, oxytocin receptor binding (OTR), which is typically increased in high LG-ABN dams, decreases to levels comparable to that of low LG-ABN dams (Champagne & Meaney, 2006). Additionally, thermal factors, access to resources, and social stressors disrupt maternal behaviors including nest building and time spent with pups and also affect the growth of dams and pups (Jans & Woodside, 1990; Nephew & Bridges, 2011; Terkel & Rosenblatt, 1971).

In humans, studies have begun to investigate the maternal brain response to infant cues such as pictures or baby cries (Lorberbaum et al., 2002; Swain, Lorberbaum, Kose, & Strathearn, 2007; Swain, 2011). The canalization process, through which maternal behavior and child behavior are reciprocal and mutually reinforcing, suggests that maternal sensitivity in perceiving infant stimuli will be shaped by the combined action of experience and biology. Using fMRI, the typical response to standard baby cry versus white noise is associated with activity in the medial thalamus, medial prefrontal cortex and right orbitofrontal cortices, hypothalamus, midbrain, dorsal and ventral striatum (Lorberbaum et al., 2002). These areas are broadly involved in key decision-making, planning, reward/punishment circuits, and are likely critically involved in the motivation circuitry defined earlier in rodent models (Numan, 2006). Importantly, infants’ neuroendocrine and behavioral response has been found to be altered dramatically by the presence or absence of supportive maternal caregiving. Just as mothers’ olfactory sensitivity to infant odors likely increases provisioning of care, infants are sensitive to odors expressed from the nipple, with increased suckling following exposure to maternal odors, and with memory for odors regulated by oxytocin released in the infant’s brain while suckling (Nagasawa et al., 2012). Theoretical and empirical traditions that highlight dyadic synchrony at behavioral levels have been extended to examine psychobiological attunement during the course of social interaction between caregivers and offspring (Feldman et al., 2010; Laurent, Ablow, & Measelle, 2012). Among a sample of 86 low-income mothers and their 18-month-old infants for example, increases in mothers’ cortisol/sAA predicted increases in infants’ cortisol/sAA across two different stressors (e.g., separation episode or LabTab fear/frustration tasks; Laurent et al., 2012). It is important to note that evidence for attunement (i.e., a moderate to strong correlation between maternal and infant trajectories of cortisol) has been found for some but not all studies (note Hibel et al., 2009, with reported correlation of r = .01 versus Atkinson et al., 2012 with reported correlations in the r = .40 to .62 range). One reason for this variability in findings may be due to the moderating role of maternal sensitivity, with significantly higher values for cortisol and sAA attunement found among highly sensitive caregivers as relative to their less sensitive counterparts (Atkinson et al., 2013). Another source of difference across these studies may be the use of alternate stress paradigms to assess child and maternal reactivity. In exploring the possible role of stressor type and the extent of stability across contexts, Atkinson et al. (2013)

Longitudinal Evidence in Support of the Experiential Canalization of Self-Regulation

report greater intraindividual variability in child HPA axis activity across multiple stressor paradigms (separation and frustration) for children of more sensitive caregivers, while children of less sensitive caregivers demonstrated weaker and less variable responses to both stressors, over time. Building on these findings, Hibel and colleagues (in press) found in a predominantly low-income prospective longitudinal sample that mother–child adrenocortical attunement is disrupted in response to a child stress task and that both maternal sensitivity during free play and child emotional reactivity elicited during a stress task moderate the strength of attunement. Whereas reductions in attunement in response to the stress task were evidenced by dyads in which mothers engaged in low amounts of sensitivity, attunement in dyads of highly sensitive mothers was moderate and stable across the stress task. This finding provides evidence that maternal sensitivity may buffer the otherwise disruptive effects of stress on attunement. LONGITUDINAL EVIDENCE IN SUPPORT OF THE EXPERIENTIAL CANALIZATION OF SELF-REGULATION AND RISK FOR THE DEVELOPMENT OF PSYCHOPATHOLOGY Given the wide variety of influences and levels of analysis that self-regulation encompasses, it is not surprising that there are numerous research literatures related to the topic, particularly in adults (e.g., Vohs & Baumeister, 2011). We have focused on relevant literatures on the early development of emotion, attention, executive functions, and stress physiology, relying on correlational and experimental work, much in nonhuman animal models. We now turn to three prospective longitudinal examples that provide some empirical evidence in humans in support of the experiential canalization model. We conclude with examples from prevention research in which efforts to foster self-regulation development and prevent the development of psychopathology provide key tests of the theory of the experiential canalization of self-regulation development. Longitudinal studies highlight the value of looking across levels of analysis and suggest the extent to which change in experience through typically occurring variation can alter developmental trajectories in specific ways. Our own work on the experiential canalization of self-regulation has been conducted through a project funded by the National Institute of Child Health and Human Development of the U.S. National Institutes of Health, known as the Family Life Project (FLP). Through the FLP, we have been following a population-based sample of 1,292 children and families recruited at birth. The sample is

509

located in predominantly rural and low-income communities in the United States (in Northern Appalachia and the Southern Black Belt) and the participating children and families have been seen in home visits for data collection at approximately annual intervals between the ages of 7 and 60 months. Children have also been seen at school from prekindergarten through the second grade. One component of the FLP is designed to assess the development of emotion, attention, and executive functions in early childhood. During home visits for data collection when children were 7, 15, and 24 months of age, data collectors implemented protocols to assess emotional reactivity and regulation in response to fear and frustration evoking challenges and collected saliva samples from the children before and after the emotion challenge procedures. The samples were later assayed for cortisol and alpha amylase in order to examine physiological reactions to stress. Cortisol, as noted above, is a glucocorticoid hormone and indicator of activity in the HPA axis. Alpha amylase is a surrogate marker of norepinephrine, and as such an indicator of the faster acting sympathetic adrenal system. With these data, we first examined the idea that the quality of early experience would shape the development of stress response physiology in ways that are consistent with the experiential canalization model. Here we focused on the quality of parenting that children received as a proximal indicator of the quality of the environment. In one analysis (Blair et al., 2008), we found a relation between the mother’s style of parenting and the infant’s cortisol levels at 7 months of age. Here, we found that infants whose mothers supported and scaffolded their child’s play, meaning that that they were neither too intrusive nor too detached, had lower levels of cortisol than infants whose primary caregivers exhibited less scaffolding and support for the child’s play. Importantly, we also found that 7-month-old infants whose mothers displayed the positive parenting style were more likely to exhibit a cortisol response to the emotion challenges. Children of mothers with the more supportive parenting style exhibited a larger increase followed by a decrease in cortisol, indicating a flexible and prototypically healthy response to the stress of the emotion challenge. As well, the amount of increase and decrease was associated with the level of emotional reactivity observed behaviorally. The physiological and behavioral responses are presumed to be one in which the organism prepares itself to deal with the challenge of the emotion inducing contingency but then returns to baseline following this challenge. In contrast, children whose mothers did not show the sensitive pattern had higher initial cortisol levels and tended to exhibit either a flat or blunted

510

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

response to the emotion challenge despite high levels of emotional reactivity to the challenge. Following the sample longitudinally, we found that at age 15 months, children receiving higher levels of positive parenting in infancy at age 7 months, maintained lower cortisol levels and were more likely to exhibit lower overall cortisol levels in response to the same emotion challenges implemented again at this later time point. These results demonstrated that typical aspects of the parenting that children receive, namely, positive scaffolding and active encouragement of children’s play behavior, are shaping or canalizing the development of the stress response system as indicated by levels of cortisol. Given expectations for relations between the regulation of cortisol and executive functioning, we next sought to chart the complete path in environmental disadvantage in the context of poverty from parenting to stress regulation to executive functions. In this analysis we examined whether the development of executive function abilities in the sample was consistent with the predictions of the experiential canalization model. That is, we examined executive function at age 3 years in relation to demographic characteristics associated with poverty (income, race, maternal education, crowding and safety in the home) as well as parenting behavior and baseline cortisol levels measured at 7, 15, and 24 months of age. Using covariance structure modeling, we found that the effects of poverty on executive function ability were mediated through both parenting and cortisol, over the child’s first two years (Blair et al., 2011). Notably, in support of the experiential canalization model of self-regulation development, we found that the effects of poverty on cortisol levels, and by extension executive function abilities, were mediated through positive aspects of parenting behavior. Primary caregivers (in almost all cases the child’s mother) living in poverty were less likely than their more affluent counterparts to exhibit the positive parenting style characterized by support for and scaffolding of their child’s play. In turn, lower levels of the positive parenting style were associated with a higher level of cortisol, which was in turn inversely related to child executive function at age 3. Positive parenting was also directly related to executive function in children. These results indicated that in addition to directly affecting child executive function, sensitive parenting also influenced the development of children’s executive function abilities through child cortisol. In a subsequent analysis with the Family Life Project data, we examined further the relation between stress physiology and executive function and also between stress physiology and early academic abilities, such as knowledge

and reasoning ability in mathematics as well as knowledge of letters, sounds and early reading ability. In the examination of stress physiology we expanded our analysis to include sAA as an indicator of activity in the sympathetic adrenal system. Consequently, we expected that baseline levels of cortisol and sAA might show an interactive pattern in relation to executive function. That is, lower cortisol levels but higher sAA levels, indicating an attentive and prepared orientation to experience might be a more optimal physiological state conducive to executive function. Conversely, simultaneously high or low levels of both cortisol and sAA might reflect an over or under aroused state. As expected, again using covariance structure analysis, we found that high levels of sAA and low levels of cortisol at each time point of measurement in infancy and early childhood, 7, 15, and 24 months of age, were associated with a higher level of executive function ability as measured by our test battery. Conversely, simultaneously high or low levels of these stress markers were associated with poor performance. There was also some indication that higher cortisol but lower sAA was also associated with a higher level of executive function than simultaneously low levels. Finally, we found that cortisol and sAA were also related to the indicator of academic ability but that this relation was fully mediated through executive function abilities (Berry et al., 2012). In the realm of emotion recognition and emotion regulation, analyses of Family Life Project data suggest that higher levels of physical aggression between parents during the period from infancy through early childhood, chronic exposure to chaos in the household and income poverty, could each be empirically distinguished as additive and statistically significant contributors to 58-month-olds’ ability to recognize, interpret, and modulate negative emotion (Raver, Blair, Garrett-Peters, and FLP Investigators, in press). Although the model of experiential canalization and self-regulation development emphasizes development early in life, from conception onward, notable longitudinal examples of the combined action of biology and experience are also seen in young adolescent high-risk samples. Analyses from these samples indicate the ways that the environment of poverty shapes physiological processes with implications for physical and mental wellbeing. For example, in a rural poverty sample, elevated stress physiology as assessed by a cumulative index of allostatic load composed of overnight levels of catecholamines (dopamine, norepinephrine) and cortisol from urine and body mass index at ages 9 and 13 years was positively related to an index of psychosocial risk in poverty in

Conclusions and Implications for Future Research and Intervention

childhood and negatively related to a measure of working memory at age 17 years (Evans & Schamberg, 2009). Findings from this sample also indicated, however, the extent to which sensitive maternal behavior can influence development. At child age 13 years, it was shown that the relation between risk from poverty and allostatic load was moderated by maternal sensitivity. In adolescent children of mothers rated as sensitively responsive, no association between risk and allostatic load was observed. It was only among children of mothers low in sensitivity that the relation between cumulative risk and allostatic load was observed (Evans, Kim, Ting, Tesher, & Shannis, 2007). Similar findings have been observed for an adolescent sample of African American youth in rural communities. Here, however, relations between risk factors associated with poverty and mental and physical health outcomes were moderated by genetic variation. Combinations of high life stress and specific genes associated with dopamine and serotonin function (e.g., DRD4, 5-HTTP) act to increase or decrease the likelihood that stress will lead to mental health problems (Simons et al., 2011). Several general findings of analyses from this sample are notable from the perspective of experiential canalization. One is that in the analysis of the moderation of life stress by DRD4 in the prediction of aggression, effects were mediated by individual’s cognitive schemas and subjective experience of emotion (hostile attribution, anger). This provides some evidence for the shaping of psychological state through the combined action of biology and experience with implications for behavior and risk for psychopathology. This finding also suggests that alterations of cognitive schemas and emotional states might act as an effective intervention for the prevention of aggression and its attendant consequences and that effects would be greatest for those with greater genetic susceptibility. Indeed, in intervention trials with this sample, the investigators demonstrated such genetic moderation of intervention effects in evaluations of programs to reduce aggression and substance use by fostering positive parent-child interaction (Brody, Chen & Beach, 2013; Brody et al., 2013). Finally, also of relevance in this sample is the indication that genetic susceptibility can also affect resilience. That is, in the sample, despite high risk, relatively few individuals exhibited high levels of behavior and substance use problems, denoting a high level of resilience. Among the proportion of the sample exhibiting resilience, however, those with the short allele of the serotonin transporter gene (associated with lower serotonin transporter availability) also tended to exhibit elevated indicators of stress physiology (high

511

allostatic load) and high levels of health problems (Brody, Yu, Chen, et al., 2013) despite healthy psychological functioning.

CONCLUSIONS AND IMPLICATIONS FOR FUTURE RESEARCH AND INTERVENTION In the previous sections of this chapter, we have drawn upon several examples from human and animal models of self-regulation that underscore the idea that individuals develop though combined biological, maturational, experiential, behavioral and cultural strands that are fused in ontogeny. Biology is found in these examples to influence and be influenced by behavior through complex interactions and pathways that have only recently become clear. In some ways, the complex integration of biology, behavior, and interpersonal interaction implies constraints on the range of developmental pathways taken. Caregivers’ provisioning for infants, and infants’ subsequent development of self-regulation, is clearly shaped not only by the best of parents’ intentions, but also by neuroendocrine processes that are breathtaking in the directness and precision of the mechanisms involved. Overall, however, we find across numerous examples that developmental paths are far more complex, circuitous, and open to environmental forces both for good and for ill than is generally assumed. That is, we have dramatically diverged from simpler arguments regarding whether nature or nurture has the upper hand in shaping development. Instead we find that exposure to multiple forms of experience shapes development through complex and conditional relations among variables at multiple levels of analysis. This complexity and conditionality signals four key conclusions. First, we are only at the beginning of the path to multidisciplinary scientific discovery needed to understand human developmental processes. At various moments it is easy to become daunted by the forest of research literatures covering the neurobiological and behavioral mechanisms related to self-regulation, across animal, infant, child, and adult literatures emphasizing development in typical, at risk, and atypical populations. To find a way through the wilderness, we will need collaborative teams of scholars from disciplines as diverse as neuroscience, biological anthropology, epidemiology, and behavioral health research among many others, united under the conceptual framework of developmental science. We will need appropriate, population-based samples of research participants that can be followed longitudinally to appropriately model processes and trajectories as they

512

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

unfold. Much can be learned from such data sets, particularly when they include comprehensive biobehavioral and psychiatric epidemiological data on adolescent and adult outcomes. Population-based data sets with a rich array of data over the life course offer opportunities to test the role of self-regulation not only for adult risk of psychopathology but also for different trajectories of both biological and economic inheritance. For example, in the Dunedin Multidisciplinary Heath Study, a birth-cohort study, self-control before age 10 predicted not only a range of better health, economic, and demographic outcomes by age 32, but also whether participants were parenting their own children in one- versus two-person households, signaling tremendous dynamic complexity on both theoretical and empirical fronts (Moffitt et al., 2011). To match this focus on development will likely require idiographic methods that explicitly model processes within the individual as well as across individuals. In the absence of an idiographic approach, multifinality and equifinality can be described conceptually but are less likely to be convincingly demonstrated empirically. Second, we conclude that our success in pursuing this ambitious intellectual course will in some ways be pinned to the clarity and precision of the key constructs in our models. To date, the term stress has been used widely and less precisely. The many disciplinary perspectives represented in this chapter will not be able to speak effectively to one another until environmentally linked terms such as stress, adversity, and neuropsychologically linked terms such as reactivity and executive attention are clearly provided and justified. Recent discussion by Odgers and Jaffee (2013) has highlighted this concern regarding the distinction between adversity and daily hassles, finding that daily hassles can actually serve to both mediate and moderate the role of adversity on child health outcomes. Here, a further implication of the approach is the need to focus on persons more so than on variables. A focus on the principles of development noted in the introduction makes clear that a given variable may take on different meanings depending on the configuration of variables with which it is embedded. Third, we conclude that we must find new ways of leveraging experience; of using interventions designed to support mental health and well being as an experimental basis from which to draw strong causal inferences regarding developmental mechanisms outlined earlier. A central point of research in the tradition of developmental psychopathology concerns the way developmental theory and preventive intervention are mutually informative. Theory informs the development of efforts to reduce or prevent the occurrence of psychopathology in individuals at risk

and in turn experimental evaluations of these preventive interventions serve as key tests of developmental theory (Cicchetti & Hinshaw, 2002). Just as pharmacological studies have manipulated doses of cortisol and oxytocin to yield evidence of neuroanatomical, behavioral, and cognitive mechanisms underlying risk for psychopathology, increased attention to the measurement of neuroanatomical, neuroendocrine, genetic, cognitive, and emotional processes in the design and evaluation of prevention trials can yield highly valuable information on processes of experiential canalization and the malleability of the multiple, interrelated processes involved. There are relatively few extant examples of prevention trials that focus on the early development of self-regulation that provide insight into experiential canalization and processes of malleability and change in self-regulation development. In infancy and toddlerhood, prevention science research has recently expanded from targeting molar dimensions of parenting, more broadly, to improving specific caregiving practices that are theorized to scaffold early regulation of attention and emotion (see, e.g., Dozier et al., 2006; Fisher et al., 2007; Neville et al., 2013). These experimental tests build upon theoretical and empirical insights provided by Izard and collaborators over a decade ago (Izard et al., 2002), outlining the ways emotion theory can inform prevention efforts. The recent generation of prevention science studies targeting processes of self-regulation provides compelling evidence of biological mediating mechanisms that have profound interpersonal consequences. In one natural experiment, for example, foster mothers’ peripheral oxytocin levels were found to become increasingly attuned, or related to ERP indicators of their brain activity in response to images of their own child versus other children over the first 2 months of caring for a foster child, providing fascinating evidence of experiential canalization processes involved in optimizing the provision of care (Bick, Dozier, Bernard, Simmons & Grazzo, 2010). There are also strong implications for the reduction of risk of psychopathology among both adults and children from this new generation of research: As another example, the Family Check-Up intervention yielded significant reductions in mothers’ depressive symptoms along with reductions in very young children’s risk of externalizing and emotional dysregulation problems (Shaw et al., 2009). In our own lab (as well as in projects recently launched by colleagues funded by the Administration for Children and Families’ Buffering Toxic Stress Consortium) we increasingly capitalize on randomized control trials that target both maternal and child indicators of physiological, emotional, and cognitive regulation

Conclusions and Implications for Future Research and Intervention

to examine processes of self-regulation development and to test models of experiential canalization. Experimental tests of the role of experience can extend to classroom contexts—one such example is a preschool-based intervention known as the Chicago School Readiness Project (CSRP). The CSRP is notable in that it focused on the early prevention of behavior problems in children by focusing on emotion regulation and did so in community-based educational settings for children in poverty, namely Head Start programs. Most notably for present purposes, it was developed within the developmental science framework, explicitly acknowledging connections among levels of analysis and the need to intervene across levels to bring about comprehensive behavior change (Raver et al., 2009, 2011). Recent results from CSRP analyses suggest that experimentally induced benefits to self-regulation gained in the context of classroom-based intervention can be undermined by later exposure to traumatic stressors such as unsafe school settings and neighborhoods with high levels of violent crime, (Sharkey et al., 2012; Raver, et al., 2013). Following children who were initially enrolled in early interventions like Family Check-Up or CSRP through early adulthood (using both biobehaviorally and psychologically anchored measures of self-regulation) will provide the empirical basis we need to test key models of experiential canalization, allowing us to use developmental theory to design preventive intervention and to use evaluation of the intervention to test developmental theory. Doing so will help to fill an important gap in research in early childhood in which knowledge of emotional development can play a central role in efforts to comprehensively foster child well-being and reduce early indicators for psychopathology. A related example of the value of preventive intervention as a test of developmental theory, also in a classroom context, is found in a program based on emotions theory. The Promoting Alternative Thinking Skills (PATHS) program is a developmental science-informed approach to reducing behavior problems in young children. It does so by focusing on emotion knowledge and emotion regulation strategies with the understanding that promoting these abilities in children will be associated with benefits to executive functions. Evaluations of PATHS have shown that it is effective at reducing behavior problems and that change in executive function abilities is associated with program effects (Riggs, Greenberg, Kusche, & Pentz, 2006). Extension of the PATHS approach to other areas of behavior management and self-regulation provide further evidence of the efficacy of the approach. Altering the focus in PATHS on emotion regulation and impulse control to

513

encompass healthy dietary choices, Riggs Sakuma, and Pentz, (2007) demonstrated changes in attitudes toward self-regulation and in choices relating to food and to physical activity. Another example of the potential for change through experience in school-based settings is our evaluation of the educational approach known as Tools of the Mind with children in kindergarten. The Tools approach is explicitly designed to provide support for child self-regulation and the development of executive functions through structured child-directed, teacher-scaffolded activities (Bodrova & Leong, 2007). We found that children in schools randomly assigned to implement the approach had moderately higher levels of executive function abilities as well as mathematics and reading achievement (ES = .10). The effects were particularly large (ES = .30–.80), however, in high poverty schools and extended to measures of children’s ability to control attention in the face of fear evoking stimuli in addition to reasoning and vocabulary ability (Blair & Raver, 2014). Notably, this demonstration of the effect of the Tools of the Mind approach followed two prior evaluations of the prekindergarten version of the curriculum, one of which demonstrated effects of the approach on children’s executive function abilities (Diamond et al., 2007), and the second of which found no effects on any aspect of children’s executive functioning or academic abilities (Wilson & Farran, 2012). Clearly, additional research is needed school-based approaches to the comprehensive promotion of self-regulation in school-based settings. In combination with the CSRP and the Tools of the Mind evaluation, longitudinal follow-ups of the Perry Preschool (Heckman et al., 2009) and Abecedarian (Campbell et al., 2014) programs, and findings from a recent meta-analysis of 213 school-based, universal social-emotional learning (SEL) programs involving 270,034 children indicating significant social-emotional and academic gains resulting from SEL (Durlak et al., 2011), provide clear empirical evidence of the potential to structure educational experiences for children in ways that comprehensively foster self-regulation development at multiple levels of analysis with the possibility of pronounced benefits into adulthood. As illustrated by CSRP, Tools of the Mind, and classic early intervention studies, capitalizing on efficacy trials of preventive intervention extends our understanding of socioeconomic adversity’s impact on self-regulation abilities. Many neuroscientists focusing on the recently established SES gradient in self-regulation may think that we can’t experimentally manipulate poverty, but experts in policy analysis know that we can and we have

514

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

experimented with populations’ experiences of material deprivation through anti-poverty programs. Were biological, neuropsychological and behavioral assays and assessments to be built into those and other experiments, we would gain additional purchase on ways that environments shape cognitions and mood states through both bottom-up and top-down processes, in ways that we can now speculate are plausible and have pieced-together empirical support. Moreover, attention to economic and demographic outcomes within efficacy trials of clinical intervention would allow us to test ways that those processes are bidirectional: Treatment of depression or other mood disorders would be expected to substantially shift not only individuals’ executive functions but also their ability to navigate key life-course decisions in areas such as employment, housing, partnership, and risk taking. In short, we might learn not only about adversity’s impact on self-regulation, but self-regulation’s impact on exposure to adversity, over the long term. There is also a host of what can be considered more proximal attempts to increase self-regulation and its consequences through interventions focusing on training working memory, modifying attention, and altering the ways individuals frame goals and intentions. Attempts to improve working memory through repeated practice reflect the top-down approach to self-regulation and represent the theory that strengthening executive control will lead to behavior change. Evaluations of this training, with children as well adults, indicate that it very clearly improves working memory abilities. Evidence to date that these improvements transfer or generalize to other aspects of behavior, however, has been limited (Melby-Lervag & Hulme, 2013; Shinaver et al., 2014). Notable examples of the effects of working memory training for children with ADHD indicate improvements in working memory tasks but limited evidence of improvements in ADHD symptoms (Klingberg et al., 2005; van Dongen-Boomsma et al., 2014). Similar to working memory training, efforts to modify attention bias to threatening, anxiety provoking, or highly appetitive, addictive stimuli indicate that the training is effective in altering the capture of attention by relevant stimuli. Effects on behavior from attention modification, however, are limited (Beard et al., 2012; Mogose et al., 2014). In contrast to limited effects on behavior from effective change at the attentional and working memory/executive cognitive levels, a relatively straightforward approach to changing problem behavior by mentally contrasting a desired goal state with one’s current condition and developing a simple action plan to implement steps

to attain the goal state, referred to as mental contrasting with implementation intentions, has demonstrated that it is effective at changing unwanted behavior (Stadler, Oettigen, & Gollwitzer, 2009) and improving academic performance (Duckworth, Kirby, Gollwitzer, & Oettigen, 2013). In planning and executing preventive intervention experiments as tests of developmental theory, investigators can remain vigilant for the ways that individuals’ responsiveness to treatment may vary as a function of genetic, cognitive, and interpersonal context. Put another way, interventions that have heretofore been understood to have relatively small average causal impact could be reexamined for large impacts in small but theoretically indicated subgroups. An interesting example concerns an analysis of the effect of working memory training on behavior among alcoholics in which problem drinkers were randomly assigned to receive working memory training or not (Houben et al., 2011). Results indicated that the training was associated with reduced alcohol intake over a 1 month period post-training for participants with strong automatic preferences for alcohol as indicated by an implicit attitudes task. The foregoing brings us to our fourth and final conclusion; namely, the way in which the experiential canalization model of self-regulation development illustrates what is in many ways the most revolutionary of the central tenets of developmental science, namely the idea that behavior is the leading edge of development (Cairns, 1991). This phrase highlights the importance of a continued focus on experience, on behavior as a dynamic rather than fixed process that represents and can itself bring about change at multiple levels of analysis in a developmental system. Recognition of this central point can orient scientific inquiry in ways that are at once both familiar and new. On the one hand, it builds on well-established traditions of the study of behavior, particularly social behavior, and the potential of a host of techniques, practices, interventions, and programs to bring about behavior change. On the other hand, it challenges our understanding of the goals of these techniques, practices, interventions, and programs to move beyond the establishment of a fixed behavioral endpoint or outcome, to an understanding of process; of behavior as the representation of an organized state of relations among variables across levels of analysis, and also as a dynamic entity that can bring about change in that organized state. Comprehensive analysis of multiple influences on behavior at multiple levels over time is not merely a scientifically interesting if daunting endeavor, it is essential for understanding human development and preventing psychopathology.

References

REFERENCES Adolphs, R., Tranel, D., Damasio, H., & Damasio, A. R. (1995). Fear and the human amygdala. Journal of Neuroscience, 15(9), 5879–5891. Agrawal, A. A. (1999). Induced responses to herbivory in wild radish: Effects on several herbivores and plant fitness. Ecology, 80(5), 1713–1723. Alexander, J. K., Hillier, A., Smith, R. M., Tivarus, M. E., & Beversdorf, D. Q. (2007). Beta-adrenergic modulation of cognitive flexibility during stress. Journal of Cognitive Neuroscience, 19, 468–478. Anderson, P. (2002). Assessment and development of executive function (EF) during childhood. Child Neuropsychology, 8(2), 71–82. Arnsten, A. F. T. (2009). Stress signalling pathways that impair prefrontal cortex structure and function. Nature Reviews Neuroscience, 10(6), 410–422. doi: 10.1038/nrn2648 Aron, A. R. (2008). Progress in executive-function research from tasks to functions to regions to networks. Current Directions in Psychological Science, 17(2), 124–129. Arsenio, W. F., Cooperman, S., & Lover, A. (2000). Affective predictors of preschoolers’ aggression and peer acceptance: Direct and indirect effects. Developmental Psychology, 36(4), 438. Atkinson, L., Gonzalez, A., Kashy, D. A., Santo Basile, V., Masellis, M., Pereira, J., . . . & Levitan, R. (2013). Maternal sensitivity and infant and mother adrenocortical function across challenges. Psychoneuroendocrinology, 38(12), 2943–2951. Barbazanges, A., Piazza, P. V., Le Moal, M., & Maccari, S. (1996a). Maternal glucocorticoid secretion mediates long-term effects of prenatal stress. Journal of Neuroscience, 16(12), 3943–3949. Barbazanges, A., Vallée, M., Mayo, W., Day, J., Simon, H., Le Moal, M., & Maccari, S. (1996b). Early and later adoptions have different long-term effects on male rat offspring. Journal of Neuroscience, 16, 7783–7790. Bar-Haim, Y., Lamy, D., Pergamin, L., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2007). Threat-related attentional bias in anxious and nonanxious individuals: A meta-analytic study. Psychological Bulletin, 133(1), 1. Barkley, R. A. (1997). Behavioral inhibition, sustained attention, and executive functions: constructing a unifying theory of ADHD. Psychological Bulletin, 121(1), 65. Barrett, J., & Fleming, A. S. (2011). Annual research review: All mothers are not created equal: Neural and psychobiological perspectives on mothering and the importance of individual differences. Journal of Child Psychology and Psychiatry, 52(4), 368–397. Barsegyan, A., Mackenzie, S. M., Kurose, B. D., McGaugh, J. L., & Roozendaal, B. (2010). Glucocorticoids in the prefrontal cortex enhance memory consolidation and impair working memory by a common neural mechanism. Proceedings of the National Academy of Sciences, 107(38), 16655–16660. Beard, C., Sawyer, A. T., & Hofmann, S. G. (2012). Efficacy of attention bias modification using threat and appetitive stimuli: A meta-analytic review. Behavior Therapy, 43(4), 724–740. Belsky, J., & Pluess, M. (2009). Beyond diathesis stress: differential susceptibility to environmental influences. Psychological Bulletin, 135(6), 885. Berry, D., Blair, C., Willoughby, M., & Granger, D. A. (2012). Salivary alpha-amylase and cortisol in infancy and toddlerhood: Direct and indirect relations with executive functioning and academic ability in childhood. Psychoneuroendocrinology, 37(10), 1700–1711. Bick, J., Dozier, M., Bernard, K., Grasso, D., & Simons, R. (2013). Foster mother–infant bonding: associations between foster mothers’

515

oxytocin production, electrophysiological brain activity, feelings of commitment, and caregiving quality. Child Development, 84(3), 826–840. Biederman, J., Petty, C. R., Fried, R., Doyle, A. E., Spencer, T., Seidman, L. J., . . . & Faraone, S. V. (2007). Stability of executive function deficits into young adult years: A prospective longitudinal follow-up study of grown up males with ADHD. Acta Psychiatrica Scandinavica, 116(2), 129–136. Bishop, S., Duncan, J., Brett, M., & Lawrence, A. D. (2004). Prefrontal cortical function and anxiety: Controlling attention to threat-related stimuli. Nature Neuroscience, 7(2), 184–188. Bishop, S. J. (2007). Neurocognitive mechanisms of anxiety: An integrative account. Trends in Cognitive Sciences, 11(7), pp. 307–316. Blair, C., Granger, D. A., Kivlighan, K. T., Mills-Koonce, R., Willoughby, M., Greenberg, M. T., . . . Family Life Project Investigators. (2008). Maternal and child contributions to cortisol response to emotional arousal in young children from low-income, rural communities. Developmental Psychology, 44, 1095–1109. doi: 10.1037/0012– 1649.44.4.1095 Blair, C., Granger, D. Willoughby, M., Mills-Koonce, R., Cox, M., Greenberg, M. T., . . . Family Life Project Investigators. (2011). Salivary cortisol mediates effects of poverty and parenting on executive functions in early childhood. Child Development, 82, 1970–1984. doi: 10.1111/j.1467–8624.2011.01643.x Blair, C., & Raver, C. C. (2012). Child development in the context of adversity: Experiential canalization of brain and behavior. American Psychologist, 67, 309–318. Blair, C. & Raver, C. C. (2014) Closing the achievement gap through modification of neurocognitive and neuroendocrine function: Results from a cluster randomized controlled trial of an innovative approach to the education of children in kindergarten. PLOS ONE 9(11): e112393 doi: 10.1371/journal.pone.0112393 Blair, C., Raver, C. C., Granger, D., Mills-Koonce, R., Hibel, L., & Family Life Project Key Investigators. (2011). Allostasis and allostatic load in the context of poverty in early childhood. Development and Psychopathology, 23, 845–857. doi: 10.1017/S0954579411000344 Blair, R. J. R., & Coles, M. (2000). Expression recognition and behavioural problems in early adolescence. Cognitive Development, 15(4), 421–434. Bodrova E, Leong DJ (2007) Tools of the mind: The Vygotskian approach to early childhood education. New York: Merrill/Prentice Hall. Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: I. An evolutionary-developmental theory of the origins and functions of stress reactivity. Development and Psychopathology, 17(2), 271–301. Bradley, B. P., Mogg, K., Millar, N., Bonham-Carter, C., Fergusson, E., Jenkins, J., & Parr, M. (1997). Attentional biases for emotional faces. Cognition & Emotion, 11(1), 25–42. Bredy, T. W., Humpartzoomian, R. A., Cain, D. P., & Meaney, M. J. (2003). Partial reversal of the effect of maternal care on cognitive function through environmental enrichment. Neuroscience, 118, 571–576. Brody, G. H., Beach, S. R., Hill, K. G., Howe, G. W., Prado, G., & Fullerton, S. M. (2013). Using genetically informed, randomized prevention trials to test etiological hypotheses about child and adolescent drug use and psychopathology. American Journal of Public Health, 103(S1), S19–S24. Brody, G. H., Chen, Y. F., & Beach, S. R. (2013). Differential susceptibility to prevention: GABAergic, dopaminergic, and multilocus effects. Journal of Child Psychology and Psychiatry, 54(8), 863–871. Brody, G. H., Yu, T., Beach, S. R., Kogan, S. M., Windle, M., & Philibert, R. A. (2013). Harsh parenting and adolescent health: A longitudinal analysis with genetic moderation. Health Psychology, 33, 401–409.

516

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Brody, G. H., Yu, T., Chen, Y. F., Kogan, S. M., Evans, G. W., Beach, S. R., . . . & Philibert, R. A. (2013). Cumulative socioeconomic status risk, allostatic load, and adjustment: A prospective latent profile analysis with contextual and genetic protective factors. Developmental Psychology, 49, 913–927. Bruce, J., Gunnar, M. R., Pears, K. C., & Fisher, P. A. (2013). Early adverse care, stress neurobiology, and prevention science: Lessons learned. Prevention Science, 14, 247–256 1–10. Bryant, R. A., Kemp, A. H., Felmingham, K. L., Liddell, B., Olivieri, G., Peduto, A., . . . & Williams, L. M. (2008). Enhanced amygdala and medial prefrontal activation during nonconscious processing of fear in posttraumatic stress disorder: An fMRI study. Human Brain Mapping, 29(5), 517–523. Buckner, J. C., Mezzacappa, E., & Beardslee, W. R. (2003). Characteristics of resilient youths living in poverty: The role of self-regulatory processes. Development and Psychopathology, 15(1), 139–162. Buss, C., Davis, E. P., Hobel, C. J., & Sandman, C. A. (2011). Maternal pregnancy-specific anxiety is associated with child executive function at 6–9 years age. Stress, 14(6), 665–676. Buss, C., Davis, E. P., Muftuler, L. T., Head, K., & Sandman, C. A. (2010). High pregnancy anxiety during mid-gestation is associated with decreased gray matter density in 6–9-year-old children. Psychoneuroendocrinology, 35(1), 141–153. Butts, K. A., Weinberg, J., Young, A. H., & Phillips, A. G. (2011). Glucocorticoid receptors in the prefrontal cortex regulate stress-evoked dopamine efflux and aspects of executive function. Proceedings of the National Academy of Sciences, 108(45), 18459–18464. Cairns, R. B. (1991). Multiple metaphors for a singular idea. Developmental Psychology, 27, 23–26. Campbell, F., Conti, G., Heckman, J. J., Moon, S. H., Pinto, R., Pungello, E., & Pan, Y. (2014). Early childhood investments substantially boost adult health. Science, 343(6178), 1478–1485. Cameron, N. M., Champagne, F. A., Parent, C., Fish, E. W., Ozaki-Kuroda, K., & Meaney, M. J. (2005). The programming of individual differences in defensive responses and reproductive strategies in the rat through variations in maternal care. Neuroscience and Biobehavioral Reviews, 29, 843–865. doi: 10.1016/j.neubiorev.2005 .03.022 Carlson, S. M. (2005). Developmentally sensitive measures of executive function in preschool children. Developmental Neuropsychology, 28(2), 595–616. doi: 10.1207/s15326942dn2802_3 Carolina Consortium on Human Development. (1996). A collaborative statement. In R. B. Cairns, G. H. Elder, & E. J. Costello (Eds.) Developmental science (pp. 1–7). New York, NY: Cambridge University Press. Caspi, A., McClay, J., Moffitt, T. E., Mill, J., Martin, J., Craig, I. W., . . . & Poulton, R. (2002). Role of genotype in the cycle of violence in maltreated children. Science, 297(5582), 851–854. Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . & Poulton, R. (2003). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301(5631), 386. Cerqueira, J. J., Mailliet, F., Almeida, O. F., Jay, T. M., & Sousa, N. (2007). The prefrontal cortex as a key target of the maladaptive response to stress. Journal of Neuroscience, 27, 2781–2787. doi: 10.1523/ JNEUROSCI.4372–06.2007 Champagne, F. A., & Meaney, M. J. (2006). Stress during gestation alters postpartum maternal care and the development of the offspring in a rodent model. Biological Psychiatry, 59(12), 1227–1235. Chen, E., & Miller, G. E. (2012). “Shift-and-persist” strategies why low socioeconomic status isn’t always bad for health. Perspectives on Psychological Science, 7(2), 135–158.

Choi, J. M., Padmala, S., & Pessoa, L. (2012). Impact of state anxiety on the interaction between threat monitoring and cognition. Neuroimage, 59(2), 1912–1923. Cicchetti, D. (1984). The emergence of developmental psychopathology. Child Development, 55(1), 1–7. Cicchetti, D., & Hinshaw, S. P. (2002). Editorial: Prevention and intervention science: Contributions to developmental theory. Development and Psychopathology, 14(4), 667–671. Cicchetti, D., & Rogosch, F. A. (2002). A developmental psychopathology perspective on adolescence. Journal of Consulting and Clinical Psychology, 70(1), 6–20. Cicchetti, D., & Toth, S. L. (1992). The role of developmental theory in prevention and intervention. Development and Psychopathology, 4, 489–493. Corter, C. M., & Fleming, A. S. (2002). Psychobiology of maternal behavior in human beings. In M. H. Bornstein (Ed.), Handbook of parenting: Biology and ecology of parenting (pp. 141–182). Hillsdale, NJ: Lawrence Erlbaum Associates. Cummings, J. S., Pellegrini, D. S., Notarius, C. I., & Cummings, E. M. (1989). Children’s responses to angry adult behavior as a function of marital distress and history of interparent hostility. Child Development, 60(5), 1035–1043. Damasio, A. (2004). Emotions and feelings. In A. S. R. Manstead & N. Frijda (Eds.), Feelings and emotions: The Amsterdam Symposium (pp. 49–57). New York, NY: Cambridge University Press. Davidson, M., Amso, D., Anderson, L., & Diamond, A. (2006). Development of cognitive control and executive functions from 4 to 13 years: Evidence from manipulations of memory, inhibition, and task switching. Neuropsychologia, 44, 2037–2078. doi: 10.1016/ j.neuropsychologia.2006.02.006 Davies, P. T., Sturge-Apple, M. L., & Cicchetti, D. (2011). Interparental aggression and children’s adrenocortical reactivity: Testing an evolutionary model of allostatic load. Development and Psychopathology, 23(3), 801–814. doi: 10.1017/S0954579411000319 de Kloet, E. R., Karst, H., & Joëls, M. (2008). Corticosteroid hormones in the central stress response: Quick-and-slow. Frontiers in Neuroendocrinology, 29(2), 268–272. De Luca, C. R., Wood, S. J., Anderson, V., Buchanan, J. A., Proffitt, T. M., Mahony, K., & Pantelis, C. (2003). Normative data from the CANTAB. I: Development of executive function over the lifespan. Journal of Clinical and Experimental Neuropsychology, 25(2), 242–254. De˛biec, J., Díaz-Mataix, L., Bush, D. E., Doyère, V., & LeDoux, J. E. (2010). The amygdala encodes specific sensory features of an aversive reinforcer. Nature Neuroscience, 13(5), 536–537. Del Giudice, M., Ellis, B. J., & Shirtcliff, E. A. (2011). The adaptive calibration model of stress responsivity. Neuroscience & Biobehavioral Reviews, 35(7), 1562–1592. Denham, S. A., McKinley, M., Couchoud, E. A., & Holt, R. (1990). Emotional and behavioral predictors of preschool peer ratings. Child Development, 61(4), 1145–1152. Derryberry, D., & Rothbart, M. K. (1997). Reactive and effortful processes in the organization of temperament. Development and Psychopathology, 9, 633–652. Desimone, R., & Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual Review of Neuroscience, 18(1), 193–222. Diamond, A. (2013). Executive functions. Annual Review of Psychology, 64, 135–168. doi: 10.1146/annurev-psych-113011–143750 Diamond, A., Barnett, W. S., Thomas, J., & Munro, S. (2007a). Preschool program improves cognitive control. Science 318(5855), 1387–1388. Diamond, D. M., Campbell, A. M., Park, C. R., Halonen, J., & Zoladz, P. R. (2007b). The temporal dynamics model of emotional memory

References processing: A synthesis on the neurobiological basis of stress-induced amnesia, flashbulb and traumatic memories, and the yerkes-dodson law. Neural Plasticity, 2007, 60803. doi: 10.1155/2007/60803 Dolan, R. J., & Vuilleumier, P. (2003). Amygdala automaticity in emotional processing. Annals of the New York Academy of Sciences, 985(1), 348–355. Dolcos, F., Iordan, A. D., & Dolcos, S. (2011). Neural correlates of emotion–cognition interactions: A review of evidence from brain imaging investigations. Journal of Cognitive Psychology, 23(6), 669–694. Dozier, M., Peloso, E., Lindhiem, O., Gordon, M. K., Manni, M., Sepulveda, S., . . . & Levine, S. (2006). Developing evidence-based interventions for foster children: An example of a randomized clinical trial with infants and toddlers. Journal of Social Issues, 62(4), 767–785. Duckworth, A. L., Grant, H., Loew, B., Oettingen, G., & Gollwitzer, P. M. (2011). Self regulation strategies improve self-discipline in adolescents: Benefits of mental contrasting and implementation intentions. Educational Psychology, 31(1), 17–26. Duncan, J., Burgess, P., & Emslie, H. (1995). Fluid intelligence after frontal lobe lesions. Neuropsychologia, 33(3), 261–268. Durlak, J. A., Weissberg, R. P., Dymnicki, A. B. Taylor, R. D., & Schellinger, K. B., (2011). The impact of enhancing students’ social and emotional learning: A metaanalysis of school based universal interventions. Child Development, 82, 405–432. El-Sheikh, M. (1994). Children’s emotional and physiological responses to interadult angry behavior: The role of history of interparental hostility. Journal of Abnormal Child Psychology, 22(6), 661–678. Espy, K. A., Bull, R., Martin, J., & Stroup, W. (2006). Measuring the development of executive control with the shape school. Psychological Assessment, 18(4), 373. Etkin, A., & Schatzberg, A. F. (2011). Common abnormalities and disorder-specific compensation during implicit regulation of emotional processing in generalized anxiety and major depressive disorders. American Journal of Psychiatry, 168(9), 968–978. Evans, G. W., & Schamberg, M. A. (2009). Childhood poverty, chronic stress, and adult working memory. Proceedings of the National Academy of Sciences of the United States of America, 106(16), 6545–6549. doi: 10.1073/pnas.0811910106 Evans, G. W., Kim, P., Ting, A. H., Tesher, H. B., & Shannis, D. (2007). Cumulative risk, maternal responsiveness, and allostatic load among young adolescents. Developmental Psychology, 43(2), 341–351. doi: 10.1037/0012–1649.43.2.341 Fair, D. A., Bathula, D., Nikolas, M. A., & Nigg, J. T. (2012). Distinct neuropsychological subgroups in typically developing youth inform heterogeneity in children with ADHD. Proceedings of the National Academy of Sciences, 109(17), 6769–6774. Fair, D. A., Nigg, J. T., Iyer, S., Bathula, D., Mills, K. L., Dosenbach, N. U., . . . Milham, M. P. (2012). Distinct neural signatures detected for ADHD subtypes after controlling for micro-movements in resting state functional connectivity MRI data. Frontiers in Systems Neuroscience, 6, 80. doi: 10.3389/fnsys.2012.00080 Fan, J. B., Zhang, C. S., Gu, N. F., Li, X. W., Sun, W. W., Wang, H. Y., . . . He, L. (2005). Catechol-O-methyltransferase gene Val/Met functional polymorphism and risk of schizophrenia: A large-scale association study plus meta-analysis. Biological Psychiatry, 57(2), 139–144. Feldman, R., Singer, M., & Zagoory, O. (2010). Touch attenuates infants’ physiological reactivity to stress. Developmental Science, 13, 271–278. Fine, S. E., Izard, C. E., Mostow, A. J., Trentacosta, C. J., & Ackerman, B. P. (2003). First grade emotion knowledge as a predictor of fifth grade self-reported internalizing behaviors in children

517

from economically-disadvantaged families. Development and Psychopathology, 15, 331–345. Fisher, P. A., Stoolmiller, M., Gunnar, M. R., & Burraston, B. O. (2007). Effects of a therapeutic intervention for foster preschoolers on diurnal cortisol activity. Psychoneuroendocrinology, 32(8), 892–905. Fleming, A. S., & Luebke, C. (1981). Timidity prevents the virgin female rat from being a good mother: emotionality differences between nulliparous and parturient females. Physiology & Behavior, 27(5), 863–868. Fleming, A. S., Ruble, D., Krieger, H., & Wong, P. Y. (1997). Hormonal and experiential correlates of maternal responsiveness during pregnancy and the puerperium in human mothers. Hormones and Behavior, 31(2), 145–158. Fleming, A. S., Steiner, M., & Anderson, V. (1987). Hormonal and attitudinal correlates of maternal behaviour during the early postpartum period in first-time mothers. Journal of Reproductive and Infant Psychology, 5(4), 193–205. Fleming, A. S., Steiner, M., & Corter, C. (1997). Cortisol, hedonics, and maternal responsiveness in human mothers. Hormones and Behavior, 32(2), 85–98. Fleming, A. S., Vaccarino, F., & Luebke, C. (1980). Amygdaloid inhibition of maternal behavior in the nulliparous female rat. Physiology & Behavior, 25(5), 731–743. Francis, D. D., Caldji, C., Champagne, F., Plotsky, P. M., & Meaney, M. J. (1999a). The role of corticotropin-releasing factor—norepinepherine systems in mediating the effects of early experience on the development of behavioral and endocrine responses to stress. Biological Psychiatry, 46, 1153–1166. Francis, D. D., Diorio, J., Liu, D., & Meaney, M. J. (1999b). Nongenomic transmission across generations in maternal behavior and stress responses in the rat. Science, 286, 1155–1158. Francis, D. D., Diorio, J., Plotsky, P. M., & Meaney, M. J. (2002). Environmental enrichment reverses the effects of maternal separation on stress reactivity. Journal of Neuroscience, 22(18), 7840–7843. Garner, P. W., Jones, D. C., & Miner, J. L. (1994). Social competence among lowincome preschoolers: Emotion socialization practices and social cognitive correlates. Child Development, 65(2), 622–637. Gianaros, P. J., & Hackman, D. A. (2013). Contributions of neuroscience to the study of socioeconomic health disparities. Psychosomatic Medicine, 75, 610–615. Gogtay, N., Giedd, J. N., Lusk, L., Hayashi, K. M., Greenstein, D., Vaituzis, A. C., . . . Thompson, P. M. (2004). Dynamic mapping of human cortical development during childhood through early adulthood. Proceedings of the National Academy of Sciences of the United States of America, 101(21), 8174–8179. doi: 10.1073/pnas.0402680101 Goldman Rakic, P. S. (1995). Architecture of the prefrontal cortex and the central executive. Annals of the New York Academy of Sciences, 769(1), 71–84. Gonzalez, A., Jenkins, J. M., Steiner, M., & Fleming, A. S. (2012). Maternal early life experiences and parenting: the mediating role of cortisol and executive function. Journal of the American Academy of Child & Adolescent Psychiatry, 51(7), 673–682. Gotlib, I. H., & Joormann, J. (2010). Cognition and depression: Current status and future directions. Annual Review of Clinical Psychology, 6, 285–312. Gottlieb, G. (1991). Experiential canalization of behavioral development: Theory. Developmental Psychology, 27, 4–13. Gottlieb, G. (1997). Synthesizing nature-nurture: Prenatal roots of instinctive behavior. Mahwah, NJ: Erlbaum. Gottlieb, G., & Kuo, Z. Y. (1965). Development of behavior in the duck embryo. Journal of Comparative and Physiological Psychology, 59(2), 183–188. doi: 10.1037/h0021831

518

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Granger, D. A., Kivlighan, K. T., El-Sheikh, M., Gordis, E. B., & Stroud, L. R. (2007). Salivary-amylase in biobehavioral research. Annals of the New York Academy of Sciences, 1098(1), 122–144. doi: 10.1196/annals.1384.008 Granger, D. A., Kivlighan, K. T., Fortunato, C., Harmon, A. G., Hibel, L. C., Schwartz, E. B., & Whembolua, G. L. (2007). Integration of salivary biomarkers into developmental and behaviorally-oriented research: Problems and solutions for collecting specimens. Physiology & Behavior, 92(4), 583–590. doi: 10.1016/j.physbeh.2007.05.004 Grant, K. E., Compas, B. E., Stuhlmacher, A., Thurm, A. E., McMahon, S., & Halpert, J. (2003). Stressors and child and adolescent psychopathology: Moving from markers to mechanisms of risk. Psychological Bulletin, 129, 447–466. Greenough, W. T., Black, J. E., & Wallace, C. S. (1987). Experience and brain development. Child Development, 58(3), 539–559. Groeneweg, F. L., Karst, H., de Kloet, E. R., & Joëls, M. (2011). Rapid non-genomic effects of corticosteroids and their role in the central stress response. Journal of Endocrinology, 209(2), 153–167. Gunnar, M., & Quevedo, K. (2007). The neurobiology of stress and development. Annual Review of Psychology, 58, 145–173. doi: 10.1146/annurev.psych.58.110405.085605halon Hebb, D. O. (1949). The organization of the brain. Brain Research Bulletin, 50, 437. Heckman, J. J., Malofeeva, L., Pinto, R., & Savelyev, P. (2008). The effect of the Perry Preschool program on cognitive and noncognitive skills: Beyond treatment effects. Unpublished manuscript, University of Chicago, Department of Economics. Henckens, M. J., van Wingen, G. A., Joëls, M., & Fernández, G. (2012). Time-dependent effects of cortisol on selective attention and emotional interference: A functional MRI study. Frontiers in Integrative Neuroscience, 6, 66. Herrenkohl, T. I., Guo, J., Kosterman, R., Hawkins, J. D., Catalano, R. F., & Smith, B. H. (2001). Early adolescent predictors of youth violence as mediators of childhood risks. Journal of Early Adolescence, 21(4), 447–469. Hibel, L., Granger, D. Blair, C., Cox, M., & FLP Investigators. (2009). Intimate partner violence moderates the association between mother-infant adrenocortical activity across an emotional challenge. Journal of Family Psychology, 23, 615–625. Hibel, L. C., Granger, D. A., Blair, C., Finegood, E. D., & Family Life Project Investigators (in press). Maternal-child adrenocortical attunement in early childhood: continuity and change. Developmental Psychobiology. Holmes, A., & Wellman, C. L. (2009). Stress-induced prefrontal reorganization and executive dysfunction in rodents. Neuroscience and Biobehavioral Reviews, 33, 773–783. doi: 10.1016/j.neubiorev.2008.11.005 Holt, E. B. (1931). Animal drive and the learning process: An essay toward radical empiricism. Oxford, UK: Holt. Houben, K., Wiers, R. W., & Jansen, A. (2011). Getting a grip on drinking behavior training working memory to reduce alcohol abuse. Psychological Science, 22(7), 968–975. Hughes, C., & Ensor, R. (2007). Positive and protective: Effects of early theory of mind on problem behaviors in at-risk preschoolers. Journal of Child Psychology and Psychiatry, 48(10), 1025–1032. Huizink, A. C., Robles de Medina, P. G., Mulder, E. J., Visser, G. H., & Buitelaar, J. K. (2002). Psychological measures of prenatal stress as predictors of infant temperament. Journal of the American Academy of Child & Adolescent Psychiatry, 41(9), 1078–1085. Husain, S. A., Allwood, M. A., & Bell, D. J. (2008). The relationship between PTSD symptoms and attention problems in children exposed to the Bosnian war. Journal of Emotional and Behavioral Disorders, 16(1), 52–62.

Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., Sanislow, C., & Wang, P. (2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167, 748–751. Izard, C. E., Fine, S., Mostow, A., Trentacosta, C., & Campbell, J. (2002). Emotion processes in normal and abnormal development and preventive intervention. Development and Psychopathology, 14(4), 761–787. Jaffee, W. B., & D’Zurilla, T. J. (2003). Adolescent problem solving, parent problem solving, and externalizing behavior in adolescents. Behavior Therapy, 34(3), 295–311. Jans, J. E., & Woodside, B. C. (1990). Nest temperature: Effects on maternal behavior, pup development, and interactions with handling. Developmental Psychobiology, 23(6), 519–534. Joëls, M., & Baram, T. Z. (2009). The neuro-symphony of stress. Nature Reviews Neuroscience, 10(6), 459–466. Kawasaki, H., Tsuchiya, N., Kovach, C. K., Nourski, K. V., Oya, H., Howard, M. A., & Adolphs, R. (2012). Processing of facial emotion in the human fusiform gyrus. Journal of Cognitive Neuroscience, 24(6), 1358–1370. Kim, J., & Cicchetti, D. (2010). Longitudinal pathways linking child maltreatment, emotion regulation, peer relations, and psychopathology. Journal of Child Psychology and Psychiatry, 51(6), 706–716. Kim, P., Leckman, J. F., Mayes, L. C., Feldman, R., Wang, X., & Swain, J. E. (2010). The plasticity of human maternal brain: Longitudinal changes in brain anatomy during the early postpartum period. Behavioral Neuroscience, 124(5), 695. Kinnunen, A. K., Koenig, J. I., & Bilbe, G. (2003). Repeated variable prenatal stress alters pre-and postsynaptic gene expression in the rat frontal pole. Journal of Neurochemistry, 86(3), 736–748. Klingberg, T., Fernell, E., Olesen, P. J., Johnson, M., Gustafsson, P., Dahlström, K., . . . & Westerberg, H. (2005). Computerized training of working memory in children with ADHD-a randomized, controlled trial. Journal of the American Academy of Child & Adolescent Psychiatry, 44(2), 177–186. Krämer, M., Schmitz, J., Heinrichs, N., & Tuschen-Caffier, B. (2011). Self-evaluation, social and cognitive performance in children with social phobia. Journal of Experimental Psychopathology, 2(4), 586–600. Krpan, K. M., Coombs, R., Zinga, D., Steiner, M., & Fleming, A. S. (2005). Experiential and hormonal correlates of maternal behavior in teen and adult mothers. Hormones and Behavior, 47(1), 112–122. Kuhl, P. K., Williams, K. A., Lacerda, F., Stevens, K. N., & Lindblom, B. (1992). Linguistic experience alters phonetic perception in infants by 6 months of age. Science, 255, 606–608. doi: 10.1126/science .1736364 Lambert, K. G., & Kinsley, C. H. (2012). Brain and behavioral modifications that accompany the onset of motherhood. Parenting, 12(1), 74–88. Landers, M. S., & Sullivan, R. M. (2012). The development and neurobiology of infant attachment and fear. Developmental Neuroscience, 34(2–3), 101–114. Laurent, H. K., Ablow, J. C., Measelle, J. (2012). Taking stress response out of the box. Stability, discontinuity, and temperament effects on HPA and SNS across social stressors in mother–infant dyads. Developmental Psychology, 48, 35–45. Lazarus, R. S., & Folkman, S. (1984). Stress, appraisal, and coping. New York, NY: Springer Publishing Company. LeDoux, J. (1996). Emotional networks and motor control: A fearful view. Progress in Brain Research, 107, 437–446. LeDoux, J. (2012). Rethinking the emotional brain. Neuron, 73(4), 653–676.

References Lewkowicz, D. J., & Ghazanfar, A. A. (2009). The emergence of multisensory systems through perceptual narrowing. Trends in Cognitive Sciences, 13, 470–478. doi: 10.1016/j.tics.2009.08.004 Liddell, B. J., Brown, K. J., Kemp, A. H., Barton, M. J., Das, P., Peduto, A., . . . & Williams, L. M. (2005). A direct brainstem–amygdala– cortical “alarm” system for subliminal signals of fear. Neuroimage, 24(1), 235–243. Liston, C., Miller, M. M., Goldwater, D. S., Radley, J. J., Rocher, A. B., Hof, P. R., . . . McEwen, B. S. (2006). Stress-induced alterations in prefrontal cortical dendritic morphology predict selective impairments in perceptual attentional set-shifting. Journal of Neuroscience, 26, 7870–7874. doi: 10.1523/JNEUROSCI.1184–06.2006 Liu, D., Diorio, J., Day, J. C., Francis, D. D., & Meaney, M. J. (2000). Maternal care, hippocampal neurogenesis, and cognitive development in rats. Nature Neuroscience, 3, 799–806. doi: 10.1038/77702 Liu, D., Diorio, J., Tannenbaum, B., Caldji, C., Francis, D., Freedman, A., . . . Meaney, M. J. (1997). Maternal care, hippocampal glucocorticoid receptors, and hypothalamic-pituitary-adrenal responses to stress. Science, 277(5332), 1659–1662. LoBue, V., & DeLoache, J. S. (2008). Detecting the snake in the grass attention to fear-relevant stimuli by adults and young children. Psychological Science, 19(3), 284–289. Lonigan, C. J., & Vasey, M. W. (2009). Negative affectivity, effortful control, and attention to threat-relevant stimuli. Journal of Abnormal Child Psychology, 37(3), 387–399. Lorberbaum, J. P., Newman, J. D., Horwitz, A. R., Dubno, J. R., Lydiard, R. B., Hamner, M. B., . . . & George, M. S. (2002). A potential role for thalamocingulate circuitry in human maternal behavior. Biological Psychiatry, 51(6), 431–445. Luciana, M., Conklin, H. M., Hooper, C. J., & Yarger, R. S. (2005). The development of nonverbal working memory and executive control processes in adolescents. Child Development, 76(3), 697–712. doi: 10.1111/j.1467–8624.2005.00872.x Lupien, S. J., King, S., Meaney, M. J., & McEwen, B. S. (2001). Can poverty get under your skin? Basal cortisol levels and cognitive function in children from low and high socioeconomic status. Development and Psychopathology, 13(3), 653–676. Luria, A. R. (1963). Restoration of function after brain injury. Oxford, UK: Macmillan. Luu, P., & Tucker, D. M. (2004). Self-regulation by the medial frontal cortex: Limbic representation of motive set-points. In M. Beauregard (Ed.), Consciousness, emotional self-regulation and the brain. Amsterdam, The Netherlands: John Benjamin. Luu, P., Tucker, D. M., & Derryberry, D. (1998). Anxiety and the motivational basis of working memory. Cognitive Therapy and Research. Special Issue: Cognition and Anxiety, 22(6), 577–594. doi: 10.1023/A:1018742120255 Lyubomirsky, S., Tucker, K., Caldwell, N. D., & Berg, K. (1999). Why ruminators are poor problem solvers: Clues from the phenomenology of dysphoric rumination. Journal of Personality and Social Psychology, 77, 1041–1060. Maccari, S., Darnaudery, M., Morley-Fletcher, S., Zuena, A. R., Cinque, C., & Van Reeth, O. (2003). Prenatal stress and long-term consequences: Implications of glucocorticoid hormones. Neuroscience & Biobehavioral Reviews, 27(1), 119–127. Magnusson, D., & Cairns, R. B. (1996). Developmental science: Toward a unified framework. In R. B. Cairns, G. H. Elder, & J. Costello (Eds.), Developmental science (pp. 2–30). New York, NY: Cambridge University Press. Mattay, V. S., Goldberg, T. E., Fera, F., Hariri, A. R., Tessitore, A., Egan, M. F., . . . Weinberger, D. R. (2003). Catechol O-methyltransferase val158-met genotype and individual variation in the brain response

519

to amphetamine. Proceedings of the National Academy of Sciences of the United States of America, 100(10), 6186–6191. Mayberg, H. S. (2003). Modulating dysfunctional limbic-cortical circuits in depression: Towards development of brain-based algorithms for diagnosis and optimised treatment. British Medical Bulletin, 65, 193–207. McCarty, C. A., Mason, W. A., Kosterman, R., Hawkins, J. D., Lengua, L. J., & McCauley, E. (2008). Adolescent school failure predicts later depression among girls. Journal of Adolescent Health: Official Publication of the Society for Adolescent Medicine, 43, 180–187. McCoy, D., Raver, C. C., & Sharkey, P. (in press). Children’s cognitive performance and selective attention following recent neighborhood violence. Journal of Health and Social Behavior. McCoy, K., Cummings, E. M., & Davies, P. T. (2009). Constructive and destructive marital conflict, emotional security and children’s prosocial behavior. Journal of Child Psychology and Psychiatry, 50(3), 270–279. McEwen, B. S. (1998). Stress, adaptation, and disease: Allostasis and allostatic load. Annals of the New York Academy of Sciences, 840(1), 33–44. McEwen, B. S. (2000). Allostasis and allostatic load: implications for neuropsychopharmacology. Neuropsychopharmacology, 22(2), 108–124. McEwen, B., & Gianoros, P. (2011). Stress- and allostasis-induced brain plasticity. Annual Review of Medicine, 62, 431–445. doi: 10.1146/ annurev-med-052209–100430 McRae, K., Hughes, B., Chopra, S., Gabrieli, J. D., Gross, J. J., & Ochsner, K. N. (2010). The neural bases of distraction and reappraisal. Journal of Cognitive Neuroscience, 22(2), 248–262. Meaney, M. J. (2001). Maternal care, gene expression, and the transmission of individual differences in stress reactivity across generations. Annual Review of Neuroscience, 24, 1161–1192. doi: 10.1146/ annurev.neuro.24.1.1161 Melby-Lervåg, M., & Hulme, C. (2013). Is working memory training effective? A meta-analytic review. Developmental Psychology, 49(2), 270–291. Mennes, M., Stiers, P., Lagae, L., & Van den Bergh, B. (2006). Long-term cognitive sequelae of antenatal maternal anxiety: Involvement of the orbitofrontal cortex. Neuroscience & Biobehavioral Reviews, 30(8), 1078–1086. Mills-Koonce, W. R., Propper, C., Gariepy, J. L., Barnett, M., Moore, G. A., Calkins, S., & Cox, M. J. (2009). Psychophysiological correlates of parenting behavior in mothers of young children. Developmental Psychobiology, 51(8), 650–661. Miller, G. E., Chen, E., Fok, A. K., Walker, H., Lim, A., Nicholls, E. F., . . . Kobor, M. S. (2009). Low early-life social class leaves a biological residue manifested by decreased glucocorticoid and increased proinflammatory signaling. Proceedings of the National Academy of Sciences, USA, 106, 14716–14721. Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., Howerter, A., & Wager, T. D. (2000). The unity and diversity of executive functions and their contributions to complex “frontal lobe” tasks: A latent variable analysis. Cognitive Psychology, 41(1), 49–100. doi: 10.1006/ cogp.1999.0734 Moffitt, T. et al., (2011). A gradient of childhood self-control predicts health, wealth, and public safety. PNAS, 108, 2693–2698. Mogg, K., & Bradley, B. P. (1998). A cognitive-motivational analysis of anxiety. Behaviour Research and Therapy, 36(9), 809–848. Mogoa¸se, C., David, D., & Koster, E. H. (2014). Clinical efficacy of attentional bias modification procedures: An updated meta-analysis. Journal of Clinical Psychology. 70, 1133–1157.

520

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Molenaar, P. (2008). On the implications of the classical ergodic theorems: Analysis of developmental processes has to focus on intra-individual variation. Developmental Psychobiology, 50, 60–69. Moos, R. H., & Holahan, C. J. (2003). Dispositional and contextual perspectives on coping: Toward and integrative framework. Journal of Clinical Psychology, 59, 1387–1403. Moriceau, S., Shionoya, K., Jakubs, K., & Sullivan, R. M. (2009). Early-life stress disrupts attachment learning: the role of amygdala corticosterone, locus ceruleus corticotropin releasing hormone, and olfactory bulb norepinephrine. Journal of Neuroscience, 29(50), 15745–15755. Nagasawa, M., Okabe, S., Mogi, K., & Kikusui, T. (2012). Oxytocin and mutual communication in mother-infant bonding. Frontiers in Human Neuroscience, 6, 31. doi: 10.3389/fnhum.2012.00031 Nakagawa, A., & Sukigara, M. (2012). Difficulty in disengaging from threat and temperamental negative affectivity in early life: A longitudinal study of infants aged 12–36 months. Behavioral and Brain Function, 8, 1–8. Nephew, B. C., & Bridges, R. S. (2008). Central actions of arginine vasopressin and a V1a receptor antagonist on maternal aggression, maternal behavior, and grooming in lactating rats. Pharmacology Biochemistry and Behavior, 91(1), 77–83. Nephew, B. C., & Bridges, R. S. (2011). Effects of chronic social stress during lactation on maternal behavior and growth in rats. Stress, 14(6), 677–684. Nesselroade, J. R., & Jones, C. J. (1991) Multi-modal selection effects in the study of adult development: A perspective on multivariate, replicated, single-subject, repeated measures designs. Experimental Aging Research, 17, 21–27. Neville, H. J., Stevens, C., Pakulak, E., Bell, T. A., Fanning, J., Klein, S., & Isbell, E. (2013). Family-based training program improves brain function, cognition, and behavior in lower socioeconomic status preschoolers. Proceedings of the National Academic of Sciences, USA, 110, 12138–12143. Nigg, J. T. (2001). Is ADHD a disinhibitory disorder? Psychological Bulletin, 127(5), 571. Nigg, J. T., & Casey, B. J. (2005). An integrative theory of attention-deficit/hyperactivity disorder based on the cognitive and affective neurosciences. Development and Psychopathology, 17(3), 785–806. Nolen-Hoeksema, S. (1991). Responses to depression and their effects on the duration of depressive episodes. Journal of Abnormal Psychology, 100(4), 569. Nolen-Hoeksema, S. (2000). The role of rumination in depressive disorders and mixed anxiety/depressive symptoms. Journal of Abnormal Psychology, 109(3), 504–511. doi: 10.1037/0021–843X.109.3.504 Numan, M. (2006). Hypothalamic neural circuits regulating maternal responsiveness toward infants. Behavioral and Cognitive Neuroscience Reviews, 5(4), 163–190. Numan, M., & Insel, T. R. (2003). The neurobiology of parental behavior. New York, NY: Springer-Verlag. O’Donnell, K., O’Connor, T. G., & Glover, V. (2009). Prenatal stress and neurodevelopment of the child: focus on the HPA axis and role of the placenta. Developmental Neuroscience, 31(4), 285–292. Ochsner, K. N., & Gross, J. J. (2005). The cognitive control of emotion. Trends in Cognitive Sciences, 9(5), 242–249. Ochsner, K. N., Bunge, S. A., Gross, J. J., & Gabrieli, J. D. (2002). Rethinking feelings: An fMRI study of the cognitive regulation of emotion. Journal of Cognitive Neuroscience, 14(8), 1215–1229. Odgers, C. L., & Jaffee, S. R. (2013). Routine versus catastrophic influences on the developing child. Annual Review of Public Health, 34, 29–48.

Öhman, A., Flykt, A., & Esteves, F. (2001). Emotion drives attention: detecting the snake in the grass. Journal of Experimental Psychology: General, 130(3), 466. Ouellet-Morin, I., Tremblay, R. E., Boivin, M., Meaney, M., Kramer, M., & Cote, S. M. (2010). Diurnal cortisol secretion at home and in child care: A prospective study of 2-year-old toddlers. Journal of Child Psychology and Psychiatry, 51, 295–303. doi: 10.1111/j.1469–7610 .2009.02167.x Oyama, S. (2000). The ontogeny of information (2nd ed.). Durham, NC: Duke University Press. Ozonoff, S., & McEvoy, R. E. (1994). A longitudinal study of executive function and theory of mind development in autism. Development and Psychopathology, 6(3), 415–431. Pessoa, L. (2009). How do emotion and motivation direct executive control? Trends in Cognitive Sciences, 13(4), 160–166. Petersen, S. E., & Posner, M. I. (2012). The attention system of the human brain: 20 years after. Annual Reviews of Neuroscience, 35, 73–89. doi: 10.1146/annurev-neuro-062111–150525 Phillips, M. L., Drevets, W. C., Rauch, S. L., & Lane, R. (2003). Neurobiology of emotion perception II: Implications for major psychiatric disorders. Biological Psychiatry, 54, 515–528. Pollak, S. D., Messner, M., Kistler, D. J., & Cohn, J. F. (2009). Development of perceptual expertise in emotion recognition. Cognition, 110, 242–247. doi: 10.1016/j.cognition.2008.10.010 Pollak, S. D., Vardi, S., Putzer Bechner, A. M., & Curtin, J. J. (2005). Physically abused children’s regulation of attention in response to hostility. Child Development, 76, 968–977. doi: 10.1111/ j.1467–8624.2005.00890.x Pons, F., Lewkowicz, D. J., Soto-Faraco, S., & Sebastian-Galles, N. (2009). Narrowing of intersensory speech perception in infancy. Proceedings of the National Academy of Sciences, USA, 106, 10598–10602. doi: 10.1073/pnas.0904134106 Posner, M. I. (1994). Attention: The mechanisms of consciousness. Proceedings of the National Academy of Sciences, 91(16), 7398–7403. Posner, M., & Rothbart, M. (2000) Developing mechanisms of self-regulation. Development and Psychopathology 12, 427–441. Posner, M. I., & Rothbart, M. K. (2007). Research on attention networks as a model for the integration of psychological science. Annual Review of Psychology, 58, 1–23. doi: 10.1146/annurev.psych .58.110405.085516 Posner, M. I., Rothbart, M. K., Sheese, B. E., & Voelker, P. (2012). Control networks and neuromodulators of early development. Developmental Psychology, 48(3), 827–835. doi: 10.1037/a0025530 Posner, M., & Petersen, S. (1990). The attention system of the human brain. Annual Review of Neuroscience, 13, 25–42. Pribram, K. H. (1960). A review of theory in physiological psychology. Annual Review of Psychology, 11, 1–40. Rabbitt, P. (Ed.). (1997). Methodology of frontal and executive function. New York NY: Psychology Press. Radley, J. J., Arias, C. M., & Sawchenko, P. E. (2006). Regional differentiation of the medial prefrontal cortex in regulating adaptive responses to acute emotional stress. Journal of Neuroscience, 26, 12967–12976. doi: 10.1523/JNEUROSCI.4297–06.2006 Ram, N., Brose, A., & Molenaar, P. C. (2013). Dynamic factor analysis: Modeling person-specific process. In T. Little (Ed.), The Oxford handbook of quantitative methods: Statistical analysis (Vol. 2, pp. 441–457). New York, NY: Oxford University Press. Ramirez, G., Gunderson, E. A., Levine, S. C., & Beilock, S. L. (2013). Math anxiety, working memory, and math achievement in early elementary school. Journal of Cognition and Development, 14(2), 187–202.

References Ramos, B. P., & Arnsten, A. F. T. (2007). Adrenergic pharmacology and cognition: Focus on the prefrontal cortex. Pharmacology & Therapeutics, 113, 523–536. doi: 10.1016/j.pharmthera.2006.11.006 Raver, C. C., Blair, C., Garrett-Peters, P., & Family Life Project Key Investigators. (in press). Chronic exposure to poverty, household chaos, and interparental aggression predict children’s emotion regulation. Development and Psychopathology. Raver, C. C., Jones, S. M., Li-Grining, C. P., Zhai, F., Metzger, M., & Solomon, B. (2009). Targeting children’s behavior problems in preschool classrooms: A cluster-randomized controlled trial. Journal of Consulting and Clinical Psychology, 77(2), 302–316. Raver, C. C., Jones, S. M., Li-Grining, C. P., Zhai, F., Bub, K, & Pressler, E. (2011). CSRP’s impact on low-income preschoolers’ pre-academic skills: Self-regulation as a mediating mechanism. Child Development, 82(1), 362–378. Ray, R. D., McRae, K., Ochsner, K. N., & Gross, J. J. (2010). Cognitive reappraisal of negative affect: Converging evidence from EMG and self-report. Emotion, 10(4), 587. Riggs, N. R., Greenberg, M. T., Kusché, C. A., & Pentz, M. A. (2006). The mediational role of neurocognition in the behavioral outcomes of a social-emotional prevention program in elementary school students: Effects of the PATHS curriculum. Prevention Science, 7(1), 91–102. Riggs, N. R., Sakuma, K. L. K., & Pentz, M. A. (2007). Preventing risk for obesity by promoting self-regulation and decision-making skills pilot results from the PATHWAYS to Health Program (PATHWAYS). Evaluation Review, 31(3), 287–310. Rive, M. M., van Rooijen, G., Veltman, D. J., Phillips, M. L., Schene, A. H., & Ruhe, H. G. (2013). Neural correlates of dysfunctional emotion regulation in major depressive disorder. A systematic review of neuroimaging studies. Neuroscience & Biobehavioral Reviews, 37, 2529–2553. Roberson-Nay, R., McClure, E. B, Monk, C. S., Nelson, E. E., Guyer, A. E.,; Fromm, S. J., . . . Pine, D. S. (2006). Increased amygdala activity during successful memory encoding in adolescent major depressive disorder: An fMRI study. Biological Psychiatry, 60(9), 966–973. Ronan, K. R., Kendall, P. C., & Rowe, M. (1994). Negative affectivity in children: Development and validation of a self-statement questionnaire. Cognitive Therapy and Research, 18(6), 509–528. Rosenblatt, J. S. (1994). Psychobiology of maternal behavior: Contribution to the clinical understanding of maternal behavior among humans. Acta Paediatrica, 83(s397), 3–8. Rueda, M. R., Posner, M. I., & Rothbart, M. K. (2004). Attentional control and self-regulation. Handbook of self-regulation: Research, Theory, and Applications, 2, 284–299. Ruhe, H. G., Booij, J., Veltman, D. J., Michel, M. C., & Schene, A. H. (2012). Successful pharmacologic treatment of major depressive disorder attenuates amygdala activation to negative facial expressions: A functional magnetic resonance imaging study. Journal of Clinical Psychiatry, 73(4), 451–459. Sapolsky, R. M., Meaney, M. J., & McEwen, B. S. (1985). The development of the glucocorticoid receptor system in the rat limbic brain. III. Negative-feedback regulation. Developmental Brain Research, 18(1), 169–173. Sapolsky, R. M., Romero, L. M., & Munck, A. U. (2000). How do glucocorticoids influence stress responses? Integrating permissive, suppressive, stimulatory, and preparative actions. Endocrine Reviews, 21(1), 55–89. Schultz, D., Grodack, A., & Izard, C. E. (2010). State and trait anger, fear, and social information processing. In M. Potegal, G. Stemmler, & C. Spielberger (Eds.), International handbook of anger (pp. 311–325). New York, NY: Springer.

521

Schultz, D., Izard, C. E., & Abe, J. A. A. (2005). The emotion systems and the development of emotional intelligence. In R. Schultz & R. D. Roberts (Eds.), Emotional intelligence: An international handbook (pp. 51–67). Gottingen, Germany: Hogrefe & Huber Publishers. Schultz, D., Izard, C. E., & Ackerman, B. P. (2000). Children’s anger attribution bias: Relations to family environment and social adjustment. Social Development, 9(3), 284–301. Schultz, D., Izard, C. E., Ackerman, B. P., & Youngstrom, E. A. (2001). Emotion knowledge in economically disadvantaged children: Self-regulatory antecedents and relations to social difficulties and withdrawal. Development and Psychopathology, 13(1), 53–67. Sharkey, P., Strayer, N., Papachristos, A., & Raver, C. C. (2012). The effect of local violence on children’s attention and impulse control. American Journal of Public Health, 102, 2287–2293. Shaw, D. S., Connell, A., Dishion, T., Wilson, M. N., & Gardner, F (2009). Improvements in maternal depression as a mediator of intervention effects on early childhood problem behavior. Development and Psychopathology, 21, 417–439. doi: 10.1017/S0954579409000236 Shin, L. M., Bush, G., Whalen, P. J., Handwerger, K., Cannistraro, P. A., Wright, C. I., . . . Rauch, S. L. (2007). Dorsal anterior cingulate function in posttraumatic stress disorder. Journal of Traumatic Stress, 20(5), 701–712. Shinaver III,, C. S., Entwistle, P. C., & Söderqvist, S. (2014). Cogmed WM training: Reviewing the reviews. Applied Neuropsychology: Child, 3(3), 163–172. Shonkoff, J. P., Boyce, W. T., & McEwen, B. S. (2009). Neuroscience, molecular biology, and the childhood roots of health disparities: Building a new framework for health promotion and disease prevention. Journal of the American Medical Association, 301, 2252–2259. doi: 10.1001/jama.2009.754 Simons, R. L., Lei, M. K., Beach, S. R., Brody, G. H., Philibert, R. A., & Gibbons, F. X. (2011). Social environmental variation, plasticity genes, and aggression: Evidence for the differential susceptibility hypothesis. American Sociological Review, 76(6), 833. Sroufe, L. A. (1997). Emotional development: The organization of emotional life in the early years. Cambridge, UK: Cambridge University Press. Sroufe, L. A., & Rutter, M. (1984). The domain of developmental psychopathology. Child Development, 55(1), 17–29. Stadler, G., Oettingen, G., & Gollwitzer, P. M. (2009). Physical activity in women: Effects of a self-regulation intervention. American Journal of Preventive Medicine, 36(1), 29–34. Steinmetz, K. R. M., & Kensinger, E. A. (2013). The emotion-induced memory trade-off: More than an effect of overt attention? Memory & Cognition, 41(1), 69–81. Störmer, V. S., Passow, S., Biesenack, J., & Li, S. C. (2012). Dopaminergic and cholinergic modulations of visual-spatial attention and working memory: Insights from molecular genetic research and implications for adult cognitive development. Developmental Psychology, 48(3), 875. Strüber, N., Strüber, D., & Roth, G. (2014). Impact of early adversity on glucocorticoid regulation and later mental disorders. Neuroscience & Biobehavioral Reviews, 38, 17–37. Sturge-Apple, M. L., Davies, P. T., Cicchetti, D., & Cummings, E. M. (2009). The role of mothers’ and fathers’ adrenocortical reactivity in spillover between interparental conflict and parenting practices. Journal of Family Psychology, 23(2), 215. Sullivan, M. W., Carmody, D. P., & Lewis, M. (2010). How neglect and punitiveness influence emotion knowledge. Child Psychiatry & Human Development, 41(3), 285–298. Swain, J. E. (2011). Becoming a parent: Biobehavioral and brain science perspectives. Current Problems in Pediatric and Adolescent Health Care, 41(7), 192–196.

522

Self-Regulation and Developmental Psychopathology: Experiential Canalization of Brain and Behavior

Swain, J. E., Kim, P., & Ho, S. S. (2011). Neuroendocrinology of parental response to baby-cry. Journal of Neuroendocrinology, 23(11), 1036–1041. Swain, J. E., Lorberbaum, J. P., Kose, S., & Strathearn, L. (2007). Brain basis of early parent–infant interactions: Psychology, physiology, and in vivo functional neuroimaging studies. Journal of Child Psychology and Psychiatry, 48(3–4), 262–287. Tan, H. Y., Callicott, J. H., & Weinberger, D. R. (2009). Prefrontal cognitive systems in schizophrenia: towards human genetic brain mechanisms. Cognitive Neuropsychiatry, 14(4–5), 277–298. Terkel, J., & Rosenblatt, J. S. (1971). Aspects of nonhormonal maternal behavior in the rat. Hormones and Behavior, 2(2), 161–171. Thelen, E. (1995). Motor development: A new synthesis. American Psychologist, 50(2), 79–95. doi: 10.1037/0003–066X.50.2.79 Thompson, L. A., & Trevathan, W. R. (2008). Cortisol reactivity, maternal sensitivity, and learning in three-month-old infants. Infant Behavior & Development, 31(1), 92. Tunbridge, E. M., Weickert, C. S., Kleinman, J. E., Herman, M. M., Chen, J., Kolachana, B. S., . . . Weinberger, D. R. (2007). Catechol-omethyltransferase enzyme activity and protein expression in human prefrontal cortex across the postnatal lifespan. Cerebral Cortex, 17(5), 1206–1212. Ursache, A., & Raver, C. C. (2014). Trait and state anxiety: Relations to executive functioning in an at-risk sample. Cognition & Emotion, 28, 845–855., 1–11. Van den Bergh, B. R., & Marcoen, A. (2004). High antenatal maternal anxiety is related to ADHD symptoms, externalizing problems, and anxiety in 8 and 9-year-olds. Child Development, 75(4), 1085–1097. Van den Bergh, B. R., Van Calster, B., Smits, T., Van Huffel, S., & Lagae, L. (2007). Antenatal maternal anxiety is related to HPA-axis dysregulation and self-reported depressive symptoms in adolescence: A prospective study on the fetal origins of depressed mood. Neuropsychopharmacology, 33(3), 536–545. van Dongen Boomsma, M., Vollebregt, M. A., Buitelaar, J. K., & Slaats Willemse, D. (2014). Working memory training in young children with ADHD: A randomized placebo controlled trial. Journal of Child Psychology and Psychiatry, 55, 886–896. van Harmelen, A. L., van Tol, M. J., Demenescu, L. R., van der Wee, N. J., Veltman, D. J., Aleman, A., . . . Elzinga, B. M. (2013). Enhanced amygdala reactivity to emotional faces in adults reporting childhood emotional maltreatment. Social Cognitive and Affective Neuroscience, 8(4), 362–369. Van Veen, J. F., Van Vliet, I. M., DeRijk, R. H., Van Pelt, J., Mertens, B., & Zitman, F. G. (2008). Elevated alpha-amylase but not cortisol in generalized social anxiety disorder. Psychoneuroendocrinology, 33(10), 1313–1321. Vohs, K., & Baumeister, R. (Eds.) (2011). Handbook of self-regulation (2nd ed.). New York, NY: Guilford. Vuilleumier, P., & Pourtois, G. (2007). Distributed and interactive brain mechanisms during emotion face perception: evidence from functional neuroimaging. Neuropsychologia, 45(1), 174–194. Vuilleumier, P., Richardson, M. P., Armony, J. L., Driver, J., & Dolan, R. J. (2004). Distant influences of amygdala lesion on visual cortical activation during emotional face processing. Nature Neuroscience, 7(11), 1271–1278. Waddington, C. H. (1942). Canalization of development and the inheritance of acquired characters: Genes as the primary source of information shaping the development of the organism. Nature, 150, 563–565. Waltz, J. A., Knowlton, B. J., Holyoak, K. J., Boone, K. B., Mishkin, F. S., de Menezes Santos, M., . . . Miller, B. L. (1999). A system for relational reasoning in human prefrontal cortex. Psychological Science, 10(2), 119–125. doi: 10.1111/1467–9280.00118

Wang, L., LaBar, K. S., Smoski, M., Rosenthal, M. Z., Dolcos, F., Lynch, T. R., . . . McCarthy, G. (2008). Prefrontal mechanisms for executive control over emotional distraction are altered in major depression. Psychiatry Research: Neuroimaging, 163(2), 143–155. Waters, A. M., Mogg, K., Bradley, B. P., & Pine, D. S. (2011). Attention bias for emotional faces in childhood social phobia. Journal of Experimental Psychopathology, 2(4), 475–489. Weaver, I. C., Cervoni, N., Champagne, F. A., D’Alessio, A. C., Sharma, S., Seckl, J. R., . . . Meaney, M. J. (2004). Epigenetic programming by maternal behavior. Nature Neuroscience, 7(8), 847–854. Weinstock, M. (1997). Does prenatal stress impair coping and regulation of hypothalamic-pituitary-adrenal axis? Neuroscience & Biobehavioral Reviews, 21(1), 1–10. Welsh, M. C. Penington, B. F. Grossier, D. B. (1991). A normative developmental study of executive function: A window on prefrontal function. Developmental Neuropsychology, 7, 131–149. doi: 10.1080/ 87565649109540483 Werker, J. F., & Tees, R. C. (1986). The effect of multilingualism on phonetic perceptual flexibility. Applied Psycholinguistics, 7, 141–155. doi: 10.1017/S0142716400007360 Werner, H. (1948). Comparative psychology of mental development. Oxford, UK: Follett Publishing Company. Wiebe, S. A., Espy, K. A., & Charak, D. (2008). Using confirmatory factor analysis to understand executive control in preschool children: I. Latent structure. Developmental Psychology, 44(2), 575. Willcutt, E. G., Betjemann, R. S., McGrath, L. M., Chhabildas, N. A., Olson, R. K., DeFries, J. C., & Pennington, B. F. (2010). Etiology and neuropsychology of comorbidity between RD and ADHD: The case for multiple-deficit models. Cortex, 46(10), 1345–1361. Willoughby, M. T., Blair, C. B., Wirth, R. J., & Greenberg, M. (2010). The measurement of executive function at age 3 years: Psychometric properties and criterion validity of a new battery of tasks. Psychological Assessment, 22(2), 306. Willoughby, M. T., Blair, C. B., Wirth, R. J., & Greenberg, M. (2012). The measurement of executive function at age 5: Psychometric properties and relationship to academic achievement. Psychological Assessment, 24(1), 226. Willoughby, M. T., Pek, J., Blair, C. B., & Family Life Project Investigators. (2013). Measuring executive function in early childhood: A focus on maximal reliability and the derivation of short forms. Psychological Assessment, 25(2), 664–670. Willoughby, M., & Blair, C. (2011). Test–retest reliability of a new executive function battery for use in early childhood. Child Neuropsychology, 17(6), 564–579. Wilson, S., & Farran, D. (2012, March). Experimental evaluation of the tools of the mind curriculum. Paper presented at the Society for Research on Educational Effectiveness Spring 2012 Conference, Washington, DC. Yang, T. T., Simmons, A. N., Matthews, S. C., Tapert, S. F., Frank, G. K., Max, J. E., . . . Martin, P. (2010). Adolescents with major depression demonstrate increased amygdala activation. Journal of the American Academy of Child & Adolescent Psychiatry, 49(1), 42–51. Yerkes, R. M., & Dodson, J. D. (1908). The relation of strength of stimulus to rapidity of habit formation. Journal of Comparative Neurology and Psychology, 18, 459–482. Zelazo, P. D. (2006). The dimensional change card sort (DCCS): A method of assessing executive function in children. Nature Protocols, 1(1), 297–301. doi: 10.1038/nprot.2006.46 Zucker, R. A., Heitzeg, M. M., & Nigg, J. T. (2011). Parsing the undercontrol–disinhibition pathway to substance use disorders: A multilevel developmental problem. Child Development Perspectives, 5(4), 248–255.

CHAPTER 14

Anxiety Regulation: A Developmental Psychopathology Perspective ERIN B. TONE, CHERYL L. GARN, and DANIEL S. PINE

INTRODUCTION 523 DEFINITIONS OF TERMS 525 Emotion Regulation, Emotion, and Feelings 525 Fear and Anxiety 527 CLASSIFICATION OF ANXIETY DISORDERS ACROSS DEVELOPMENT 529 Evidence for Anxiety Disorder Specificity 531 Shared Characteristics Among the Anxiety Disorders 534 ANXIETY REGULATION 535 Cognitive Mechanisms of Anxiety Regulation 537

Neural Mechanisms of Anxiety Regulation 540 The Role of Context in Anxiety Regulation 545 Translational Implications of Research on Anxiety Regulation 546 CONCLUSIONS AND RECOMMENDATIONS FOR FUTURE WORK 546 REFERENCES 548

INTRODUCTION

functions and concepts, questions remain regarding how it is best conceptualized and the ways it develops, evolves, and manifests across contexts (Aldao, 2013; Thompson & Goodman, 2010; Thompson, Lewis, & Calkins, 2008). Indeed, some researchers and theorists even debate its fundamental validity as a distinct entity that can stand independently from other facets of emotion (e.g., Gross & Feldman, 2011; Kappas, 2011). Notably, although Thompson’s (1994) framework emphasized processes involved in modulating emotions in general, recently burgeoning literatures also examine the biological and psychological processes involved in the regulation of specific emotions, including happiness (Quoidbach et al., 2010; Urry et al., 2012), sadness (e.g., Cassano & Zeman, 2010; Sheppes & Meiran, 2007), anger (e.g., Kim & Deater-Deckard, 2011; Novin et al., 2011), and anxiety/fear (e.g., Bowers, Choi, & Ressler, 2012; Cisler et al., 2010; Hartley & Phelps, 2010). These emotion types share multiple regulatory commonalities; indeed, findings regarding shared mechanisms have driven interest in transdiagnostic approaches to treating psychopathology (Aldao & Nolen-Hoeksema, 2010; Kring & Sloan, 2010). However, across the life span there also appear to be distinctive cognitive and neural patterns associated with the modulation of individual emotions (e.g., Amstadter, 2008; Kovacs, Joormann, & Gotlib, 2008), which provide

It has been roughly 20 years since Ross A. Thompson (1994) published a highly influential monograph outlining the challenges associated with defining emotion regulation and embedding the construct accurately in the broader matrix of knowledge about social behavior and cognition. This work synthesized and expanded on converging strands of research about a topic that had received a growing amount of attention in academic journals and books in the preceding years, including a widely cited special section of Developmental Psychology on development of emotion regulation that Kenneth A. Dodge edited in 1989. Around the same time, Gross and Muñoz (1995) clearly laid out various ways emotion regulation and dysregulation contribute to changes in mental health. During the intervening decades since these works appeared, the phrase emotion regulation has become a central component in our definitions of most psychopathology (Dillon, Deveney, & Pizzagalli, 2011), and researchers have made great strides in clarifying the processes by which we modulate our emotional experiences across the life span in both adaptive and maladaptive ways (e.g., Tamir, 2011; Thompson, 2011). Not surprisingly, however, given that the term emotion regulation comprises a complex, nuanced, and dynamically interactive range of 523

524

Anxiety Regulation

striking evidence of the complexity, flexibility, and nuanced functioning of our emotion regulation systems. Together, these patterns of overlapping and discrete functions serve to support a broad range of responses to emotional experience and expression. Many recent reviews cogently summarize the literature on the umbrella construct of emotion regulation as it develops and changes across the life span (e.g., Denham, 2010; Jacob et al., 2011; Nolen-Hoeksema, 2012; Webb, Miles, & Sheeran, 2012). These works approach the topic from behavioral, cognitive, and psychophysiological perspectives and reflect in particular our growing understanding of the ways the brain mediates the experience and modulation of emotions. To extend and complement this body of research, the present chapter aims to describe and critically review the literature that focuses specifically on the regulation of anxiety from infancy through older adulthood. We use the developmental psychopathology perspective, which emphasizes several core tenets, as a guiding framework for this chapter (Cicchetti & Cohen, 1995). Our review of the literature thus carries within it a number of implicit assumptions about the ways psychopathology emerges and evolves over the life span. First among these is the idea that multiple factors interact in a dynamic, transactional fashion to affect the ways a disorder might initially manifest, how it changes over time, and whether and how it resolves. Consistent with this idea, clinically significant anxiety can emerge as an endpoint of an array of developmental pathways, in accordance with the principle of equifinality (Ollendick & Hirshfeld-Becker, 2002). For example, varied combinations of genetic and environmental risk factors can precipitate the development of a particular anxiety disorder (McGrath et al., 2012). However, in keeping with the related principle of multifinality, none of those pathways is necessarily guaranteed to result in that condition, or, for that matter, in any anxiety disorder. Indeed a given etiological factor (e.g., genetic vulnerability) can produce any of several different effects depending on the system or context in which it operates. Findings that genetic characteristics, such as serotonin transporter polymorphisms, interact with stressful life events to differentially predict anxious symptoms (along with depressive and other symptom types) nicely illustrate this principle (Caspi et al., 2003; Petersen et al., 2012). Thus, the probability that an anxiety disorder will emerge at a given point in development varies depending on the evolving interplay among vulnerabilities and protective factors for any particular child, adolescent, or adult. Based on the idea that an accurate understanding of dysfunction necessitates a comparable understanding of

normal or successfully adaptive developmental paths, an additional core assumption of the developmental psychopathology perspective is that typical and atypical development must be studied in concert. This assumption is particularly salient to anxiety, which constitutes an integral, adaptive, and normative part of human existence throughout the life span (Barlow, 2002). Indeed, the absence of anxiety, a goal that clinically anxious adults in one qualitative study articulated with some regularity, has been aptly conceptualized as mythical normativity (Lloyd & Moreau, 2011). It is only under certain circumstances that anxiety shifts from a normative to a dysfunctional response; considerable evidence suggests that convergence of salient biological and environmental factors leads to neural hypersensitivity that manifests behaviorally as hypervigilance and overresponsiveness to stimuli perceived as threats, or pathological anxiety (Rosen & Schulkin, 1998). Precisely how and why this transition occurs in some individuals, but not others, is a focus of considerable research attention; if we are to obtain accurate answers to these questions, it is critical that we comprehensively map the terrain that surrounds the transition. An additional core tenet of the developmental psychopathology perspective informing this chapter is that the developmental context must be considered when evaluating whether or not a disorder is emerging or subsiding. Symptoms vary in their meaning and import at different points in development. A sizable proportion of children between the ages of 4 and 12 years, for example, report nighttime fears (i.e., realistic or unrealistic fears that something bad will happen during the night), with prevalence reaching its peak between seven and nine years (Muris et al., 2001). In contrast, such fears are less common among adolescents (Gordon et al., 2007). Nighttime fears thus appear more likely to be pathological, at least in terms of deviance from typical developmental patterns, at some points than at others. Similarly, whereas being separated from primary caregivers commonly provokes fear among infants and toddlers (Thompson & Limber, 1990), comparable separation-related distress is less typical among older children and adolescents (Compton, Nelson, & March, 2000). Its occurrence among older youths may thus indicate the presence of pathology. Taking these tenets as a guiding structure, we aim in this chapter to examine the mechanisms and methods by which humans regulate anxiety over the course of the life span. In the first section, we define salient terms and concepts that are at the heart of our review. We clarify the models we rely on to describe emotion and anxiety regulation, and

Definitions of Terms

we then offer a short overview of common manifestations of anxiety across development. In the second section, we shift focus to the regulation of anxiety and fear per se. In particular, we examine whether and how cognitive and neural mechanisms underlying the regulation of anxiety vary across development. The third section of this chapter turns to the contexts, both intrapersonal and interpersonal, that may facilitate or inhibit the use of different regulatory mechanisms and strategies among individuals experiencing acute or chronic anxiety across the life span.

DEFINITIONS OF TERMS Emotion Regulation, Emotion, and Feelings The human capacity to modulate emotional experience and expression in intentional ways has been a topic of scholarly interest for centuries. Physicians, philosophers, and psychologists have long acknowledged both the power of emotions to both elevate and disrupt human experience and the many ways people try variously to rein in or amplify these internal states. In a succinct review of influential ideas on these topics, Cicchetti, Ackerman, and Izard (1995) traced the path along which our understanding of emotion and emotional control has evolved from the time of Hippocrates and Aristotle to the late twentieth century. In particular, they highlighted two core themes that pervade thinking on these topics throughout history: first, that uncontrolled emotion is a key player in the development of psychopathology; and second, that reason is necessary to keep the emotions under control. As early as the mid- to late 1800s and the early 1900s, while some continued to explore these themes by delving into the relationship between emotion and the broad constructs of reason, sensibility, and will (e.g., Bain, 1859; Jastrow, 1915; Moore, 1852), other researchers and theorists moved away from treating these constructs as unitary forces that one could willfully exercise to keep emotions in line. These scholars instead, or concomitantly, examined specific cognitive mechanisms that might participate in or facilitate the regulation of emotion. In his book on principles of mental physiology, for instance, the prominent British physiologist William B. Carpenter (1876) speculated about the fundamental role that attention plays in the regulatory process, noting that until our attention is drawn to our emotional reactions, we are unable to intentionally direct those reactions; indeed, we may not even be aware of whether and how they are changing. Motoric inhibition of emotional displays, if not experiences, also received mention in

525

the literature (e.g., Bain, 1859), as did executive functions, such as cognitive or mental inhibition (e.g., Clouston & Folson, 1884). Early in the twentieth century, references to emotional regulation or control also began to appear in the educational literature, reflecting a surging interest in expanding the role of schools in the processes of behavioral and characterological development, in concert with their ongoing efforts to encourage intellectual growth and achievement (Henry, 1917). Further, educators were acknowledging with increasing frequency the impact of emotional factors on academic performance, which appeared to heighten interest in helping children acquire self-regulatory skills (e.g., Arthur, 1922; Johnson, 1928). This theme recurred in the education and child-rearing literatures through the mid-twentieth century (e.g., Arlitt, 1938; Stendler, 1950), and it continued to translate more broadly, particularly to developmental psychology outlets, in the latter decades of the twentieth century. By the late 1980s and early 1990’s, the term emotion regulation appeared regularly in the general psychological and psychiatric literatures; recent searches of online databases (PsycINFO, PubMed) yield nearly 2,000 publications since 1980 that list it as a key search term or phrase, and thousands of additional articles and chapters reference the topic. In these many publications, emotion regulation has been used to label a diverse array of clinical and psychological phenomena, some very broad and others more narrow in scope. For example, as Grolnick, McMenamy, and Kurowski (2006) summarized, some definitions focus on general changes in emotions, which involve altering or maintaining patterns of emotional valence, intensity, or course in the context of interpersonal transactions or salient cultural norms (Campos, Mumme, Kermoian & Campos, 1994; Cole, Martin, & Dennis, 2004, Fox, 1994). Others, however, maintain that emotion regulation encompasses a broader set of processes; Campos, Frankel, and Camras (2004), for example, defined it as “the modification of any process in the system that generates emotion or its manifestation in behavior” (p. 380). For the purposes of this chapter, we rely primarily on Thompson’s (1994) definition of emotion regulation. More precisely, we use the term specifically to describe “processes responsible for monitoring, evaluating, and modifying emotional reactions . . . to accomplish one’s goals” (pp. 27–28). Of note, our chosen definition identifies emotional reactions as the target of regulatory processes. These reactions are multifaceted and may manifest in one or more different ways, both within and out of awareness. Indeed, as Jacob and colleagues (2011) underscore,

526

Anxiety Regulation

emotion regulation involves an array of dynamic processes; while inhibition or suppression are the processes most often intuitively linked to emotion regulation, other regulatory processes, including exaggeration, substitution, avoidance, or neutralization of affective responses, warrant attention. We consider affectively valenced states that one experiences subjectively and can describe reliably to be feelings, consistent with Damasio’s writings (Damasio, 2001; Damasio et al., 2000). Feelings may or may not be associated with emotions, which we define as patterns of neural response (along with their broader physiological or information processing correlates) associated with the experience of rewards (any stimuli that animals or humans work to procure) or punishments (any stimuli that animals or humans make effort to avoid; Rolls, 1999). Changes in emotion are thus necessarily reflected in neurophysiological and, in many cases, information-processing measures; they can be, but do not have to be, reflected in measures of subjectively reported experience. As such, the relationship between feeling and emotion is unclear, and will remain so until a suitable theory of consciousness exists to define subjective feeling states in the same types of neuroscience-based terminology used currently to define emotions. Nevertheless, theorists and researchers have articulated ways the two may interact. For example, scholars who approach the topic from developmental, neuroscience, and translational perspectives contend that feelings act to influence emotional states, both by monitoring variations that occur at the level of emotions and by feeding information back to higher cognitive processes, which in turn modulate the ongoing stream of emotion (e.g., Damasio, 2000; Hoeksma, Oosterlaan, & Schipper, 2004; Panksepp, 1998; Scherer, 2000). Carefully distinguishing feelings from emotions is important; although the terms are often used interchangeably in lay language, the psychological/psychiatric literatures specify subtle distinctions between the two and use them in slightly different contexts. In particular, feelings are at the core of many currently conceptualized psychiatric disorders included in the fifth edition of the American Psychiatric Association’s Diagnostic and Statistical Manual for Mental Disorders (DSM-5; APA, 2013) in that subjectively reported changes in emotional states (e.g., depressed mood most of the day, nearly every day) or behaviors (e.g., distractibility) form the basis for a variety of diagnostic criteria. Objectively measured behaviors that are putatively associated with emotional states (e.g., pressured speech; avoidance) also constitute a handful of criteria. Notably, no current diagnostic criteria incorporate data from directly measured changes in brain states, although research under

the auspices of the National Institute of Mental Health’s Research Domain Criteria (RDoC) initiative aims to reorganize our understanding of psychopathology and its diagnostic classification according to a neuroscience-based framework (Insel & Wang, 2010; Morris & Cuthbert, 2012). This initiative, if successful, will facilitate the use of biomarkers in addition to or in place of phenomenological observations for diagnostic purposes; it already played an influential role in shaping the fifth edition of the DSM (Kupfer & Regier, 2011). Until the literature better supports a biomarker-driven method for identifying and classifying psychopathology, however, a continued focus on feelings in our efforts to understand and distinguish among most psychiatric diagnoses will be necessary. This reliance on subjective data is problematic at a number of levels. First, correlations between subjective reports and performance on objective measures tapping the same domain are often relatively weak, even for cognitive and behavioral functions that observers can readily assess with high reliability (e.g., Burdick, Endick, & Goldberg, 2005; Gosling, John, Craik, & Robins, 1998; Svendsen et al., 2012). It thus remains unclear how well and how consistently self-reports regarding internal, affective states, which are currently hard to evaluate objectively, describe lived experience. Notably, some evidence suggests that even very young children are able to provide valid and meaningful information regarding their feelings, based on findings of convergence between 3- and 6-year-old youths’ self-reports and objective coding of their facial expressions during affectively evocative tasks (Durbin, 2010). However, the magnitude of convergence between these measures was small to moderate across emotions, with self-report accounting for no more than 15% of the variance in objective coding scores. Second, reliance on subjective reports of feelings in our conceptualizations of mental disorders introduces marked challenges into the conduct of translational work that integrates research across human and nonhuman species. Even though many mammalian species show evidence of consistencies and continuities in terms of experienced emotion (Davis & Whalen, 2001; LeDoux, 1994, 1998, 2000), it is difficult to measure the presence of feelings in nonhuman organisms. This difficulty, at least in part, reflects our almost exclusive reliance on verbal reports of feelings, which we cannot obtain from nonhuman species. It is further confounded by the lack of an accepted model of consciousness and subjective reports, which could support cross-species investigations of feelings. Panksepp (2011; 2012), however, argued that the current approach is unnecessarily limiting; instead, he contends that by

Definitions of Terms

directly stimulating brain networks that have consistently been shown to mediate emotional behavior, it is possible to access at least some types of emotional experience without need for verbal report. His ideas suggest a path out of the current impasse, in which nonhuman species’ lack of verbal language appears to prohibit insight into their mental and emotional processes. However, until empirical data point us toward a broadly accepted and widely used, reliable, and valid approach to accessing mental and emotional states in the absence of language, translational studies will continue to grapple with the interface between feelings and emotions, which remains imprecisely specified across species. Our understanding of regulatory mechanisms is similarly complicated by the challenges inherent in identifying boundaries between emotions and feelings. Indeed, in light of our emphasis on the distinction between emotions and feelings, one could argue that in addition to attention to emotion regulation, a parallel focus on feeling regulation is warranted. However, despite the availability of cogent hypotheses about the dynamic interplay between feelings and emotions (e.g., Hoeksma et al., 2004), the term emotion regulation is broadly operationally defined, so that it encompasses the processes that serve to modulate both constructs. Thus, in the interest of simplicity and consistency with much of the literature that we review, and in recognition of the inextricable links between emotions and feelings, we have adopted a comparably broad and inclusive use of the term emotion regulation. Fear and Anxiety Just as we tend to conflate emotions and feelings in popular discourse, we also commonly treat fear and anxiety as indistinguishable. Notably, however, as far back as 45 BC, Cicero wrote about the divergence between anxietas, a state of general fearfulness, and angor, a passing state of fear provoked by a specific stimulus (Cicero, 1927). Consistent with Cicero’s ideas, the scientific literature has come to recognize anxiety and fear as representing different constructs (e.g., Sylvers, Lilienfeld, & LaPrairie, 2011), and in this chapter, accordingly, we treat them distinctly. For our purposes, the term anxiety denotes a frequently enduring phenomenon that may arise in response to a range of nonspecific, overt, or covert stimuli that an individual perceives as representing potential threat (Dias, Banerjee, Goodman, & Ressler, 2013). The perception of possible impending threat is commonly accompanied by feelings of apprehension, as well as increased arousal and vigilance. The term fear, in contrast, represents a transient state

527

that typically evokes defensive responses in the presence of specific overt stimuli and subsides when those stimuli are gone (Dias et al., 2013). Thus, whereas fear is tied to relatively clear, unambiguous threats, the dangers that give rise to anxiety are less clearly demarcated. Such a conceptualization of anxiety and fear as distinct constructs finds support in empirical work on the structure of mental disorders. Krueger (1999) analyzed comorbidity patterns in a large epidemiological sample of adults and found that the best fitting factor structure included distinct, although highly correlated, first-order anxious-misery and fear factors that loaded on a higher order internalizing factor. This pattern of findings has been widely replicated in adult samples (e.g., Slade & Watson, 2006; Vollebergh et al., 2001; Watson, 2005). In a recent effort to extend Krueger’s (1999) work to an adolescent and young adult sample, however, Beesdo-Baum and colleagues (2009) found a slightly different factor structure to yield the best fit. Like Krueger’s (1999), their data were consistent with the presence of anxious-misery and fear first-order factors; notably, however, the correlation between the two factors was much weaker, and a model including the higher order internalizing factor yielded a weak fit. This structural variation raises questions about the developmental stability of anxiety and fear constructs and points to the need for more research in this area, ideally with developmentally sensitive methods that permit examination of within-subject changes over time. Neurobiological studies in both nonhuman and human species also point to clear distinctions between anxiety and fear. As Sylvers and colleagues (2011) summarized, research on varied rodent species, as well as on nonhuman primates, converges to suggest that discrete neural structures play key roles in fear and anxiety. Whereas the lateral and central nuclei of the amygdala (LA and CeA) appear to be critical players in the acquisition of fear and the emergence and expression of fear-conditioned responses (Kalin, Shelton, & Davidson, 2004; Ledoux et al., 1988, 1990; Wilensky, Schafe, Kristensen, & Ledoux, 2006), additional regions, including the bed nucleus of the stria terminalis, the septum, and hippocampus may figure in the emergence of anxious responses (e.g., Degroot & Nomikos, 2005; Edinger & Frye, 2006; Gray & McNaughton, 2000). Of note, despite their distinctions, anxiety and fear share a number of common features, which fuels the common tendency to treat them as interchangeable. Both, for example, can be categorized as either emotions or feeling states that humans experience when they encounter stimuli that they perceive as threatening or as capable of producing harm and that elicit avoidance. Although both

528

Anxiety Regulation

can manifest in atypical or dysfunctional ways, neither necessarily connotes a pathological state, as we discuss below in further detail in our overview of anxiety disorders. In addition, both anxiety and fear elicit a number of common physiological responses including autonomic arousal (Cook, Hawk, Davis, & Stevenson, 1991), as well as activation from a number of overlapping neural structures that include the amygdala and areas within the medial prefrontal cortex in human and nonhuman species (Blair, 2003). Regulation of Fear and Anxiety When applied to fear and anxiety, our definition of emotion regulation demands attention to two distinct, but related, processes. First, anxiety/fear regulation is initiated by stimuli and circumstances that elicit fearful or anxious reactions, due to their real or perceived capacity to harm the organism. Second, the individual enacts a range of behavioral, cognitive, and neural response that serve to monitor, evaluate, and modify the elicited fearful or anxious reactions that the stimulus elicits. The sequence as a whole thus necessitates a triggering stimulus or circumstance and a regulatory response to it. Taken together with the increasing tendency in the developmental and general psychopathology literatures to conceptualize anxiety disorders as centrally characterized by maladaptive emotion regulation (e.g., Cisler, Olatunji, Feldner, & Forsyth, 2010; Esbjørn, Bender, Reinholdt-Dunne, Munck, & Ollendick, 2012; Jacob, Thomassin, Morelen, & Suveg, 2011; Rodebaugh & Heimberg, 2008), our definition suggests that anxious pathology lies less in the circumstances that elicit emotion or the magnitude of an initial emotional reaction and more in the degree to which an individual can engage resources to regulate this reaction once it emerges. Within the course of the regulatory process, however, there are also nuances that warrant attention. Davidson (1998) laid some of these out clearly in his description of a model of emotion regulation that emphasizes affective chronometry, referring to the time course of emotional responding. According to this model, it is important to view the processing of emotionally evocative stimuli in the context of their temporal dynamics rather than as a unitary block. In other words, the processes that are engaged as an affective reaction rises to its peak may differ from those engaged as the reaction subsides and disruption of processes during either period may contribute to the emergence of different forms of pathology. Further, the trajectory of the rise and fall of a response is likely to vary considerably both across individuals and across biological and cognitive systems within an individual. We thus risk

inadequately characterizing a regulatory response if we fail to consider its temporal course along with its content. A range of such temporal features, including the continuity of a feeling or emotion from one moment to the next (Koval & Kuppens, 2012) and the process of transition within an emotional response (Filipowicz, Barsade, & Melwani, 2011), have received increasing research attention since Davidson’s (1998) article first appeared. From an affective chronometry perspective then, conceptual distinctions between emotional reactions and emotion regulation may be arbitrary. Instead, it may be more useful to consider the two constructs as representing conceptually unified, but temporally distinct phases within an individual’s response to an emotional stimulus at behavioral, cognitive, and neural levels. Such a perspective gives rise to the possibility that some disorders may relate to perturbations that occur early in the course of a response, while others may relate to perturbations that are evident later during the processing of emotional stimuli. Davidson (1998) reviewed data, for example, that suggest that differences between anxiety and depression, which share many features and often emerge comorbidly (Angold & Costello, 1993; Brown, Campbell, Lehman, Grisham, & Mancill, 2001), may, in part, reflect variability in the functioning of neural regions that participate in regulatory processes (e.g., prefrontal cortices) that leads to differential influences on the dynamic unfolding of the emotion regulation sequence. For example, in a later publication focused explicitly on anxiety, Davidson (2002) underscored the important roles that ventromedial and orbital regions of the prefrontal cortex play in modulating the time course of responses in subcortical neural structures such as the amygdala, which converging data have shown to function atypically in the context of anxiety disorders. Taken together, both content- and process-oriented perspectives provide a broad range of intersecting hypotheses about how abnormalities in either emotional reactions or emotion regulation, as they are measured in the laboratory, may relate to the emergence and maintenance of developmental psychopathologies, including anxiety disorders. Research over the past few decades has made advances in testing these hypotheses and we review, in subsequent sections of this chapter, the findings that are emerging within this growing literature and their interpretation through a developmental psychopathology lens. Many questions in this domain, however, remain unanswered and we attempt to point out gaps in the existing body of work on the regulation of anxiety that merit further attention, particularly in the context of anxiety that is severe or pervasive enough to qualify as pathological.

Classification of Anxiety Disorders Across Development

Pathological Anxiety As a group, anxiety disorders represent the most severe and extreme manifestations of anxiety. These conditions are common throughout development, with recent lifetime prevalence estimates for any anxiety disorder in the United States appearing as high as 32.4% for adolescents to 33.7% for adults (Kessler, Petukhova, Sampson, Zaslavsky, & Wittchen, 2012). The same study also found twelve-month prevalence rates for individual anxiety disorders to be high in adolescent and adult populations, with, for example, up to 7.4% of individuals meeting criteria for social phobia and 12.1% meeting criteria for a specific phobia in a given year. Prevalence rates for 12 months among children under 12 also appear high; a review of epidemiological research using both DSM-III-R and DSM-IV criteria found 12-month estimates of the presence of any anxiety disorder to range from a low of 3.05% to a high of 20.2%, with a median of 4.02% (Cartwright-Hatton, McNicol, & Doubleday, 2006). Evidence also suggests that at least some anxiety disorders are stable over time; one study, for example, found 56.7% of adolescents and adults diagnosed with social anxiety disorder to continue to meet criteria 10 years later (Beesdo-Baum et al., 2012). Stability rates appear particularly high in individuals with comorbid anxiety and depressive disorders (Fichter, Quadflieg, Fischer, & Kohlboeck, 2010). Surprisingly, however, in light of their frequent occurrence and persistence, anxiety disorders received relatively little research attention until the past few decades, perhaps because they have historically been perceived as less severe than other forms of psychopathology. Research since the late 1980s indicates decisively that perceptions of anxiety as minimally impairing and rarely associated with adverse outcomes are inaccurate. Anxiety disorders are associated with a variety of adverse outcomes, including development of other psychiatric problems such as major depression and suicidality (Fehm, Beesdo, Jacobi, & Fiedler, 2008; Fichter et al., 2010; Pine, Cohen, Gurley, Brook, & Ma, 1998). Indeed, most adults who suffer from a mood or anxiety disorder will have first developed signs of their illness, manifest as an anxiety disorder, during childhood or adolescence (e.g., Costello et al., 2002; Pine et al., 1998; Ramsawh, Weisberg, Dyck, Stout, & Keller, 2011). Anxiety disorders are also costly to society in terms of both financial and social consequences. Medical costs and utilization rates associated with this diagnostic category are high (Berger et al., 2011; Martin & Leslie, 2003), especially when treatment is concordant with practice guidelines (Prins et al., 2011). School dropout rates are also elevated among adolescents who meet criteria for anxiety disorders

529

(Leach & Butterworth, 2012), and affected adults have also been found to miss work more frequently and to show decreased output when they do attend work (Erickson et al., 2009; Revicki et al., 2012). Finally, anxiety disorders are a source of impairment in multiple domains for children, particularly those who meet criteria for more than one diagnosis (Alfano, 2012; Klein & Pine, 2002; Langley, Bergman, McCracken, & Piacentini, 2004). Research has found evidence, for example, that sleep-related problems (Alfano, Ginsburg, Kingery, & Newman, 2007), interpersonal difficulties (Weeks, Coplan, & Kingsbury, 2009), and school absenteeism (Dube & Orpinas, 2009) are associated with the presence of anxiety, particularly when it manifests at clinically significant levels. In light of their negative impact on individuals across the course of development, it is critically important that we understand what the anxiety disorders are, how they emerge and evolve across the course of development, and how they converge with and diverge from normal manifestations of anxiety.

CLASSIFICATION OF ANXIETY DISORDERS ACROSS DEVELOPMENT In May 2013, the American Psychiatric Association published the fifth edition of its Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5; American Psychiatric Association, 2013) with the aim of updating the diagnostic nosology based on changes in the research base since the fourth edition of the manual (DSM-IV; APA, 1994) was published in the mid-1990s and revised in 2000 (DSM-IV-TR; APA, 2000). The authors’ intensive attention to the integration of novel research findings with solidly supported older data resulted in a number of changes to diagnostic criteria and, in some cases, disorder definitions (American Psychiatric Association, 2013). However, the magnitude of change in the anxiety disorders is relatively minimal, particularly when compared with the changes that occurred more than 30 years ago with the publication of DSM-III (APA, 1980), where anxiety classification was radically transformed. Given that the DSM-5 was not yet published when this chapter was in preparation, the research data on anxiety disorders that we cite come from studies that used the nosology of the DSM-IV or earlier editions of the DSM. These editions recognized eight distinct anxiety disorders: separation anxiety disorder (SAD), social phobia, generalized anxiety disorder (GAD), specific phobia, posttraumatic stress disorder (PTSD), acute stress disorder (ASD), panic disorder/agoraphobia,

530

Anxiety Regulation

and obsessive-compulsive disorder (OCD), along with a less specific diagnosis, anxiety disorder not otherwise specified (NOS), for clinically significant conditions that did not fully meet criteria for other disorders. With the transition to the fifth edition of the DSM, however, this grouping changed in two key ways. First, PTSD and ASD were moved to a separate section on trauma- and stressor-related disorders; this section also includes two forms of attachment disorder that were both classified in DSM-IV under the rubric of reactive attachment disorder (RAD). Second, OCD was shifted to a section centered on obsessive-compulsive and related disorders (DSM-5; APA, 2013). These changes reflect ample evidence, described below, that the longitudinal outcomes, comorbidities, familial aggregation patterns, and biological correlates of these conditions are distinct from those of other disorders traditionally considered to be anxiety disorders. In this chapter, accordingly, we focus only on conditions classified in the DSM-5 as anxiety disorders; we further limit our focus to three that commonly aggregate and are often treated as a cluster (RUPP, 2003), especially in the developmental literature: SAD, social phobia/social anxiety disorder, and GAD. Notably, although social anxiety disorder and GAD were previously classified as anxiety disorders in DSM-IV, SAD was identified as a child-specific disorder along with such conditions as attention-deficit/hyperactivity disorder (ADHD). The section devoted exclusively to childhood disorders was eliminated from DSM-5 due to evidence that it encompassed conditions that manifested in adults as well as youths and thus set artificial age-related boundaries. In place of a section devoted explicitly to children and adolescents, a developmentally influenced perspective helped shape the diagnostic criteria and considerations for all of the anxiety disorders. Nevertheless, SAD was placed in the anxiety disorders section for DSM-5, as opposed to the neurodevelopmental disorders, and readers were informed that it was suitable to apply the diagnosis of SAD to adults. Although the importance of a developmental mindset is overtly recognized with the publication of DSM-5, developmental considerations have been important, if subtler, components of the DSM diagnostic process for the past two decades. With the DSM-IV, the American Psychiatric Association adopted a nosological convention in which identical diagnostic criteria were to be applied to individuals across the life span, in the absence of strong data that supported using a developmentally tailored approach. Thus, as noted already, in the DSM-IV, all anxiety disorders but one (SAD—which was conceptualized as occurring exclusively in childhood) were diagnosed

using the same criteria for children, adolescents, and adults. In addition, although the GAD criterion items were identical for youths and adults, juveniles had to show fewer symptoms than adults to cross the threshold for the disorder. In the DSM-5, however, based on data suggesting that some elements of this developmental tailoring lacked empirical support, SAD is recognized as a condition that can manifest after the age of 18 (Marnane & Silove, 2013). To meet criteria for any anxiety disorder, individuals must exhibit a specific set of symptoms; in the third and fourth editions of the DSM, they were also required to show signs of either functional impairment or significant distress. The latter requirement, commonly termed the impairment/distress criterion, was initially added to diagnoses that had first appeared in the DSM-III in 1980, with the aim of ensuring that individuals who experienced normal fluctuations in psychiatric symptoms were not inappropriately diagnosed with and treated for categorically distinct disorders. However, despite the intuitive appeal of this conceptual distinction, it has proved difficult to apply in practice. First, many core symptoms of anxiety disorders consist of variants of distress, which Rapee and colleagues (2012) define as “a personal and subjective sense of malaise and negativity” (p. 455). Second, it is simply easier to disentangle symptoms and the impairment associated with them in some conditions than in others (Rapee et al., 2012). For example, even highly anxious individuals can function adequately or even well in many domains (Keyes, 2005), unlike most individuals with some other forms of severe psychopathology, such as psychosis or dementia, particularly at advanced stages. Indeed, given the long-noted lack of specific functional impairment criteria for the anxiety disorders (e.g., Beidel, Silverman, & Hammond-Laurence, 1996), it is remarkably challenging to determine reliably where the line is between adequate and problematic adaptation to a given environment. This challenge is compounded by the fact that diagnostic decisions about the presence or absence of anxiety disorders, especially those made regarding youths, must often be made against the background of normative developmental changes in behavior, cognition, and affect. As we noted earlier, some anxious thoughts, feelings, and behaviors are normal at different stages of development—fears of negative evaluation from others, for example, become increasingly common as children transition into adolescence (Vasey, Crnic, & Carter, 1994). Such fears, in contrast, are relatively unusual among three year olds (Spence, Rapee, McDonald, & Ingram, 2001); thus, the presence of negative evaluation fears might be considered normal in adolescents but atypical in young children.

Classification of Anxiety Disorders Across Development

Because of this narrow and constantly moving line between typical and atypical anxiety across the life span, as well as the fact that symptoms that lead to impairment at one age may not do so at others, the developmental psychopathology perspective’s emphasis on considering normal and deviant processes in conjunction is particularly appropriate to the study of anxiety disorders. In recognition of these and other related issues, the authors of the DSM-5 considered removing the impairment/distress criterion from individual disorders, thus shifting from a view in which impairment/distress are integral features of different conditions to one in which they are consequences of psychopathology. Ultimately, however, they retained this criterion for most of the anxiety disorders, while electing to provide on the DSM-5 Web site (http://www.psychiatry.org/practice/dsm/dsm5/online -assessment-measures#Disorder) dimensional disorderspecific scales that can be used to describe severity. The subtlety of the distress/impairment threshold for anxiety disorders clearly poses problems for clinicians, who are charged with distinguishing between normal range and clinically significant conditions. It also creates issues in the epidemiological literature, in that it complicates the determination of population-level prevalence rates. This set of problems is particularly salient for anxiety-related conditions; indeed, studies that set different impairment thresholds tend to find dramatically different prevalence rates for anxiety disorders, for which many symptoms are internal and accessible only by self-report. In contrast, the threshold for impairment has a much smaller influence on rates for other psychiatric conditions, such as depression or conduct disorder, whose symptoms may be more readily observable to others (Costello, Mustillo, Erkanli, Keeler, & Angold, 2003; Shaffer et al., 1996). The challenges associated with differentiating normal from abnormal anxiety, in terms of both impairment and distress, reflect the limitations of our current approach to identifying and categorizing most psychopathology primarily via clinical evaluation (e.g., interview, self-report). Alternate, more objective approaches, such as assessments of physiology, genetics, or brain function have generated considerable excitement among researchers and enjoy some empirical support (e.g., Siegle et al., 2012). They have only been used to date clinically, however, for the diagnosis of mental disorders that result from underlying nonpsychiatric medical illnesses (e.g., genetic disorders such as Fragile X syndrome, disorders such as AIDS that can be associated with dementia). Thus until we can move beyond a reliance on clinical evaluation and incorporate a broader range of converging measures in our diagnostic

531

decisions, consistent with the NIMH’s RDoC initiative (Insel & Wang, 2010; Morris & Cuthbert, 2012), our definitions of most mental disorders must be regarded as preliminary. Identification of reliable and effective biological predictors of mental illness may also contribute to the already active shift away from an exclusively categorical nosology for mental disorders and toward the incorporation of dimensional models that help more precisely capture the multifaceted and complex nature of most psychiatric illness (Morris & Cuthbert, 2012). Indeed, by clarifying the ways observable behaviors and neurobiological characteristics relate to the presence or absence of psychopathology, we may ultimately alter the criteria we use to distinguish health from disease broadly and normal from abnormal anxiety more specifically. Moreover, some psychiatric conditions, including certain anxiety disorders, may eventually be recognized as extremes on continua with normality, just as some forms of hypertension are thought to represent the extreme end of the blood pressure continuum. Indeed, efforts have been under way for several years to develop dimensional measures for use in diagnostic assessments for anxiety (e.g., LeBeau et al., 2012; Niles, Lebeau, Liao, Glenn, & Craske, 2012) disorders and many of these are now accessible on the DSM-5 Web site (http://www.psychiatry.org/practice/dsm/dsm5/online -assessment-measures). Evidence for Anxiety Disorder Specificity The conditions classified as anxiety disorders, both historically and currently, are broadly related, in that each has anxiety of some sort as a core feature. Moreover, virtually all of the individual anxiety disorders have been linked in longitudinal studies to later onset of another anxiety disorder. However, these conditions do not form a neatly unified group; indeed, as noted earlier, in recognition of their marked distinctions, several (PTSD, acute stress disorder, OCD) were removed from the anxiety disorders cluster with the transition to the DSM-5. These decisions were based in part on five sets of standards that researchers have used to distinguish among the anxiety disorders; notably, relatively few data are available that use these criteria to validate specific anxiety disorders in children and adolescents and most of the research base is thus focused on adults. First, both the DSM-IV and the DSM-5 define anxietyrelated conditions as distinct by setting forth different sets of diagnostic criteria for each disorder. These criterion sets identify symptoms that most frequently cluster in patients

532

Anxiety Regulation

TABLE 14.1 Social Anxiety Disorder, Generalized Anxiety Disorder, and Separation Anxiety Disorder: Changes From DSM-IV-TR to DSM-5 DSM-IV-TR Social phobia

DSM-5 Social anxiety disorder

Cardinal symptoms

Marked and persistent fear of one or more social or performance situations involving potential scrutiny or exposure to unfamiliar people.

Marked fear or anxiety about one or more social situations involving potential scrutiny.

Differences between Editions

Adults (but not children) must recognize that their fear of social situations is excessive or unreasonable. Recognized two forms of the disorder—isolated fear of public speaking versus anxiety in a range of social situations.

Recognition that fear is unreasonable or excessive is no longer required for individuals aged 18 years or older. Anxiety must be out of proportion to actual risks, after cultural contextual factors are taken into account. Most cases treated as generalized, with specifiers added to cases restricted to public speaking or involving selective mutism.

Disorder name Cardinal symptoms

Generalized anxiety disorder Excessive and hard-to-control anxiety and worry about a number of events or activities.

Generalized anxiety disorder Excessive and hard-to-control anxiety and worry about a number of events or activities.

Differences between Editions

—

No substantive changes from DSM-IV-TR.

Disorder name Cardinal symptoms

Separation anxiety disorder Developmentally inappropriate and excessive anxiety concerning separation from home or from attachment figures.

Separation anxiety disorder Developmentally inappropriate and excessive fear or anxiety concerning separation from those to whom an individual is attached.

Differences between Editions

Classified as a disorder usually first diagnosed in infancy, childhood, or adolescence. Onset before the age of 18. Symptoms last at least 4 weeks.

Classified as an anxiety disorder.

Disorder name

with one or another condition. Although there is considerable consistency in the diagnostic criteria that the DSM-IV and DSM-5 use for the disorders on which this chapter focuses (SAD, social anxiety disorder, GAD), there have also been several changes, as summarized in Table 14.1. The remaining four standards focus on characteristics that can vary across condition types. Thus, the second standard centers on clarifying distinctions in the specific longitudinal trajectories and associated pattern of comorbidities for the anxiety disorders, and the third takes into account variations in familial patterns of aggregation within families for different anxiety-related conditions and their comorbidities. The fourth standard shifts focus to the specific risk factors or perturbations in biological systems reflective of underlying pathophysiology that research has shown to relate to individual disorders. Treatment response is at the heart of the fifth standard, which has led to evidence that at least some adult anxiety disorders show differential patterns of improvement following specific psychotherapeutic or psychopharmacologic interventions. OCD and PTSD Within the DSM-IV anxiety disorders, the strongest evidence for specificity based on these five standards derives

No cutoff set for age of onset. Symptoms last at least 4 weeks in children and adolescents and, typically, 6 months or more in adults.

from studies of disorders that were moved, due to cumulating findings suggesting that they were better classified elsewhere, to new sections of the DSM-5. The case is particularly clear for OCD, which is characterized by recurrent, time-consuming, and impairing compulsions or obsessions. With respect to longitudinal trajectory and associated comorbidity, OCD exhibits relatively strong cross sectional and longitudinal associations with disorders of impulse control, including tic disorders and ADHD (Grados et al., 2001; Peterson, Pine, Cohen, & Brook, 2001). Similarly, OCD that occurs early in development predicts an increased risk for further OCD either later in adolescence or during adulthood (Peterson et al., 2001). Family-genetic studies show similar patterns of coaggregation, with associations among OCD, tics, and ADHD emerging frequently within families (Pauls, Alsobrook, Goodman, Rasmussen, & Leckman, 1995; Pauls, Leckman, Towbin, Zahner, & Cohen, 1986; Peterson et al., 2001) and work is also under way to identify gene variants associated with the condition (Grados, 2010). Such patterns of comorbidity are thought to reflect dysfunction in a common neural circuit encompassing the prefrontal cortex, striatum, and thalamus, forming so-called cortico-striato-thalamo-cortical (CSTC) loops (Saxena & Rauch, 2000).

Classification of Anxiety Disorders Across Development

533

Relatively consistent findings also differentiate PTSD and acute stress disorder (ASD) from other anxiety disorders (but see Zoellner, Rothbaum, & Feeny, 2011, for a dissenting view); in particular, both conditions stand out from most other DSM disorders, in that a specific causal factor, in the form of a frightening event or stressor, is explicitly tied to their onset. The pattern of associated symptoms and comorbidities is also distinctive in PTSD (e.g., Brett, 1997; Sullivan & Gorman, 2002), and a few studies show associations between PTSD in children and signs of stress-related psychopathology in their parents (Yehuda, Halligan, & Bierer, 2001; Yehuda, Halligan, & Grossman, 2001). A growing body of research provides evidence that PTSD is characterized in adults by a specific pattern of neural dysfunction, with excessive activation evident in the amygdala and hippocampus and underactivation apparent in medial prefrontal regions (see Patel, Spreng, Shin, & Girard, 2012 for a review). Less evidence exists to distinguish childhood PTSD from other childhood conditions at a neurobiological level, although some findings suggest that youths with PTSD show structural anomalies in a set of brain regions (corpus callosum, prefrontal cortical areas) that is distinct from areas documented as functioning atypically in adults (Jackowski, de Araújo, de Lacerda, Mari, & Kaufman, 2009). Trauma exposure itself, however, has been associated in a number of pediatric studies with a range of neurodevelopmental consequences, and further work is needed to disentangle these from sequelae or correlates of PTSD (Watts-English, Fortson, Gibler, Hooper, & DeBellis, 2006). Finally, optimal psychotherapeutic treatment of PTSD also differs from that for other anxiety disorders, in that it often involves a heavy focus on repeated exposure to traumatic events (Smith et al., 2013). Evidence for the specificity of the three conditions on which we focus in this chapter—separation anxiety disorder, GAD, and social anxiety disorder—is weaker than for OCD and PTSD. Indeed, these three conditions are frequently, particularly in pediatric research, treated as a cluster, given their frequent comorbidity (RUPP, 2001). Nonetheless, for each of the three disorders, data suggest that at least some of the five standards are met, including distinct patterns of symptoms.

view has held that this condition is closely related to panic disorder, as defined in the DSM-IV, and recent findings support a genetic link (Roberson-Nay, Eaves, Hettema, Kendler, & Silberg, 2012). Panic disorder is characterized by sudden, unexpected paroxysms of extreme anxiety or terror. While panic attacks can occur at any point during the life span, the experience of spontaneous, recurrent panic attacks rarely begins before adolescence. In severe cases, the disorder typically follows a developmental course in which isolated, spontaneous panic attacks begin during adolescence and are followed later in life by multiple, recurrent panic attacks that meet criteria for panic disorder (e.g., Pine, Cohen, Gurley, Brook, & Ma, 1998). Agoraphobia, or anxiety about being in places or circumstances from which escape would be difficult or embarrassing in the case of a panic attack, is a common ensuing complication. The conceptualization of SAD as related to panic disorder arose largely from clinical data in adults with panic disorder, who report high rates of SAD during childhood (Bandelow et al., 2001; Silove, Manicavasagar, Curtis, & Blaszczynski, 1996). Such data are problematic, however, in that they are vulnerable to referral biases and retrospective distortions. In general, data from family and longitudinal studies provide some support for the view that SAD relates to panic disorder; however, the data are not unequivocal (Biederman et al., 2007; Pine et al., 1998; Roberson-Nay et al., 2010). The strongest support for the link between SAD and panic disorder comes from studies of respiratory function, where panic attacks and SAD have been tied to similar breathing perturbations (e.g., Atli, Bayin, & Alkin, 2012; Goodwin & Pine, 2002; Pine, Klein, et al., 2000). Notably, however, high rates of SAD are also found in offspring of parents with major depression (Biederman et al., 2001), in individuals who have experienced traumatic stress (Goenjian et al., 1995; Pine & Cohen, 2002), and even in the context of streptococcal infection (Swedo et al., 1998). In light of these findings, as well as evidence that childhood SAD is a strong predictor not only of adult panic disorder, but also of adult depression (Lewinsohn, Holm-Denoma, Small, Seeley, & Joiner, 2008), there may be utility in alternate models that conceptualize SAD as a specific pattern of responses to general distress (Lipsitz et al., 1994).

Separation Anxiety Disorder

Generalized Anxiety Disorder

Investigators have adopted conflicting conceptualizations of separation anxiety disorder (SAD) during the past few decades and only recently have they begun to recognize it as a condition that can manifest across the life span (Marnane & Silove, 2013). For many years, the prevailing

GAD, which is characterized by a pattern of pervasive and recurrent worries about a range of topics, along with avoidance of emotional experience or processing, has also been conceptualized as a set of responses to real or anticipated distress (Newman & Llera, 2011). The diagnosis of GAD

534

Anxiety Regulation

has evolved since its initial appearance in the DSM-III (APA, 1980). In its first incarnation, the diagnosis was assigned only to adults who presented with a pattern of recurrent worries; children with similar symptoms were instead diagnosed with overanxious disorder (OAD). With the publication of DSM-IV, despite mixed evidence regarding the overlap of GAD and OAD (Beesdo-Baum et al., 2011), the GAD diagnosis subsumed overanxious disorder and age was eliminated as a defining diagnostic criterion, although lower symptom thresholds were still used for youths than for adults (Andrews et al., 2010). The DSM-5 conceptualization of GAD, which emerged after substantive debate regarding multiple issues, including the relative merits of dropping some associated symptoms (Comer, Pincus, & Hofmann, 2012) and the most appropriate minimum duration of symptoms and appropriate symptom thresholds (Beesdo-Baum et al., 2011), continues to allow youths to display fewer symptoms than adults to cross the diagnostic threshold. Of the conditions classified in DSM-5 as anxiety disorders, GAD has generated perhaps the most debate, because it virtually always presents with another concurrent anxiety disorder (Klein & Pine, 2002; RUPP, 2001). Indeed, some have contended that GAD represents a complication of other anxiety disorders rather than an independent disorder in its own right (Klein & Pine, 2002). Further, GAD shows unusually tight links with major depressive disorder (MDD), in children as well as adults (Kendler, 2001; Pine et al., 1998; Silberg, Rutter, & Eaves, 2001), leading some prominent theorists to suggest that GAD be grouped conceptually with other distress disorders marked by high levels of negative affectivity, including MDD and PTSD (Watson, 2005). Social Anxiety Disorder In individuals with social phobia or social anxiety disorder, anxiety symptoms emerge in anticipation of embarrassment in social or performance situations. Although this construct was historically labeled social phobia, evidence that social anxiety disorder differs dramatically from specific phobias on several fronts (impairment, treatment-seeking behavior, longitudinal outcome, and familial aggregation; e.g., Fyer, 1998; Fyer, Mannuzza, Chapman, Martin, & Klein, 1995; Pine et al., 1998) led to a preference for a new label: social anxiety disorder. Whereas DSM-IV recognized two forms of social anxiety disorder, one characterized by an isolated fear of public speaking and the other characterized by high levels of anxiety in a range of social situations, DSM-5 treats most cases as generalized, with a specifier assigned to those cases

that are restricted to public speaking fears. This change reflects considerable evidence that although individuals with exclusive performance/speaking fears share other socially anxious individuals’ core concern about being negatively evaluated, they show distinctive etiological and treatment response profiles (Bögels et al., 2010). Of note, prior to the publication of DSM-IV, children who showed severe avoidance of social situations could be diagnosed with avoidant disorder of childhood instead of social phobia/social anxiety disorder. In light of evidence, however, that the distinction between the two conditions was minimal (e.g., Francis, Last, & Strauss, 1992), the authors of the DSM-IV elected to fold avoidant disorder of childhood into the social anxiety disorder category, just as they decided that GAD would subsume overanxious disorder of childhood. This change not only eliminated age as a diagnostic criterion, but it also decreased redundancy among diagnoses, given that avoidant disorder appeared to be rare and to overlap considerably with social anxiety disorder. Shared Characteristics Among the Anxiety Disorders A large and growing body of evidence identifies distinctive symptom and familial aggregation patterns as well as markers of neurophysiological dysfunction, within individual anxiety diagnoses. As interpreted most recently by the DSM-5’s authors, this literature supports the view that anxiety disorders can be defined as a family of related, but discrete, conditions. However, although individual disorders clearly show specific profiles, there is also ample and comparably strong evidence of commonalities among at least some conditions. For example, common to virtually all of the anxiety disorders is a persistent tendency to experience any of a wide range of anxious symptoms. These can be somatic (e.g., racing heart, sweating, nausea, trembling), cognitive (e.g., worry, fear), or behavioral (e.g., avoidance, tearfulness). These symptoms may occur discretely or in clusters; when four or more somatic or cognitive symptoms suddenly co-occur and rapidly peak in the context of intense fear, they constitute a panic attack. In individuals with panic disorder, another condition classified as an anxiety disorder (but not a focus of this chapter), panic attacks occur spontaneously and without apparent provocation. In this context, such attacks sometimes lead to agoraphobia, or a fear of experiencing panic symptoms or attacks in a public setting. However, in the presence of specific feared stressors, individuals with other anxiety disorders can also experience panic attacks. For example, individuals with

Anxiety Regulation

social anxiety disorder may experience panic when they must speak to a group or mingle at a party. Children with separation anxiety disorder may experience panic when leaving for school in the morning, and adults with the disorder may have such attacks while heading to work. Worry also represents a core feature of most anxiety disorders for youths and adults alike (Borkovec, 1994; Fialko, Bolton, & Perrin, 2012). Conceptualized as hard-to-control repetitive thoughts or images about future events perceived as uncertain and likely negative, worry represents an effort to reduce the likelihood that those future events will occur (Borkovec, Robinson, Pruzinski, & DePree, 1983). As we discuss later in this chapter, it can function as a type of emotion regulation strategy, albeit one that is less adaptive than others. Worry can be observed in children as young as preschoolers and it appears to become an increasingly central factor in anxiety disorders over the course of development. This pattern of repetitive thinking appears to emerge as a function of multiple processes—cognitive avoidance, proclivity to interpret problems as difficult to address, beliefs that the process of worry has value, and intolerance of uncertainty (Fialko, Bolton, & Perrin, 2012). Because the process of worry differs minimally across disorders, it is the content and precipitating environmental cues that serve as distinctive characteristics. In GAD, for example, worry and excessive anxiety develop about a variety of events or circumstances and can float from target to target. Objects of worry may also be circumscribed items or situations, as in SAD or social anxiety disorder. Thus, although SAD, GAD, and social anxiety disorder each demonstrate distinctive profiles in numerous ways, they also share a number of core features. Further, because these three disorders commonly co-occur, particularly in childhood and adolescence, it is difficult to disentangle and study individual disorders in isolation. For these reasons, our review of the literature on mechanisms of anxiety regulation will focus largely on research that encompasses one or more of these three conditions. Recent findings from both developmental psychology and cognitive neuroscience studies of these disorders and of normal-range anxiety converge to provide an increasingly clear picture of the mechanisms of anxiety regulation, at behavioral, cognitive, and neural levels.

ANXIETY REGULATION The relevance of the construct emotion regulation to our understanding of anxiety, in both clinical and normative

535

manifestations, has been a topic of intense interest in the past several years, leading to the publication of numerous reviews and analytical papers. These put forth a number of ideas about how emotion regulation processes may figure in the onset and maintenance of anxiety disorders in childhood and adulthood (e.g., Cisler, Olatunji, Feldner, & Forsyth, 2010; Rodebaugh & Heimberg, 2008; Suveg et al., 2010). Models vary in the degree to which they emphasize cognitive versus neural versus behavioral processes, but most recognize the importance of all three to our management of anxiety. Further, theorists and researchers are acknowledging with increasing regularity the value of multilevel models for understanding the regulation of such constructs as anxiety; such models are designed to integrate knowledge regarding multiple constructs, processes, and modulatory factors, including genetics, neural function and structure, cognition, and context (e.g., Philippot et al., 2004). Finally, in recent years, more models of anxiety regulation have emerged that pay explicit attention to broad developmental processes and contexts and the ways they may figure in an individual’s acquisition of skills for managing and responding to anxiety. Cognitive-process-oriented models, such as Gross’s widely cited (1998) process model of emotion regulation, emphasize the varied cognitive activities that people enact in the course of an emotional event with the aim of modulating their emotional responses. Gross distinguishes between antecedent-focused activities, which include selecting whether to enter given situations or contexts, modifying those situations, deploying attention toward or away from emotionally salient features, and changing one’s appraisal of those features, and response-focused activities, which encompass attentional, evaluative, and behavioral processes that occur after an emotionally charged event has occurred. The adaptive value of these strategies varies—results of recent meta-analyses suggest that reappraisal strategies have the greatest impact on emotional responses or psychopathology, with weaker evidence available to support the use of attentionally centered strategies (e.g., distraction or concentration) or efforts to suppress or amplify emotions (Aldao, NolenHoeksema, & Schweizer, 2010; Webb, Miles, & Sheeran, 2012). Notably, however, some strategies within the attentional deployment and response modulation categories appeared more effective than others, which may have obscured significant effects. Specifically, active distraction and control of emotional expression were more closely associated with successful regulation than were concentration or effort to control emotional experience (Webb et al., 2012).

536

Anxiety Regulation

With respect to anxiety in particular, findings regarding cognitive regulatory strategy use are consistent with those from studies focused on emotion regulation more broadly. Reappraisal appears to be an especially effective strategy for modulating anxiety, and youths and adults who report elevated or clinically significant anxiety show atypically weak skill at implementing this strategy and make infrequent use of it (Carthy, Horesh, Apter, Edge, & Gross, 2010; Carthy, Horesh, Apter, & Gross, 2010; Werner et al., 2011). Further, recent findings suggest that even when anxious adults engage in cognitive reappraisal, it may not have the impact on daily life experiences that it appears to have for nonanxious peers (Farmer & Kashdan, 2012). The process model’s focus on cognitive activities aimed at modulating emotion is complemented by other perspectives that aim more specifically to understand how such activities, along with behavioral actions, may be put into use in the context of anxiety. Rodebaugh and Heimberg (2008), for example, framed their model of anxiety regulation in terms of Carver and Scheier’s (1998) self-regulatory theory, in which both approach-related or appetitive systems and defensive or avoidant systems engage to facilitate attainment of goals. According to Rodebaugh and Heimberg, anxiety disorders are marked by an excessive focus on ensuring safety and avoidance of negative outcomes, which engages the prevention system, as well as relatively weak activation of the promotion system, which mediates the pursuit of positive or desired goals. Thus, anxious individuals preferentially enact behaviors and thoughts that presumably help them avoid negative outcomes; in the example that Rodebaugh and Heimberg used, individuals with GAD engage in worry behaviors to avoid experiencing negative affect, leaving few resources available for instead initiating behaviors that might help them reach goals that would reduce the negative affect that they are avoiding. Models also differ in their focus on anxiety versus fear. Anxiety-focused models (e.g., Rodebaugh & Heimberg, 2008) emphasize the ways emotion regulation may modulate the cognitive and neural correlates of the anxious state. Fear-focused models (e.g., Cisler et al., 2010), in contrast, consider the ways emotion regulation may modulate the process of fear conditioning and its outcomes. Each model type informs our understanding in distinctive ways, and both hold value for helping us to better understand the role that emotion regulation plays in the emergence and resolution of both anxiety and fear. Cisler and colleagues’ (2010) proposed model suggests that capacity for emotion regulation influences one’s range of response to fear cues after the fear conditioning process is complete. These authors contend that, first, an

individual’s access to and facility with different emotion regulation strategies will affect the ways that an individual responds in the moment when facing a fear cue. These online responses, in turn, can either weaken or strengthen the learned fear association; in the latter instance, they would increase risk for onset of an anxiety disorder. For example, two people who have been conditioned to fear public speaking are likely to employ different regulation strategies when they find themselves standing in front of an audience and experience the conditioned fear response. One may flexibly employ adaptive regulation strategies such as reappraisal, thus increasing the likelihood that she will be able to proceed with her talk and thus to gain exposure to experiences likely to weaken the learned association. In contrast, the other might use a less adaptive strategy, such as suppression, which paradoxically heightens her fear in the moment and leads her to leave the podium. She thus avoids the experiences that might have decreased her fear and instead chooses an experiential path that maintains or amplifies it. In the models discussed so far, developmental considerations are implicit, if incorporated at all. In recent years, however, additional models of the anxiety regulation process have been proposed that explicitly assume a central role for developmental processes. Suveg and colleagues (e.g., Suveg, Morelen, Brewer, & Thomassin, 2010), for example, have described the emotion dysregulation model of anxiety (EDMA), in which child temperament and family emotional environment interact to predict capacity for emotion regulation and risk for anxiety. According to this model, youths who exhibit high emotional reactivity, manifested as a behaviorally inhibited temperamental style, and who also inhabit a family whose emotional environment is weighted toward the expression of negative affect rather than positive affect, are at risk for developing a maladaptive emotion regulation style. Consequent failure to manage arousal in the context of anxiety-provoking stimuli, in turn, is predicted to maintain anxiety. This model, which focuses primarily on factors that may contribute to poor emotion regulation skills, complements other models, which place emphasis instead on the cognitive, neural, or behavioral processes implemented in efforts to upregulate or downregulate emotional reactions. Another recent developmentally informed model of anxiety regulation is embedded in attachment theory and underscores the key role that early attachment experiences may play in determining the emergence of adaptive or maladaptive approaches to managing stress and, by extension, the development of anxiety disorders (Nolte, Guiney, Fonagy, Mayes, & Luyten, 2011). This model is notable

Anxiety Regulation

in that its authors have made great effort to integrate data across multiple levels of analysis, with attention to the putative roles of genetic, neural/physiological, behavioral, cognitive, and contextual variables. In brief, Nolte and colleagues (2011) propose that children who encounter an early life environment that is characterized by inconsistent or unpredictable caregiving, particularly if they are temperamentally vulnerable to hyperarousal, may develop a hyperactive response style for coping with stress. While this style is initially adaptive, it begins to become problematic as the child generalizes it to circumstances outside the initial attachment relationship, eventually consolidating a maladaptive style marked by avoidance of threat cues and consequent failure to develop realistic expectations and perceptions of the environment. Consistent with the developmental psychopathology perspective, this model also takes into account the probability that salient variables interact in different ways at different points in development, and thus that multiple anxiety-related outcomes are possible. Recent reviews of the literature suggest that many elements of this model enjoy considerable empirical support (Esbjørn et al., 2012; Jacob et al., 2011); clearly there is value in a perspective regarding anxiety regulation that incorporates attention to developmental factors. Further integrating this kind of developmentally informed model, which takes a wide-angle view of the big picture, with more elemental models that revolve around the discrete cognitive and neural processes involved in regulating anxiety and fear is a logical next step in the literature. Such broadly integrative efforts will need to take into account emerging data regarding the viability of different strategies and processes at different points in the life span, as well as the neural structures and functions and cognitive skills and capacities that must be in place to support them. In the next sections of this chapter, we summarize recent findings regarding the cognitive and neural processes and structures that support and facilitate the regulation of anxiety. Where possible, we also examine ways individual differences may influence the paths along which these factors evolve from childhood through the life span. Cognitive Mechanisms of Anxiety Regulation Much of the research on cognitive mechanisms of anxiety regulation as they emerge and evolve across the life span has focused on strategies that are commonly characterized as explicit or consciously and intentionally implemented. As Amstadter (2008) summarized in a recent review, the use and effectiveness of two strategies, reappraisal and

537

suppression, have been especially common targets of study in samples across much of the life span. As we briefly noted earlier in this chapter, reappraisal, or cognitively reframing an anxiety-provoking situation in ways that are calming and constructive, enjoys perhaps the most empirical support as a useful approach to regulating anxiety in youths and adults (Webb et al., 2012). It is also a strategy that anxious individuals from middle childhood to early adolescence and onward appear to implement less often and less effectively than their less-anxious peers (Carthy, Horesh, Apter, Edge, & Gross, 2010; Carthy, Horesh, Apter, & Gross, 2010; Farmer & Kashdan, 2012; Legerstee, Garnefski, Jellesma, Verhulst, & Utens, 2010; Werner et al., 2011; but see Decker et al., 2008, and Tan et al., 2012, for conflicting findings), at least in part because they doubt the utility of this approach (Sung et al., 2012). In light of these findings, it is not surprising that quick mastery of cognitive reappraisal in the context of CBT shows promise as a predictor of positive treatment outcomes for anxious adults (Moscovitch et al., 2012). Effective implementation of cognitive reappraisal places demands on a number of cognitive systems, some of which mature slowly. For example, cognitive control, or the ability to intentionally use effort to manipulate attention, memory, and other mental functions, appears to be a critical foundation for successful reappraisal (e.g., Vanderhasselt, Baeken, Van Schuerbeek, Luypaert, & De Raedt, 2013). This capacity does not appear to be fully developed until adulthood (Kar, Vijay, & Mishra, 2013), which is not surprising given that its neural substrates show a long, slow, and complex developmental trajectory that can extend into an individual’s mid-twenties (Raznahan et al., 2011). Further, some evidence suggests that as the brain ages, the neural underpinnings of cognitive control show waning efficiency, as reflected in increased difficulty implementing reappraisal strategies (Opitz, Rauch, Terry, & Urry, 2012). Thus developmental status is an important consideration in evaluations of whether an individual shows deficits or strengths in capacity for cognitive control. Ability to control and direct attention in the context of emotion appears to be particularly important for cognitive reappraisal, which demands that an individual move flexibly from one set of perceptions and interpretations to another. Anxious adults, for example, even those who do not cross diagnostic thresholds for any disorders, appear to have more difficulty than their less-anxious peers in consistently implementing this skill, even when they show typical ability to redirect attention to nonemotional cues (Johnson, 2009). Similar patterns of difficulty controlling attention to emotional stimuli characterize youths with

538

Anxiety Regulation

anxiety, which suggests that even rudimentary or immature capacity for attentional control may be critical to effective regulation of anxious thoughts and feelings (White, Helfinstein, Reeb-Sutherland, Degnan, & Fox, 2009). Numerous approaches to treating anxiety disorders have begun to incorporate explicit training in cognitive reappraisal as a core component. In particular, cognitive reappraisal’s compatibility with traditional cognitive behavioral therapy (CBT) for anxiety, which integrates techniques and approaches from cognitive and behavioral perspectives and has been successfully implemented with anxious children, adolescents, and adults, has generated considerable interest among researchers and clinicians. Traditional CBT focuses on exposure to anxiety-provoking stimuli and challenges to maladaptive thinking patterns; while a large proportion of individuals with anxiety disorders respond well, a sizable number (up to 30 or 40% in some studies) do not (Hofmann, Asnaani, Vonk, Sawyer, & Fang, 2012). The merits of introducing cognitive reappraisal training to boost treatment effects, particularly in those who show weak responses to traditional CBT have received increasing attention, and some have even suggested making cognitive reappraisal and other emotion regulation techniques core elements of anxiety treatment for both adults (e.g., Mennin, 2006) and children (e.g., Hannesdottir & Ollendick, 2007). Although data are still emerging from studies that test the hypothesis that training in cognitive reappraisal and other adaptive emotion regulation techniques may augment treatment response among anxious individuals, findings to date are mixed and suggest that close attention to individual differences in patterns of response is warranted (Newman et al., 2011). A second emotion regulation strategy that has received considerable attention in the context of research on anxiety is suppression of emotional experience or, more commonly, expression. Unlike cognitive reappraisal, suppression has been associated with a range of problematic consequences in anxious and healthy adults, including slowed reduction of negative emotion, decreased experience of positive affect and events, and, in some cases, heightened physiological arousal (Campbell-Sills et al., 2006; Farmer & Steger, 2006; Feldner, Zvolensky, Stickle, Bonn-Miller, & Leen-Feldner, 2006). In both adults and adolescent girls, use of suppression to regulate emotions has also been found to relate to a lack of emotional awareness (Eastabrook, Flynn, & Hollenstein, 2014; Gross & John, 2003). A few studies have suggested that in both youths and adults, suppression is positively associated with the presence of anxiety, particularly social anxiety. For adolescent girls, for example, implementation of this emotion

regulation strategy appears to mediate associations between limited emotional awareness and social anxiety (Eastabrook et al., 2014). Adults with social anxiety disorder, as well as those who endorse having avoidant attachment styles, which have been linked to a pattern of anxious and avoidant behavior, endorse more use of this strategy than do healthy controls, and avoidantly attached adults have been shown to evidence a distinctive neural profile when using suppression to regulate emotion (Vrtiˇcka, Bondolfi, Sander, & Vuilleumier, 2012; Werner et al., 2012). Notably, in healthy youths, the use of suppression, which is more common among boys than girls, appears to decline with increasing age (Gullone, Hughes, King, & Tonge, 2010), which suggests that as complex cognitive skills emerge and mature, healthy individuals are likely to switch to the use of more adaptive strategies. John and Gross (2004) hypothesized that this normative developmental shift may also reflect the acquisition of life experience, which provides individuals with direct feedback about the pros and cons of using different emotion regulation strategies. Further, the rewards and negative consequences associated with the use of suppression may change across the life span, thus differentially reinforcing its use at different developmental stages. Other emotion regulation strategies that are prominently associated with anxiety include worry and rumination, which constitute distinct, but related, patterns of repetitive negative thinking that an individual finds hard to control (McEvoy & Brans, 2013). Worry, which characterizes many anxiety disorders and serves as a core feature of GAD, is thought to function by helping individuals avoid negative affect. At least two ideas about how worry may facilitate emotional avoidance have received empirical support; first, it may constitute an effort to prevent or prepare for negative experiences, second, it may blunt bodily responses to stimuli that elicit fear. Thus, at both cognitive and physiological levels, worry prevents the processing that is critical for extinguishing anxious responses to perceived threats (Borkovec, 1994). Worry has also been conceptualized, on the basis of accumulating findings, as a means of maintaining a state of chronic distress in order to avoid having to shift from positive or neutral to negative emotional states in the face of negative events (Newman & Llera, 2011). Tendencies to worry emerge early in life; however, although worries have been documented among preschoolers, they appear to increase in prevalence around the age of eight years (Vasey, Crnic, & Carter, 1994). The ability to worry depends on being able to anticipate future events

Anxiety Regulation

and to anticipate catastrophic outcomes, thus children are less likely to employ this as a primary regulatory strategy until these reasoning skills reach at least a rudimentary level of development. As early as elementary school, girls endorse more worrying than do boys (Silverman, LaGreca, & Wasserstein, 1995), and this pattern of sex difference appears to continue into adulthood (Zalta & Chambless, 2008). Rumination, which differs from worry in that it focuses more heavily on past events than on the future and on content that is oriented toward issues of self-worth and loss rather than anticipated negative experiences, is also more common among women than among men (see Nolen-Hoeksema, 2012, for a review). Like worry, it also appears to account in part for sex differences in rates of anxiety symptoms and disorders from at least middle childhood onward (Bender et al., 2012; Zalta & Chambless, 2008; Zlomke & Hahn, 2010). As Nolen-Hoeksema (2012) pointed out, much less is known about the emotion regulation strategies that men, particularly anxious men, employ, than is known about women’s preferred strategies. Further research is clearly needed across the life span to better characterize the male developmental trajectory for adaptive and maladaptive approaches to managing anxiety. Research in both children and adults tends to find rumination to relate more strongly to depression than anxiety (McEvoy & Brans, 2013; Verstraeten, Bijttebier, Vasey, & Raes, 2011). Not surprisingly, however, given the high rates of comorbidity between anxious and depressive conditions, ruminative thinking also appears common in individuals identified as primarily anxious. Individuals with social anxiety, for example, frequently engage in postevent processing, which consists of repetitive reflection on negative aspects of a social interaction. Further, some evidence suggests that a ruminative response to anxiety may increase risk for depressive disorders and, conversely, rumination both in and outside of the context of depression may increase risk for anxiety (McLaughlin & Nolen-Hoeksema, 2011; Starr & Davila, 2012). These associations between rumination and psychopathology may stem, at least in part, from negative interpersonal correlates of rumination, such as excessive support or reassurance seeking (McLaughlin & Nolen-Hoeksema, 2012). In adults with social anxiety, CBT appears to decrease tendencies for post-event processing or anxious rumination (Price & Anderson, 2011). Limited research has examined treatment effects on this emotion regulation strategy in youths; however, at least one influential theoretical model of child anxiety and its treatment explicitly

539

identifies postevent attributions as a key element that warrants attention (Kendall, 2012). Further, CBT has been used successfully to target worry in children as young as 7 years of age (Suveg, Sood, Comer, & Kendall, 2009), which suggests that further efforts to modify repetitive negative thought patterns in youths are likely to yield positive results. In addition to the anxiety regulation strategies outlined above, individuals may also use a range of other adaptive and maladaptive approaches to modulate their emotional state in the context of an anxiety-provoking circumstance. No one strategy is likely optimal in every situation; even those identified as typically maladaptive may sometimes be the best option an individual has available. Further, the range of accessible strategies is likely to vary across development, as will the contexts that demand their implementation; it is thus likely that strategies that are less effective at some developmental stages may hold more utility at others. In fact, rather than simply being skilled at a few adaptive regulation strategies, possession of a broad and flexible emotion regulation repertoire may serve as a stronger predictor of psychological health. In at least one study, individuals who reported using an array of regulatory strategies exhibited less internalizing psychopathology than did peers who endorsed a more limited range of strategies (Lougheed & Hellenstein, 2012). A growing body of research also points to the value of mindfulness, or fostering awareness and acceptance of one’s emotional experience, for regulating anxiety. Together with self-reported difficulties with emotion regulation, limited mindfulness predicts both anxiety symptoms and disorders in adults (Roemer et al., 2009). Treatment approaches that explicitly teach and encourage mindfulness have shown efficacy in reducing anxiety, along with its neural correlates, in adults (Goldin & Gross, 2010). Anxious children and adolescents, too, have been found to respond positively to mindfulness-based therapies in both open and randomized controlled treatment trials, and clinicians are incorporating mindfulness training into child treatment with increasing frequency (Semple & Burke, 2012). Notably, however, at least one study found the use of mindfulness or acceptance techniques to perpetuate, rather than mitigate negative affect in adults with GAD, which suggests that individual differences that predict a positive response to such approaches warrant further investigation (Aldao & Mennin, 2012). Research into cognitive strategies for regulating emotion broadly and anxiety specifically has made great strides in the past few decades, with particularly notable advances made in understanding normative patterns of strategy

540

Anxiety Regulation

use across development. Integration of the cognitive, developmental, and clinical literatures has also yielded useful information about what strategies are most effective and whether and how such strategies, particularly cognitive reappraisal and mindfulness, can be efficiently incorporated into evidence-based treatment approaches. Perhaps the most striking developments in the knowledge base regarding anxiety regulation, however, come from translational work that spans neuroscience, clinical, and developmental domains. In the next section, we review recent research findings regarding the neural substrates of anxiety regulation as they evolve across the life span. Neural Mechanisms of Anxiety Regulation Research in neuroscience has made fundamental contributions to the conceptualization of the emotion regulation construct, broadly speaking. Moreover, due to cross-species conservation in brain-behavior relationships for fear and anxiety, neuroscience research in this area has provided a particularly influential perspective in shaping conceptualizations of both emotion and emotion regulation. This neuroscience perspective has clear advantages when it is assumed in attempts to understand emotion regulation in typically and atypically developing children and adolescents, as well as later in the life course; it also, however, holds several inherent disadvantages. Because the brain is the ultimate arbiter of behavior and cognition, neuroscience research has the advantage of examining factors that may ultimately cause emotions and emotion regulation, providing a mechanistic understanding of the constructs. However, because considerable research in neuroscience is conducted using a cross-species approach, a certain degree of reduction is necessitated in the experimental approach if researchers are to maintain experimental control. The reductive aspects of neuroscience carry marked disadvantages for clinical application, in that findings from neuroscience-based research can be difficult to extrapolate to the complexities of the clinical scenario, particularly for children, where development adds additional complications and considerations. Again, research on anxiety has been at the forefront of attempts to integrate findings from neuroscience into current thinking about developmental psychopathology. Accordingly, research on the neuroscience of the regulation of normal and pathological anxiety is important not only to our understanding of anxiety per se, but also to our broader understanding of individual differences in other clinical domains and for emotion regulation more generally.

Consistent with the definitions that we outlined earlier in this chapter, the somewhat reductionist neuroscience perspective on emotion regulation considers emotion to constitute a brain state associated with stimuli, such as threat or rewards, that the organism will expend effort to approach or avoid. In this context, a threat is a dangerous stimulus, capable of harming the organism, which creates fear (the brain state created by an immediately present threat) or anxiety (the brain state created by a more distal or potential threat). For both sets of stimuli, the associated distinctive brain state can be identified in various species based on the organism’s tendency to avoid the threatening stimulus. Of course, these definitions can be applied to both normal and abnormal forms of fear or anxiety, with the distinction hinging on the degree to which the brain state ultimately facilitates or hinders adaptation. In the neuroscience perspective, emotion regulation refers to any attempt by the organism to alter or modulate this initial emotional response. For both fear and anxiety, regulation is viewed typically from this perspective as attempts to modulate levels of the emotion to facilitate ongoing adaptive behavior. In this context, normal emotion regulation can be differentiated from pathological regulation, based on the organism’s ability to reduce the level of fear or anxiety that is expressed when signs of danger have dissipated. In some sense, this distinction between an emotion and the attempt to regulate it is arbitrary. This is because the attempt to regulate an emotion can begin virtually instantaneously with onset of the emotion. However, emotion and emotion regulation can be viewed as co-existing on a dimension. In this perspective, the term emotion refers most directly to the organism’s immediate response, and the term emotion regulation becomes increasingly relevant in describing this response at later points in time. At these later points in time, regulation of emotion is expressed as the organism engages various cognitive and associated behavioral processes to alter the initial emotional reaction. Considerable research has translated work in rodents and nonhuman primates to clinical contexts. Interest has focused particularly deeply on brain-imaging research examining the substrates of emotion regulation. This interest has emerged, at least in part, because such research extends to humans the considerable range of findings in other species regarding the brain substrates of emotion generation and the associated structures that regulate generated emotions. As is typical in many domains of brain imaging research, more studies of emotion regulation examine adults than children or adolescents. This state of affairs reflects the methodological and ethical

Anxiety Regulation

complexities of conducting brain imaging research in vulnerable individuals, especially children, who may find the scanning environment frightening and who may have difficulty monitoring and managing their own movement and distress (Bookheimer, 2000). These issues are particularly salient in studies focused on the experience of negative emotions, such as fear and anxiety, the induction of which may require the application of gentler stimuli than those that are used with adults (e.g., Pine, Fyer, et al., 2001). Nevertheless, after many years of successful application of imaging technology in research on adults, a growing brain imaging literature also has arisen that focuses on normal and abnormal child and adolescent development across multiple emotional and cognitive domains. Over the past 10 years, imaging research on emotion regulation typically has adopted one of two experimental approaches, consistent with the dual-process framework that Gyurak and colleagues (2011) outlined. In one approach, the research participant is directly instructed to regulate any emotion that arises during a brain-imaging experiment. While regulatory acts could involve increases or decreases in either positive or negative emotions, the most consistent results have emerged in studies examining attempts to reduce fear, an observation that again follows from the cross-species conservation in brain-behavior relationships for this brain state. This first approach has been termed explicit emotion regulation, reflecting the fact that the research participant is explicitly told the goal of the experiment, which is to change an initial emotional response. Moreover, in some such experiments, the participants are directly taught methods for regulating their emotions. These methods vary across studies, but most research focuses on reappraisal, or changing the way in which one thinks about an emotionally charged situation or topic with the aim of increasing or decreasing its emotional impact. In other experiments, no instructions are provided regarding whether or how to employ particular regulatory strategies. The other approach to emotion regulation does not provide any instructions to the participant regarding emotion; indeed, no reference is made to emotion at all. Rather, the participant is required to complete a neutral task, such as target identification or memory encoding. In this context, task-irrelevant emotional and neutral items are embedded, and emotion regulation is indexed based on the degree to which research participants can stay focused on and engaged in the neutral task when they are confronted by task-irrelevant emotional stimuli. Studies using this second approach are considered to be examining implicit

541

emotion regulation, or regulatory processes that occur out of awareness. A number of fMRI studies map the neural correlates of re-appraisal and other explicit emotion regulation strategies in healthy adults, with regulatory efforts focused on negative or positive emotions in general (e.g., McRae et al., 2012) or on anxiety in particular (e.g., Ball, Ramsawh, Campbell-Sills, Paulus, & Stein, 2012). This work in adults also has been extended to various anxiety states, including normal individual differences in healthy populations on anxiety rating scales and clinical diagnoses of social phobia or generalized anxiety disorder (Ball et al., 2012; Blair et al., 2012; Brühl et al., 2012). Moreover, studies of this type have examined differences in the neural correlates of emotion regulation between healthy adolescents and adults, and although fMRI studies of cognitive reappraisal in children are limited, researchers have successfully measured neural activity during reappraisal tasks using electroencephalographic (EEG) techniques in youths as young as 5 years (DeCicco, Solomon, & Dennis, 2012). The research base is thus growing and expanding in ways that will allow a more developmentally informed understanding of the neural substrates of emotion and anxiety regulation than was possible initially. An additional factor renders a developmental approach to this area of study particularly challenging. Not only do researchers studying explicit emotion regulation face the general methodological and ethical issues inherent in neuroimaging work with youths, but they also must address one more specific set of complications. Because research participants are directed to attend to the study goals on each experimental trial, expectancies about what they should do can heavily influence their patterns of task responding. The relatively common use of self-report assessments of emotional change as the core measure of emotion regulation competency gives further weight to expectancies, and it remains unclear how tightly real life efforts to regulate emotion in the absence of instruction to do so or measurement of one’s success map onto those produced during explicit emotion regulation tasks. In part because the effects of expectancy might vary among children, adolescents, and adults (Brenner, 2000), thus introducing a confound that is difficult to control for, relatively few studies compare the neural correlates of reappraisal in healthy adolescents and adults (but see McRae et al., 2012). None of these studies to date have made comparable comparisons across developmentally stratified groups with pathological forms of anxiety. The most consistent findings from fMRI studies of explicit emotion regulation focus on the functioning of

542

Anxiety Regulation

areas within the association cortices, regions of the brain that are not directly involved in perception, motor execution, or even initial emotional responding, but instead support more integrative brain-based processes. Typically these areas operate in the context of brain networks known to support relatively complex forms of decision making, and include areas of the dorso-medial prefrontal cortex (PFC), the ventro-lateral PFC, and the dorsal parietal cortex. In healthy adults, these regions typically show enhanced activation during successful emotion regulation—both explicit and implicit, in tandem with reduced activation in brain regions such as the amygdala that are thought to support aspects of emotion reactions (e.g., Bush, Luu, & Posner, 2000; Somerville et al., 2013). Moreover, studies of adults with anxiety disorders, as well as those with temperamental or personality characteristics associated with anxiety generally find reduced activations in these parietal and PFC regions, suggesting that deficient emotion regulation in these individuals reflects an inability to engage regulatory neural structures (e.g., Ball et al., 2012; Vrtiˇcka et al., 2012). Although the base of studies that take a developmental perspective to understanding the neural correlates of anxiety regulation is still growing, emerging evidence from comparisons of neural correlates of emotion regulation between healthy adolescents and adults suggests that the regions perturbed in adult anxiety disorders continue to mature through adolescence (e.g., McRae et al., 2012). Few studies have yet used neuroimaging to examine reappraisal in children prior to adolescence; one fMRI study, however, in which 5 to 10 year old children were asked to reappraise sad images, yielded evidence of bilateral activation in lateral and medial PFC regions, along with right-lateralized activity in the ACC and ventral lateral PFC (Levesque et al., 2004). Further, evidence from research that has measured evoked response potentials in preadolescent children during cognitive reappraisal of emotion also supports the hypothesis that the neural substrates of emotion and anxiety regulation are not yet functioning at an adult level (Dennis & Hajcak, 2009; DeCicco et al., 2012). Thus, maturation in these regions during development may explain why some forms of anxiety wane during adolescence. Our understanding of the neural mechanisms of explicit anxiety regulation has become increasingly sophisticated in recent years, with a shift of focus from the activation of individual structures to the ways neural regions interact during an emotional event. This literature has yielded evidence that the degree to which prefrontal brain regions engage interactively and responsively with the amygdala is an important factor underlying variability in adult capacity

to effectively down-regulate or decrease negative emotion. Lee and colleagues (2012) for example, recently found evidence that the PFC more effectively and efficiently modulates amygdala activity in adults who are more adept at using a reappraisal strategy than in those who show less skill at implementing this strategy. Considerable fMRI research examines implicit emotion regulation in both healthy and anxious adults; this extensive database facilitates comparisons of the neural correlates of implicit and explicit forms of emotion regulation in such samples. Moreover, while relatively few studies examine developmental aspects of explicit emotion regulation (e.g., McRae et al., 2012), a more extensive database exists regarding implicit emotion regulation among children and adolescents. This body of research encompasses considerable work regarding the regulation of emotions in general, as well as the regulation of anxiety in particular. Research on implicit emotion regulation generally has relied on three different sets of processes, which can be characterized based on the neutral tasks that research participants complete. These tasks variously require participants to regulate emotion in the context of cognitive conflict, to orient attention, or to engage in stimulus-reinforcement learning. In fMRI research on the brain response to conflict, participating individuals are presented with neutral stimuli that they are instructed to classify; for example, a subject may be asked to identify the direction in which an arrow is pointing or to label the gender of each model in a series of neutral faces. To make this classification more difficult, the neutral stimulus is surrounded with other, distracting stimuli, such as multiple task-irrelevant arrows pointing in opposite directions. The inclusion of task-irrelevant stimuli introduces cognitive conflict, in that the participant must stay focused on task-relevant stimuli and resist being derailed by the task-irrelevant distracters. Finally, to make the otherwise neutral task emotional, researchers embed features that connote threats and rewards in at least a subset of trials. For example, on neutral trials, a subject may be asked to perform difficult classifications of neutral faces, whereas on emotional trials, the subject must classify faces making emotional expressions. In these tasks, successful performance requires the subject to reduce responding to emotional stimuli in order to maintain task-relevant goals. One particularly interesting example of such research examines the brain regions that implement both the monitoring and the control of cognitive conflict. Etkin and colleagues (2006) asked healthy adults undergoing fMRI scans to indicate the affect presented on fearful and happy faces; to introduce conflict, each face was overlaid with the word HAPPY or the word FEAR so that some face/word

Anxiety Regulation

combinations were congruent and others were incongruent. Their findings indicated that many of the brain regions involved in explicit emotion regulation, particularly prefrontal cortical areas and the anterior cingulate cortex (ACC), also play a role in conflict monitoring and control during performance of tasks that employ emotional stimuli. Further, taken together with the results of research focused on explicit processes, as well as those from additional studies that have found associations between state anxiety in healthy adults and patterns of function in prefrontal and ACC regions during cognitive conflict tasks (e.g., Bishop, Duncan, Brett, & Lawrence, 2004), these findings suggest that implicit and explicit forms of emotion regulation may draw on overlapping brain regions. Moreover, the same group also implicates perturbed functions in these same regions in adults with GAD (Etkin et al., 2010; Etkin & Schatzberg, 2011). Using the same facial emotion conflict task, the authors of both studies found study participants with GAD to show difficulty engaging regulatory processes and their mediating neural structures that healthy individuals appear to activate spontaneously in the absence of explicit instruction. Blair and colleagues (2012) obtained strikingly similar findings in research directly comparing implicit and explicit forms of emotion regulation in patients with GAD and patients with social phobia. Notably, no significant differences in patterns of activation emerged between patient groups, suggesting that at least some shared neural substrates characterize the two disorders. An important next step will be clarifying the neurochemistry of the anxiety regulation process as it is mediated in the PFC and other salient regions of the brain; studies that simultaneously probe neurochemical activity and neural function hold promise for advancing knowledge in this area. In one recent study, for example, Sripada and colleagues (2013) found evidence that by manipulating levels of allopregnalone, a neurosteroid that modulates the function of gamma-aminobutyric acid (GABA) type A receptors in healthy adults, they could enhance neural activity in the PFC and amygdala during an emotionally charged cognitive conflict task. When linked with the considerable data indicating that dysregulation of the GABA neurotransmitter system may play a key role in the development of anxiety (Kalueff & Nutt, 2007), these findings move us closer to understanding the anxiety regulation process as it plays out at multiple levels of analysis. Because implicit emotion regulation tasks create different expectancies than explicit tasks, there has been interest in employing them in research that takes a developmental approach to understanding psychopathology, including

543

anxiety. In healthy children as young as 9 years of age, evidence is emerging that suggests that prefrontal cortical activation, particularly dorsolateral regions, already differs between emotional and neutral task conditions (Lamm, White, McDermott, & Fox, 2012). Initial attempts to extend this work to individuals with or at risk for anxiety, however, have generated findings that are difficult to reconcile with the extensive research on normal and pathological emotion regulation, in the service of cognitive control. For example, two studies that have examined patterns of neural activation in adults or adolescents who were characterized during childhood as at high risk for anxiety due to a behaviorally inhibited (BI) temperamental style (Jarcho et al., 2013; McDermott et al., 2009), have yielded mixed findings. Whereas McDermott and colleagues (2009) found evidence that patterns of neural response to emotional cues presented within a conflict task moderated associations between childhood BI and adolescent anxiety disorders, Jarcho and colleagues’ (2013) study of adults with and without a history of BI did not find brain activity patterns to function as either a moderator or a mediator in such associations (although BI history did predict atypical activation in prefrontal cortical regions). Research on attention orienting, much like that on conflict, typically involves performance of a neutral task. However, participants tend to perform more quickly and accurately on orienting tasks than on cognitive conflict tasks, which suggests that the former are easier. In some orienting tasks, the participant is simply required to identify, as quickly as possible, the spatial locale of an easy-to-detect target. In others, participants might have to classify the target much as they might in cognitive conflict tasks; however, the classification is typically far easier due to the absence of explicitly conflicting cues. Orienting tasks present such stimuli in different contexts—some involving no other stimuli, others involving proximal neutral stimuli, and a third set involving proximal emotional stimuli. For example, in a hypothetical orienting task, a letter or symbol might serve as the target; in some trials it will appear alone (no other stimuli), in others it will appear with neutral faces (proximal neutral stimuli), and in still others it will appear with emotional faces (proximal emotional stimuli). In the context of this kind of task, emotion regulation is conceptualized as the capacity to rapidly orient across these different contexts at a consistent rate over time. Typically, orienting ability is quantified based on reaction time (difference scores are calculated for emotional vs. nonemotional trials), and various task arrays have been used.

544

Anxiety Regulation

The most frequently employed orienting paradigm in developmentally informed studies of normal and abnormal aspects of anxiety is the dot-probe task (e.g., Mogg & Bradley, 1998); numerous versions exist, which variously use emotional words, faces, and pictures as target stimuli. In healthy youths and adults, successful orienting to target cues in the context of the dot-probe task and other similar orienting tasks requires engagement of the ventro-lateral PFC, encompassing associated expanses of the anterior insular cortex (e.g., Monk et al., 2006) as well as parietal regions (Pourtois et al., 2006). Interestingly the ventro-lateral PFC region implicated in orienting processes overlaps with regions implicated in reappraisal (e.g., Campbell-Sills et al., 2011), which echoes evidence from conflict studies regarding parallels in neural correlates of explicit and implicit emotion regulation tasks. The most consistent finding in studies of pathological anxiety, including trauma-related distress, is enhanced activation in this PFC region, which often correlates negatively with levels of anxiety (Fani et al., 2012; Monk et al., 2006). This pattern of results suggests that engagement of the ventro-lateral PFC facilitates maintenance of orienting, particularly in contexts where threats can disrupt it. Findings regarding both normal and abnormal variations in adolescent anxiety extend this framework to developmental psychopathology, representing a particularly broad application of the emotion regulatory framework. The third area of fMRI research aimed at elucidating the neural correlates of implicit emotion regulation focuses on aspects of stimulus-reinforcement learning. Much as research on cognitive control and attention orienting does, studies of stimulus-reinforcement learning consider how responses to stimuli that initially are neutral change when those stimuli are linked to emotionally salient information. Of the three task types, those that involve stimulus-reinforcement learning place the fewest demands on research participants, whose learning can be assessed in the absence of any overt behavioral response. The most basic form of stimulus-reinforcement learning examined in efforts to clarify the neural substrates of anxiety and its regulation focuses on fear conditioning (Dunsmoor & LaBar, 2013). In fear conditioning tasks, a neutral stimulus, such as a tone or a light, is paired with an aversive unconditioned stimulus (the UCS). Research with rodents, nonhuman primates, and adult humans typically uses electric shock, a highly potent stimulus, as the UCS; studies in children and adolescents, however, have relied for ethical and practical reasons on milder UCS, such as loud sounds or annoying air blasts (e.g., Monk et al., 2003; Pine et al., 2001). Following pairing of the neutral stimulus and UCS, the neutral stimulus acquires the capacity to evoke the same

emotion as the UCS, in part because the participant learns to associate the two together. As a function of its newly acquired evocative capacity, the neutral stimulus is reclassified as a conditioned stimulus (CS+). In fMRI research, this initial form of learning is typically viewed as part of an emotional reaction that unfolds over time as the association between the UCS and the CS+ emerges, strengthens, and subsides. In the context of most research on fear conditioning, emotion regulation is operationally defined in terms of changes that occur in this initial emotional reaction, following other experimental manipulations. Probably the most extensive body of research in this area considers changes in extinction, or the reduction or disappearance of a conditioned response. Extinction occurs when an organism repeatedly encounters a CS+ in the absence of the UCS, which appears to lead to a new learning process in which the CS+ is linked with safety, or the absence of threat (Bouton, 2004). This leads the organism to reappraise and reclassify the CS+ as an ambiguous stimulus, one that could evoke fear in one context, but not in others, including the UCS-free context in which the stimulus is currently encountered. Many theories of anxiety, particularly those that emanate from the behaviorist tradition, emphasize the role of deficient extinction in sustaining individual differences in anxiety. Empirical data lend support to this notion; Lissek and colleagues (2005) found, in a meta-analysis of the fear conditioning literature, that not only does fear conditioning occur more rapidly in individuals with anxiety disorders, but extinction also occurs more slowly. Such impairment of extinction processes may reflect underlying genetic diatheses. Indeed, one study found that individuals with a particular genetic polymorphism (COMT met/met) that has been implicated in anxiety and other negative mood states were unable to extinguish conditioned responses at all (Lonsdorf et al., 2009). How these diatheses interact with experience and context to influence the developmental course of anxiety remains unclear; however, recent work takes a first step toward answering this question by suggesting that extinction failures may serve as a link between childhood anxiety (which is common but typically resolves) and the less common outcome of developing an adult anxiety disorder (Britton et al., 2011). Some of the strongest support for these theories derives from research on the therapeutics of anxiety disorders, where CBT has emerged as one of the more effective treatments across the life span (Hofmann, Asnaani, Vonk, Sawyer, & Fang, 2012). CBT for anxiety holds as central aims the modification of maladaptive thinking and the reduction of avoidance or facilitation of extinction via

Anxiety Regulation

exposure to feared or anxiety-provoking stimuli. Along with training in cognitive reappraisal, as discussed earlier, exposure has proven to be one of the most important elements in successful CBT treatment of a range of anxiety disorders (Moscovitch, Anthony, & Swenson, 2009), and neuroscience research into the mechanisms that underlie the exposure-mediated extinction process is yielding clues regarding both how extinction works and how it can be best achieved in the context of CBT. In healthy adults, fMRI research has demonstrated that the amygdala is a primary mediator of fear learning and that the medial prefrontal cortex and the dorsal ACC play key roles in implementing extinction via inhibition of amygdala activation (Quirk, Garcia, & Gonzales-Lima, 2006; Sehlmeyer et al., 2009). There is some evidence that the ability to extinguish conditioned threats is still maturing during adolescence (Pattwell et al., 2012), with decreased capacity for extinction learning relative to children and adults observed both in adolescent humans and adolescent mice. Although the neuroimaging literature on fear conditioning and extinction in youths is still too limited to permit definitive statements about neural mechanisms, presumably this developmental pattern maps onto the maturation of relevant brain regions, particularly the PFC. Notably, however, related studies have yielded evidence of developmental changes in the fear conditioning process, as well as its psychophysiological correlates from age 3 years onward (e.g., Gao, Raine, Venables, Dawson, & Mednick, 2010), which suggests that similar attention to neural correlates across development may be fruitful. Clearly our knowledge about the neural correlates and substrates of anxiety regulation has grown exponentially in the past several years and as researchers continue to find new ways to bridge neuroscience, clinical, and developmental domains, this literature is likely to continue to expand at a rapid pace. One particularly exciting direction that this work is taking builds on ideas that Tucker (1989) articulated many years ago by considering the interpersonal contexts in which emotion and anxiety regulation processes unfold. In the final section of this chapter, we describe recent advances in understanding how contexts, intrapersonal and interpersonal, may reciprocally influence the neural and cognitive processes involved in anxiety regulation across development. The Role of Context in Anxiety Regulation As Aldao (2013) astutely pointed out in a recent article, understanding contextual influences on emotion regulation in individuals with psychopathology is especially important, given that an inflexible set of responses to environmental cues serves as a core characteristic of many

545

disorders. Many of the cognitive and behavioral biases that are commonly observed among individuals with anxiety disorders constitute precisely this kind of response set, and it will thus be critical to clarify how both intrapersonal and extrapersonal features may reinforce or discourage their implementation. A broad range of such features is likely to be relevant, with different contextual factors weighing more heavily into the anxiety regulation process at different points in development. In infancy and early childhood, for example, interpersonal interactions with parents constitute a large proportion of an individual’s daily life, and thus those interactions will typically play a large role shaping in the individual’s responses to emotional cues. A large literature examines the ways parent–child interactions facilitate or interfere with healthy emotion regulation (Scaramella & Leve, 2004), and increasingly sophisticated research takes into account parent and child characteristics and the interplay between them as predictors of child responses to distressing circumstances (e.g., Aktar, Majdandži´c, de Vente, & Bögels, 2013). In adolescence, in contrast, an individual’s peer group assumes increasing salience for both adaptive and maladaptive responses to emotional cues, and while family relationships remain important, their patterns of influence on regulatory processes evolve (Adrian et al., 2009; Larsen et al., 2012). Studying how these interaction patterns figure in the development of regulatory strategies and capacities is complicated, particularly if one aims to gather data at multiple levels (e.g., neural, cognitive, and behavioral). A number of recent neuroimaging studies focused on adolescents have addressed this challenge by incorporating clever behavioral paradigms that involve real or perceived social interchanges that carry varied affective charges (see Pfeifer & Blakemore, 2012 for a brief overview). For example, tasks that lead participants to believe that they are being rejected by peers tend to elicit activity consistently in ventral prefrontal regions (e.g., Guyer et al., 2012; Masten et al., 2009) that map onto those engaged during implicit and explicit emotion and anxiety regulation tasks. Such findings suggest that the neural activity associated with decontextualized experimental tasks parallels that evoked by real-life (or close to real-life) experiences. Other studies have begun to experiment with real-time interaction with an actual person during fMRI scanning (e.g., Redcay et al., 2010), an approach that may prove particularly fruitful for studying the neural correlates of efforts to regulate social anxiety. Just as studies of the enactment of anxiety regulation and its neural correlates in the context of ecologically valid interpersonal interactions are important, research that

546

Anxiety Regulation

focuses on the process of selecting a regulatory strategy in response to environmental cues will be useful. Very little work to date has examined neural activity while an individual chooses among available regulatory approaches— typically, either participants are instructed to use a particular strategy or they are permitted to accomplish the regulatory task in any way they prefer to do so. Research that presents individuals with a range of regulatory options and then tracks patterns of choice and their correlates (e.g., neural, psychophysiological) may help elucidate optimal points of intervention for individuals who tend to rely inflexibly on maladaptive strategies. At least one study has found evidence that the ability to flexibly implement different regulatory strategies predicts later psychological health in college students (Bonnano, Papa, Lalande, Westphal, & Coifman, 2004); clarifying intrapersonal and situational factors that facilitate such flexibility and how those factors may or may not vary across development will help us better understand ways to encourage and support adaptive coping with anxiety. At an even broader level, examining anxiety regulation in comparable interactions across varied interpersonal contexts (e.g., family, friends, romantic partners, unfamiliar peers) may provide clues about the ways specific patterns of interaction and different levels of intimacy may modulate both the available and the selected strategies for regulating anxiety. Engaging with a trusted person, for example, might facilitate more flexible use of regulatory approaches than would interaction with a novel individual, with whom a more circumscribed set of reactions might feel safer. Sociological approaches to the study of emotion, which provide useful insights into how we might better take into account the background against which anxiety and its subsequent regulation occurs may help inform work exploring such questions (e.g., Barbalet, 2011). Translational Implications of Research on Anxiety Regulation As the research we have described in this chapter illustrates, work conducted from multiple perspectives, at varied levels of analysis (e.g., genetic, neural, behavioral), and with a broad range of measurement tools is converging to yield a rich picture of emotion and its regulation. Calls for translational studies that facilitate communication between basic and applied scientists who study emotion and its regulation have been increasing over the past decade, driven in part by NIMH initiatives aimed at supporting research that transforms our understanding of mental illness and its neurodevelopmental origins (see Austin, 2013;

DelCarmen-Wiggins, 2008). Increasingly, researchers are responding to these calls by designing studies of emotion regulation that transcend a single-level-of-analysis approach and that target ways biological processes, viewed through varied lenses, both influence and are shaped by behavior and context (e.g., Oberle, Schonert-Reichl, Lawlor, & Thomson, 2012). In addition, the importance of translating basic-science-informed clinical knowledge into both effective and efficacious practices in the community is informing scholarship on emotion regulation (e.g., Ferrell, 2009). Anxiety represents a domain of normal and abnormal behavior and psychology that is particularly ripe for such a translational perspective. This is because there is particularly marked cross-species conservation in research on anxiety, relative to other areas. Thus, the increasingly widespread implementation of translational and multilevel approaches to research on anxiety regulation holds numerous implications for both our understanding of anxiety and our efforts to intervene with and prevent its maladaptive manifestations. One key implication is that scientists and clinicians will need to engage with each other in creative ways to bridge the terminological and methodological divides that commonly impede their collaboration (Cicchetti & Toth, 2006). Multidisciplinary teams, as well as cross-site collaborations, are of growing importance in the study of anxiety and its regulation, as is research that looks for underlying biological commonalities among superficially distinct psychopathologies (e.g., Meyer-Lindenberg, Domes, Kirsch, & Heinrichs, 2011). As a function of broader interactions among scientists and clinicians who bring multiple perspectives to the table, it is likely that diagnostic boundaries for the anxiety disorders will continue to shift and evolve. As this process unfolds, it will be essential for scholars to hold assumptions lightly and loosely, and to be prepared to revise or jettison long-favored conceptions of anxiety and its management in the face of new evidence about its biological correlates and the ways they interact with environmental contingencies to yield varied patterns of response. However, such research on nosology is likely to unfold relatively slowly, compared with research on novel therapeutics. In fact, translational research in this area already has begun to influence anxiety treatment (Pine et al., 2009; Pine & Fox, 2015). CONCLUSIONS AND RECOMMENDATIONS FOR FUTURE WORK This chapter reviews the literature on the cognitive and neural mechanisms that allow humans to regulate anxiety

Conclusions and Recommendations for Future Work

over the course of the life span. We provide an integrative perspective on research from clinical psychopathology, developmental psychology, and cognitive neuroscience, with the goal of illustrating the ways these related, but distinct, fields inform each other. Such translational engagement among researchers is critical if we are to identify optimal, personalized approaches to enhancing anxiety regulation repertoires, especially among those who are vulnerable to internalizing psychopathology broadly, and to anxiety disorders specifically. Clearly, this need has been widely recognized among anxiety researchers, who have devised a range of creative approaches to bridging relevant literatures that historically have operated in parallel. In reviewing their work, we find that several themes emerge, each of which suggests directions for future study. First, although anxiety and its regulation across development have historically been understudied, the past decade has seen an explosion of research focused on this topic. The historical neglect of these topics reflects, in part, a clearly outdated belief that anxiety disorders constitute mild psychopathology. It is now widely acknowledged that anxiety, particularly if it emerges early in life, predicts a variety of negative outcomes, including major depressive disorder and suicide, as well as marked functional impairment across the life span. Further, researchers and clinicians typically recognize that enhancement and expansion of anxious individuals’ emotion regulation repertoires has strikingly positive effects on their internal experience and their ability to perform effectively in social, academic, and occupational domains. This new clinical vision resonates with the availability of novel translational neuroscience methods and findings from basic neuroscience to raise particular interest in clinical applications of research on anxiety regulation. To maintain the momentum of research on anxiety and its regulation, it will be critical that we produce studies that integrate a developmental psychopathology perspective with some of the elegant theoretical models of emotion and anxiety regulation that have been put forth in the past few decades. In particular, attention to developmental variation in patterns of affective chronometry (Davidson, 1998) as they manifest across different anxiety regulation strategies might help identify salient points for targeted intervention. Such work should, ideally, also gather data at multiple levels of analysis (e.g., genetic, neural, and cognitive), to facilitate generation of a more integrative and comprehensive picture of the anxiety regulation process. In addition, continued effort to move beyond subjective reports of experience as the core markers of anxiety

547

disorders and of the processes that become dysregulated in the context of these disorders is key to advancing the field. Consistent with the RDoC project (Insel & Wang, 2010), which aims to move the diagnostic process away from its reliance on difficult to quantify self-reports and toward the use of readily observable and measurable behavioral and biological data, researchers have begun to identify neural and psychophysiological correlates of both typical and atypical anxiety regulation processes. Extending this work to very young children is an important, if difficult, next step—notably, functional imaging studies have been completed that demonstrate atypical activation in emotion regulation circuitry in depressed preschoolers as young as 3 years of age (e.g., Gaffrey et al., 2011), but no published work has yet emerged to describe neural correlates of anxiety regulation in children this young. Further, continuing a process that figured prominently in the most recent revision of the DSM, it is critical that we draw on novel clinical, genetic, and neuroscientific data to identify alternative classification schemes for anxiety. Given increasingly clear evidence that anxiety disorders represent the end result of complex interplay among many risk factors, it will be important to precisely describe the dynamic processes that can, under some circumstances, lead to a pathological outcome. If we can elucidate both the potential cascading effects of, for instance, genetic perturbations, that may interact with other genes or environmental factors to increase risk at neural and cognitive levels, we will be better poised to identify effective, precisely targeted interventions that can be implemented before an individual crosses the threshold into a full-blown clinical syndrome. Notably, such knowledge may also upend our current nosology for anxiety disorders and diagnostic boundaries may blur or disappear, possibly in different ways at different points in development. Already, research findings have led to the reclassification of OCD and PTSD; as the research base expands, further changes are likely. Additionally, as Aldao (2013) clearly articulated with regard to emotion regulation research in general, there has not yet been enough research on anxiety regulation that fully considers the contexts in which strategies are learned and implemented. Integrative research that better characterizes the complex interplay among genetics, neurobiology, family environment, peer relations, and other influences is needed and is beginning to emerge in the literature. Research that combines genetic approaches with neuroimaging techniques, for example, has become increasingly common, and fMRI studies that are designed to maximize ecological validity are also appearing frequently in the literature.

548

Anxiety Regulation

Finally, further characterizing the elemental processes that constitute various explicit and implicit anxiety regulation processes in terms that are amenable to research in experimental psychology and neuroscience is likely to advance the field notably. To achieve this goal, it will be useful to move beyond broad constructs such as reappraisal by breaking them down into more fundamental cognitive and neural processes, each of which may have a distinct developmental trajectory. By defining salient constructs in language that bridges neuroscience, clinical, and developmental domains, we will be better poised to capitalize on the important findings emerging in each area and, ultimately, to develop novel, highly personalized, and rapidly effective treatments for anxiety and the regulatory impairments that maintain and perpetuate it.

REFERENCES Adrian, M., Zeman, J., Erdley, C., Lisa, L., Homan, K., & Sim, L. (2009). Social contextual links to emotion regulation in an adolescent psychiatric inpatient population: Do gender and symptomatology matter? Journal of Child Psychology and Psychiatry, 50(11), 1428–1436. doi: 10.1111/j.1469–7610.2009.02162.x Aktar, E., Majdandži´c, M., de Vente, W., & Bögels, S. M. (2013). The interplay between expressed parental anxiety and infant behavioural inhibition predicts infant avoidance in a social referencing paradigm. Journal of Child Psychology and Psychiatry, 54(2), 144–156. doi: 10.1111/j.1469–7610.2012.02601.x Aldao, A. (2013). The future of emotion regulation research: Capturing context. Perspectives on Psychological Science, 8(2), 155–172. doi: 10.1177/1745691612459518 Aldao, A., & Mennin, D. S. (2012). Paradoxical cardiovascular effects of implementing adaptive emotion regulation strategies in generalized anxiety disorder. Behaviour Research and Therapy, 50(2), 122–130. doi: 10.1016/j.brat.2011.12.004 Aldao, A., & Nolen-Hoeksema, S. (2010). Specificity of cognitive emotion regulation strategies: A transdiagnostic examination. Behaviour Research and Therapy, 48(10), 974–983. doi: 10.1016/j.brat.2010 .06.002 Aldao, A., Nolen-Hoeksema, S., & Schweizer, S. (2010). Emotion- regulation strategies across psychopathology: A meta-analytic review. Clinical Psychology Review, 30, 217–237. doi: 10.1016/j.cpr.2009.11.004 Alfano, C. (2012). Are children with “pure” generalized anxiety disorder impaired? A comparison with comorbid and healthy children. Journal of Clinical Child and Adolescent Psychology, 41(6), 739–745. Alfano, C. A., Ginsburg, G. S., & Kingery, J. (2007). Sleep-related problems among children and adolescents with anxiety disorders. Journal of the American Academy of Child & Adolescent Psychiatry, 46(2), 224–232. doi: 10.1097/01.chi.0000242233.06011.8e American Psychiatric Association. (1980). Diagnostic and statistical manual of mental disorders (3rd ed.). Washington, D.C.: Author. American Psychiatric Association. (1994). Diagnostic and statistical manual of mental disorders (4th ed.). Washington, D.C.: Author. American Psychiatric Association. (2000). Diagnostic and statistical manual of mental disorders (4th ed., Text Revision). Washington, D.C.: Author.

American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Arlington, VA: American Psychiatric Publishing. Amstadter, A. (2008). Emotion regulation and anxiety disorders. Journal of Anxiety Disorders, 22(2), 211–221. doi: 10.1016/j.janxdis .2007.02.004 Andrews, G., Hobbs, M. J., Borkovec, T. D., Beesdo, K., Craske, M. G., Heimberg, R. G., . . . Stanley, M. A. (2010). Generalized worry disorder: A review of DSM-IV generalized anxiety disorder and options for DSM-V. Depression and Anxiety, 27(2), 134–147. doi: 10.1002/da.20658 Angold, A., & Costello, E. J. (1993). Depressive comorbidity in children and adolescents: Empirical, theoretical, and methodological issues. American Journal of Psychiatry, 150(12), 1779–1791. Arlitt, A. H. (1938). Training and emotional control. In The adolescent (pp. 82–100). New York, NY: Whittlesey House/McGraw-Hill Book Company. doi: 10.1037/13505–005 Arthur, G. G. (1922). Eliminating first grade failure through the control of intellectual, physical and emotional factors. School & Society, 15, 474–484. Atlı, Ö., Bayın, M., & Alkın, T. (2012). Hypersensitivity to 35% carbon dioxide in patients with adult separation anxiety disorder. Journal of Affective Disorders. 141(2–3), 315–323] doi: 10.1016/j.jad.2012.03.032 Austin, C. P. (2013). NCATS director statement: Institute of medicine report on the CTSA program at NIH. Retrieved from http://www .ncats.nih.gov/research/cts/ctsa/about/iom/iom.html Bain, A. (1859). Control of feelings and thoughts. In The emotions and the will (pp. 398–419). London, UK: John W Parker & Son, West Strand. doi: 10.1037/10617–019 Ball, T., Ramsawh, H., Campbell-Sills, L., Paulus, M., & Stein, M. (2012). Prefrontal dysfunction during emotion regulation in generalized anxiety and panic disorders. Psychological Medicine, 43(7), 1–12. Bandelow, B., Tichauer, G., Späth, C., Broocks, A., Hajak, G., Bleich, S., & Rüther, E. (2001). Separation anxiety and actual separation experiences during childhood in patients with panic disorder. La Revue Canadienne De Psychiatrie, 46(10), 948–952. Barbalet, J. (2011). Emotions beyond regulation: Backgrounded emotions in science and trust. Emotion Review, 3(1), 36–43. doi: 10.1177/1754073910380968 Barlow, D. (2002). Anxiety and its disorders. New York, NY: Guilford. Beesdo-Baum, K., Höfler, M., Gloster, A. T., Klotsche, J., Lieb, R., Beauducel, A., . . . Wittchen, H. (2009). The structure of common mental disorders: A replication study in a community sample of adolescents and young adults. International Journal of Methods in Psychiatric Research, 18(4), 204–220. Beesdo-Baum, K. K., Knappe, S. S., Fehm, L. L., Höfler, M. M., Lieb, R. R., Hofmann, S. G., & Wittchen, H. U. (2012). The natural course of social anxiety disorder among adolescents and young adults. Acta Psychiatrica Scandinavica, 126(6), 411–425. doi: 10.1111/j.1600 –0447.2012.01886.x Beesdo-Baum, K., Winkel, S., Pine, D. S., Hoyer, J., Höfler, M., Lieb, R., & Wittchen, H. (2011). The diagnostic threshold of generalized anxiety disorder in the community: A developmental perspective. Journal of Psychiatric Research, 45(7), 962–972. doi: 10.1016/j .jpsychires.2010.12.007 Beidel, D. C., Silverman, W. K., & Hammond-Laurence, K. (1996). Overanxious disorder: subsyndromal state or specific disorder? A comparison of clinic and community samples. Journal of Clinical Child Psychology, 25(1), 25–32. doi: 10.1207/s15374424jccp2501_3 Bender, P. K., Reinholdt-Dunne, M. L., Esbjørn, B. H., & Pons, F. F. (2012). Emotion dysregulation and anxiety in children and

References adolescents: Gender differences. Personality and individual Differences, 53(3), 284–288. doi: 10.1016/j.paid.2012.03.027 Berger, A., Edelsberg, J., Bollu, V., Alvir, J. J., Dugar, A., Joshi, A. V., & Oster, G. (2011). Healthcare utilization and costs in patients beginning pharmacotherapy for generalized anxiety disorder: A retrospective cohort study. BMC Psychiatry, 11, 193 doi: 10.1186/1471–244X-11–193 Biederman, J., Faraone, S. V., Hirshfeld-Becker, D. R., Friedman, D., Robin, J. A., & Rosenbaum, J. F. (2001). Patterns of psychopathology and dysfunction in high-risk children of parents with panic disorder and major depression. American Journal of Psychiatry, 158(1), 49–57. Biederman, J., Petty, C., Hirshfeld-Becker, D., Henin, A., Faraone, S., Fraire, M., & . . . Rosenbaum, J. (2007). Developmental trajectories of anxiety disorders in offspring at high risk for panic disorder and major depression. Psychiatry Research, 153(3), 245–252. Bishop, S., Duncan, J., Brett, M., & Lawrence, A. D. (2004). Prefrontal cortical function and anxiety: Controlling attention to threat-related stimuli. Nature Neuroscience, 7(2), 184–188. doi: 10.1038/nn1173 Blair, R. J. R. (2003). Neurobiological basis of psychopathy. British Journal of Psychiatry, 182, 5–7. Blair, K. S., Geraci, M., Smith, B. W., Hollon, N., DeVido, J., Otero, M., . . . Pine, D. S. (2012). Reduced dorsal anterior cingulate cortical activity during emotional regulation and top-down attentional control in generalized social phobia, generalized anxiety disorder, and comorbid generalized social phobia/generalized anxiety disorder. Biological Psychiatry, 72(6), 476–482. doi: 10.1016/j.biopsych.2012.04.013 Bögels, S. M., Alden, L., Beidel, D. C., Clark, L., Pine, D. S., Stein, M. B., & Voncken, M. (2010). Social anxiety disorder: Questions and answers for the DSM-V. Depression and Anxiety, 27(2), 168–189. doi: 10.1002/da.20670 Bookheimer, S. Y. (2000). Methodological issues in pediatric neuroimaging. Mental Retardation and Developmental Disabilities Research Reviews, 6(3), 161–165. doi: 10.1002/1098–2779(2000)6:33.0.CO;2-W Borkovec, T. D. (1994). The nature, functions, and origins of worry. In G. L. Davey & F. Tallis (Eds.), Worrying: Perspectives on theory, assessment and treatment (pp. 5–33). Oxford, UK: John Wiley & Sons. Bouton, M. E. (2004). Context and behavioral processes in extinction. Learning & Memory, 11(5), 485–494. doi: 10.1101/lm.78804 Bowers, M. E., Choi, D. C., & Ressler, K. J. (2012). Neuropeptide regulation of fear and anxiety: Implications of cholecystokinin, endogenous opioids, and neuropeptide Y. Physiology & Behavior, 107(5), 699–710. doi: 10.1016/j.physbeh.2012.03.004 Brenner, E. (2000). Mood induction in children: Methodological issues and clinical implications. Review of General Psychology, 4(3), 264–283. doi: 10.1037/1089–2680.4.3.264 Brett, E. A. (1997). The classification of posttraumatic stress disorder. In B. A. van der Kolk, A. C. McFarlane, & L. Weisaeth (Eds.), Traumatic stress. The effects of overwhelming experience on mind, body, and society (pp. 117–128). New York, NY: Guilford. Britton, J., Lissek, S., Grillon, C., Norcross, M., & Pine, D. (2011). Development of anxiety: the role of threat appraisal and fear learning. Depression and Anxiety, 28(1), 5–17. doi: 10.1002/da.20733 Brown, T. A., Campbell, L. A., Lehman, C. L., Grisham, J. R., & Mancill, R. B. (2001). Current and lifetime comorbidity of the DSM-IV anxiety and mood disorders in a large clinical sample. Journal of Abnormal Psychology, 110(4), 585–599. Brühl, A., Herwig, U., Delsignore, A., Jäncke, L., & Rufer, M. (2012). General emotion processing in social anxiety disorder: Neural issues of cognitive control. Psychiatry Research: Neuroimaging, 212(2),108–15. doi: 10.1016/j.pscychresns.2012.05.006

549

Burdick, K. E., Endick, C. J., & Goldberg, J. F. (2005). Assessing cognitive deficits in bipolar disorder: Are self-reports valid? Psychiatry Research, 136(1), 43–50. doi: 10.1016/j.psychres.2004.12.009 Bush, G., Luu, P., & Posner, M. I. (2000). Cognitive and emotional influences in anterior cingulate cortex. Trends in Cognitive Sciences, 4(6), 215–222. doi: 10.1016/S1364–6613(00)01483–2 Campbell-Sills, L., Barlow, D. H., Brown, T. A., & Hofmann, S. G. (2006). Acceptability and suppression of negative emotion in anxiety and mood disorders. Emotion, 6(4), 587–595. doi: 10.1037/1528 –3542.6.4.587 Campbell-Sills, L., Simmons, A. N., Lovero, K. L., Rochlin, A. A., Paulus, M. P., & Stein, M. B. (2011). Functioning of neural systems supporting emotion regulation in anxiety-prone individuals. Neuroimage, 54(1), 689–696. doi: 10.1016/j.neuroimage.2010.07.041 Campos, J. J., Campos, R. G., & Barrett, K. C. (1989). Emergent themes in the study of emotional development and emotion regulation. Developmental Psychology, 25(3), 394–402. doi: 10.1037/0012–1649.25.3.394 Campos, J. J., Frankel, C. B., & Camras, L. (2004). On the nature of emotion regulation. Child Development, 75(2), 377–394. doi: 10.1111/j .1467–8624.2004.00681.x Campos, J. J., Mumme, D. L., Kermoian, R., & Campos, R. (1994). A functionalist perspective on the nature of emotion. Monographs of the Society for Research in Child Development, 59(2–3), 284–303. doi: 10.2307/1166150 Carpenter, W. B. (1876). Of attention. In, Principles of mental physiology with their applications to the training and discipline of the mind and the study of its morbid conditions (4th ed., pp. 130–147). London, UK: Henry S. King & Co. doi: 10.1037/12426–003 Carthy, T., Horesh, N., Apter, A., Edge, M. D., & Gross, J. J. (2010). Emotional reactivity and cognitive regulation in anxious children. Behaviour Research and Therapy, 48(5), 384–393. doi: 10.1016/j .brat.2009.12.013 Carthy, T., Horesh, N., Apter, A., & Gross, J. J. (2010). Patterns of emotional reactivity and regulation in children with anxiety disorders. Journal of Psychopathology and Behavioral Assessment, 32(1), 23–36. doi: 10.1007/s10862–009–9167–8 Cartwright-Hatton, S., McNicol, K., & Doubleday, E. (2006). Anxiety in a neglected population: Prevalence of anxiety disorders in pre-adolescent children. Clinical Psychology Review, 26(7), 817–833. doi: 10.1016/j.cpr.2005.12.002 Carver, C. S., & Scheier, M. F. (1998). On the self-regulation of behavior. New York, NY: Cambridge University Press. Cassano, M. C., & Zeman, J. L. (2010). Parental socialization of sadness regulation in middle childhood: The role of expectations and gender. Developmental Psychology, 46(5), 1214–1226. doi: 10.1037/a0019851 Caspi, A., Sugden, K., Moffitt, T. E., Taylor, A., Craig, I. W., Harrington, H., . . . Poulton, R. (2003, July 18). Influence of life stress on depression: Moderation by a polymorphism in the 5-HTT gene. Science, 301, 386–389. doi: 10.1126/science.1083968 Cicchetti, D., Ackerman, B. P., & Izard, C. E. (1995). Emotions and emotion regulation in developmental psychopathology. Development and Psychopathology, 7(1), 1–10. doi: 10.1017/S0954579400006301 Cicchetti, D., & Cohen, D. J. (1995). Perspectives on developmental psychopathology. In D. Cicchetti & D. J. Cohen (Eds.), Developmental Psychopathology: Theory and methods (Vol. 1, pp. 3–20). New York, NY: Wiley. Cicero, M. T. (1927). Tusculan disputations (Loeb Classical Library, Cicero Volume 18; translated by J. E. King). Cambridge, MA: Harvard University Press. Cisler, J. M., Olatunji, B. O., Feldner, M. T., & Forsyth, J. P. (2010). Emotion regulation and the anxiety disorders: An integrative review.

550

Anxiety Regulation

Journal of Psychopathology and Behavioral Assessment, 32(1), 68–82. doi: 10.1007/s10862–009–9161–1

Developmental Psychology, 25(3), 339–342. doi: 10.1037/0012–1649 .25.3.339

Clouston, T. S., & Folson, C. F. (1884). Clinical lectures on mental diseases. Philadelphia, PA: Henry C. Lea’s Son & Company.

Dunsmoor, J. E., & LaBar, K. S. (2013). Neural basis of human fear learning. In J. Armony & P. Vuilleumier (Eds.), The Cambridge handbook of human affective neuroscience (pp. 419–443). New York, NY: Cambridge University Press. doi: 10.1017/CBO9780511843716 .023

Cole, P. M., Martin, S. E., & Dennis, T. A. (2004). Emotion regulation as a scientific construct: Methodological challenges and directions for child development research. Child Development, 75(2), 317–333. doi: 10.1111/j.1467–8624.2004.00673.x Compton, S. N., Nelson, A. H., & March, J. S. (2000). Social phobia and separation anxiety symptoms in community and clinical samples of children and adolescents. Journal of the American Academy of Child & Adolescent Psychiatry, 39(8), 1040–1046. doi: 10.1097/00004583 –200008000–00020 Cook, E. W., Hawk, L. W. Davis, T. L., & Stevenson, V. E. (1991). Affective individual differences and startle reflex modulation. Journal of Abnormal Psychology, 100(1), 5–13. Costello, E. J., Mustillo, S., Erkanli, A., Keeler, G., & Angold, A. (2003). Prevalence and development of psychiatric disorders in childhood and adolescence. Archives of General Psychiatry, 60(8), 837–844. Costello, E. J., Pine, D. S., Hammen, C., March, J. S., Plotsky, P. M., Weissman, M. M., et al. (2002). Development and natural history of mood disorders. Biological Psychiatry, 52(6), 529–542. Damasio, A. R. (2000). The feeling of what happens: Body and emotion in the making of consciousness. New York, NY: Harcourt Brace. Damasio, A. R. (2001). Fundamental feelings. Nature, 413(6858), 781. Damasio, A. R., Grabowski, T. J., Bechara, A., Damasio, H., Ponto, L. L., Parvizi, J., & Hichwa, R. D. (2000). Subcortical and cortical brain activity during the feeling of self-generated emotions. Nature Neuroscience, 3(10), 1049–1056. Davidson, R. J. (1998). Affective style and affective disorders: Perspectives from affective neuroscience. Cognition and Emotion, 12(3), 307–330. doi: 10.1080/026999398379628 DeCicco, J. M., Solomon, B., & Dennis, T. A. (2012). Neural correlates of cognitive reappraisal in children: An ERP study. Developmental Cognitive Neuroscience, 2(1), 70–80. doi: 10.1016/j.dcn.2011.05.009 Decker, M. L., Turk, C. L., Hess, B., & Murray, C. E. (2008). Emotion regulation among individuals classified with and without generalized anxiety disorder. Journal of Anxiety Disorders, 22(3), 485–494. doi: 10.1016/j.janxdis.2007.04.002 Degroot, A., & Nomikos, G. G. (2005). Fluoxetine disrupts the integration of anxiety and aversive memories. Neuropsychopharmacology, 30, 391−400. DelCarmen-Wiggins, R. (2008). Introduction to the special section: Transformative research on emotion regulation and dysregulation. Child Development Perspectives, 2(3), 121–123. doi: 10.1111/j.1750 –8606.2008.00053.x Denham, S. (2010). Emotion regulation: Now you see it, now you don’t. Emotion Review, 2(3), 297–299. doi: 10.1177/1754073909357448 Dennis, T. A., & Hajcak, G. (2009). The late positive potential: A neurophysiological marker for emotion regulation in children. Journal of Child Psychology and Psychiatry, 50(11), 1373–1383. doi: 10.1111/j .1469–7610.2009.02168.x Dias, B., Banerjee, S., Goodman, J., & Ressler, K. (2013). Towards new approaches to disorders of fear and anxiety. Current Opinion in Neurobiology. 23(3):346–52. doi: 10.1016/j.conb.2013.01.013 Dillon, D. G., Deveney, C. M., & Pizzagalli, D. A. (2011). From basic processes to real-world problems: How research on emotion and emotion regulation can inform understanding of psychopathology, and vice versa. Emotion Review, 3(1), 74–82. doi: 10.1177/1754073910380973 Dodge, K. A. (1989). Coordinating responses to aversive stimuli: Introduction to a special section on the development of emotion regulation.

Eastabrook, J. M., Flynn, J. J., & Hollenstein, T. (2014). Internalizing symptoms in female adolescents: Associations with emotional awareness and emotion regulation. Journal of Child and Family Studies. 23, 487–496. doi: 10.1007/s10826–012–9705-y Edinger, K. L., & Frye, C. A. (2006). Intrahippocampal administration of an androgen receptor agonist, flutamide, can increase anxiety-like behavior in intact and DHT-replaced male rats. Hormones and Behavior, 50, 216−222. Erickson, S. R., Guthrie, S., VanEtten-Lee, M., Himle, J., Hoffman, J., Santos, S. F., . . . Abelson, J. L. (2009). Severity of anxiety and work-related outcomes of patients with anxiety disorders. Depression and Anxiety, 26(12), 1165–1171. doi: 10.1002/da.20624 Esbjørn, B. H., Bender, P. K., Reinholdt-Dunne, M. L., Munck, L. A., & Ollendick, T. H. (2012). The development of anxiety disorders: Considering the contributions of attachment and emotion regulation. Clinical Child and Family Psychology Review, 15(2), 129–143. doi: 10.1007/s10567–011–0105–4 Etkin, A., Egner, T., Peraza, D., Kandel, E., & Hirsch, J. (2006). Resolving emotional conflict: A role for the rostral anterior cingulate cortex in modulating activity in the amygdala. Neuron, 51(6), 871–882. Etkin, A., Prater, K. E., Hoeft, F., Menon, V., & Schatzberg, A. F. (2010). Failure of anterior cingulate activation and connectivity with the amygdala during implicit regulation of emotional processing in generalized anxiety disorder. American Journal of Psychiatry, 167(5), 545–554. doi: 10.1176/appi.ajp. 2009.09070931 Etkin, A., & Schatzberg, A. F. (2011). Common abnormalities and disorder-specific compensation during implicit regulation of emotional processing in generalized anxiety and major depressive disorders. American Journal of Psychiatry, 168(9), 968–978. doi: 10.1176 /appi.ajp. 2011.10091290 Fani, N., Jovanovic, T., Ely, T. D., Bradley, B., Gutman, D., Tone, E. B., & Ressler, K. J. (2012). Neural correlates of attention bias to threat in post-traumatic stress disorder. Biological Psychology, 90(2), 134–142. doi: 10.1016/j.biopsycho.2012.03.001 Farmer, A., & Kashdan, T. B. (2012). Social anxiety and emotion regulation in daily life: Spillover effects on positive and negative social events. Cognitive Behaviour Therapy, 41(2), 152–162. doi: 10.1080/16506073 .2012.666561 Fehm, L., Beesdo, K., Jacobi, F., & Fiedler, A. (2008). Social anxiety disorder above and below the diagnostic threshold: Prevalence, comorbidity and impairment in the general population. Social Psychiatry and Psychiatric Epidemiology, 43(4), 257–265. doi: 10.1007/s00127 –007–0299–4 Feldner, M. T., Zvolensky, M. J., Stickle, T. R., Bonn-Miller, M. O., & Leen-Feldner, E. W. (2006). Anxiety sensitivity-physical concerns as a moderator of the emotional consequences of emotion suppression during biological challenge: An experimental test using individual growth curve analysis. Behaviour Research and Therapy, 44(2), 249–272. doi: 10.1016/j.brat.2005.02.003 Ferrell, C. B. (2009). Reengineering clinical research science: A focus on translational research. Behavior Modification, 33(1), 7–23. doi: 10.1177/0145445508322616 Fialko, L., Bolton, D., & Perrin, S. (2012). Applicability of a cognitive model of worry to children and adolescents. Behaviour Research and Therapy, 50(5), 341–349. doi: 10.1016/j.brat.2012.02.003

References Fichter, M. M., Quadflieg, N. N., Fischer, U. C., & Kohlboeck, G. G. (2010). Twenty-five-year course and outcome in anxiety and depression in the Upper Bavarian longitudinal community study. Acta Psychiatrica Scandinavica, 122(1), 75–85. doi: 10.1111/j.1600–0447 .2009.01512.x Filipowicz, A., Barsade, S., & Melwani, S. (2011). Understanding emotional transitions: The interpersonal consequences of changing emotions in negotiations. Journal of Personality and Social Psychology, 101(3), 541–556. doi: 10.1037/a0023545 Fox, N. A. (1994). Dynamic cerebral processes underlying emotion regulation. Monographs of the Society for Research in Child Development, 59(2–3), 152–166, 250–283. doi: 10.2307/1166143 Francis, G., Last, C. G., & Strauss, C. C. (1992). Avoidant disorder and social phobia in children and adolescents. Journal of the American Academy of Child & Adolescent Psychiatry, 31(6), 1086–1089. doi: 10.1097/00004583–199211000–00014 Fyer, A. J. (1998). Current approaches to etiology and pathophysiology of specific phobia. Biological Psychiatry, 44(12), 1295–1304. Fyer, A. J., Mannuzza, S., Chapman, T. F., Martin, L. Y., & Klein, D. F. (1995). Specificity in familial aggregation of phobic disorders. Archives of General Psychiatry, 52(7), 564–573. Gaffrey, M. S., Luby, J. L., Belden, A. C., Hirshberg, J. S., Volsch, J., & Barch, D. M. (2011). Association between depression severity and amygdala reactivity during sad face viewing in depressed preschoolers: An fMRI study. Journal of Affective Disorders, 129(1–3), 364–370. doi: 10.1016/j.jad.2010.08.031 Gao, Y., Raine, A., Venables, P., Dawson, M., & Mednick, S. (2010). The development of skin conductance fear conditioning in children from ages 3 to 8 years. Developmental Science, 13(1), 201–212. doi: 10.1111/j.1467–7687.2009.00874.x Garnefski, N., Kraaij, V., & van Etten, M. (2005). Specificity of relations between adolescents’ cognitive emotion regulation strategies and internalizing and externalizing psychopathology. Journal of Adolescence, 28(5), 619–631. doi: 10.1016/j.adolescence.2004.12.009 Goenjian, A. K., Pynoos, R. S., Steinberg, A. M., Najarian, L. M., Asarnow, J. R., Karayan, I., Ghurabi, M., & Fairbanks, L. A. (1995). Psychiatric comorbidity in children after the 1988 earthquake in Armenia. Journal of the American Academy of Child and Adolescent Psychiatry, 34(9), 1174–1184. Goldin, P. R., & Gross, J. J. (2010). Effects of mindfulness-based stress reduction (MBSR) on emotion regulation in social anxiety disorder. Emotion, 10(1), 83–91. doi: 10.1037/a0018441 Gordon, J., King, N., Gullone, E., Muris, P., & Ollendick, T. H. (2007). Nighttime fears of children and adolescents: Frequency, content, severity, harm expectations, disclosure, and coping behaviours. Behaviour Research and Therapy, 45(10), 2464–2472. doi: 10.1016 /j.brat.2007.03.013 Gosling, S. D., John, O. P., Craik, K. H., & Robins, R. W. (1998). Do people know how they behave? Self-reported act frequencies compared with on-line codings by observers. Journal of Personality and Social Psychology, 74(5), 1337–1349. doi: 10.1037/0022–3514.74 .5.1337 Grados, M. A. (2010). The genetics of obsessive-compulsive disorder and Tourette syndrome: An epidemiological and pathway-based approach for gene discovery. Journal of the American Academy of Child & Adolescent Psychiatry, 49(8), 810–819. doi: 10.1016/j.jaac.2010.04.009 Grados, M. A., Riddle, M. A., Samuels, J. F., Liang, K. Y., Hoehn-Saric, R., Bienvenu, O. J., . . . Nestadt, G. (2001). The familial phenotype of obsessive-compulsive disorder in relation to tic disorders: The Hopkins OCD family study. Biological Psychiatry, 50(8), 559–565. Gray, J. A., & McNaughton, N. (2000). The neuropsychology of anxiety (2nd ed.). New York, NY: Oxford University Press.

551

Grigorenko, E. L., & Cicchetti, D. (2012). Genomic sciences for developmentalists: The current state of affairs. Development and Psychopathology, 24(4), 1157–1164. doi: 10.1017/S0954579412000612 Grolnick, W. S., McMenamy, J. M., & Kurowski, C. O. (2006). Emotional self-regulation in infancy and toddlerhood. In L. Balter & C. S. Tamis-LeMonda (Eds.), Child psychology: A handbook of contemporary issues (2nd ed., pp. 3–25). New York, NY: Psychology Press. Gross, J. J. (2013). Emotion regulation: Taking stock and moving forward. Emotion. 13(3):359–65. doi: 10.1037/a0032135 Gross, J. J., & Barrett, L. (2011). Emotion generation and emotion regulation: One or two depends on your point of view. Emotion Review, 3(1), 8–16. doi: 10.1177/1754073910380974 Gross, J. J., & Muñoz, R. F. (1995). Emotion regulation and mental health. Clinical Psychology: Science and Practice, 2(2), 151–164. doi: 10.1111/j.1468–2850.1995.tb00036.x Gullone, E., Hughes, E., King, N., & Tonge, B. (2010). The normative development of emotion regulation strategy use in children and adolescents: A 2-year follow-up study. Journal of Child Psychology and Psychiatry, and Allied Disciplines, 51(5), 567–574. doi: 10.1111/j .1469–7610.2009.02183.x Gyurak, A., Gross, J. J., & Etkin, A. (2011). Explicit and implicit emotion regulation: A dual-process framework. Cognition and Emotion, 25(3), 400–412. doi: 10.1080/02699931.2010.544160 Hannesdottir, D., & Ollendick, T. H. (2007). The role of emotion regulation in the treatment of child anxiety disorders. Clinical Child and Family Psychology Review, 10(3), 275–293. doi: 10.1007/s10567 –007–0024–6 Henry, T. S. (1917). The education and control of the emotions. Journal of Educational Psychology, 8(7), 407–415. doi: 10.1037/h0072203 Hartley, C. A., & Phelps, E. A. (2010). Changing fear: The neurocircuitry of emotion regulation. Neuropsychopharmacology, 35(1), 136–146. doi: 10.1038/npp. 2009.121 Hoeksma, J. B., Oosterlaan, J., & Schipper, E. M. (2004). Emotion regulation and the dynamics of feelings: A conceptual and methodological framework. Child Development, 75(2), 354–360. doi: 10.1111/j.1467–8624.2004.00677.x Hofmann, S. G., Asnaani, A., Vonk, I. J., Sawyer, A. T., & Fang, A. (2012). The efficacy of cognitive behavioral therapy: A review of meta-analyses. Cognitive Therapy and Research, 36(5), 427–440. doi: 10.1007/s10608–012–9476–1 Hofstatter, L. L., Smolik, E. A., & Busch, A. K. (1945). Prefrontal lobotomy in treatment of chronic psychoses with special reference to section of the orbital areas only. Archives of Neurology & Psychiatry, 53(2), 125–130. Insel, T.R., & Wang, P.S. (2010). Rethinking mental illness. Journal of the American Medical Association, 303(19), 1970–1971. Jackowski, A., de Araújo, C., de Lacerda, A., Mari, J., & Kaufman, J. (2009). Neurostructural imaging findings in children with post-traumatic stress disorder: Brief review. Psychiatry and Clinical Neurosciences, 63(1), 1–8. doi: 10.1111/j.1440–1819.2008.01906.x Jacob, M. L., Thomassin, K., Morelen, D., & Suveg, C. (2011). Emotion regulation in childhood anxiety. In D. McKay & E. A. Storch (Eds.), Handbook of child and adolescent anxiety disorders (pp. 171–185). New York, NY: Springer Science + Business Media. doi: 10.1007/978 –1–4419–7784–7_12 Jarcho, J. M., Fox, N. A., Pine, D. S., Etkin, A., Leibenluft, E., Shechner, T., & Ernst, M. (2013). The neural correlates of emotionbased cognitive control in adults with early childhood behavioral inhibition. Biological Psychology, 92(2), 306–314. doi: 10.1016/j .biopsycho.2012.09.008 Jastrow, J. (1915). The emotions and conduct. In, Character and temperament (pp. 104–173). New York, NY: D Appleton & Company. doi: 10.1037/10899–003

552

Anxiety Regulation

Jeffrey, G. R. (1911). On the mechanism of mental processes, with special reference to emotional control. Journal of Mental Science, 57(236), 85–93. John, O., & Gross, J. (2004). Healthy and unhealthy emotion regulation: personality processes, individual differences, and life span development. Journal of Personality, 72(6), 1301–1333. Johnson, B. B. (1928). The three-year-old at school. Childhood Education, 5, 57–60. Johnson, D. R. (2009). Emotional attention set-shifting and its relationship to anxiety and emotion regulation. Emotion, 9(5), 681–690. doi: 10.1037/a0017095 Kalin, N. H., Shelton, S. E., & Davidson, R. J. (2004). The role of the central nucleus of the amygdala in mediating fear and anxiety in the primate. Journal of Neuroscience, 24, 5506−5515. Kalueff, A. V., & Nutt, D. J. (2007). Role of GABA in anxiety and depression. Depression and Anxiety, 24(7), 495–517. doi: 10.1002/da.20262 Kappas, A. (2011). Emotion and regulation are one! Emotion Review, 3(1), 17–25. doi: 10.1177/1754073910380971 Kar, B., Vijay, N., & Mishra, S. (2013). Development of cognitive and affective control networks and decision making. Progress in Brain Research, 202, 347–368. doi: 10.1016/B978–0–444–62604–2.00018–6 Kashdan, T. B., & Steger, M. F. (2006). Expanding the topography of social anxiety: An experience-sampling assessment of positive emotions, positive events, and emotion suppression. Psychological Science, 17(2), 120–128. doi: 10.1111/j.1467–9280.2006.01674.x Katzelnick, D. J., Kobak, K. A., DeLeire, T., Henk, H. J., Greist, J. H., Davidson, J. R., . . . Helstad, C. P. (2001). Impact of generalized social anxiety disorder in managed care. American Journal of Psychiatry, 158(12), 1999–2007. Kendall, P. C. (2012). Guiding theory for therapy with children and adolescents. In P. C. Kendall (Ed.), Child and adolescent therapy: Cognitive-behavioral procedures (4th ed., pp. 3–24). New York, NY: Guilford Press. Kendler, K. S. (2001). Twin studies of psychiatric illness: an update. Archives of General Psychiatry, 58(11), 1005–1014. Kessler, R. C., Petukhova, M., Sampson, N. A., Zaslavsky, A. M., & Wittchen, H. (2012). Twelve-month and lifetime prevalence and lifetime morbid risk of anxiety and mood disorders in the United States. International Journal of Methods in Psychiatric Research, 21(3), 169–184. doi: 10.1002/mpr.1359 Keyes, C. M. (2005). Mental illness and/or mental health? Investigating axioms of the complete state model of health. Journal of Consulting and Clinical Psychology, 73(3), 539–548. doi: 10.1037/0022–006X.73.3.539 Kim, J., & Deater-Deckard, K. (2011). Dynamic changes in anger, externalizing and internalizing problems: Attention and regulation. Journal of Child Psychology and Psychiatry, 52(2), 156–166. doi: 10.1111/j.1469–7610.2010.02301.x Klein, R. G., & Pine, D. S. (2002). Anxiety disorders. In M. Rutter, E. Taylor, & L. Hersov (Eds.), Child and adolescent psychiatry: Modern approaches (4th ed., pp. 486–509). London, UK: Blackwell Publishers. Kovacs, M., Joormann, J., & Gotlib, I. H. (2008). Emotion (dys)regulation and links to depressive disorders. Child Development Perspectives, 2(3), 149–155. doi: 10.1111/j.1750–8606.2008.00057.x Koval, P., & Kuppens, P. (2012). Changing emotion dynamics: Individual differences in the effect of anticipatory social stress on emotional inertia. Emotion, 12(2), 256–267. doi: 10.1037/a0024756 Kring, A. M., & Sloan, D. M. (2010). Emotion regulation and psychopathology: A transdiagnostic approach to etiology and treatment. New York, NY: Guilford Press. Krueger, R. F. (1999). The structure of common mental disorders. Archives of General Psychiatry, 56(10), 921–926. doi: 10.1001/archpsyc .56.10.921

Kupfer, D. J., & Regier, D. A. (2011). Neuroscience, clinical evidence, and the future of psychiatric classification in DSM-5. American Journal of Psychiatry, 168(7), 2–3. Lamm, C., White, L. K., McDermott, J., & Fox, N. A. (2012). Neural activation underlying cognitive control in the context of neutral and affectively charged pictures in children. Brain and Cognition, 79(3), 181–187. doi: 10.1016/j.bandc.2012.02.013 Langley, A. K., Bergman, R. L., McCracken, J., & Piacentini, J. C. (2004). Impairment in childhood anxiety disorders: Preliminary examination of the child anxiety impact scale-parent version. J Child Adolesc Psychopharmacol, 14(1), 105–114. Larsen, J. K., Vermulst, A. A., Eisinga, R., English, T., Gross, J. J., Hofman, E., & . . . Engels, R. E. (2012). Social coping by masking? Parental support and peer victimization as mediators of the relationship between depressive symptoms and expressive suppression in adolescents. Journal of Youth and Adolescence, 41(12), 1628–1642. doi: 10.1007/s10964–012–9782–7 Leach, L., & Butterworth, P. (2012). The effect of early onset common mental disorders on educational attainment in Australia. Psychiatry Research, 199(1), 51–57. doi: 10.1016/j.psychres.2012.03.040 Lebeau, R. T., Glenn, D. E., Hanover, L. N., Beesdo-Baum, K., Wittchen, H., & Craske, M. G. (2012). A dimensional approach to measuring anxiety for DSM-5. International Journal of Methods in Psychiatric Research, 21(4), 258–272. doi: 10.1002/mpr.1369 LeDoux, J. E., Cicchetti, P., Xagoraris, A., & Romanski, L. M. (1990). The lateral amygdaloid nucleus: Sensory interface of the amygdala in fear conditioning. Journal of Neuroscience, 10, 1062−1069. LeDoux, J. E., Iwata, J., Cicchetti, P., & Reis, D. J. (1988). Different projections of the central amygdaloid nucleus mediate autonomic and behavioral correlates of conditioned fear. Journal of Neuroscience, 8, 2517−2529. Lee, H., Heller, A. S., van Reekum, C. M., Nelson, B., & Davidson, R. J. (2012). Amygdala–prefrontal coupling underlies individual differences in emotion regulation. Neuroimage, 62(3), 1575–1581. doi: 10.1016/j.neuroimage.2012.05.044 Legerstee, J., Garnefski, N., Jellesma, F., Verhulst, F., & Utens, E. (2010). Cognitive coping and childhood anxiety disorders. European Child & Adolescent Psychiatry, 19(2), 143–150. doi: 10.1007/s00787 –009–0051–6 Levesque, J., Joanette, Y., Mensour, B., Beaudoin, G., Leroux, J-M., Bourgouin, P., & Beauregard, M. (2004). Neural basis of emotional self-regulation in childhood. Neuroscience, 129, 361–369. Lewinsohn, P. M., Holm-Denoma, J. M., Small, J. W., Seeley, J. R., & Joiner, T. R. (2008). Separation anxiety disorder in childhood as a risk factor for future mental illness. Journal of the American Academy of Child & Adolescent Psychiatry, 47(5), 548–555. doi: 10.1097/CHI .0b013e31816765e7 Lipsitz, J. D., Martin, L. Y., Mannuzza, S., Chapman, T. F., Liebowitz, M. R., Klein, D. F., & Fyer, A. J. (1994). Childhood separation anxiety disorder in patients with adult anxiety disorders. American Journal of Psychiatry, 151(6), 927–929. Lissek, S., Powers, A. S., McClure, E. B., Phelps, E. A., Woldehawariat, G., Grillon, C., & Pine, D. S. (2005). Classical fear conditioning in the anxiety disorders: A meta-analysis. Behaviour Research and Therapy, 43(11), 1391–1424. doi: 10.1016/j.brat.2004.10.007 Lloyd, S., & Moreau, N. (2011). Pursuit of a “normal life”: Mood, anxiety, and their disordering. Medical Anthropology, 30(6), 591–609. doi: 10.1080/01459740.2011.596175 Lonsdorf, T. B., Weike, A. I., Nikamo, P., Schalling, M., Hamm, A. O., & Öhman, A. (2009). Genetic gating of human fear learning and extinction: Possible implications for gene–environment interaction in anxiety disorder. Psychological Science, 20(2), 198–206. doi: 10.1111/j.1467–9280.2009.02280.x

References Lougheed, J. P., & Hollenstein, T. (2012). A limited repertoire of emotion regulation strategies is associated with internalizing problems in adolescence. Social Development, 21(4), 704–721. doi: 10.1111/j.1467 –9507.2012.00663.x MacLeod, C., & Bucks, R. S. (2011). Emotion regulation and the cognitive-experimental approach to emotional dysfunction. Emotion Review, 3(1), 62–73. doi: 10.1177/1754073910380970 McGrath, L. M., Weill, S., Robinson, E. B., MacRae, R., & Smoller, J. W. (2012). Bringing a developmental perspective to anxiety genetics. Development and Psychopathology, 24(4), 1179–1193. doi: 10.1017 /S0954579412000636 McDermott, J. M., Perez-Edgar, K., Henderson, H. A., Chronis-Tuscano, A., Pine, D. S., & Fox, N. A. (2009). A history of childhood behavioral inhibition and enhanced response monitoring in adolescence are linked to clinical anxiety. Biological Psychiatry, 65(5), 445–448. doi: 10.1016/j.biopsych.2008.10.043 McEvoy, P. M., & Brans, S. (2013). Common versus unique variance across measures of worry and rumination: Predictive utility and mediational models for anxiety and depression. Cognitive Therapy and Research, 37(1), 183–196. doi: 10.1007/s10608–012–9448–5 McLaughlin, K. A., & Nolen-Hoeksema, S. (2011). Rumination as a transdiagnostic factor in depression and anxiety. Behaviour Research and Therapy, 49(3), 186–193. doi: 10.1016/j.brat.2010.12.006 McLaughlin, K., & Nolen-Hoeksema, S. (2012). Interpersonal stress generation as a mechanism linking rumination to internalizing symptoms in early adolescents. Journal of Clinical Child and Adolescent Psychology, 41(5), 584–597. McRae, K., Gross, J. J., Weber, J., Robertson, E. R., Sokol-Hessner, P., Ray, R. D., . . . Ochsner, K. N. (2012). The development of emotion regulation: An fMRI study of cognitive reappraisal in children, adolescents and young adults. Social Cognitive and Affective Neuroscience, 7(1), 11–22. doi: 10.1093/scan/nsr093 Marnane, C., & Silove, D. (2013). DSM-5 allows separation anxiety disorder to grow up. Australian and New Zealand Journal of Psychiatry, 47(1), 12–15. doi: 10.1177/0004867412461285 Martin, A., & Leslie, D. (2003). Psychiatric inpatient, outpatient, and medication utilization and costs among privately insured youths, 1997–2000. American Journal of Psychiatry, 160(4), 757–764. Masten, C.L., Eisenberger, N.I., Borofsky, L.A., et al. (2009). Neural correlates of social exclusion during adolescence: understanding the distress of peer rejection. Social Cognitive and Affective Neuroscience, 4(2), 143–57. Mennin, D. S. (2006). Emotion regulation therapy: An integrative approach to treatment-resistant anxiety disorders. Journal of Contemporary Psychotherapy, 36(2), 95–105. doi: 10.1007/s10879–006 –9012–2 Mennin, D. S., McLaughlin, K. A., & Flanagan, T. J. (2009). Emotion regulation deficits in generalized anxiety disorder, social anxiety disorder, and their co-occurrence. Journal of Anxiety Disorders, 23(7), 866–871. doi: 10.1016/j.janxdis.2009.04.006 Meyer-Lindenberg, A., Domes, G., Kirsch, P., & Heinrichs, M. (2011). Oxytocin and vasopressin in the human brain: Social neuropeptides for translational medicine. Nature Reviews. Neuroscience, 12(9), 524–538. doi: 10.1038/nrn3044 Mogg, K., & Bradley, B. (1998). A cognitive-motivational analysis of anxiety. Behaviour Research and Therapy, 36(9), 809–848. Monk, C. S., Grillon, C., Baas, J. P., McClure, E. B., Nelson, E. E., Zarahn, E., & . . . Pine, D. S. (2003). A neuroimaging method for the study of threat in adolescents. Developmental Psychobiology, 43(4), 359–366. doi: 10.1002/dev.10146 Monk, C. S., Nelson, E. E., McClure, E. B., Mogg, K., Bradley, B. P., Leibenluft, E., . . . Pine, D. S. (2006). Ventrolateral prefrontal cortex activation and attentional bias in response to angry faces in

553

adolescents with generalized anxiety disorder. American Journal of Psychiatry, 163(6), 1091–1097. doi: 10.1176/appi.ajp. 163.6.1091 Moore, G. (1852). The power of the soul over the body, considered in relation to health and morals. New York, NY: Harper & Brothers. doi: 10.1037/12046–029 Morris, S. E., & Cuthbert, B. N. (2012). Research domain criteria: Cognitive systems, neural circuits, and dimensions of behavior. Dialogues in Clinical Neuroscience, 14(1), 29–37. Moscovitch, D. A., Antony, M. M., & Swinson, R. P. (2009). Exposure-based treatments for anxiety disorders: Theory and process. In M. M. Antony & M. B. Stein (Eds.), Oxford handbook of anxiety and related disorders (pp. 461–475). New York, NY: Oxford University Press. Moscovitch, D. A., Gavric, D. L., Senn, J. M., Santesso, D. L., Miskovic, V., Schmidt, L. A., . . . Antony, M. M. (2012). Changes in judgment biases and use of emotion regulation strategies during cognitive-behavioral therapy for social anxiety disorder: Distinguishing treatment responders from nonresponders. Cognitive Therapy and Research, 36(4), 261–271. doi: 10.1007/s10608–011–9371–1 Muris, P., Merckelbach, H., Ollendick, T. H., King, N. J., & Bogie, N. (2001). Children’s nighttime fears: Parent–child ratings of frequency, content, origins, coping behaviors and severity. Behaviour Research and Therapy, 39(1), 13–28. doi: 10.1016/S0005–7967(99)00155–2 Newman, M., Castonguay, L., Borkovec, T., Fisher, A., Boswell, J., Szkodny, L., & Nordberg, S. (2011). A randomized controlled trial of cognitive-behavioral therapy for generalized anxiety disorder with integrated techniques from emotion-focused and interpersonal therapies. Journal of Consulting and Clinical Psychology, 79(2), 171–181. doi: 10.1037/a0022489 Newman, M. G., & Llera, S. J. (2011). A novel theory of experiential avoidance in generalized anxiety disorder: A review and synthesis of research supporting a contrast avoidance model of worry. Clinical Psychology Review, 31(3), 371–382. doi: 10.1016/j.cpr.2011.01.008 Niles, A., Lebeau, R., Liao, B., Glenn, D., & Craske, M. (2012). Dimensional indicators of generalized anxiety disorder severity for DSM-V. Journal of Anxiety Disorders, 26(2), 279–286. doi: 10.1016/j .janxdis.2011.12.007 Nolen-Hoeksema, S. (2012). Emotion regulation and psychopathology: The role of gender. Annual Review of Clinical Psychology, 8, 61–87. doi: 10.1146/annurev-clinpsy-032511–143109 Nolte, T., Guiney, J., Fonagy, P., Mayes, L. C., & Luyten, P. (2011). Interpersonal stress regulation and the development of anxiety disorders: An attachment-based developmental framework. Frontiers in Behavioral Neuroscience, 5, 55. doi: 10.3389/fnbeh.2011.00055 Novin, S., Rieffe, C., Banerjee, R., Miers, A. C., & Cheung, J. (2011). Anger response styles in Chinese and Dutch children: A socio-cultural perspective on anger regulation. British Journal of Developmental Psychology, 29(4), 806–822. doi: 10.1348/2044–835X.002010 Oberle, E., Schonert-Reichl, K. A., Lawlor, M., & Thomson, K. C. (2012). Mindfulness and inhibitory control in early adolescence. Journal of Early Adolescence, 32(4), 565–588. doi: 10.1177/0272431611403741 Ollendick, T. H., & Hirshfeld-Becker, D. R. (2002). The developmental psychopathology of social anxiety disorder. Biological Psychiatry, 51(1), 44–58. doi: 10.1016/S0006–3223(01)01305–1 Opitz, P., Rauch, L., Terry, D., & Urry, H. (2012). Prefrontal mediation of age differences in cognitive reappraisal. Neurobiology of Aging, 33(4), 645–655. doi: 10.1016/j.neurobiolaging.2010.06.004 Panksepp, J. (1998). Affective neuroscience: The foundations of human and animal emotions. New York, NY: Oxford University Press. Panksepp, J. (2011). Toward a cross-species neuroscientific understanding of the affective mind: Do animals have emotional feelings? American Journal of Primatology, 73(6), 545–561. doi: 10.1002/ajp. 20929

554

Anxiety Regulation

Panksepp, J. (2012). What is an emotional feeling? Lessons about affective origins from cross-species neuroscience. Motivation and Emotion, 36(1), 4–15. doi: 10.1007/s11031–011–9232-y Pattwell, S., Duhoux, S., Hartley, C., Johnson, D., Jing, D., Elliott, M., . . . Lee, F. (2012). Altered fear learning across development in both mouse and human. Proceedings of the National Academy of Sciences of the United States of America, 109(40), 16318–16323. doi: 10.1073/pnas.1206834109 Pauls, D. L., Alsobrook, J. P., 2nd,, Goodman, W., Rasmussen, S., & Leckman, J. F. (1995). A family study of obsessive-compulsive disorder. American Journal of Psychiatry, 152(1), 76–84. Pauls, D. L., Leckman, J. F., Towbin, K. E., Zahner, G. E., & Cohen, D. J. (1986). A possible genetic relationship exists between Tourette’s syndrome and obsessive-compulsive disorder. Psychopharmacology Bulletin, 22(3), 730–733. Petersen, I. T., Bates, J. E., Goodnight, J. A., Dodge, K. A., Lansford, J. E., Pettit, G. S., & . . . Dick, D. M. (2012). Interaction between serotonin transporter polymorphism (5-HTTLPR) and stressful life events in adolescents’ trajectories of anxious/depressed symptoms. Developmental Psychology, 48(5), 1463–1475. doi: 10.1037/a0027471 Peterson, B. S., Pine, D. S., Cohen, P., & Brook, J. S. (2001). Prospective, longitudinal study of tic, obsessive-compulsive, and attention-deficit/hyperactivity disorders in an epidemiological sample. Journal of the American Academy of Child and Adolescent Psychiatry, 40(6), 685–695. Pfeifer, J. H., & Blakemore, S. (2012). Adolescent social cognitive and affective neuroscience: Past, present, and future. Social Cognitive and Affective Neuroscience, 7(1), 1–10. doi: 10.1093/scan/nsr099 Philippot, P., Baeyens, C., Douilliez, C., & Francart, B. (2004). Cognitive regulation of emotion: Application to clinical disorders. In P. Philippot & R. S. Feldman (Eds.), The regulation of emotion (pp. 71–97). Mahwah, NJ: Lawrence Erlbaum Associates Publishers. Pine, D. S., & Cohen, J. A. (2002). Trauma in children and adolescents: risk and treatment of psychiatric sequelae. Biological Psychiatry, 51(7), 519–531. Pine, D. S., Cohen, P., Gurley, D., Brook, J., & Ma, Y. (1998). The risk for early-adulthood anxiety and depressive disorders in adolescents with anxiety and depressive disorders. Archives of General Psychiatry, 55(1), 56–64. Pine, D. S., & Fox, N. A. (2015). Childhood antecedents and risk for adult mental disorders. Annual Review of Psychology, 66, 459–485. Pine, D. S., Fyer, A., Grun, J., Phelps, E. A., Szeszko, P. R., Koda, V., . . . Bilder, R. M. (2001). Methods for developmental studies of fear conditioning circuitry. Biological Psychiatry, 50(3), 225–228. doi: 10.1016/S0006–3223(01)01159–3 Pine, D. S., Helfinstein, S. M., Bar-Haim, Y., Nelson, E. & Fox, N. A. (2009). Neuropsychopharmacology Reviews, 34, 213–228. doi: 10.1038/npp. 2008.113 Pitskel, N. B., Bolling, D. Z., Kaiser, M. D., Crowley, M. J., & Pelphrey, K. A. (2011). How grossed out are you? The neural bases of emotion regulation from childhood to adolescence. Developmental Cognitive Neuroscience, 1(3), 324–337. doi: 10.1016/j.dcn.2011.03.004 Prins, M., Bosmans, J., Verhaak, P., van der Meer, K., van Tulder, M., van Marwijk, H., & . . . Bensing, J. (2011). The costs of guideline-concordant care and of care according to patients’ needs in anxiety and depression. Journal of Evaluation in Clinical Practice, 17(4), 537–546. doi: 10.1111/j.1365–2753.2010.01490.x Quirk, G. J., Garcia, R., & González-Lima, F. (2006). Prefrontal mechanisms in extinction of conditioned fear. Biological Psychiatry, 60(4), 337–343. doi: 10.1016/j.biopsych.2006.03.010 Quoidbach, J., Berry, E. V., Hansenne, M., & Mikolajczak, M. (2010). Positive emotion regulation and well-being: Comparing the impact

of eight savoring and dampening strategies. Personality and Individual Differences, 49(5), 368–373. doi: 10.1016/j.paid.2010.03.048 Ramsawh, H. J., Weisberg, R. B., Dyck, I., Stout, R., & Keller, M. B. (2011). Age of onset, clinical characteristics, and 15-year course of anxiety disorders in a prospective, longitudinal, observational study. Journal of Affective Disorders, 132(1–2), 260–264. doi: 10.1016/j.jad .2011.01.006 ˝ Rapee, R. M., Bogels, S. M., van der Sluis, C. M., Craske, M. G., & Ollendick, T. (2012). Annual Research Review: Conceptualising functional impairment in children and adolescents. Journal of Child Psychology and Psychiatry, 53(5), 454–468. doi: 10.1111/j.1469 –7610.2011.02479.x Raznahan, A., Shaw, P., Lalonde, F., Stockman, M., Wallace, G. L., Greenstein, D., . . . Giedd, J. N. (2011). How does your cortex grow? Journal of Neuroscience, 31(19), 7174–7177. doi: 10.1523 /JNEUROSCI.0054–11.2011 Redcay, E., Dodell-Feder, D., Pearrow, M. J., Mavros, P. L., Kleiner, M., Gabrieli, J. E., & Saxe, R. (2010). Live face-to-face interaction during fMRI: A new tool for social cognitive neuroscience. Neuroimage, 50(4), 1639–1647. doi: 10.1016/j.neuroimage.2010.01.052 Revicki, D. A., Travers, K., Wyrwich, K. W., Svedsäter, H., Locklear, J., Mattera, M., . . . Montgomery, S. (2012). Humanistic and economic burden of generalized anxiety disorder in North America and Europe. Journal of Affective Disorders, 140(2), 103–112. doi: 10.1016/j.jad.2011.11.014 Roberson-Nay, R., Klein, D. F., Klein, R. G., Mannuzza, S., Moulton, J., Guardino, M., & Pine, D. S. (2010). Carbon dioxide hypersensitivity in separation-anxious offspring of parents with panic disorder. Biological Psychiatry, 67(12), 1171–1177. doi: 10.1016/j.biopsych.2009.12.014 Roberson-Nay, R., Eaves, L., Hettema, J., Kendler, K., & Silberg, J. (2012). Childhood separation anxiety disorder and adult onset panic attacks share a common genetic diathesis. Depression and Anxiety, 29(4), 320–327. doi: 10.1002/da.21931 Rodebaugh, T. L., & Heimberg, R. G. (2008). Emotion regulation and the anxiety disorders: Adopting a self-regulation perspective. In A. Vingerhoets, I. Nyklíˇcek, & J. Denollet (Eds.), Emotion regulation: Conceptual and clinical issues (pp. 140–149). New York, NY: Springer Science + Business Media. doi: 10.1007/978–0–387–29986–0_9 Roemer, L., Lee, J. K., Salters-Pedneault, K., Erisman, S. M., Orsillo, S. M., & Mennin, D. S. (2009). Mindfulness and emotion regulation difficulties in generalized anxiety disorder: Preliminary evidence for independent and overlapping contributions. Behavior Therapy, 40(2), 142–154. doi: 10.1016/j.beth.2008.04.001 Rolls, E. T. (1999). The brain and emotion. Oxford, UK: Oxford University Press. Rosen, J. B., & Schulkin, J. (1998). From normal fear to pathological anxiety. Psychological Review, 105(2), 325–350. doi: 10.1037/0033 –295X.105.2.325 Rusch, S., Westermann, S., & Lincoln, T. M. (2012). Specificity of emotion regulation deficits in social anxiety: An Internet study. Psychology and Psychotherapy: Theory, Research and Practice, 85(3), 268–277. doi: 10.1111/j.2044–8341.2011.02029.x Saxena, S., & Rauch, S. (2000). Functional neuroimaging and the neuroanatomy of obsessive-compulsive disorder. Psychiatric Clinics of North America, 23(3), 563–586. Scaramella, L. V., & Leve, L. D. (2004). Clarifying parent–child reciprocities during early childhood: The early childhood coercion model. Clinical Child and Family Psychology Review, 7(2), 89–107. doi: 10.1023/B:CCFP.0000030287.13160.a3 Scherer, K. R. (2000). Emotions as episodes of subsystem synchronization driven by non linear appraisal processes. In M. D. Lewis & I. Granic (Eds.), Emotion, development, and self-organization: Dynamic

References systems approaches to emotional development (pp. 70–99). Cambridge, UK: Cambridge University Press. Sehlmeyer, C., Schöning, S., Zwitserlood, P., Pfleiderer, B., Kircher, T., Arolt, V., & Konrad, C. (2009). Human fear conditioning and extinction in neuroimaging: a systematic review. Plos One, 4(6), e5865. doi: 10.1371/journal.pone.0005865 Semple, R. J., & Burke, C. A. (2012). Mindfulness-based treatment for children and adolescents. In P. C. Kendall (Ed.), Child and adolescent therapy: Cognitive-behavioral procedures (4th ed., pp. 411–426). New York, NY: Guilford Press. Shaffer, D., Fisher, P., Dulcan, M. K., Davies, M., Piacentini, J., Schwab-Stone, M. E., . . . Regier, D. A. (1996). The NIMH Diagnostic Interview Schedule for Children Version 2.3 (DISC-2.3): Description, acceptability, prevalence rates, and performance in the methods for the epidemiology of child and adolescent mental disorders study. Journal of the American Academy of Child and Adolescent Psychiatry, 35(7), 865–877. Sheppes, G., & Meiran, N. (2007). Better late than never? On the dynamics of online regulation of sadness using distraction and cognitive reappraisal. Personality and Social Psychology Bulletin, 33(11), 1518–1532. doi: 10.1177/0146167207305537 Siegle, G. J., Thompson, W. K., Collier, A., Berman, S. R., Feldmiller, J., Thase, M. E., & Friedman, E. S. (2012). Toward clinically useful neuroimaging in depression treatment: Prognostic utility of subgenual cingulate activity for determining depression outcome in cognitive therapy across studies, scanners, and patient characteristics. JAMA Psychiatry, 69(9), 913–924. Silberg, J. L., Rutter, M., & Eaves, L. (2001). Genetic and environmental influences on the temporal association between earlier anxiety and later depression in girls. Biological Psychiatry, 49(12), 1040–1049. Silove, D., Manicavasagar, V., Curtis, J., & Blaszczynski, A. (1996). Is early separation anxiety a risk factor for adult panic disorder? A critical review. Comprehensive Psychiatry, 37(3), 167–179. doi: 10.1016/S0010 –440X(96)90033–4 Silverman, W. K., La Greca, A. M., & Wasserstein, S. (1995). What do children worry about? Worries and their relation to anxiety. Child Development, 66(3), 671–686. doi: 10.2307/1131942 Slade, T., & Watson, D. (2006). The structure of common DSM-IV and ICD-10 mental disorders in the Australian general population. Psychological Medicine, 36(11), 1593–1600. doi: 10.1017/S0033291706008452 Smith, P., Perrin, S., Dalgleish, T., Meiser-Stedman, R., Clark, D., & Yule, W. (2013). Treatment of posttraumatic stress disorder in children and adolescents. Current Opinion in Psychiatry, 26(1), 66–72. doi: 10.1097/YCO.0b013e32835b2c01 Somerville, L. H., Wagner, D. D., Wig, G. S., Moran, J. M., Whalen, P. J., & Kelley, W. M. (2013). Interactions between transient and sustained neural signals support the generation and regulation of anxious emotion. Cerebral Cortex, 23(1), 49–60. doi: 10.1093/cercor/bhr373 Spence, S. H., Rapee, R., McDonald, C., & Ingram, M. (2001). The structure of anxiety symptoms among preschoolers. Behaviour Research and Therapy, 39(11), 1293–1316. doi: 10.1016/S0005–7967(00)00098-X Sripada, R. K., Marx, C. E., King, A. P., Rampton, J. C., Ho, S., & Liberzon, I. (2013). Allopregnanolone elevations following pregnenolone administration are associated with enhanced activation of emotion regulation neurocircuits. Biological Psychiatry. 73(11):1045–1053. doi: 10.1016/j.biopsych.2012.12.008 Starr, L. R., & Davila, J. (2012). Responding to anxiety with rumination and hopelessness: Mechanism of anxiety-depression symptom co-occurrence? Cognitive Therapy and Research, 36(4), 321–337. doi: 10.1007/s10608–011–9363–1 Stendler, C. B. (1950). Sixty years of child training practices: Revolution in the nursery. Journal of Pediatrics, 36(1), 122–134.

555

Sullivan, G. M., & Gorman, J. M. (2002). Finding a home for posttraumatic stress disorder in biological psychiatry. Is it a disorder of anxiety, mood, stress, or memory? Psychiatric Clinics of North America, 25(2), 463–468. Sung, S. C., Porter, E., Robinaugh, D. J., Marks, E. H., Marques, L. M., Otto, M. W., . . . Simon, N. M. (2012). Mood regulation and quality of life in social anxiety disorder: An examination of generalized expectancies for negative mood regulation. Journal of Anxiety Disorders, 26(3), 435–441. doi: 10.1016/j.janxdis.2012.01.004 Suveg, C., Morelen, D., Brewer, G. A., & Thomassin, K. (2010). The emotion dysregulation model of anxiety: A preliminary path analytic examination. Journal of Anxiety Disorders, 24(8), 924–930. doi: 10.1016/j.janxdis.2010.06.018 Suveg, C., Sood, E., Comer, J., & Kendall, P. (2009). Changes in emotion regulation following cognitive-behavioral therapy for anxious youth. Journal of Clinical Child and Adolescent Psychology, 38(3), 390–401. doi: 10.1080/15374410902851721 Svendsen, A. M., Kessing, L. V., Munkholm, K., Vinberg, M., & Miskowiak, K. W. (2012). Is there an association between subjective and objective measures of cognitive function in patients with affective disorders? Nordic Journal of Psychiatry, 66(4), 248–253. doi: 10.3109/08039488.2011.626870 Swedo, S. E., Leonard, H. L., Garvey, M., Mittleman, B., Allen, A. J., Perlmutter, S., et al. (1998). Pediatric autoimmune neuropsychiatric disorders associated with streptococcal infections: clinical description of the first 50 cases. American Journal of Psychiatry, 155(2), 264–271. Sylvers, P., Lilienfeld, S. O., & LaPrairie, J. L. (2011). Differences between trait fear and trait anxiety: Implications for psychopathology. Clinical Psychology Review, 31(1), 122–137. doi: 10.1016/j.cpr.2010.08.004 Tamir, M. (2011). The maturing field of emotion regulation. Emotion Review, 3(1), 3–7. doi: 10.1177/1754073910388685 Tan, P. Z., Forbes, E. E., Dahl, R. E., Ryan, N. D., Siegle, G. J., Ladouceur, C. D., & Silk, J. S. (2012). Emotional reactivity and regulation in anxious and nonanxious youth: A cell-phone ecological momentary assessment study. Journal of Child Psychology and Psychiatry, 53(2), 197–206. doi: 10.1111/j.1469–7610.2011.02469.x Thompson, R. A. (1994). Emotion regulation: A theme in search of a definition. Monographs of the Society for Research in Child Development, 59, 25–52. Thompson, R. A. (2011). Emotion and emotion regulation: Two sides of the developing coin. Emotion Review, 3(1), 53–61. doi: 10.1177 /1754073910380969 Thompson, R. A., & Goodman, M. (2010). Development of emotion regulation: More than meets the eye. In A. M. Kring & D. M. Sloan (Eds.), Emotion regulation and psychopathology: A transdiagnostic approach to etiology and treatment (pp. 38–58). New York, NY: Guilford Press. Thompson, R. A., Lewis, M. D., & Calkins, S. D. (2008). Reassessing emotion regulation. Child Development Perspectives, 2(3), 124–131. doi: 10.1111/j.1750–8606.2008.00054.x Thompson, R. A., & Limber, S. P. (1990). ‘Social anxiety’ in infancy: Stranger and separation reactions. In H. Leitenberg (Ed.), Handbook of social and evaluation anxiety (pp. 85–137). New York, NY: Plenum Press. Tortella-Feliu, M., Balle, M., & Sesé, A. (2010). Relationships between negative affectivity, emotion regulation, anxiety, and depressive symptoms in adolescents as examined through structural equation modeling. Journal of Anxiety Disorders, 24(7), 686–693. doi: 10.1016/j.janxdis.2010.04.012 Trosper, S. E., Buzzella, B. A., Bennett, S. M., & Ehrenreich, J. T. (2009). Emotion regulation in youth with emotional disorders: Implications

556

Anxiety Regulation

for a unified treatment approach. Clinical Child and Family Psychology Review, 12(3), 234–254. doi: 10.1007/s10567–009–0043–6 Trosper, S. E., & May, J. (2011). The relationship between trait, expressive, and familial correlates of emotion regulation in a clinical sample of anxious youth. Journal of Emotional and Behavioral Disorders, 19(2), 117–128. doi: 10.1177/1063426609353763 Tucker, D. M. (1989). Neural and psychological maturation in a social context. In D. Cicchetti (Ed.), The emergence of a discipline: Rochester symposium on developmental psychopathology (Vol. 1, pp. 69–88). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Urry, H. L., Roeser, R. W., Lazar, S. W., & Poey, A. P. (2012). Prefrontal cortical activation during emotion regulation: Linking religious/spiritual practices with well-being. In A. Warren, R. M. Lerner, & E. Phelps (Eds.), Thriving and spirituality among youth: Research perspectives and future possibilities (pp. 19–29). Hoboken, NJ: John Wiley & Sons Inc. Vanderhasselt, M., Baeken, C., Van Schuerbeek, P., Luypaert, R., & De Raedt, R. (2013). Inter-individual differences in the habitual use of cognitive reappraisal and expressive suppression are associated with variations in prefrontal cognitive control for emotional information: An event related fMRI study. Biological Psychology, 92(3), 433–439. doi: 10.1016/j.biopsycho.2012.03.005 Vasey, M. W., Crnic, K. A., & Carter, W. G. (1994). Worry in childhood: A developmental perspective. Cognitive Therapy and Research, 18(6), 529–549. doi: 10.1007/BF02355667 Verstraeten, K., Bijttebier, P., Vasey, M., & Raes, F. (2011). Specificity of worry and rumination in the development of anxiety and depressive symptoms in children. British Journal of Clinical Psychology/British Psychological Society, 50(4), 364–378. doi: 10.1348 /014466510X532715 Vollebergh, W. M., Iedema, J., Bijl, R. V., de Graaf, R., Smit, F., & Ormel, J. (2001). The structure and stability of common mental disorders: The NEMESIS study. Archives of General Psychiatry, 58(6), 597–603. doi: 10.1001/archpsyc.58.6.597 Vrtiˇcka, P., Bondolfi, G., Sander, D., & Vuilleumier, P. (2012). The neural substrates of social emotion perception and regulation are modulated by adult attachment style. Social Neuroscience, 7(5), 473–493. doi: 10.1080/17470919.2011.647410 Walkup, J. T., Laballarte, M. J., Riddle, M. A., Pine, D. S., Greenhill, L., Klein, R. . . . Roper, M. (2001). Fluvoxamine for the treatment of anxiety disorders in children and adolescents. New England Journal of Medicine, 344, 1279–1285.

Watson, D. (2005). Rethinking the mood and anxiety disorders: A quantitative hierarchical model for DSM-V. Journal of Abnormal Psychology, 114(4), 522–536. doi: 10.1037/0021–843X.114.4.522 Webb, T. L., Miles, E., & Sheeran, P. (2012). Dealing with feeling: A meta-analysis of the effectiveness of strategies derived from the process model of emotion regulation. Psychological Bulletin, 138(4), 775–808. doi: 10.1037/a0027600 Weeks, M., Coplan, R. J., & Kingsbury, A. (2009). The correlates and consequences of early appearing social anxiety in young children. Journal of Anxiety Disorders, 23(7), 965–972. doi: 10.1016/j.janxdis .2009.06.006 Werner, K. H., Goldin, P. R., Ball, T. M., Heimberg, R. G., & Gross, J. J. (2011). Assessing emotion regulation in social anxiety disorder: The emotion regulation interview. Journal of Psychopathology and Behavioral Assessment, 33(3), 346–354. doi: 10.1007/s10862–011–9225-x White, L. K., Helfinstein, S. M., Reeb-Sutherland, B. C., Degnan, K. A., & Fox, N. A. (2009). Role of attention in the regulation of fear and anxiety. Developmental Neuroscience, 31(4), 309–317. doi: 10.1159/000216542 Wilensky, A. E.,Schafe, G. E., Kristensen, M. P.,& LeDoux, J. E. (2006). Rethinking the fear circuit: The central nucleus of the amygdala is required for the acquisition, consolidation, and expression of Pavlovian fear conditioning. Journal of Neuroscience, 26, 12387−12396. Yehuda, R., Halligan, S. L., & Bierer, L. M. (2001). Relationship of parental trauma exposure and PTSD to PTSD, depressive and anxiety disorders in offspring. Journal of Psychiatric Research, 35(5), 261–270. Yehuda, R., Halligan, S. L., & Grossman, R. (2001). Childhood trauma and risk for PTSD: relationship to intergenerational effects of trauma, parental PTSD, and cortisol excretion. Development and Psychopathology, 13, 733–753. Zalta, A. K., & Chambless, D. L. (2008). Exploring sex differences in worry with a cognitive vulnerability model. Psychology of Women Quarterly, 32(4), 469–482. doi: 10.1111/j.1471–6402.2008.00459.x Zlomke, K. R., & Hahn, K. S. (2010). Cognitive emotion regulation strategies: Gender differences and associations to worry. Personality and Individual Differences, 48(4), 408–413. doi: 10.1016/j.paid.2009 .11.007 Zoellner, L. A., Rothbaum, B. O., & Feeny, N. C. (2011). PTSD not an anxiety disorder? DSM committee proposal turns back the hands of time. Depression and Anxiety, 28(10), 853–856. doi: 10.1002/da .20899

Trim Size: 8.5in x 11in

Cicchetti

c15.tex printer - Volume II - 01/12/2016

7:23pm

CHAPTER 15

Typical and Atypical Brain Development Across the Life Span in a Neural Network Model of Psychopathology BARBARA GANZEL and PAMELA MORRIS

INTRODUCTION 557 DIATHESIS–STRESS MODELS OF THE ORIGINS OF PSYCHOPATHOLOGY 559 Diathesis Meets Stress 559 Biological Sensitivity to Context 561 DIATHESIS AND STRESS IN THE BRAIN 562 The Stress Concept 562 Allostasis Situates the Brain in Diathesis–Stress Models 564 THE CASE FOR A CONSTANTLY CHANGING BRAIN 569 Early Human Neurodevelopment 569 Early Functional Specialization of Local Neural Networks: Focus on Gray Matter 570 Late-Maturing Structural and Functional Integration: Focus on White Matter 577 A Life Span View of Neural Networks 581 ALLOSTASIS IN A CONSTANTLY CHANGING BRAIN 583 Windows of Vulnerability; Windows of Opportunity 584 Animal Models of Early Stress-Sensitive Periods 585 Early Stress Sensitivity in Humans 586

Stress Sensitivity in Adolescence 587 Stress-Related Neural Plasticity and/or Damage in Adults 590 Incorporating Modulated Allostasis Into Diathesis–Stress Models 594 TRIPLE NETWORK ALLOSTASIS AND PSYCHOPATHOLOGY 597 Dysfunctional Large-Scale Neural Networks in Psychopathology 598 Triple Network Allostasis: A Novel Diathesis–Stress Model 601 Stress Vulnerability Within the Functional Architecture of the Brain 602 Triple Network Allostasis in Context 605 Moderators of Triple Network Allostasis 605 FUTURE DIRECTIONS 606 Triple Network Allostasis as a Dynamic System 606 Is Triple Network Allostasis Universal? 607 Can Triple Network Allostasis Serve as an Organizing Concept? 608 Is Triple Network Allostasis Only Human? 609 CONCLUSION 610 REFERENCES 613

INTRODUCTION

1988) and (2) new thinking about the role of the neural connections and networks in psychopathology (Menon, 2011). The brain is central to the concept of diathesis (e.g., Zuckerman, 1999) and, under allostasis, the brain is also central to the stress process (Ganzel, Morris, & Wethington, 2010; McEwen, 2007; Sterling, 2004; Sterling & Eyer, 1988). With the brain as a core feature of both sides of any diathesis–stress interaction, integration of these concepts is facilitated. This has two main benefits. First, bringing research on allostasis and large scale neural networks (connectomics2 ) together in the context of a diathesis–stress

In this chapter, we present a new diathesis–stress model of the onset of psychopathology across the life span. We highlight the role of the brain in the diathesis–stress processes through the integration of two paradigm shifts in the fields of stress research and psychopathology, respectively: (1) the role of allostasis in stress research (Sterling & Eyer,

We would like to thank Dr. Nathan Spreng of the Department of Human Development and the Human Neuroscience Institute at Cornell University for his expert advice on the dynamics of large-scale brain networks. 1 Color versions of Figures 15.2, 15.8, 15.9, 15.12–15.14, 15.17, 15.21–15.24 are available at http://onlinelibrary.wiley.com /book/10.1002/9781119125556

2

Connectomics is the study of the connectome, the functional and structural connections in the brain (e.g., Johansen-Berg, 2013). In this chapter, we focus on the large-scale connectomics of the human brain. 557

Page 557

558

Life Span Diathesis–Stress Model of Psychopathology

model enables us to identify stress-vulnerable brain processes that are likely to be implicated in the onset of psychopathology. We note that this effectively reframes the debate about whether the root cause of psychopathology is internal or external to the individual. Second, the central role of the brain in the diathesis–stress process mandates a consideration of the role of brain development; given evidence of a continually changing brain, we posit a developmental model of the onset of psychopathology. In this way, we advance understanding of the neural processes involved in the onset of mental illness and, in so doing, squarely embrace the core principles of developmental psychopathology. We begin our discussion with a history of the diathesis– stress model in the study of psychopathology. Fifty years ago, clinicians regularly eschewed the role of physiological processes in the onset of psychopathology. More recently, though, the field of psychopathology has slowly come to embrace the joint roles of the internal (diathesis) and external (stress) in the onset of psychopathology. Questions now focus on the form of the diathesis–stress interaction and the processes by which they together produce psychopathology. Even so, it has been only recently that the tools of neuroimaging have allowed the neural processes underlying psychopathology to be explored in depth. A key turning point in the history of stress research for this discussion is the introduction of the concept of allostasis. The value of the allostatic model is that it explicitly brings that which was considered external (stress) into the brain itself. Our discussion of allostasis builds on our prior work (Ganzel et al., 2010), which used an integrative human neuroscience approach (Cacioppo, Berntson, Sheridan, & McClintock, 2000) to develop theoretical models about the relationship between stress and both physical and mental health. Our own work has highlighted the brain’s survival/salience circuitry and associated modulatory systems3 as the neural front lines of allostasis, making this the specific neural circuitry that is likely to be most vulnerable to the accumulated wear and tear of allostatic load (the physiological cost of stress; Ganzel et al., 2010). This has allowed us to develop an allostatic model of the onset of psychopathology in response to current and prior stressors. Thus, under the concept of allostasis, stress and diathesis are intersecting processes that both take place in the brain. 3 The neural systems underlying the central motivational “states of the brain” (Rosen & Schulkin, 1998, p. 326) formerly termed the core emotion systems of the brain in our work (Ganzel & Morris, 2011; Ganzel et al., 2010).

Because the brain is the central mediator of allostasis, and allostasis is key in our diathesis–stress model of psychopathology, we have also argued that brain development across the life span must play an essential role in the diathesis/stress process. To inform this discussion, we review the evidence for brain maturation and senescence across this life span. Doing so allows us to extend our model of diathesis–stress to include an explicit developmental component (Ganzel et al., 2010). As such, we explicitly examine how life span brain development intersects with developmental psychopathology (Cicchetti, 1993; Cicchetti & Posner, 2005) to inform diathesis–stress models of psychopathology across development. To a diathesis–stress process taking place in the developing brain, we add a final level of complexity: the role of large-scale brain networks in psychopathology (Menon, 2011; Uhlhaas & Singer, 2012). Much of the prior work linking stress, psychopathology, and the brain has focused on individual regions of the brain thought to underlie the deficits in cognitive and affective functioning that characterize mental conditions, such as schizophrenia and depression. However, given the complexity of psychopathological disorders in their affective, cognitive, and behavioral presentations, it is perhaps not surprising that mapping of disorders onto local differences in brain regions has proved insufficient. Instead, current thinking suggests that neural circuitry might be critical here, i.e., it is networks of brain regions (and not just individual regions) that matter. Thus, building from a model in which stress can drive long-term structural and functional change in survival/salience-related brain regions (via allostasis; Ganzel et al., 2010), we highlight the role of stress sensitivity in large-scale neural networks in our development of a new diathesis–stress model of psychopathology. In this work, we strongly espouse a number of key principles of the developmental psychopathology perspective (Cicchetti & Toth, 2009). First, we recognize that the study of normality and pathology are mutually informative, with normal development providing an understanding of the baseline processes on which there is deviation in the context of psychopathology. As such, our discussion of allostasis builds from a discussion of typical (nonpathological) neurophysiological responses to stress and our discussion of brain plasticity draws from what we know about typical (nonpathological) brain development. Second, as discussed previously, we develop a life span model of the diathesis/stress process, arguing for a model that maps brain development across age and stage of development with the development of psychopathology across this same period. Third, in keeping with the field’s

Diathesis–Stress Models of the Origins of Psychopathology

focus on the interplay between the internal and the external, and the goal of reducing biological/environmental dualism, we highlight how external processes such as stressor exposure come to be instantiated in the brains of typical (nonpathological) individuals as neural processes and long-term changes in brain structure and function. This approach moves the entire diathesis/stress interaction into the brain. Finally, we draw on information from developmental neurobiology, from stress research in the social sciences and in neuroscience, and from new work on large-scale neural networks in psychopathology to bring cross-disciplinary perspectives to bear on the development of our diathesis–stress model.

DIATHESIS–STRESS MODELS OF THE ORIGINS OF PSYCHOPATHOLOGY Diathesis Meets Stress The classic diathesis–stress model was not born whole. Instead, it grew incrementally over time during the sturm und drang in psychiatry and clinical psychology that accompanied the decline of psychodynamics in the mid-twentieth century and the emergence of modern, more integrative perspectives on mental health. The 1972 inaugural volume of the International Journal of Mental Health provides a case in point. This was a special issue on the topic of genetics and mental disorders, which was a highly divisive topic at the time. A number of the authors in this volume describe the resistance of their clinical colleagues to a consideration of genetics in the origins of psychopathology. For example, the journal’s editor, Martin Gittelman, wrote: . . . Bear in mind that one of the major theories of the etiology of mental illness is “parentogenesis,” i.e., that family characteristics, interaction, milieu, etc., are largely responsible for mental illness, and that long-term hospitalization is therefore considered therapeutic since it removes the patient from the noxious influences of the family. However, if we can entertain the possibility that endogenous factors play an important role in severe mental illness, then one of the principal justifications for long-term hospitalization is made less tenable. . . . The implications for the organization of mental health services are obvious. (Gittleman, 1972, p. 4).

Other authors in this volume expressed parallel sentiments: Clinicians typically react negatively to genetic formulations on the ground that a genetic emphasis contradicts what we know about psychodynamics; that, as a corollary, it makes

559

psychological intervention (e.g., psychotherapy, behavior modification, milieu therapy, vigorous combating of the hospitalization syndrome) theoretically incomprehensible or pragmatically useless; and, finally, that the overall result would be the development among professionals of a malignant therapeutic nihilism. (Meehl, 1972, P. 10)

This latter author, Meehl (1962, 1972, 1989), is credited with one of the one of the earliest diathesis4 –stress models of the development of a mental disorder—in this case, a gene–environment model predicting the origins of schizophrenia (also see Gottesman & Shields, 1967; although they emphasized the polygenic nature of the diathesis more so than its interaction with environmental influence). Meehl’s model included a dichotomous diathesis (Zuckerman, 1999), in that schizophrenia was portrayed as the result of a patient’s dominant schizogene interacting with an inconsistent and aversive home environment generated by the unfortunate patient’s battle-ax mother. This rearing condition was hypothesized to drive the cognitive slippage and soft neurological and psychometric signs that, over time, resulted in the symptoms and inappropriate behaviors associated with schizophrenia. The model broke new ground in that it explicitly embraced the role of biology in mental illness (a contentious point with many mental health professionals of the day) and the role of social and environmental factors in mental health (which was an equally sore point with much of the rest of medicine at that time: Engel, 1977). In an article in the journal Science, George Engel articulated the impasse at which the field of psychiatry found itself (Engel, 1977). Psychoanalysis and psychodynamics were still the dominant paradigms in psychiatry, but these approaches were increasingly criticized on a number of grounds (e.g., Delprato & McGlynn, 1985; Grunbaum, 1979, 1982). Psychiatry, as a field, was struggling with the claim that mental illness was a myth (i.e., that many mental disorders did not achieve the medical standard of being identifiable disease states because they had no basis in biological brain dysfunctions; see, e.g., Szasz, 1960). The psychosocial elements of human malfunction were deemed by some in the medical establishment to be the province of the theologian and the philosopher and needed to be disentangled from the organic elements of disease 4 The concept of a diathesis has a lengthy history in medicine. The term is used in the present day to refer to a biological vulnerability to disorder that is either hereditary or acquired (Monroe & Simons, 1991), although in the more distant past it was also used to refer to the character of a disorder (e.g., hemorrhagic [yellow] fever versus pneumonia; Brown, 1813; LeConte, 1810).

560

Life Span Diathesis–Stress Model of Psychopathology

(Engel, 1977; Szasz, 1960). In short, mental illness did not fit the medical model of the time, which emphasized the biological components of disease and eschewed the underbrush of psychosocial variables. But Engel cautioned psychiatrists against embracing the medical model of the 1970’s, which he believed to be reductionistic, dogmatic, and irrationally supportive of mind-body dualism (1977). Other clinicians of the time shared this frustration, stating that “reductionism is particularly harmful when it neglects the impact of nonbiological circumstances upon biological processes” (Holman, 1976, p. 134). In response to this debate, Engel (1977) proposed an inclusive biopsychosocial medical model of psychopathology that defied Western medicine’s traditional separation of the mental from the somatic, which he argued had a practical basis in medieval Christian church politics and had thereafter devolved into unexamined sociocultural dogma (also see Rasmussen, 1975). Engel’s biopsychosocial model included the patient as a biological individual in a social context with a history of early, past, and present life experiences (Engel, 1977). Engel’s work was strongly informed by Ludwig von Bertalanffy’s (1952) influential work on general systems theory, which stood in opposition to the reductionism that had become embedded in the dominant models of biology. Von Bertalanffy argued that the organism is better explained as a functionally interdependent whole rather than the sum of its independent biological parts. Under systems theory, the linkages among all levels of organization are revealed by holistic approaches to scientific inquiry. Following this lead, Engel included consideration of the doctor–patient relationship and the social context of the health care system into his ideas for a biopsychosocial model, thus articulating the interdependent levels of analysis within a patient–doctor system that ranged from biology to society. The breadth of Engel’s vision may have defeated the usefulness of his model, which failed to find practical application in research designs of the day (McCutcheon, 2006). Detractors argue that Engel can be credited with proposing the biopsychosocial model, but he failed to adequately specify its form (McLaren, 1998). Others have pointed out that science has only recently progressed to the point where the form of such a model could be specified, since explicit testing of the associations across a multilevel and interactive biopsychosocial system has been possible only with the recent development of new technologies supporting neuroimaging, genetics, and psychosocial measurement (McCutcheon, 2006) and the statistical modeling of multilevel and dynamic data in complex systems (e.g., Sporns & Chialvo, 2004).

Following Meehl’s lead, Zuckerman (1999) is often credited with providing one of the first comprehensive discussions of the onset of psychopathological conditions deriving from two (not one) primary determinants— diathesis and stress. Zuckerman’s models draw extensively on work by Meehl (1962), Gottesman and Shields (1967), and Monroe and Simons (1991), who made the case for an interactive diathesis–stress relationship in the onset of schizophrenia and depression, respectively. As with his predecessors, Zuckerman’s framework makes it clear that a diathesis is a predisposition, and as such creates a vulnerability to stress, but that vulnerability alone is insufficient to result in expression of the disorder. Rather, psychopathology results from the combination of that predisposition (which he argued is an enduring trait of the individual) and the experience of an environmental stressor (which he argues are objective and external to the individual; Figure 15.1). Notably, for Zuckerman, the term diathesis refers to “genetic and biological factors in the nervous system” (Zuckerman, 1999, p. 7), a point that we will return to later. Zuckerman’s contributions to the diathesis–stress concept are extensive. His contributions expand the relevance of the diathesis–stress model to a broader set of psychopathological outcomes, and make explicit what constitutes stress, diathesis, and their relationship. In doing so, he highlights two aspects of Monroe and Simons’s (1991) work that enlarged upon the conceptualization of the diathesis–stress relationship relative to prior formulations (e.g., by Meehl and others). First, he defines a fan-shaped interactive relationship between stress and diathesis in predicting the onset of disorder. While it might seem like only a small step forward from Meehl’s (1962) dichotomous-diathesis–stress interactive model, Monroe and Simons (1991) and Zuckerman (1999) present a quasi-continuous model, in which a diathesis-threshold

Stressor

Diathesis

Inside the brain

Figure 15.1

Mental Health

Traditional diathesis–stress model.

Diathesis–Stress Models of the Origins of Psychopathology

Chrousos & Gold, 1992; Goldstein, 1995a, 1995b; Levine & Ursin, 1991; McEwen & Stellar, 1993; Sterling & Eyer, 1988; Walker & Diforio, 1997; Zuckerman, 1999). Indeed, the entire theory volume of the 2006 Handbook of Developmental Psychology is devoted to the implications of von Bertalanffy’s (1952) general systems theory as applied to developmental science, which argues for mutually influencing and interactive processes across a multileveled system (Damon, Lerner, & Eisenberg, 2006). This increased consideration of multilevel and interactive perspectives has spanned biology, medicine, psychology, and sociology to foster diverse and comprehensive accounts of an organism’s health and development. Biological Sensitivity to Context As the previous discussion makes clear, consideration of the multilevel and interactive perspectives inherent in diathesis–stress models of psychopathology is in fact quite young—having emerged over the last few decades from relatively one-sided views of the sources of psychopathology (environment-only versus biology-only). Yet these earliest accounts represent a relatively straightforward formulation of the interaction between diathesis and stress (Figure 15.2a.). As we discuss later, the last 5 years or so have seen a reconsideration of the form of that interaction. Gone are discussions of whether it is biology or environment that is the source of psychopathology. However, there is continued debate about precisely how stress and diathesis interact to jointly produce psychopathology, as well as about the mediating processes of that relationship. In Belsky’s differential susceptibility theory (Belsky, Bakermans-Kranenberg, & van IJzendoorn, 2007), genetic factors do not make children vulnerable; instead, such children are susceptible or plastic to environmental influences.

presence of diathesis

High

high biological sensitivity

Negative Outcome

Psychopathology

still exists (the level of the diathesis above which there is a positive relation between stress and disorder) but the strength of the positive relation now varies as a function of the level of the diathesis above this threshold. This is important because it allows for the conceptualization of a continuous diathesis, which was heretofore unacknowledged. Second, these models suggest that diathesis may predict the onset of stress. In this way, the experience of stress is not randomly distributed across persons, such that some (but indeed not all) stressors may be the result of the diathesis. For example, certain personality characteristics may elicit negative behavior from others, and these resulting negative interactions may be the daily stressors that, in interaction with the diathesis, cause the onset of disorder. In this way, the internal becomes external. As we discuss later, the newest models suggest the reverse may be true as well. Importantly, Zuckerman (1999) also presents diathesis as having two components—the role of personality and the timing of stressor exposure (early in development versus at the onset of the disorder). In so doing, he argues for a dual-diathesis model—one aspect of which is mediated by personality and is the result of the effect of genes on the CNS, and a second aspect that has a biological basis in the nervous system that can originate from the environment. In other words, this second aspect of diathesis derives from the ongoing impact of the environment on the central nervous system (CNS). Such a conceptualization—in which environmental stressor exposure can become the basis of later vulnerability—is highly consistent with the model of stress and allostasis that we discuss later. The mid- to late twentieth century saw a sweeping upsurge in interest in multilevel approaches that opposed one-sided accounts of biological versus psychosocial functioning (Belsky, 1997; Bronfenbrenner & Morris, 1998;

absense of diathesis

None

Low

High Stress (a)

561

low biological sensitivity Positive

Positive

Negative Environment (b)

Figure 15.2 Diathesis–stress models. Traditional diathesis–stress model (a). Diathesis–stress model based on differential susceptibility theory/biological sensitivity to context (b). See footnote 1. Source: Adapted from Belsky et al. (2007).

562

Life Span Diathesis–Stress Model of Psychopathology

As such, the same vulnerability factor that conferred risk in the context of poor environments confers intellectual or social-emotional benefits in positive environments. Statistically, this requires a crossover interaction (rather than a fan-shaped interaction), in which the slope for the susceptible group is different from zero and steeper than that for the nonsusceptible group (which is ideally zero; Figure 15.2b.). In this way, susceptible children are more strongly influenced by environmental influences, but how those children fare depends strongly on the valence of that environment’s influences. Thus, the diathesis is less of a precondition or threat to psychopathology and more of a measure of plasticity. Belsky draws heavily on evolutionary theory to suggest that having children varying in susceptibility to different environmental contexts is advantageous because it creates heterogeneity within the population. He cites both experimental and nonexperimental evidence to suggest that some children previously considered vulnerable may actually be the most plastic and benefit the most from positive rearing conditions (Belsky et al., 2007). Boyce and Ellis’s (2005) biological sensitivity to context theory provides a rendition of the diathesis–stress interaction that seems empirically similar to Belsky’s, but has critical conceptual distinctions. In Boyce and Ellis’ formulation, stress reactivity has the potential for negative effects under conditions of adversity and positive effects under conditions of supportive environments. But unlike Belsky’s formulation, these highly reactive phenotypes are not highly malleable, in that they are inflexible in the face of a variety of ecological conditions. Rather, they are highly responsive to both the threats and dangers in high stress environments, as well as to the social supports in highly supportive environments. It is this increased vigilance or awareness to environmental influence that results in their highly divergent outcomes across contexts. If we are proposing to model the etiology of psychopathology (the dark side of adaptation), one might think there is little difference between diathesis–stress theories and their more optimistic cousins, differential susceptibility theory (e.g., Belsky et al., 2007) and biological sensitivity to context theory (Boyce & Ellis, 2005). For negative outcomes, all of these theories make precisely the same statistical prediction, that is, that an interaction between a preexisting vulnerability and a stressor will drive maladaptation (Roisman et al., 2012). However, these theories do differ with regard to the underlying process by which the interaction between diathesis and stress occurs, and the nature of what happens in high-quality environments. The interactive nature of stress and diathesis is not the only contribution that Boyce and Ellis (2005) bring

to this discussion. As they describe, these biological predispositions are partly the result of adaptations of stressresponse systems to the environment (although genes and gene–environment interactions are also sources of such predispositions). In the context of low-stress environments, the high reactivity phenotype serves to support children in eking out the good in the environment, allowing them to benefit especially well. But, in the context of highly stressful environments, high reactivity serves to ready the individual to respond to stress, even if it has physiological costs. But to fully understand those physiological costs, we turn to a discussion of how the brain plays a role in this diathesis–stress model of psychopathology.

DIATHESIS AND STRESS IN THE BRAIN In diathesis–stress models of psychopathology, the brain has consistently been portrayed in the literature as a key mediator of diatheses (e.g., Meehl, 1962; Zuckerman, 1999). Not so the stress process. Stressors are modeled as external (e.g., Zuckerman, 1999) and the stress process itself has largely been conceived as having bottom-up effects on the brain through homeostatic processes and hypothalamic-adrenal-pituitary (HPA) axis effects. We argue here that placement of the brain as the primary mediator of the stress process is potentially transformative for thinking on diathesis–stress models of psychopathology. To fully understand this work, we begin with a brief history of the stress process with attention to the inclusion over time of the role of the brain. The Stress Concept Cannon and Selye Stress research shares common origins in the work of Walter Cannon (e.g., Cannon, 1932) and Hans Selye (e.g., Selye, 1956). In Cannon’s conceptualization, the sympatheticadrenal medullary system mobilizes the body’s energy resources by increasing epinephrine (adrenaline), which in turn increases blood pressure, heart rate, and blood sugar, as well as hastening blood coagulation, clearing fatigue products from muscles, and decreasing digestion (Cannon, 1920). In Cannon’s view, these processes occurred locally in the body, independent of central nervous system (CNS) control (and hence did not involve the brain). He used the term homeostasis to refer to the way this array of independent physiological systems works together to reestablish initial conditions when the system is perturbed (Cannon, 1932).

Diathesis and Stress in the Brain

By the mid-twentieth century, Hans Selye demonstrated that an organism has an adaptive response to adversity that includes both Cannon’s sympathetic (adrenaline-driven) responses plus the actions of hormones from the pituitary gland, which globally affect the major organs of the body in indirect but important ways (e.g., Selye, 1956). Selye proposed the existence of a generalized physiological syndrome that occurs in response to a great diversity of threats to the integrity of the organism (e.g., Selye, 1956). The syndrome included three stages of physiological response to stressor exposure, which he called the General Adaptation Syndrome or GAS (Selye, 1956). The first stage is the sequential actions of the hypothalamic-pituitary-adrenal (HPA) axis, including production of corticotropin releasing hormone (CRH), the resulting release of adrenocorticotropin releasing hormone (ACTH), and the production of glucocorticoids (cortisol, in humans) (alarm stage). In the second stage, overt symptoms of stress are often reduced or disappear. In the final (exhaustion) stage, physiological defenses are depleted and the organism will die if the stressor is not released. For Selye, stress was “that which stimulated the GAS response” (1956, p. 54). Selye’s claims for the scope of the application of the GAS were very broad and his work prompted thousands of articles across medicine, psychology, sociology, and biology (Goldstein, 1995b; Mason, 1975). However, while expanding the concept of homeostasis (Bernard, 1878; Cannon, 1932), his work retained many of the omissions inherent in Cannon’s original conceptualization, including inattention to the significance of psychological state, anticipation, coping, life history, and environmental context beyond the specific stressor (Burchfield, Woods, & Elich, 1980; Mason, 1975; Toates, 1995). Moreover, Selye wrote of a hypothetical neural or endocrine common mediator that lies between application of the noxious stressor and the GAS response (1950, p. 105). This unknown common mediator would serve to receive and integrate local inputs and then transmit generalized messages to all parts of the body to activate the GAS response, including facilitating changes in regulatory (i.e., homeostatic) set points during the stage of resistance (e.g., Selye, 1950, 1956). Selye never identified this critical common mediator. Work following Selye balked at these gaps with increasing intensity. For example, in psychosocial research, it was noted the measurement of the valence or the level of threat posed by a given stressor often improved the amount of variance explained in the outcomes ( Brown & Harris, 1978; Dohrenwend, Ashkenasy, Krasnoff, & Dohrenwend, 1978; Holmes & Rahe, 1967; Pearlin & Schooler, 1978; Wheaton,

563

1999). Changes in stressor context over time also made significant independent contributions to the outcomes of stress models (e.g., Baltes & Baltes, 1990; Furstenberg, Brooks-Gunn, & Morgan, 1987; Hobfoll, Johnson, Ennis, & Jackson, 2003; Holahan, Moos, Holahan, & Cronkite, 1999; Lazarus, 1993; 1999; Masten et al., 1988; Rutter, 1979; Sameroff, Seifer, Barocas, Zax, & Greenspan, 1987), as did individual differences in cognitive and emotional response (e.g., anticipation, appraisal, coping, learning, and other types of information processing: Foa & Kozak, 1986; Holahan, Moos, Holahan, & Cronkite, 1999; Ironson et al., 2000; Lazarus, 1991; Siegal & Allan, 1998; Sterling & Eyer, 1988; Stone, 1995; Toates, 1995; Tolin & Foa, 2002; Wheaton, 1985). Taken together, this body of work has stretched the psychosocial requirements for models of the stress process well beyond the classical homeostatic perspective or the GAS. Similarly, within biomedical research, evidence of the influence of cognitive and affective processing on the stress process was also prompting reconsideration of classical homeostasis and Selye’s GAS (Mason, 1971, 1975; Pacák & Palkovits, 2001; Schulkin, 2004). For example, Mason (1971, 1975) documented substantial variation in elements of the GAS response as a function of the stressor context, the individual’s history, and the individual’s perception of the noxiousness of the stressor, thus raising questions about the specificity of the “nonspecific” GAS response (also see Pacák & Palkovits, 2001). “The knowledge that the psyche is superimposed upon the humoral machinery for endocrine regulation drastically complicates our whole view . . . ” (Mason, 1975, p. 177). Thus, the biomedical stream of stress research also found itself in need of a model that included appraisal, cognition, and affective (or emotional) state as primary elements of the stress process (Frankenhaeuser, 1980; Moore-Ede, 1986; Weinberg & Levine, 1979). The Brain Enters the Stress Process This work set the stage for Sterling and Eyer (1988) to propose the concept of allostasis, which holds that the central nervous system (CNS) exerts executive control over all physiological regulatory responses to environmental demand in the alert, intact organism (see Figure 15.3). CNS control of the stress response, in turn, allows the regulatory set points of the organism to vary in response to environmental demand (McEwen & Wingfield, 2003; Sterling & Eyer, 1988). This placed the CNS in general, and the brain in particular, squarely in the role of Selye’s common mediator between environmental demand and physiological response (1950).

564

Life Span Diathesis–Stress Model of Psychopathology

Stressor

Diathesis

Inside the brain

“Stress Response”

Mental Health

Figure 15.3 Bringing stress into the brain. The most basic diathesis–stress model of response to a current stressor is presented as a traditional psychosocial stressor → stress response → distress model of the stress process; with the exception that the processes are represented as occurring within the brain (the solid gray line represents operations that occur within the brain, within which all processes are biological).

This work went largely unnoticed at the time of its publication. Three years later, Levine and Ursin (1991) proposed that the sensory input from all types of stressors is gated through the brain before affecting any other physiological regulatory response, and that this input is modified by expectation and evaluation. They argued that this is true even for stressors that do not appear overtly psychological, such as cold exposure and tissue damage, because novelty, expectation, and efforts to avoid noxious stimuli will all activate brain responses (Levine & Ursin, 1991). Soon after, Chrousos and Gold (1992) and Goldstein (1995a, 1995b) independently proposed a model in which homeostasis resets itself in response to stress exposure, and Goldstein (1995b) argued that “distress invariably resets homeostasis.” P. 41. Thus, by the end of the twentieth century, it was clear that homeostasis required revision to account for the influences of environment and the CNS on physiological regulation. None of these proposed modifications of homeostatic theory addressed the accumulating consequences of physiological accommodation to environmental challenge over the life span. This prompted McEwen and Stellar (1993) to introduce the concept of allostatic load, which they defined as the physiological cost of making long-term adaptive shifts across a broad range of systems to match internal functioning to environmental demand. With the inclusion of allostatic load as a key element of the theory, allostasis became the most comprehensive account of regulatory accommodation to environmental demands and accumulated physiological cost over time. The adaptation and dissemination of allostatic theory by McEwen and

an expanding cohort of colleagues is lending allostasis the status of “a new conceptual framework” (Schulkin, Gold, & McEwen, 1998, p. 220) for the study of stress (e.g., McEwen, 1998, 2000, 2001, 2003; McEwen & Seeman, 2003; McEwen & Stellar, 1993; McEwen & Wingfield, 2003; Schulkin, 2004; Schulkin, McEwen, & Gold, 1994; Singer, Ryff, & Seeman, 2004). The concept of allostasis provides us with a new approach to specifying the mechanisms of the diathesis– stress model of mental disorder. In particular, it suggests a need to update the classic diathesis–stress model to include the role of the brain in the stress process. In the following sections, we will argue that the process by which diathesis and stress result in psychopathology occur through the central mechanisms of allostasis. A key consequence of model is that diathesis is no longer static, but is dynamically changing over time. Articulating this process requires that we briefly specify these mechanisms. Allostasis Situates the Brain in Diathesis–Stress Models Allostasis and Response to a Current Stressor In prior work, we used the theory of allostasis as the foundation for a model of the ongoing stress process (Ganzel et al., 2010). The most basic model of allostatic accommodation to a current stressor (Figure 15.4) is presented as a classic psychosocial stressor → stress response → distress model of the stress process. In this model, the stress response allows the resetting of homeostatic set points during the process of reestablishing physiological equilibrium during and after stressor exposure (i.e., allostasis). From the previous discussion ( Figure 15.3), recall that the brain is the central mediator of allostasis (e.g., McEwen, 2007; Sterling, 2004; Sterling & Eyer, 1988). Allostatic response and adaptation to a current stressor can be viewed as a two-stage process, wherein a brain-based central allostasis has executive control over peripheral allostasis in stress-responsive physiological systems external to the central nervous system (which in turn have critical feedback on the brain: for a full discussion, see Ganzel et al., 2010). These effects include, but are not limited to, production of stress hormones and catecholamines, inhibition of the reproductive system; and alterations of metabolism, cardiovascular activity, and gastrointestinal and immune system function (Davis, Walker, & Lee, 1997; Heinrichs, Menzaghi, Pich, Britton, & Koob, 1995; Maier & Watkins, 1998; Stratakis & Chrousos, 1995). These processes feed back on the brain via interoceptive, hormonal, catecholamine, and immune signals (Damasio, 2003; Davidson, Maxwell, & Shackman, 2004; Sapolsky,

Diathesis and Stress in the Brain

Current Context Risk proximal, distal

565

Current Environmental Stressor

Current Contextual Resources proximal, distal

Sensory Processing Complex Representation of Sensory Input Inside the brain

Central Allostasis Specific and general survival circuits of the brain

Current Mental Health

Peripheral Allostatic Accommodation HPA axis, immune, cardiovascular, etc.

Figure 15.4 An allostatic model of response to a current stressor. Here, the brain is shown as a dynamically adapting interface between ecological challenge (in context) and the biological self. This model shows the role of central and peripheral allostasis in predicting mental health. Dotted lines indicate active feedback. Interaction would be expected between allostatic load and major pathways of allostatic accommodation (not shown). The role of genes is not shown here. Source: Adapted from B. L. Ganzel, P. A. Morris, & E. Wethington, Allostasis and the human brain: Integrating models of stress from the social and life sciences, Psychological Review, 117(1), 134, 2010.

1998) and create iterative sequelae of physiological and psychological effects over time in response to the trigger stimulus, as discussed later. This central-peripheral distinction allows us to specify the neural mechanisms underlying the moment-to-moment physiological response to a current stressor. Genes serve as an additional source of input to these processes (Figure 15.5). As with a classical diathesis–stress model, there is a role for both biological and social/ environmental processes. But stress gets under the skin and into the brain/body system, so our model posits that diathesis is not biological and stress is not external—but rather both are deeply embedded within convergent neural processes. Previously, we highlighted the brain’s survival circuitry (LeDoux, 2012) and its closely linked domain-general modulatory processes5 as the neural front lines of allostasis, and we reviewed the evidence that these brain regions 5

Termed the core emotion systems of the brain in our prior work (Ganzel & Morris, 2011; Ganzel et al., 2010). Also see discussion on this topic by LeDoux (2012) and Barrett and Satpute (2013).

are particularly vulnerable to the accumulated wear and tear of allostatic load (Ganzel et al., 2010). There are multiple, overlapping survival circuits in the brain, each of which detects a particular class of trigger stimulus (e.g., defense–avoidance, reward, energy and nutritional maintenance, fluid balance, thermoregulation) and then integrates relevant sensory and motor information to drive adaptive response behavior (LeDoux, 2012). The triggers for each survival circuit are salient, either because they have intrinsic biological value (stimulating stereotypical approach or avoidance behaviors) or because they have learned, personal value (for discussions, see LeDoux, 2012). These specific survival circuits, together with a general-purpose modulatory system, form an interconnected network that serves to represent salient stimuli and produce and regulate behavioral response. This extended (specific + general) survival network constitutes the central motivational states of the brain (Rosen & Schulkin, 1998) that underlie human behavioral adaptation to the environment (also see core affect: Barrett, Mesquita, Ochsner, & Gross, 2007; Bechara, Damasio, Damasio, & Anderson, 1994; Damasio, 1999; LeDoux, 2012). We have argued that

566

Life Span Diathesis–Stress Model of Psychopathology

Current Context Risk proximal, distal

Current Environmental Stressor

Current Contextual Resources proximal, distal

Sensory Processing

Genes

Complex Representation of Sensory Input

Inside the brain

Central Allostasis Specific and general survival circuits of the brain

Current Mental Health

Peripheral Allostatic Accommodation HPA axis, immune, cardiovascular, etc.

Figure 15.5 The direct effect of genes on central allostasis. Here, we emphasize the main effect of original genetic compliment on central allostasis, but we note that genes are also likely to serve as a moderator of associations within central and peripheral allostasis. We also anticipate an effect of genes on input to the model—that is, on current contextual risk, current contextual resources, and the current stressor itself.

this extended survival circuitry is also the main organizing factor in allostasis (i.e., in the translation of stressor exposure into a behavioral or physiological stress response) and thereafter into stress-related changes in mental health (Ganzel et al., 2010). Variously called the limbic system (but see LeDoux, 1996, pp. 98–102) or the emotional brain (but see Barrett & Satpute, 2013), this extended survival system is at the heart of the diathesis–stress model. We will focus on the defense circuit (LeDoux, 2012) as our exemplar of a specific survival circuit. The defense circuit detects and responds to threat; it includes the amygdala and periamygdaloid regions, hippocampus, thalamus, hypothalamus, midbrain periaqueductal gray, and the basal forebrain/brainstem nuclei that underlie stress-related neurotransmitter production, all which work together to process threatening stimuli and facilitate behavioral response (e.g., Armony & LeDoux, 1997; Davis et al., 1997; LeDoux, 1996; Phan, Wager, Taylor, & Liberzon, 2002; Phelps, 2004, 2006). Other key survival circuits exist. Notable among these is the reward circuitry of the brain, within which the ventral striatum plays a key role. The amygdala and ventral striatum are highly connected with

one another and with the anterior insula, with which they share many functional similarities. The more domain-general modulatory regions (e.g., insula, anterior cingulate cortex, orbitofrontal cortex) integrate information from multiple survival circuits with interoceptive information and prior experience to provide threat evaluation or response modulation (e.g., Bremner & Vermetten, 2001; Kaufman & Charney, 2001; Sanchez, Ladd, & Plotsky, 2001) (also see supporting human neuroimaging data : e.g., Dolcos, LaBar, & Cabeza, 2004; H. Kim et al., 2004; Urry et al., 2006; Wager, Phan, Liberzon, & Taylor, 2003). The high level of structural and functional connectivity between these brain regions and the rest of the cortex allows this extended defense circuitry to recruit and integrate a wide range of the brain’s computational resources to process threatening or stressful stimuli (Pessoa, 2008; Stephan et al., 2000). Such stimuli effectively engage attention, inhibit other activities (Choi & Brown, 2003; Gray, 1987; Petrovich, Setlow, Holland, & Gallagher, 2002), and are prioritized in the competition for processing resources in the brain (e.g., Anderson, 2005; Ohman, Flykt, & Esteves, 2001; Vuilleumier, Armony, Driver, & Dolan,

Diathesis and Stress in the Brain

Prior Context Incl. prior life events, prior health behaviors, prior coping behaviors prior SES

567

Current Context Risk, resources

Current Stressor

Current Contextual Resources proximal, distal

Sensory Processing

Prior Mental Health

Central Allostatic Load modification within core emotional regions first

Complex Representation of Sensory Input Central Allostasis Specific and general survival circuits of the brain

Current Mental Health

Inside the brain Peripheral Allostatic Load chronic changes in cortisol, DHEA, NE, epi, etc., damage to cardiac, pulmonary, pancreatic and other peripheral systems

Peripheral Allostatic Accommodation HPA axis, immune, cardiovascular, etc.

Figure 15.6 Including allostatic load in the model. The allostatic model, including the role of central and peripheral allostatic load in the response to a current stressor. Interaction would be expected between allostatic load and major pathways in central allostatic accommodation (not shown). The role of genes is not shown here. Source: Adapted from B. L. Ganzel, P. A. Morris, & E. Wethington, Allostasis and the human brain: Integrating models of stress from the social and life sciences, Psychological Review, 117(1), 134, 2010.

2001). This provides salient (in our exemplar, threatening) stimuli with a privileged status6 in the brain (Davidson et al., 2004; LeDoux, 2012) and underscores the critical role that the survival circuits play in allostasis. Allostasis Across Time: Allostatic Load The expanded version of this model (Figure 15.6) shows the accumulation of physiological wear and tear in the form of allostatic load (McEwen & Stellar, 1993). Under chronic and repeated exposure to stressors, the short-term gains of allostatic accommodation are reduced and associated physiological adaptations may become entrenched and automatic (e.g., chronic high blood pressure: Sterling & Eyer, 1988). In this figure, previous stressors have also contributed to the need for allostatic accommodation, which, over time, has accumulated in the form of allostatic load. Acquisition of load, in turn, affects the individual’s ability to respond to a current stressor. In this way, the body carries with it the long-term results of prior experiences 6

We note that this privileged status is vulnerable to high attentional load and strongly competing sensory input (e.g., Pessoa, Japee, Sturman, & Ungerleider, 2006; Pessoa, McKenna, Gutierrez, & Ungerleider, 2002).

into its current responses to a stressful event (Geronimus, Hicken, Keene, & Bound, 2006). The relationship between stressor exposure and adaptive physiological outcomes takes the form of an inverted U-shaped curve (e.g., McEwen & Lasley, 2002). With mild stressors, the accumulation of load that results from ongoing allostatic accommodation is likely to be negligible—well within the elastic limit of human resilience described by Cannon (1932) in his engineering model of stress. In the adaptive portion of the U-shaped curve, the benefits of allostatic accommodation outweigh the costs. Moderate amounts of stress (good stress or eustress: Lazarus, 1966; Selye, 1974) may be associated, for example, with enhanced neurogenesis (e.g., Kirby et al., 2013; Rhodes et al., 2003), improved brain metabolism and enhancements of neuronal architecture (e.g., Kempermann, Kuhn, & Gage, 1998; Sirevaag & Greenough, 1988), improved executive function (Colcombe et al., 2003), and improved immune system function (Sapolsky, Romero, & Munck, 2000). For example, moderate amounts of exercise are associated with increased cortical plasticity and improved executive function, even in adults (Colcombe et al., 2003). However, allostatic load will increase over time when allostatic accommodation to a

568

Life Span Diathesis–Stress Model of Psychopathology

stressor is sustained, when it is repeated, or when it is large (McEwen, 1998, 2000; McEwen & Stellar, 1993). Studies of peripheral allostatic load using a composite index of biological risk factors (e.g., cortisol, norepinepherine, epinephrine, glycosated hemoglobin, systolic and diastolic blood pressure, cholesterol, waist-to-hip ratio) suggest that load is accumulated according to stress dosage, such that increased stressor exposure is associated with increased load (e.g., Brody et al., 2013; Evans, Kim, Ting, Tesher, & Shannis, 2007; Glover, Stuber, & Poland, 2006). In addition, different types of stressors are likely to accumulate allostatic load somewhat differently, and accommodation to some types of stressors may be more costly than others (Alleva & Santucci, 2001; Pacák & Palkovits, 2001). The tipping-point in the inverted U-shaped curve is the point at which a healthy challenge becomes a progressively unhealthy stressor. Healthy exercise, for example, can segue to being worked to death. Within the body, there is a point at which stressor exposure begins to compromise immunity, neurogenesis, dendritic growth in the hippocampus, and the host of other negative physiological consequences of increasing stressor exposure (for a review of these, see Ganzel et al., 2010). In terms of the present model, this is the point at which the process of allostatic accommodation begins to generate deleterious amounts of allostatic load. Under chronic, repeated, or severe stress, the short-term gains of allostatic accommodation dwindle, while its physiological adaptations become entrenched and automatic (fixed automatisms, see: Sterling & Eyer, 1988) and the long-term physiological costs of that sustained accommodation continue to accumulate. Under extreme stress, this process may happen very quickly. Fast or slow, continuing accrual of load-related symptoms will be physiologically overwhelming and ultimately lethal, as described in Selye’s general adaptation syndrome (GAS; Selye, 1950, 1956). On the way to the deadly end of stage three of the GAS, it is increasingly likely that the individual will experience physiological and/or psychological failure to compensate for intolerable levels of stressor exposure (e.g., Seeman et al., 2004; Seeman, McEwen, Rowe, & Singer, 2001; Seeman, Singer, Rowe, Horwitz, & McEwen, 1997). That individual’s diathesis will determine when decompensation is reached, and how, but no one is exempt. Vulnerable individuals simply reach decompensation earlier and perhaps in a nonstandard way. Allostasis Across the Life Span and Generations: Epigenetics Finally, recent work suggests that continuity in load and its associated physiological consequences may reflect

intergenerational contributions of environmental experience through epigenesis, or environmental effects on gene expression (Moffitt, Caspi, & Rutter, 2006). What is important for this discussion is that stress in the environment may result in changes in DNA expression that can have implications for an individual’s physiology and mental health throughout that individual’s own life span and for subsequent generations (Harper, 2005). As such, environmental conditions present in one generation are remembered in the behavioral and physical responses of the subsequent generation even after the precipitating environmental conditions have ended (see Harper, 2005, for a discussion). Some of the most compelling empirical evidence driving this theory comes from studies of rodents in Meaney’s lab (Champagne, 2010; for reviews, see Diorio & Meaney, 2007; e.g., Francis, Diorio, Liu, & Meaney, 1999; Weaver et al., 2004). This work demonstrates that rat pups who experienced low levels of maternal care during the first week of lactation go on to show more fear of novelty as adults and increased HPA-axis reactivity to stressors, relative to pups who experienced higher levels of care from their mothers. In this example, DNA methylation patterns appear to be the key mediating mechanism (Weaver et al., 2004). That is, alterations in the methylation of a glucocorticoid receptor gene affect the hippocampus’ ability to regulate glucocorticoid negative feedback on the HPA axis, producing a lasting increase in HPA-axis response in offspring who received lower levels of maternal care (Weaver et al., 2004). This is an example of epigenetics serving as the mechanism through which environmental conditions produce central allostatic accommodation to environmental challenge (for a discussion, see Ganzel & Morris, 2011). This environmentally-mediated epigenetic alteration in hippocampal function carries forward in the life of the rat pup as allostatic load. These changes appear to have long-lasting implications to future generations (Champagne, 2010). Notably, parallel processes may take place among humans, as demonstrated by evidence that severe food shortages during pregnancy have similar lasting effects on later health of the child (physical growth; Susser & Stein, 1994). In this way, epigenetics serves as a key process by which early experiences may carry forward to bring lasting modifications in the individuals’ response to stressors at later points in development and in future generations. These discoveries require diathesis–stress models of psychopathology to reflect past history of adversity. Future models might benefit from including past history of the parent generation, as well (for a first foray, see Ganzel et al., 2010).

The Case for a Constantly Changing Brain

THE CASE FOR A CONSTANTLY CHANGING BRAIN In a prior theory paper (Ganzel & Morris, 2011), we presented a set of models that were a first step toward a more developmentally informed understanding of allostasis and stress-related health outcomes. In that work, we touched on how stress-sensitive periods may manifest as windows of plasticity, on one hand, and as windows of vulnerability on the other. Our aim in the present chapter is to expand on this previous work by applying it to diathesis–stress models of the origins of psychopathology. To best do so, we will first provide an overview of early brain development. As with other complex systems, brain development is observable across multiple levels of analysis. So to lay the groundwork for our allostatic diathesis–stress model, we weave together information from multiple levels of analysis, from molecules to genes to large-scale neural networks. Activity across these levels of analysis can be observed across a wide range of time spans (milliseconds to years) using multiple methods of measurement. Electroencephalography (EEG) measures millisecond changes in voltage from the scalp that directly indicate electrical current flow in neurons. Magnetic resonance imaging (MRI) gives a static measure of tissue densities in the brain, thus providing a spatially detailed map of brain structure that we discuss here as gray matter and white matter. Diffusion tensor imaging (DTI) allows the mapping of white matter fiber tracts and their integrity. Functional MRI provides a dynamic (although indirect and somewhat slowed) measure of brain activity—usually in response to a stimulus or task (fMRI) but also in the resting state (RSFC). These are all techniques that can be used to study the brains of living people; we will also refer to findings from studies of dead brains (e.g., immunohistochemistry of autopsy tissue), which provide very detailed information about brain structure that cannot be obtained from living humans. Our goal here is to provide relevant snapshots of brain development at these several levels of analysis to illustrate our point that the brain undergoes constant nonlinear change across the life span. We then develop our allostatic diathesis–stress model within this framework. We will begin with a focus on the relatively earlymaturing processes that underlie increasing functional specificity within local (several cubic millimeters) ensembles of neurons or nodes. We follow with a discussion of the slower-growing white matter tracts that link these nodes into mature functionally integrative large-scale neural networks that span the entire brain (Hagmann et al., 2010; Honey et al., 2009; Stevens, 2009; van den Heuvel,

569

Mandl, Kahn, Pol, & Hilleke, 2009). The vastly enlarged capabilities of the mature human brain are thought to be supported by synchronous neural activity within these complex large-scale networks (e.g., Fair et al., 2007; Fries, 2005) and there is increasing evidence that dysfunction within these networks plays a prominent role in psychopathology (Menon, 2011; Uhlhaas & Singer, 2012). The emergence and senescence of these nodes and networks across the life span creates a constantly changing neural landscape within which diathesis and stress intersect to predict mental health. This sets the stage for the subsequent section, in which we integrate this information into new diathesis–stress models. Early Human Neurodevelopment The central nervous system (CNS) begins to form during the first few months of fetal development in humans (the first few weeks of gestation in rodents; Rice & Barone, 2000). The neural plate is formed first via neural induction, then a further series of inductive processes leads to the expression of different developmental programs within distinct subregions of the new neural plate. These dictate the formation of the major subdivisions of the CNS and the early progenitors of its major structures (e.g., Rakic, 1995; Rubenstein, 2009). Subsequently, region-specific early precursor cells proliferate outward following radial glial processes. In the process, they undergo differentiation, axogenesis, synaptogenesis, pruning, and myelination as they form the mature neural circuits that underlie normal physiological function and complex behavior (Caviness & Takahashi, 1995; Oda & Huttenlocher, 1974; for a comprehensive review, see Rubenstein & Rakic, 2013). The wiring of the human nervous system is accomplished through the construction and sculpting of connections that foster information exchange among multiple cell types (e.g., Courchesne, Chisum, & Townsend, 1994). The two primary cell types within the brain are neurons and neuroglia (glia). In neuroimaging terms, neurons make up the gray matter of the brain, which is composed of neuronal cell bodies, dendrites, and short axons, while neuroglia (oligodendrocytes in particular) make up the white matter of the brain, the most visible element of which is the pale-colored myelination on long neural axons that connect regions of gray matter (e.g., Wonders & Anderson, 2006). In the period between the third prenatal trimester and the third year of life in humans, a period of heightened apoptosis (programmed cell death) eliminates about half of previously generated neurons (e.g., Davies, 2013; Lossi & Merighi, 2003). In the mature brain, the ratio of

570

Life Span Diathesis–Stress Model of Psychopathology

neurons to glia varies by brain region. For example, the adult human cerebral cortex has approximately four times more glia than neurons, while the cerebellum has four times more neurons than glia (Pelvig, Pakkenberg, Stark, & Pakkenberg, 2008). Neurons have long been considered the core components of the nervous system, and the mechanisms that guide neuronal maturation and senescence are well studied (e.g., Rubenstein & Rakic, 2013). Until relatively recently, however, neuroglia have been thought to play passive structural or housekeeping roles in the brain (e.g., myelination, ionic homeostasis, metabolic support of neighboring neurons and glia: Kurth & Huber, 2012; Parpura & Verkhratsky, 2012; Wang & Bordey, 2008). It has become clear, however, that different classes of neuroglia variously (1) interact with neurons to allow signal transfer at synapses, (2) play a critical role in brain development (for a review, see Wang & Bordey, 2008), (3) function as neural stem cells (e.g., Cameron, Woolley, McEwen, & Gould, 1993; Doetsch, Caillé, Lim, García-Verdugo, & Alvarez-Buylla, 1999) and as precursor cells for other types of glia (e.g., Alvarez-Buylla & Kriegstein, 2013). There is also preliminary evidence that glia (e.g., oligodendrocyte precursor cells) may use neurotransmitters to signal to other glia to allow ongoing plasticity in white matter structure and thus promote ongoing plasticity in the functional capabilities of neural networks (Alvarez-Buylla & Kriegstein, 2013). The fact that both neurons and glia play unique and critical functions in the brain argues for the consideration of both in a developmentally informed diathesis–stress process. Early Functional Specialization of Local Neural Networks: Focus on Gray Matter Early Cortical Development via Proliferation and Pruning In humans, much of the basic adult form of the brain (overall size, weight, cortical folding, and rudimentary functional specialization) is in place by early to middle childhood (Armstrong, Schleicher, Omran, Curtis, & Zilles, 1995; Giedd et al., 1999). During this time, local brain areas begin to develop varying levels of specialization, i.e., they learn to process specific types of information in specific ways. Refinement of functional specificity in local information processing involves the formation and experience-dependent remodeling of synapses. This process is largely complete by young adulthood. This early phase of adaptive organization/re-organization is instantiated in developmental changes in gray matter thickness, which are measurable using magnetic resonance imaging (MRI: e.g., Good et al., 2001). Cortical (gray matter) thickness

rises and falls across the life span; it is greatest around 10 years of age in girls and around 12 years of age in boys (e.g., Giedd, 2004; Sowell et al., 2003; Westlye et al., 2010b). Different brain regions vary as to when they reach maximum thickness. In general, the human brain develops from posterior to anterior and from the medial regions out to the lateral portions of the brain. For example, the occipital pole (the primary visual processing area at the very back of the brain) is one of the first brain regions to reach its maximum gray matter thickness (at about age 7.5 years; e.g., Shaw & Gross, 2008), while the superior and lateral prefrontal regions are among the last to reach their maximum thickness at ages 10–12 (between 10 and 12 years of age: Giedd, 2004; Shaw & Gross, 2008). Maximum thickness does not, however, define the maximum adult cognitive capacity that is yet to come. In humans, neural synaptogenesis and the expression of synaptic genes in the prefrontal cortex begin to rise steeply before birth; this continues until the fifth postnatal year, after which rates increase more slowly until about age 10 (Liu et al., 2012). Increased synaptogenesis is associated with enhanced cortical plasticity and roughly coincides with increasing gray matter thickness and increases in cognitive capacity (McArdle, Ferrer-Caja, Hamagami, & Woodcock, 2002). This early proliferation of synapses is countered by bouts of accelerated synaptic pruning, the first of which occurs in early childhood and the second of which occurs during later adolescence (Gogtay et al., 2004; Rakic, Bourgeois, & Goldman-Rakic, 1994). During pruning, the neurons themselves are preserved but any neural connections (synapses and dendrites) that have not undergone experience-dependent strengthening are eliminated, leaving a more efficient circuit, as well as room in the skull for continuing myelination (for a discussion, see Paus, Keshavan, & Geidd, 2008). Pruning is argued to streamline the neural circuitry of the brain to do best what it has been doing, and it thus defines a period when experience may profoundly shape brain structure and function (i.e., the spontaneous or evoked activity of a neural system specifies its long-term connectivity: e.g., Changeaux & Danchin, 1976; Huttenlocher & Dabholkar, 1997). The effects of synaptic pruning during adolescence are profound (see Figure 15.7). Postmortem studies allow direct count of synapse density. Such studies of humans and macaques have found a 50 to 55% decrease in the number of synapses across the entire cortex between late childhood and early adulthood (Good et al., 2001; Huttenlocher & Dabholkar, 1997; Rakic, Ayoub, Breunig, & Dominguez, 2009). These findings are reflected in MRI research, which has identified a 40% thinning in cortical

The Case for a Constantly Changing Brain 20

4 Years

Synopses/mm3 × 100

1 Month

571

15

10

5 NB

0.5

1

5

10

15 20 YEARS

40

60

80

100

Figure 15.7 Pruning in multiple waves. These figures illustrate the acceleration of neuronal synaptogenesis and dendritic arborization in early childhood, followed by a rapid decline (via pruning) during adolescence (note expansion of time scale in early years). This These waves of synaptic pruning are associated with brain maturation and improved cognition. Another decline in synapse number and cortical thickness occurs late in life and is associated with brain senescence and impaired cognition. Source: Adapted from I. Feinberg & I. G. Campbell, Sleep EEG changes during adolescence: An index of fundamental brain reorganization, Brain and Cognition, 72(1), 59, 2009.

gray matter during adolescence (Shaw & Gross, 2008; Sowell et al., 2003). This cortical thinning is a marker of brain maturation and is associated with enhanced cognitive ability. It is considered to be a distinctly different process from the cortical thinning in the latter half of the life span, which is associated with the atrophy of large neurons, loss of neuropil (mostly dendritic arbors and synapses), and white matter degeneration (Bertoni-Freddari et al., 2002; Terry, DeTeresa, & Hansen, 1987; Westlye et al., 2010b). As with cortical thickening and synaptic proliferation, synaptic pruning and cortical thinning/maturation proceed from posterior to anterior and from medial regions out to the lateral portions of the brain. Thus, the occipital pole (shown in black in Figure 15.8) is one of the first brain regions to mature, while the prefrontal and lateral temporal regions are among the last (shown in lighter shades in Figure 15.8; Giedd, 2004; Westlye et al., 2010b). Regional differences in the development of brain structure can be mapped onto maturational differences in brain function, leading to the progressive maturation of cognitive and perceptual capabilities. For example, the early maturation of the primary visual cortex (occipital pole) can be demonstrated through the early emergence of visual acuity (Teller, 1981). Complex motor skills mature later, during adolescence, and executive skills (associated with anterior and lateral PFC function) are not fully developed

Medial

Lateral >32 20 60 years) with major depression show an increase in astroglia, combined with a notable reduction in pyramidal neurons. Because of this, Rajkowska and Miguel-Hidalgo (2007) proposed that many years of stress-related elevation in extracellular glutamate does substantial neuronal damage in aging depressives (Load Three). This exemplifies the life span development of load and an individual’s diathesis (genes + load). To explain the age-related increase in glia, Rajkowska and Miguel-Hidalgo invoked the well-established tendency for glial cells to proliferate in response to neuronal damage—in this case, damage due to glutamate excitotoxicity (2007). We note that this model requires repeated stressor exposure (e.g., Stressor Four) to drive the ongoing, changing burden of allostatic load (e.g., Load Four). However, exposure to stress and trauma is relatively common in the general population (Copeland, Keeler, Angold, & Costello, 2007; Kessler, Sonnega, Bromet, Hughes, & Nelson, 1995) and psychopathology itself can serve as a stressor (e.g., Bruce, 1998). Hypothesis Generation Here, we depart from our reconceptualization of Rajkowska and Miguel-Hidalgo’s (2007) glial model of depression to explore some of the ways this model—and our own overall approach to diathesis–stress modeling— might be used in hypothesis generation. For example, earlier we noted that astrocytes regulate non–rapid eye movement (NREM) sleep, as well as modulating sleep need, sleep intensity, and EEG slow waves, i.e., the slow neuronal oscillations that occur during NREM sleep (Fellin et al., 2009; Halassa et al., 2009). If stress-related astrocyte pathology is integral to depression, then these properties of NREM sleep should also be dysregulated in depression. There is a growing body of evidence for this. Dysregulated sleep is a core symptom of depression and there is a recent research focus on the interplay between astrocyte function, sleep (esp. NREM sleep), and depression in humans and in animal models (e.g., Cao et al., 2013; Duric et al., 2010; Florian, Vecsey, Halassa, Haydon, & Abel, 2011; Hines, Schmitt, Hines, Moss, & Haydon, 2013). Also from earlier, recall that EEG slow-wave activity during NREM sleep is argued to play a crucial role in synaptic homeostasis by downscaling LTP-related increases in synaptic weight, thus preserving memory and optimizing

597

learning (e.g., Tononi & Cirelli, 2006). We would therefore expect that learning and memory would be impaired in depressives in direct relationship to their degree of astrocyte pathology and compromised LTP-related synaptic homeostasis. While astrocytes are observed to modulate cognitive declines associated with sleep loss (e.g., Halassa et al., 2009), to our knowledge, Tononi & Cirelli’s (2003, 2006, 2012) synaptic homeostasis hypothesis has not been explored as a possible mechanism associating astrocyte pathology and the cognitive impairments observed in depression. The R-MH model of cellular pathology in depression (Figure 15.20) embraces a life span perspective. Specifically, this model predicts pathological increases in the density of astrocytes and a parallel decrease in pyramidal neurons in key brain areas in aging depressives. It follows from the earlier discussion that we would expect to see progressive changes in NREM sleep biology, sleep-related symptoms, and cognitive deficits as depressed individuals grow old. Similarly, if age-related variation in glial and neuronal density is driven by stressor exposure, then we would expect people who have experienced more stress and trauma in lifetime to show an earlier transition to the aged scheme of depression neurobiology (more glia, fewer neurons), relative to depressed individuals with less lifetime stressor exposure. We leave these questions to future research. In exploring this example, we do not suggest that astrocyte pathology is the only important factor in diathesis– stress models of depression. Our goal has been to demonstrate a specific use of our diathesis–stress model for hypothesis generation, as well as to emphasize that life span development is likely to underlie at least some of the observed heterogeneity in the neurobiology of psychopathology. We next turn to broader questions raised by our allostatic diathesis–stress model.

TRIPLE NETWORK ALLOSTASIS AND PSYCHOPATHOLOGY In this section, we incorporate current research on the large-scale organization of the human brain into our allostatic model of response to a current stressor (Figure 15.4). This effort incorporates a neural network approach to psychopathology into our allostatic model of the stress process, discussed in detail in Sections Two and Four. Incorporated into a diathesis–stress framework, this approach serves as the basis for hypotheses about allostatic load within the large-scale neural networks the brain, and

598

Life Span Diathesis–Stress Model of Psychopathology

its role in the onset and maintenance of psychopathology. We review emerging evidence for these effects and segue to directions for future research. Dysfunctional Large-Scale Neural Networks in Psychopathology A Paradigm Shift in Psychopathology Research The intrinsic neuronal signaling within the human brain has been compared to the dark energy of the universe (Zhang & Raichle, 2010). New astrophysical evidence suggests that dark energy makes up approximately 68% of the mass energy of the universe, so this recently acknowledged (and largely hidden) component of the universe in fact constitutes most of the universe (Ade et al., 2013). This discovery is paralleled by recent findings in human neuroscience that less than 5% of the brain’s resources are devoted to processing external stimuli—thus leaving the vast majority of the brain’s energy budget devoted to previously uncharted intrinsic activity within brain systems (Raichle & Mintun, 2006). Exploration of this intrinsic activity has resulted in a functional network model of the neural function that is based on stable large-scale networks of synchronized neuronal activity. This model constitutes a paradigm shift in systems neuroscience (Sylvester et al., 2012), with a consequent paradigm shift under way now in clinical neuroscience and neuropsychiatric research (Menon, 2011; Sylvester et al., 2012; Uhlhaas & Singer, 2012). Neuroimaging studies of psychopathology are increasingly focused on dysfunctions within large-scale neural networks. These dysfunctions can occur via damage to the nodes of a functional network, to the connections (“edges”) between nodes, or to a combination of these. For example, changes in the excitation inhibition balance (Figure 15.21) at the local or node level can disrupt the synchronized neuronal activity in adjacent nodes. Thus, damage to a node can propagate to other nodes to which it is linked in a functional network (Carter, Shulman, & Corbetta, 2012; Ewers, Sperling, Klunk, Weiner, & Hampel, 2011). This may also lead to compensatory changes in other parts of the network (e.g., strengthening connectivity between between other nodes; Crofts et al., 2011). Even if the nodes of a network are healthy, network function is compromised if the connections between nodes are damaged. This can happen, for example, through age-related degeneration of white matter integrity. Damage to axonal myelin results in increased signal degradation between nodes and this, in turn, is associated with disruptions in neuronal synchrony (e.g., Mander et al., 2013; Rajkowska

& Miguel-Hidalgo, 2007; Ringli & Huber, 2011). Aberrations in nodes or edges can derive from initial genetic loading or they can be acquired through experience, insult, or disease at any point in the life span. Analysis of abnormalities in functional neural networks provides a common quantitative framework for a novel synthesis of findings in psychopathology research, as well as highlighting fruitful directions for clinical research. This, in turn, has been productive in providing new insights into the neural underpinnings of schizophrenia, autism, anxiety disorders, and a wide range of other major forms of psychopathology, as we discuss later. A Triple Network Model of Psychopathology There is a groundswell of new research suggesting that some, if not all, psychopathologies are associated with more or less unique abnormalities in the functional connectivity of large-scale neural networks (for reviews, see Menon, 2011; Seeley, Crawford, Zhou, Miller, & Greicius, 2009; Uhlhaas & Singer, 2012; Zhang & Raichle, 2010). This suggests the alluring prospect that functional connectivity MRI (fcMRI) may be diagnostic at some point in the future. Meanwhile, research in fcMRI is driving new theorizing about neuropsychiatric disorders as pathologies of networks rather than brain regions. We will focus our discussion on an integrative model of large-scale neural network dysfunction in psychopathology, the triple network model (Menon, 2011). The core elements of this model are the three canonical neurocognitive networks discussed in Section Three—the default mode network (DMN), the salience network (SN), and the central executive network (CEN). As discussed, these three networks serve as the framework of the brain’s intrinsic functional architecture, and there is growing consensus that disturbance in the function or organization of these three networks are fundamental to many neuropsychiatric disorders (e.g., Menon, 2011; Seeley et al., 2009; Uhlhaas & Singer, 2012). Menon (2011) placed the salience network in general— and the right anterior insula in particular—at the core of triple network theory. As such, the salience network has a pivotal role in the onset of affective and cognitive dysfunction. In support of this hypothesis, he points to a robust pattern of structural or functional abnormalities in bilateral or right insula in schizophrenia (Glahn et al., 2008; Palaniyappan & Liddle, 2012; Palaniyappan, Mallikarjun, Joseph, White, & Liddle, 2011; Sommer et al., 2008; White, Joseph, Francis, & Liddle, 2010), bipolar disorder (Bora, Fornito, Yücel, & Pantelis, 2010), frontotemporal dementia (Seeley et al., 2006; Zhou et al., 2010), addiction (Naqvi & Bechara, 2009; Scott & Hiroi, 2011), anxiety

Triple Network Allostasis and Psychopathology

Current Context Risk

599

Current Environmental Stressor

proximal, distal

Current Contextual Resources Proximal, distal

Sensory Processing

Central Allostasis

Inside the brain

Current Mental Health

Peripheral Allostasis HPA axis, immune, cardiovascular, etc.

(a)

Prior Context Incl. prior life events, prior health behaviors, prior coping behaviors prior SES

Current Context Risk, resources

Current Stressor

Current Contextual Resources proximal, distal

Sensory Processing

Prior Mental Health

Central Allostatic Load

Central Allostasis Current Mental Health

modification within core emotional regions first

Peripheral Allostatic Load

Peripheral Allostatic Accommodation

chronic changes in cortisol, DHEA, NE, epi, etc., damage to cardiac, pulmonary, pancreatic and other peripheral systems

HPA axis, immune, cardiovascular, etc.

(b)

Figure 15.21 Triple network allostasis, including (a) the role of the canonical large-scale neural networks in the response to a current stressor and (b) the role of prior experience and prior allostatic load. Interaction would be expected between allostatic load and major pathways in central allostasis (not shown). DMN = default mode network; CEN = central executive network; SN = salience network; Survival = survival circuits (LeDoux, 2012). See footnote 1. Source: Adapted from B. L. Ganzel, P. A. Morris, & E. Wethington, Allostasis and the human brain: Integrating models of stress from the social and life sciences, Psychological Review, 117(1), 134, 2010; V. Menon, Large-scale brain networks and psychopathology: A unifying triple network model, Trends in Cognitive Sciences, 15(10), 483–506, 2011.

600

Life Span Diathesis–Stress Model of Psychopathology

(Etkin, Prater, Schatzberg, Menon, & Greicius, 2009), autism (Allman, Watson, Tetreault, & Hakeem, 2005; Santos et al., 2011), and pathological pain processing (e.g., chronic migraine; Hadjikhani et al., 2013; irritable bowel syndrome; Hong et al., 2013). Menon’s (2011) proposal is based on the role of the anterior insula and the rest of the salience network in (1) detecting salient stimuli and (2) recruiting the central executive network or the default network to generate adaptive behavioral response, as discussed earlier. In his triple network model, abnormal salience detection of external stimuli or internal mental events is a core feature of psychopathology. In anxiety disorders, for example, overactive salience detection may result in misattribution of salience to relatively neutral or minor events. Atypical salience mapping could also stimulate atypical engagement of activity in the other canonical neurocognitive networks, including attention and executive control via the CEN or self-referential processing via the DMN (e.g., autobiographical memory; assignment of personal value to a stimulus). In depressed individuals, atypical salience control signals may prompt pathological rumination through hyperactivation of the default network (Berman et al., 2011; Menon, 2011; Sheline et al., 2009). In individuals with addictions, salience may be incorrectly mapped, resulting in heightened salience of drug-related cues and decreased CEN-based cognitive control of response to these cues. There is now evidence for this model of addiction through lesion studies (Naqvi & Bechara, 2009, 2010; Naqvi, Rudrauf, Damasio, & Bechara, 2007) and functional connectivity studies (Cisler et al., 2013), which have identified a crucial and unique role for the right insula in addiction. Salience mapping may also play a key role in psychotic disorders. In schizophrenia, bilateral anterior insula volume is reduced relative to healthy controls, and these volume reductions are correlated with increasing symptoms of psychosis (Palaniyappan et al., 2011). Node-level dysfunction within the anterior insula may result in impaired connectivity within the salience network itself, and this may produce hallmark symptoms of schizophrenia. For example, disruptions in the synchrony between anterior insula and the dorsal ACC/presupplementary motor area (preSMA) could disregulate motor planning and motor execution, resulting in the psychomotor poverty or motor overflow associated with schizophrenia (Menon, 2011). Atypical activation of anterior insula is also associated with auditory verbal hallucinations, suggesting that abnormal processing of salience plays a role in these symptoms. It has been suggested that salience network

dysfunction in schizophrenia underlies both the positive and negative symptoms of this disorder through disrupted recruitment of the DMN and the CEN, which together engender hallucinations and delusions and deficits in information processing (Palaniyappan et al., 2011). In the triple network model, abnormal salience detection may have reverberating consequences not only within the salience network, but within either or both of the other canonical networks as well (Menon, 2011). Depending on diatheses and life experience, these downstream effects in other networks may be larger than the initial deficit in salience processing. Moreover, the small-world structure of each network will uniquely constrain the manifestation of these effects. In this way, the distinctive constellations of symptoms associated with particular neuropsychiatric disorders may arise out of characteristic perturbations/malfunctions within the functional architecture of the brain. While there is some disagreement regarding whether the amygdala and ventral striatum are core elements of individual survival circuits or if they belong to the general-purpose salience network, Menon (2011) included them in the salience network (which we are calling the survival/salience network for clarity).15 Menon contended that these nodes are particularly relevant to psychopathology. Amygdala dysfunction plays a consistent role in anxiety disorders (e.g., Etkin & Wager, 2007) and shows atypical function or structure in a number of other psychiatric disorders, while ventral striatum function is clearly aberrant in addiction and schizophrenia, among others (e.g., Barch & Dowd, 2010; R. W. Morris et al., 2011; Volkow, Wang, 15

Recall from earlier that the amygdala and ventral striatum are key elements of two major survival circuits (LeDoux, 2012) associated with defense/avoidance and reward/approach, respectively. We also previously noted that the literature has been ambiguous about distinguishing the survival circuitry from the salience network. The distinction between these circuits depends on the type of study paradigm used. As discussed previously, activity in the salience network can be studied at rest but co-activation of survival circuitry is sporadic and highly variable. Reliable activation of any given survival circuit requires exposure to an appropriate circuit-specific trigger stimulus (e.g., Bachis et al., 2008; Ganzel et al., 2008; Kang et al., 2012; LeDoux, 2012; Phan et al., 2004; Phan et al., 2002). In experimental paradigms that do include these trigger stimuli, there will be activation of relevant survival circuitry in addition to the salience network. For clarity, then, we have elected to use the term survival/salience network in discussing co-activation of the salience network and one or more of the survival circuits in salience processing. Under stress, there will activation of this combined network.

Triple Network Allostasis and Psychopathology

Fowler, Tomasi, & Telang, 2011). As previously discussed, the amygdala and ventral striatum are highly connected with one another and with the anterior insula, with which they share many functional similarities. For example, Pessoa and Adolphs (2010) argued that the function of the amygdala in many cortical networks is primarily modulatory, in that it processes the biological value of stimuli (including ambiguity, salience, unpredictability) and serves to allocate and prioritize processing resources to generate adaptive behavior. If this sounds similar to the functions described for the anterior insula, discussed previously, it should be noted that studies of nonhuman primates indicate that the insula and amygdala are tightly linked, as well as being physiologically adjacent. The amygdala consists of multiple separate nuclei, and the insula has reciprocal connections with nearly all of them (Mufson, Mesulam, & Pandya, 1981), which helps to explain why the two structures appear functionally similar. Like the anterior insula, the amygdala is highly interconnected with much of the cortex, particularly medial and orbital PFC, but also much of the rest of the PFC. Studies from macaques indicate that amygdala inputs are able to reach approximately 90% of the PFC after a single connection (Averbeck & Seo, 2008) and that the amygdala constitutes a core brain area in terms of global brain connectivity (Modha & Singh, 2010). As previously discussed, however, the differing circumstances under which the insula, amygdala, and ventral striatum are functionally connected suggest that these three brain regions each play very different roles in the onset and maintenance of psychopathology. As previously noted, functional connectivity between anterior insula and the rest of the salience network is observable all of the time, even at rest. On the other hand, the amygdala and ventral striatum and their respective survival circuits join robustly with the salience network only upon exposure to an appropriate circuit-specific trigger stimulus (e.g., Bachis et al., 2008; Ganzel et al., 2008; Kang et al., 2012; LeDoux, 2012; Phan et al., 2002, 2004). Menon (2011) argued for the unique integrative role of the right anterior insula in linking interoceptive/homeostatic information with the central processing of salient/survival related stimuli, including its command role in switching between the default and central executive networks. For this reason, the right anterior insula is the fulcrum, the pivot point, of the triple network model. Triple Network Allostasis: A Novel Diathesis–Stress Model Menon’s (2011) triple network model of psychopathology provides a valuable heuristic for conceptualizing

601

the process by which dysfunction within the large-scale neural networks of the brain predicts mental disorder. We incorporate this approach in developing an allostatic diathesis–stress model that takes the large-scale functional architecture of the brain into account. We refer to the resulting model as triple network allostasis, the core components of which are the canonical large-scale neural networks discussed earlier. This is an expansion of Menon’s (2011) approach to include current contextual risk and resources, prior experience, life span development, peripheral stress physiology, and a dynamic diathesis into a working model that predicts mental state, psychopathology onset, and future mental health. Triple Network Allostasis Previously, we argued for the core affective16 systems of the brain as the central mediator of allostasis (Ganzel & Morris, 2011; Ganzel et al., 2010). This followed from a growing body of evidence indicating that these regions are the main organizing factor in the translation of environmental stimulus into behavioral or physiological response (e.g., LeDoux, 1996; Ohman & Mineka, 2001; Rosen & Schulkin, 1998). In light of recent research on large-scale neural networks, we update this prior model by placing the survival/salience network in this key mediating role. As with Menon’s (2011) triple network model, this proposal is based on the role of the survival/salience network in detecting the biological significance of external stimuli or internal events. The salience network is positioned to integrate information from the survival circuits (Cameron et al., 1993) with interoceptive information (e.g., Damasio, 2003; Rakic, 2009) to facilitate threat evaluation, selection of motor and homeostatic response, as well as response modulation (Touroutoglou et al., 2012). Moreover, the ability of the anterior insula/dorsal ACC hub to switch among canonical networks (Kempermann et al., 1998) provides flexible access to autobiographical memory, prospection, and semantic memory via the default network, and to executive control and goal-directed planning via the central executive network. Thus, the survival/salience network is the most current and precise representation of an organizing force linking salient stimuli with behavioral and physiological outcome. As such, the survival/salience network necessarily constitutes the primary regulator of allostasis, to which all other physiological and behavioral allostatic accommodation is secondary. 16 Previously referred to as the “core emotional regions of the brain” (Ganzel & Morris, 2011; Ganzel et al., 2010).

602

Life Span Diathesis–Stress Model of Psychopathology

It follows from this that the survival/salience network will be the first and primary site to show evidence of long-term wear and tear (central allostatic load) as a part of the cost of physiological accommodation to environmental demand. Accrual of allostatic load in the brain will, in turn, have consequences for peripheral physiology, psychosocial adaptation, and mental health. Identification of the specific neural circuitry that is most likely to undergo long-term change or outright damage from ongoing allostasis will allow better understanding of the physiological and behavioral consequences of environmental adversity. This will, in turn, allow more precision in intervention strategies to aid populations at risk. The survival/salience network is a subset of synchronized nodes within the core affective systems of the brain, as those systems are currently delineated (Barrett & Satpute, 2013; Lindquist & Barrett, 2012), so that this update does not change the fundamental relationships within our prior model (Figure 15.4). In arguing that the survival/salience network is the central mediator of allostasis, we gain greater specificity in our understanding of the mechanisms of allostasis. We also acquire new and broader perspectives on the neural underpinnings of diathesis–stress processes in psychopathology. This model replaces our previous schematic of the organization of information flow and influence during moment-to-moment allostasis. Compared with the life span models discussed previously (e.g., Figure 15.19 ), triple network allostasis as depicted in Figure 15.21 represents a moment in the ongoing stress process—one point in time in which the salience of a current stressor is being computed, and processing resources are being mobilized in response. In a world with perfect graphic representation, progressive developmental variations of triple network allostasis would be incorporated as individual columns of a life span model like that we discussed in the previous section (e.g., Figure 15.19). With mental health as our outcome of interest, this becomes a novel articulation of the classic diathesis–stress model. Peripheral Allostasis and Allostatic Load In triple network allostasis, the salience network is the central mediator of allostasis, so that this neural network would be expected to accrue allostatic load at a faster pace than surrounding tissue (Ganzel et al., 2010). Said differently, the function and structure of the salience network is likely to be critically influenced by individual differences in life experience, which would, in turn, drive differences in accumulated central and peripheral allostatic load. This is an emerging area of stress research; to our knowledge, there are no studies that link prior life stress with

canonical network integrity and autonomic, endocrine, or immune function in nonclinical humans (although see Eisenberger, Taylor, Gable, Hilmert, & Lieberman, 2007; Ganzel, Kim, Altemus, Voss, & Temple, 2007; Gianaros, Marsland, Sheu, Erickson, & Verstynen, 2012). Yet from our allostatic model, we would predict that individual differences in central allostatic load influence almost all peripheral allostatic regulatory functions (McEwen, 2004), and we would expect this to be a key predictor of long-term health and mortality. It may also be a key predictor of emotion, feelings, and sense of self, as it has been argued that these fundamentally derive from interoceptive information conveyed up from the physiological periphery, through the spinal cord and the thalamus, to the anterior insula—which then disseminates this information to the functional networks of the brain (e.g., Damasio, 1999; Damasio & Carvalho, 2013). This is the substance of the iterative feedback loop depicted in Figure 15.21a between central and peripheral allostasis. In this perspective, the irate boss is seen with the eye, perceived in the brain, felt in the gut, and acted upon using integrated brain/body input. Under allostasis, the physiological mechanisms underlying perception and feeling in this model are subject to central and peripheral allostatic load respectively (Figure 15.21b). Stress Vulnerability Within the Functional Architecture of the Brain We previously reviewed evidence for selective stress-related damage or long-term change in individual nodes of the survival/salience network and the default mode network in nonclinical adults and juvenile humans and animals. In this section, we focus on the impact of stress on large-scale neural network function and structure in nonclinical adults, with some reflection on the role of development. In the Laboratory In the laboratory, healthy/nonclinical adults show robust reconfiguration of salience network activity during exposure to highly aversive cinematographic stimuli (Hermans et al., 2011). When viewing aversive stimuli relative to neutral ones, study subjects’ functional connectivity robustly increased in anterior insula, dorsal ACC, amygdala, ventral striatum, ventromedial PFC, and other regions associated with the survival/salience network. Increased synchronized activity within this network was correlated with heightened cortisol and alpha amylase production, as well as with increases in negative affect (Hermans et al., 2011; also see van Marle, Hermans, Qin, & Fernández, 2010). Interestingly, pharmacological manipulation during

Triple Network Allostasis and Psychopathology

exposure to these aversive stimulus showed that these effects were mediated by stress-related production of norepinephrine, but not cortisol. Norepinephrine (NE) neurons project from the locus coeruleus to salience-related brain regions, and firing in these projections varies to support focal attention during mild arousal, and hypervigilance and distraction during extreme arousal (Arnsten, 2009; Corbetta, Patel, & Shulman, 2008; Valentino & Van Bockstaele, 2008). This is consistent with a model of norepinephrine function as the means through which neural network activity—particularly salience network activity—is quickly co-opted in response to a stimulus and then reorganized to support adaptive response (Bouret & Sara, 2005; also see Olson et al., 2011). Also in the laboratory, a similar network reconfiguration favoring salience processing has been observed in response to sustained deep-tissue pain in humans (J. Kim et al., 2013). Moreover, there is evidence in animal models that exposure to uncontrollable stressors can produce generalized sensitization of NE-producing neurons, so that future stressors of different types will drive exaggerated production of NE and behavioral disturbance (for the potential roles of serotonin and BDNF, see Anisman, Merali, & Hayley, 2008; Olson et al., 2011). This suggests one mechanism through which stressor exposure may generate allostatic load (sensitization of NE-producing neurons) that in turn influences future stress reactivity. Laboratory stressors also impact resting state functional connectivity within the default mode network (e.g., Vaisvaser et al., 2013; Veer et al., 2011). In one study, nonclinical adults exposed to a social stress test showed immediate, but transient, increases in resting state functional connectivity between the posterior cingulate cortex (PCC) and medial PFC, inferior parietal lobule, caudate, and thalamus, along with decreased connectivity between the PCC and bilateral posterior insula and lingual gyrus (Vaisvaser et al., 2013). Immediate post-test increases in amygdala-hippocampus functional connectivity were also observed. Among the regions of interest selected for this study, only the stress-related alterations in amygdala-hippocampal connectivity persisted for more than 90 minutes after the stress test. Self-reports of increased subjective stress during testing were related only to increases in amygdala-hippocampal connectivity, and not to alterations within the default network (Vaisvaser et al., 2013). In a RSFC study employing a similar social stress test but different regions of interest, persistent (>90 minutes) increases in functional connectivity were observed between amygdala and medial PFC, and between amygdala and PCC/retrosplenial cortex/precuneus (Veer et al.,

603

2011). This latter finding is consistent with research identifying a direct ascending anatomical projection between the basal nucleus of the amygdala and the retrosplenial cortex in the macaque brain (Buckwalter, Schumann, & Van Hoesen, 2008). Direct reciprocal anatomical connections between amygdala and medial PFC have been previously established (e.g., Ghashghaei, Hilgetag, & Barbas, 2007). Brain development will also play a key role in the interplay between the survival/salience network and the default network during stress. This, in turn, has implications for the accrual of allostatic load and the developmental timing of the emergence of psychopathology. For example, there is a full developmental reversal in functional connectivity between the amygdala (of the survival/salience network) and the medial PFC (of the default network) (Gee et al., 2013). Children younger than ten years of age demonstrate positive connectivity between amygdala and medial PFC during a laboratory stressor, such as viewing fearful faces; However, there is a linear decrease in the valence of this connectivity between the ages of 10 and 22, and amygdala-medial PFC connectivity in nonclinical young adults is robustly negative (Gee et al., 2013). Change in the valence of this association has been shown to mediate the decreases in separation anxiety that normally occur with increasing age, and this is argued to follow from increasing top-down control of amygdala reactivity by the maturing medial PFC (Gee et al., 2013). We note that this is yet another profound alteration in brain function that occurs between 10 and 20 years of age, along with the reversal of the global excitatory/inhibitory (E/I) balance of the brain, the winnowing of synapse density during the massive second wave of pruning, and the abrupt downward shift in brain energy metabolism and amplitude of slow wave EEG activity during NREM sleep that accompanies the migration of the peak amplitude from the back to the front of the brain (and which may underlie the optimization of learning and memory). While these considerable changes are all complex outcomes of brain maturation, their interrelationship is only loosely understood and their impact on the diathesis–stress process is, as yet, unclear. In the Wild There are very few ecologically valid studies of the impact of stress/trauma on functional neural connectivity in nonclinical human adults. Of the studies that do exist, most highlight aberrant connectivity within regions of interest in the default network. Exceptions do exist, though. One such study examined resting state functional connectivity in nonclinical adults with recent exposure to the great Sichuan earthquake of 2008 (Lui et al., 2009). Relative to a

604

Life Span Diathesis–Stress Model of Psychopathology

Figure 15.22 Survival/salience network disruption after the great Sichuan earthquake of 2008. In earthquake survivors, decreased functional connectivity was observed in salience-related circuitry (light-colored lines). Increased regional activity at the nodes of the salience circuit was correlated with increased distress (more symptoms of anxiety and depression). See footnote 1. Source: From S. Lui, X. Huang, L. Chen, H. Tang, T. Zhang, X. Li, . . . A. Mechelli, High-field MRI reveals an acute impact on brain function in survivors of the magnitude 8.0 earthquake in China. Proceedings of the National Academy of Sciences of the United States of America, 106(36), 15412–15417, 2009.

control group scanned prior to the earthquake, nonclinical adult earthquake survivors showed strong evidence of disrupted survival/salience network activity. Specifically, survivors showed decreased global resting state functional connectivity across bilateral insula, ACC, amygdala, hippocampus, striatum, and cerebellum (see Figure 15.22). Global decreases in functional connectivity across the salience network were accompanied by local increases in the amplitude of low frequency [synchronized] activity at nodes of the salience network, including insula, amygdala, hippocampus, and striatum. In the survivor group, this increased regional activity correlated with increased distress (more symptoms of anxiety and depression; Lui et al., 2009). This research provides convergent behavioral and neuroimaging evidence for stress-related change in the salience network in nonclinical adults. (We note that this convergence of findings helps to support the study’s conclusions in the face of recent statistical concerns about group comparisons in resting state studies; e.g., Tononi & Cirelli, 2003.) Functional connectivity studies of nonclinical adults who have experienced early life stress/trauma are also rare. One recent effort, however, used a graph analysis approach to examine the impact of early life stress on an emotion processing/emotion regulation network, which included

bilateral insula, amygdala, hippocampus, and striatum, and PFC (Cisler et al., 2012). This was a rare three-group study that included women with a history of early life stress and a history of psychopathology (depression-susceptible individuals), women with early life stress and no psychopathology (resilient individuals), and a control group without histories of early life stress or psychopathology. Overall, this study found that more severe early life stress was associated with fewer connections between nodes, and fewer hubs. Interestingly, one of the network traits that best differentiated the three groups was the hub-like nature (the betweenness centrality) of left amygdala connectivity. Women in the nonclinical/resilient group had amygdalae that were less hub-like than controls, whereas women in the susceptible group had amygdalae that were more hub-like than controls. Increased centrality of the amygdala in information processing may lead to more affectively biased processing in susceptible women with early life stress (Cisler et al., 2013) and a greater vulnerability to affective disorders over time. If so, then a question for future research concerns the social, contextual, or biological factors that may moderate the hub-like qualities of the amygdala in children and adults who have experienced early life stress. This, in turn, may pave the way for more effective interventions. Reverberating Effects The survival/salience network is the neural front-line responder in initiating and supporting adaptive behavioral response to environmental challenge and (as argued here) the central mediator of allostasis (Ganzel et al., 2010). Earlier, we reviewed some of the early evidence that suggests this network is selectively vulnerable to stress, and that this influences its ability to respond to future stressors (allostatic load). We hypothesize that this stress-related damage or plasticity in the survival/salience network will drive disruptions in the other two canonical networks over time. For example, we have noted that the amygdala is extensively connected with other brain regions, serving as a connector hub that links other hubs (Pessoa, 2008). This provides ample means for the salience network, via the amygdala, to influence nodes in other networks. For example, the amygdala has direct ascending connections to the PCC and the mPFC, as previously discussed. These are the primary nodes of the default network, so that stress-related amygdala hyperactivity has a direct means of influencing default network function. Default activity is anticorrelated with activity in the dorsal attention network (which makes up half of the CEN), suggesting that stress-related disruption in one may well propagate to the

Triple Network Allostasis and Psychopathology

other via the amygdala. Most tellingly, the anterior insula and the dACC are included in the frontoparietal control network (which makes up the other half of the CEN), so that stress-related change in the salience network equals stress-related change in these key nodes of the CEN. Thus, we argue that stressor exposure can impact neural network function on a global scale, in a manner that (following our models) can be expected to be deeply influenced by individual diathesis. Central allostasis drives peripheral allostasis (Ganzel et al., 2010). Stress-related disruptions in central allostasis can therefore be expected to drive disruptions in the stress response in physiological systems peripheral to the CNS. It is difficult to overstate the scope of these responses. Stress-related activation of both the HPA axis and the sympathetic nervous system produces glucocorticoids (cortisol, in humans) and catecholamines (e.g., epinepherine and norepinepherine; e.g., Feldman, Conforti, Itzik, & Weidenfeld, 1995; Pacák & Palkovits, 2001; Schulkin, 2003). These have profound preparative and modulating effects throughout the body, creating extensive short- and long-term stress-related alterations in peripheral physiology (e.g., McEwen, 2004, 2007; Sapolsky et al., 2000). These, in turn, feed back on brain function and structure via multiple pathways, including interoceptive input to the salience network (e.g., Critchley et al., 2004; Farb, Segal, & Anderson, 2013) and the effects of cortisol on hippocampal function and structure (McEwen, 2004, 2007; Pruessner et al., 2008). Individual differences will modulate this feedback, providing yet another avenue through which diathesis may influence the long-term impact of stress on brain and behavior. Triple Network Allostasis in Context The individual is embedded in a multilevel dynamical system that includes a current context, which influences the relationship between the current stressor and sensory processing, allostasis, and, ultimately, mental health (Figure 15.21). Consistent with our interest in specifying the multiple ecological levels of influence on mental health outcomes, we draw from psychosocial stress research to conceptualize the special conditions of contextual risk in this model. We also draw from dynamic systems theories in modeling the context of stress as an important source of input into physiological allostasis. Dynamic systems theory helps us understand the structure of context, how context interacts with other ecological and individual levels, and how contextual effects may play out across time. First, as shown in Figure 15.21, we conceptualize

605

contextual risk as including both proximal and distal risk factors. While not shown on the figure for the purpose of efficiency, we do consider them as nested and interacting influences on the stress process. The direct arrow from current context to environmental stressor reflects the recognition that low-resourced environments are likely to result in higher levels of stressor exposure (Brooks-Gunn, Duncan, & Aber, 1997; Earls & Buka, 2000; Shonkoff & Phillips, 2000). In some cases, there is feedback (as indicated by the reverse arrow) from stressor to the current social or physical context. This accounts for the selection of individuals into environments, as well as the constraint on opportunities that stressors create in the choice of such contexts. Given the interactive nature of this system, we posit that prior and current risk will moderate, or shape, the effects of environmental stressors on sensory processing, as well as on the canonical large-scale neural networks and downstream peripheral physiology. In other words, the meaning of the stressor to the individual interacts with information about the physical and social context in which the individual is embedded (e.g., Hobfoll, 2001) to drive perception of the emotional intensity of the stressor and the resulting behavioral response. In short, the individual (mind, brain, peripheral physiology, and behavior) is highly influenced by extraindividual context. Moderators of Triple Network Allostasis Triple network allostasis also suggests approaches for predicting important moderators of the stress process. If the brain is the central mediator of allostasis, the biological capacities and organization of the brain can help to determine the properties of a stressor that are likely to influence mental and physical health outcomes in the allostatic model. For example, a stressor must have sufficient magnitude to activate the emotional circuitry of the brain or the stress response will not be invoked by the organism; conversely, stressors that are of a magnitude sufficient to overwhelm the mechanisms of allostatic accommodation will produce greater allostatic load. A stressor’s duration or chronicity is also likely to modify the allostatic response; a stressor of sufficient persistence to exhaust (or render toxic) the biological processes of accommodation can produce profound effects on allostatic load, as previously discussed. We have also noted the expanding evidence that stressors of different types may be processed differently within the emotional circuitry of the brain, as, for example, with the neural processing of social pain versus physical threat (for reviews, see Alleva & Santucci, 2001; Dedovic, Duchesne, Andrews, Engert, & Pruessner, 2009; Depue

606

Life Span Diathesis–Stress Model of Psychopathology

& Morrone-Strupinsky, 2005; Eisenberger & Lieberman, 2004; Lieberman, 2007). Likewise, there is evidence that the developmental timing of stressor exposure can moderate central allostasis (e.g., Liu et al., 1997; Susser & Stein, 1994). We have reviewed evidence here that significant amounts of central and peripheral allostatic load may be acquired in adulthood; thus, when long-term biological effects of stressor exposure are observed in children, it remains to be determined what is different about the mechanisms of allostasis and allostatic load in early human development and how the long-term health consequences of allostatic load acquired in childhood may be distinct from the consequences of load acquired in adulthood. This special role of developmental timing in allostasis may derive from the development of knowledge of emotion (Tottenham, Leon, & Casey, 2006), from differences in maturation time across the emotional circuitry of the brain (Casey, Getz, & Galvan, 2008), from a biological stress-sensitive period (analogous to the well-established stress sensitivity in the first ten days of a rat pup’s life; e.g., Levine, 2001; Liu et al., 1997), or a combination of these. It remains for future research to distinguish the impact of all of these potential moderators of triple network allostasis. Distinction between the differing impact of developmental timing and stressor type is of particular importance for studies of the consequences of early stressor exposure. The most frequently studied stressors of childhood involve separation from (or conflict with) the primary attachment figure (e.g., parental loss, neglect, rejection, or conflict). These are stressors that fit the description of social pain. As previously discussed, there is evidence for qualitative physiological differences in the neural processing of physical threat relative to that of social pain, regardless of age at time of stressor exposure (e.g., Dedovic et al., 2009; Eisenberger & Lieberman, 2004). If the processing of social pain and physical threat are biologically distinguishable within the core emotional systems of the brain, then the allostatic model would predict that central and peripheral allostatic load accrues in distinguishable physiological systems for each type of stressor. This, in turn, should lead to different long-term health outcomes for each type of stressor (there is preliminary evidence for this, e.g., Post, Leverich, Xing, & Weiss, 2001). If stressor type indeed moderates the stress process in a way that is independent of developmental timing, it will be important to include stressor type as a separate variable in models of early stress. Furthermore, the effects of stressors of different types may change across the course of development, so that developmental timing and stressor type interact with one another (a three-way interaction).

Together, these considerations point to the need for attention to likely moderators of triple network allostasis in studies of stress, as well as highlighting opportunities for new research. These may include, for example, further research into differences over time in the physiological processes underlying central and peripheral allostasis and allostatic load in response to stressors of different intensity, chronicity, type, and developmental timing, which in turn would be expected to drive multifinality in outcomes associated with long-term health and behavior. Longitudinal studies of amygdala responsiveness over time in samples of different ages and with different types and levels of stressor exposure would shed light on these points. Such work would also clarify the time frames for the development of stress-related alteration in brain structure and function, and their possible reversal. This is precisely the kind of work that can be built from the model we present.

FUTURE DIRECTIONS Triple Network Allostasis as a Dynamic System When we undertook this theoretical work (Ganzel et al., 2010), we aimed to model a dynamic bioecological system that fully penetrates physiology (gets under the skin). In doing so, we followed Bronfenbrenner and Morris (1998) in requiring this model to include “the joint, interactive, mutually reinforcing effects” of person, proximal process, context, and time (Bronfenbrenner & Morris, 1998, p. 996). We also incorporated many of the principles of dynamic systems theories (Bronfenbrenner & Morris, 1998, 2006; Ford & Lerner, 1992; Sameroff & Chandler, 1975; Thelen & Smith, 1998; Waddington, 1957) that are integral to a multilevel approach to understanding behavior and development (Cacioppo & Berntson, 1992). For example, two principles of dynamic systems theories helped us to consider the relations between the levels of a multilevel system. The first is the principle of multiple nested and interacting levels, which states that the person is embedded in an ecological system that represents the influences of more distal systems through the most proximal system (e.g., a current stressor, as modified by contextual resources and risks). Individual behavior is moderated, or shaped, by the context in which that individual is embedded (interaction across systems). The second of these principles is the transactional relationships between levels of the system, which states that the relations between individuals and their environment, or between proximal and distal contexts, are transactional instead of unidirectional. That is, causality

Future Directions

(a) Salience network

(b) Schematic of the medial PFC brainstem axis

(c) Evidence for stressrelated “wear and tear”

(d) Neural reference space for discrete emotions

607

Figure 15.23 Triple network allostasis as an organizing concept. (a) Salience network. (b) Schematic illustration of the brain–body axis that runs along the medial PFC and into the brainstem. (c) Regional gray matter decreases in nonclinical adults exposed to the World Trade Center disaster, 3 years after September 11, 2001. (d) The core affective systems of the brain. See footnote 1. Source: Adapted from W. W. Seeley, V. Menon, A. F. Schatzberg, J. Keller, G. Glover, H. Kenna, . . . M. D. Greicius, Dissociable intrinsic connectivity networks for salience processing and executive control, Journal of Neuroscience, 27, 2349–2356, 2007; (b) R. D. Lane & T. D. Wager, The new field of brain–body medicine: What have we learned and where are we headed? NeuroImage, 47(3), 1135–1140, 2009; (c) B. L. Ganzel, P. Kim, G. Glover, & E. Temple, Resilience after 9/11: Multimodal neuroimaging evidence for stress-related change in the healthy adult brain, NeuroImage, 40(2), 788–795, 2008; (d) K. A. Lindquist, T. D. Wager, H. Kober, E. Bliss-Moreau, & L. Barrett, The brain basis of emotion: A meta-analytic review, Behavioral and Brain Sciences, 35(3), 121–143, 2012.

works in both directions (e.g., as in modulated allostasis; Figures 15.23 and 15.24). Three more principles of dynamic systems helped us to consider how life span development plays out across time. The first, multifinality and equifinality, indicates that there are many paths to a single outcome and a single pathway can lead to a diverse set of outcomes. The second, attractor states constrain possible pathways, suggests that development proceeds on a small number of trajectories and is subject to a small number of possible influences, with biological and ecological constraints that limit possible outcomes. The third, continuous and discontinuous change, posits that change can occur through continuous

Figure 15.24 Primary location of von Economo neurons in the human brain. See footnote 1. Source: Adapted from P. C. Williamson & J. M. Allman, The human illnesses: Neuropsychiatric disorders and the nature of the human brain. New York, NY: Oxford University Press, 2011. Color plate 3.1 (page 116)]

developmental processes as well as through large changes (e.g., threshold effects). These concepts have provided structure for our application of life span developmental neuroscience to the allostatic models we have presented here. These concepts may also be used in future research to aid in hypothesis generation about the form and nature of model outcomes. Integration of life span development and triple network allostasis into a diathesis–stress model suggests several avenues for future research, along with a number of predictions. Is Triple Network Allostasis Universal? Within a developmental psychopathology perspective, the study of normality and pathology are understood to be mutually informative. Nonetheless, study of the long-term effects of environmental and social stressors on the nonclinical human brain is still a newly emerging

608

Life Span Diathesis–Stress Model of Psychopathology

topic of investigation in human neuroscience. The nonclinical human brain is of critical interest to these questions because this is where the basic mechanisms of allostasis will be the least confounded with disease processes. If allostatic load is to be a general mechanism underlying the “long-term carry forward of stress and adversity” (Rutter, 1996, p. 37), it must also have significant consequences for nonclinical adults and children, i.e., if allostasis is to be a useful replacement of classical homeostasis, it must apply to everyone (Ganzel & Morris, 2011; Ganzel et al., 2010). In preceding sections, we have reviewed evidence for the accrual of central allostatic load from stressors that occur in nonclinical humans both earlier and later in development. This is a pivotal concept in defining our model of the stress process because it suggests that allostasis and costs of allostatic load are not confined to a special clinical population or a particular developmental sensitive period. Nevertheless, studies linking neural networks and long-term effects of prior stress have typically focused on clinical populations or early childhood. In addition, few studies link life stress and large-scale neural network integrity/function with measures of socioemotional adaptation/mental health in nonclinical humans. Still fewer studies blend these factors with a developmental component (although see Thomason et al., 2013; Thomason et al., 2011) and none, to our knowledge, include a broader life span perspective. However, our life span model of modulated allostasis (Figure 15.19) combined with the novel triple network approach to allostasis presented here (Figure 15.21), leads us to predict that the salience/survival network will be the first brain areas to show evidence of long-term plasticity or wear and tear as a consequence of accumulating stressor exposure, and that this will be the case for all populations. Via the salience network, individual differences in central allostatic load can be expected to robustly influence the other canonical neural networks in an age-dependent manner across the life span and in all populations, regardless of clinical status. Happily, these predictions are testable and results will lead to important refinements of the model. We also argued here for inclusion of central and peripheral allostatic load in an individual’s diathesis, thus rendering diathesis to be a dynamic force arising out of the confluence of an individual’s initial genetic loading and their ongoing life experience. This allowed us to propose an allostatic diathesis–stress model that has triple network allostasis at its core. In this model, past and current stressor exposure interacts with a constantly changing diathesis to predict psychopathology onset, with robust implications for future mental health. As previously, we predict that

this model is applicable regardless of an individual’s age or current clinical status, and earnestly hope that this work prompts expansion of the research on stress, neural networks, and mental health beyond the circumscribed study of clinical populations or childhood stressors. Can Triple Network Allostasis Serve as an Organizing Concept? Next, we call attention to the considerable overlap between the salience network (Figure 15.22a; Seeley et al., 2007) and the brain areas that have key brain-body regulatory functions (Figure 15.22b; Lane & Wager, 2009). Identification of remarkable commonalities in the brain mechanisms that regulate the human autonomic, endocrine, and immune systems has been an exciting step forward in brain–body medicine (e.g., Gianaros & Sheu, 2009; Ohira et al., 2010; Urry, van Reekum, Johnstone, & Davidson, 2009; Wager et al., 2009). The neural loci of these brain–body interfaces fall along an axis that starts at the superior medial PFC and descends along the midline of the PFC and down into the brainstem. Thus, these neural brain–body regulatory mechanisms share significant common ground with the survival/salience network. Moreover, both the salience network and the brain–body “medial PFC-brainstem axis” (Lane & Wager, 2009, p. 1136) overlap with brain areas showing evidence of long- and short-term stress-related change in humans (Figure 22c; e.g., Ganzel et al., 2008; Gianaros, Horenstein, et al., 2007; Gianaros, Jennings, et al., 2007; Liston et al., 2009) and animals (Cerqueira et al., 2005; Vyas et al., 2002). While this three-way overlap is highly consistent with our theoretical models, there is relatively little research that examines the linkage between central and peripheral allostasis in the context of large-scale neural networks. We commend this topic to future research. In the context of a constantly changing brain and our life span model of modulated allostasis (Figure 15.19), we argue that all allostatic functions will be influenced by life span development, i.e., they would be expected to vary with age. In triple network allostasis, central allostasis (as initiated and directed by the survival/salience network) is hypothesized to drive short-term stress-related alterations in autonomic, endocrine, and immune function, and peripheral/visceral physiology (peripheral allostasis), which are modified by prior accumulations of peripheral allostatic load (Ganzel et al., 2010). This somatic/visceral response feeds back via the insula to influence salience network function and the experience of emotion (Damasio & Carvalho, 2013) or affect (Figure 15.23d; Lindquist,

Future Directions

Wager, Kober, Bliss-Moreau, & Barrett, 2012). In the context of a constantly and systematically changing brain, age- and stress-related change in central allostasis can be expected to reverberate through this feedback loop to generate age- and stress-related change in the peripheral stress response (and in the quality and quantity of peripheral allostatic load), with subsequent feedback effects on the salience/survival network itself. Thus, this model predicts that there will be age-related change in the experience of salience, and indeed of affect. As a corollary to this, we would predict that life stress interacts with age to impact the experiences of salience and affect, and the activities of emotional cognition (e.g., emotion regulation, emotional decision making; Burri et al., 2013; Marsh, Nagengast, & Morin, 2012; Tsolaki et al., 2010). Ongoing work in psychology, psychiatry, and human neuroscience tie together many of these themes in pairs (stress and emotional cognition; stress and development; development and functional neural networks; neural networks and psychopathology, neural networks and affect). The models presented here suggest novel combinations of these factors for future investigation, as well as the additional need for deployment of these factors as statistical controls. Is Triple Network Allostasis Only Human? We have argued that the triple network model of allostasis is common to all humans, regardless of age or clinical status. We have also suggested ways in which this model may serve as an organizing concept to unite theories of stress and diathesis–stress with brain–body medicine and research on emotion. In this section, we present the suggestion that this organizing concept may apply uniquely to humans and perhaps some of the higher primates and other big-brained species. This is not for grandiosity, but rather to outline possibilities—including possible limitations—for future research. The pivotal role of the salience network in our allostatic triple network model raises some broad questions about the nature and origins of psychopathology. We note that this point has implications for how appropriate traditional animal models are for studying some major mental disorders. Von Economo Neurons As discussed previously, the right anterior insula/dorsal ACC subnetwork is a hub of information traffic within the salience network and it is argued to play a unique and pivotal role in switching between all three canonical neural networks (Kempermann et al., 1998). It is therefore notable

609

that the right anterior insula and ACC are the two primary locations of a unique type of neuron—the von Economo neuron or spindle neuron (Maroun & Richter-Levin, 2003; Vyas et al., 2004; Williamson & Allman, 2011). These large projection neurons have one apical axon and one primary dendrite, giving them a spindle shape (Olson et al., 2011). The von Economo neurons appear to be exclusive to the brains of humans, other great apes, elephants, and cetaceans, leading to the hypothesis that they are an adaptation necessary for the rapid flow of communication within large brains (e.g., Cui et al., 2008; Evrard, Forro, & Logothetis, 2012; Stimpson et al., 2011). Concentrations of von Economo neurons are lateralized in the human brain, with populations being denser on the right side of the brain than the left. These neurons are also highly localized, occurring almost exclusively in layer five of anterior insula and anterior cingulate, with much smaller concentrations recently observed in human dorsolateral PFC (Hill et al., 2012; Vyas et al., 2006; see Figure 15.24). In keeping with their large and streamlined structure, von Economo neurons provide rapid relay of simple signals from anterior insula and dACC to target regions that, in animal models, include PFC, orbitofrontal cortex, temporal cortex, and subcortical survival systems (Hill et al., 2012). A recent resting state functional connectivity study has observed that von Economo-containing regions in the human brain activate in synchrony with individual nodes in the salience network, the dorsal attention network, the sensorimotor network, and portions of the default mode network (Hill, McLaughlin, et al., 2010) suggesting that von Economo neurons may project to nodes in each of these networks (also see Fischer et al., 2005). Thus, the network switching capabilities of the right anterior insula/dorsal ACC subnetwork may be, at least in part, due to the unique properties of von Economo neurons (Sridharan et al., 2008). Von Economo Neurons Across the Life Span In humans, von Economo neurons begin to appear by the thirty-fifth week of gestation; at birth, the density of these neurons is approximately 15% of the eventual total (Hill et al., 2012). The full complement and typical rightward anatomical distribution of this neuronal population is reached by the age of four years (Cui et al., 2008; Tsolaki et al., 2010). It is argued that this developmental timing renders von Economo neurons vulnerable to disruption during the third trimester of pregnancy, and in the perinatal and early postnatal period (Hill et al., 2012). Moreover, the immunocytochemistry of these neurons suggests that they play a role in social bonding (expression

610

Life Span Diathesis–Stress Model of Psychopathology

of the vasopressin 1a receptor), reward under uncertainty (dopamine D3 receptor expression), monitoring of visceral response and possible amplification of threat-related signals (unique serotonin 2b receptor expression within the human CNS), and cognition (robust expression of DISC1, the primary product of the disrupted in schizophrenia [DISC] gene; Allman et al., 2011; Allman et al., 2005). Due to their unique cytochemistry and distinctive phylogeny, von Economo neurons have been proposed to play a role in human capabilities, such as rapid intuitive decision making in complex sociocultural contexts (Chick & Reyna, 2012; Cui et al., 2008; Williamson & Allman, 2011). Abnormal development of von Economo neurons is therefore hypothesized to underlie multiple neurodevelopmental disorders in which aberrant social conduct or emotion regulation are diagnostic features, including autism (e.g., Campolongo et al., 2009; Freund et al., 2003), and schizophrenia (e.g., Cui et al., 2008; Tsolaki et al., 2010). Less is known about age-related change past midlife in von Economo neurons, except in the context of Alzheimer’s disease and behavioral variant frontotemporal dementia (bvFTD; Tsolaki et al., 2010). In Alzheimer’s disease, there appears to be a relative sparing of von Economo neurons, whereas bvFTD is characterized by a profound and specific early loss of these neurons (66 to 75% of the expected total; Cui et al., 2008; Fischer et al., 2005; Gross et al., 1997; Tessitore et al., 2005). The degree of selective degeneration of von Economo neurons in bvFTD is correlated with decreases in behavioral and functional indicators of empathy, self-control, and social awareness (e.g., Gross et al., 1997). This, in turn, is hypothesized to result from a catastrophic failure to integrate visceral-autonomic, social-affective, and cognitive cues as the anterior insula/dorsal ACC network hub degenerates (Gross et al., 1997; Tsolaki et al., 2010). Allostatic Load and von Economo Neurons This suggests questions for future research. As we reviewed previously, recent trauma exposure appears to selectively reduce gray matter volume in the anterior insula and ACC (Blair et al., 2013; Ganzel et al., 2008; Monohan et al., in preparation) suggesting that von Economo neurons may be uniquely vulnerable to life stress.17 However, when traumas are farther in the past, the neural functional and structural correlates of trauma exposure vary more widely (Andersen et al., 2008; Blair et al., 2013; Monohan et al., in preparation). These data are consistent with the 17 This does not necessarily mean that they are a target of glucocorticoids, although they may be.

hypothesis that von Economo neurons are at or near the initial locus of impact of stress and trauma in the human brain, and that stress-related disruption subsequently propagates through the salience network, the default and central executive network, and the systems of peripheral allostasis. Developmental stage and individual differences in diathetic vulnerability would determine the direction and extent of the propagation of stress-related change in neural networks or nodes over time. Under this hypothesis, von Economo neurons would play a pivotal role in triple network allostasis and the onset of psychopathology. If so, further research into the stress-vulnerability of von Economo neurons may shed new light on the mechanisms of allostasis and allostatic load, and the origins of psychopathology. Investigation of possible age- and sex-specific effects in stress vulnerability of these neurons would then also be fruitful in this regard.

CONCLUSION Diathesis–stress models of the origins of psychopathology have been a major organizing framework in mental health research and practice for the past half-century (e.g., Meehl, 1962). At their inception, these models were open to criticism on the grounds that they lacked specificity (McLaren, 1998) and considered stress and diathesis as a distinct (even if interacting) processes. However, such models have benefitted greatly from the development of new scientific tools that allow specification and measurement of processes that span levels of analysis from genes and brains, to cognition and behavior, and to broader contextual influences. Simultaneously, this work has benefited from disparate lines of inquiry into the origins of psychopathology and the sequelae of stressful experiences that spans the fields of biology, medicine, social neuroscience, and psychology. Together, these allow us to build a developmentally informed diathesis–stress model of psychopathology onset that reflects the individual’s past history of adversity and that strongly draws on the core tenets of a developmental psychopathology perspective (e.g., Cicchetti & Toth, 2009). We stand on the shoulders of three major contributions to the study of stress and psychopathology in the development of our new model. First, the last two decades has seen a paradigm shift in thinking about how humans respond physiologically to stress. As a major revision of the classical theory of homeostasis, allostasis identifies the brain as the executor and central mediator of ongoing system-wide adjustment to meet environmental challenge (Sterling & Eyer, 1988).

Conclusion

Two points are critical from this work: (1) the brain’s survival/salience network becomes the primary, physiological targets of stress, bringing stress well into the brain (and not merely as an external force influencing cognition or behavior); and (2) stress has long-term physiological costs in the form of allostatic load that affect later responses to stress, allowing stress to become a diathesis of sorts for the next stressful experience. Second, we build on new research on the constantly changing brain, citing evidence for brain development and senescence across the life span. The central premise of our thesis is as follows: if the brain is the central mediator of a stress/diathesis-psychopathology relationship, and the brain is not static but represents a constantly changing neural landscape across the life span, then the stress/diathesis-psychopathology process itself must vary across development as well. That is, variations in stressor timing across the life span must have very different implications for brain processes if the neural circuitry of the salience network that is serving as the front lines of stress is changing as well (said another way, stressors at different ages may have different neural sequelae). And if the central mediators of allostasis are changing, then the accretion of allostatic load must be changing as well (i.e., stressors at different ages would have different effects on the development of this form of diathesis). With differential physiological consequences both for the sequelae of stress and the development of diathesis to future stress, it is not hard to see why psychopathological outcomes may also differ across the life course. This thinking has implications for a number of findings, including why the effects of stress and trauma might vary for children and adults; why a critical period approach to stress during early life might be insufficient; and why there are effects of trauma exposure in adulthood even in nonclinical/low risk individuals, to name a few. Making this argument allows us to build a theory of modulated allostasis, in which both the stress process and the diathesis (in part due to load from prior periods) contribute to different psychopathological outcomes, across developmental time from the prenatal period to older adulthood. In this way, diathesis–stress becomes a developmental model. Third, Menon’s (2011) triple network model of psychopathology is enormously useful addition to our diathesis–stress model. With this addition, deficits in the functional connectivity of neural networks (rather than specific brain regions alone) are key predictors of psychopathology. For example, while the right anterior insula in the salience network is key in the emergence of psychopathology, Menon adds the reasons why—for

611

example, that the insula is critical in the recruitment of the executive network and the default network in generating adaptive behavioral responses. As a result, downstream effects on other networks may be even larger than those in the initial deficits in the salience network. In prior work, we argued that the core affective systems of the brain were central to the translation of stress into physiological and then behavioral response. Menon allows us to improve the neural specificity of this model—that it is synchronized nodes in the affective core system (i.e., the survival/salience network)—that produces allostatic responses in the context of stress, and results in the deficits in processing of information that results in deficits in mental health. The resulting model is termed triple network allostasis and represents our new proposed diathesis–stress model of psychopathology. Having built a new diathesis–stress model of psychopathology onset, the question now arises: can it be used to better design intervention approaches aimed at reducing psychopathology? Indeed, a key principle of a developmental psychopathology perspective is its commitment to translation (Cichetti & Toth, 2009), and a central benefit of identifying the neural underpinnings of the stress–psychopathology relation is to support the development of more effective interventions for individuals at risk (Ganzel & Morris, 2011). However, research in this area is only in its infancy. The question is whether population-scalable, cost-effective, social and behavioral interventions can be developed that target the large-scale neural networks as key mediators of the diathesis/stress-psychopathology relationship. While research on brain architecture has strongly informed pharmacological treatment strategies, it has less often been utilized to inform social and behavioral interventions. The hope here is that if we were to design and effectively implement promising social and behavioral interventions targeted at relevant brain processes, they would drive longer lasting changes in psychopathological outcomes. But the challenge is that we do not yet have a body of research to draw on about the physiological effects of community-based interventions. A number of research teams are forerunners in this research in their inclusion of physiological assessments in intervention efforts (e.g., Cicchetti, Rogosch, Toth, & Sturge-Apple, 2011; Dozier, Peloso, Lewis, Laurenceau, & Levine, 2008; Fisher, Stoolmiller, Gunnar, & Burraston, 2007). However, we are a far cry from being certain about which interventions are most likely to mitigate the risks of psychopathology because of their effects on neural networks. That is the next frontier for this area of research—translating

612

Life Span Diathesis–Stress Model of Psychopathology

neuroscience from the lab and the clinic into the community, thus fully realizing the translational commitment of the developmental psychopathology perspective. One promising, but still nascent, area of intervention includes mindfulness-based interventions. Mindfulness interventions are aimed at altering the experience of the stress process for the individual by addressing the individual’s response to stressors, which should, in turn, result in alterations in the underlying stress–response system. Mindfulness-based interventions have been most often implemented with adults, and have shown some successes with moderate effect sizes (although more methodologically rigorous work is warranted; Baer, 2003; Grossman, Niemann, Schmidt, & Walach, 2004). Another approach to these interventions aims at improving the regulatory capacity of youth exposed to stress and risk. These use mindfulness-based training to focus attention and awareness and reducing rumination that can contribute to the physiological and psychological costs of stress as we have described here. The first work using these approaches for children and adolescents has demonstrated that these interventions are feasible to implement and are well received, but efficacy evidence is only now emerging (Burke, 2010; Mendelson et al., 2010). Further work in this area is currently under way (Greenberg & Harris, 2012) and is needed to determine if this approach can effectively build resilience in children exposed to stress and trauma by reducing the physiological and psychological costs of stress. In contrast to the relatively new work on mindfulnessbased strategies, social-emotional learning programs in schools have a much longer history and these sometimes also target the self-regulatory capacity of children and youth at risk. As such, these programs may have implications for building the social-emotional skills that underlie successful coping in the context of stress. As we have discussed, the brain’s survival/salience network underlies the processing of salient (survival-related) information and serves as the central mediator of allostatic accommodation to environmental demand (Menon, 2011). We have argued that these systems of the brain underlie the translation of a stressor into a behavioral or physiological response, and are thus key to the onset of mental disorder. Then it stands to reason that interventions that target these same systems might be important in the prevention of such disorders. An extensive meta-analysis of over 200 social-emotional learning programs studied in the context of experimental and quasi-experimental studies shows moderate effect sizes on outcomes like emotional distress and conduct problems that may be the precursors to psychopathology (Durlak, Weissberg, Dymnicki, Taylor, & Schellinger, 2011). But

because these studies rarely if ever include physiological assessments, it is not at all clear if such interventions support the theory we have proposed here, or whether they effectively support children’s development through other neural mechanisms. While not typically considered as interventions aimed at psychopathology per se, a number of policy experiments have been conducted that are intended to directly reduce stressors in the external environment by targeting the contextual precipitators of that stress process. Some of the most well-studied of these are employment-based antipoverty programs, that directly target family income (Morris, Duncan, Huston, Crosby, & Bos, 2001; Morris, Gennetian, & Duncan, 2005). These programs provide low-income families with resources either from within or outside the welfare system. They typically tie that increased income to work by providing monthly cash or in-kind supplements to earnings, or by allowing welfare-recipient parents to keep more of their welfare benefits when they transition from welfare into employment. A relatively large and promising research base shows that these programs (known as Conditional Cash Transfers, or CCTs; Fiszbein, Schady, & Ferreira, 2009) can effectively raise income of families in ways that should reduce stressor exposure (Bloom & Michalopoulos, 2001; Morris et al., 2001). These condition cash assistance on parents’ participation in some program or institution (historically, school or health visits). Indeed, findings demonstrate that CCTs might be especially good at encouraging participation in services and institutions that might benefit high-risk individuals, even while the increased resources going to families might reduce that risk exposure in the first place. In this way, CCTs could be effectively paired with supply side services to boost the emotional skills of high-risk children (e.g., mindful meditation or social-emotional learning programs). At the same time, they support key demand side adjustments (parents’ and children’s take-up of those services, which can be notoriously low for voluntary programs). These programs have the added benefit of simultaneously reducing family stress through a key contextual target—increasing income. Future work should certainly explore these opportunities to reduce the risk for psychopathology, although rarely do large-scale randomized trials include the careful measurement of physiological processes that would enable us to move forward in this line of work. As we have discussed elsewhere (Ganzel & Morris, 2011), we strongly support the inclusion of physiological measurement in community intervention studies that hold promise for reducing psychopathology. Measuring change

References

in neural processes and in concomitant mental health outcomes in randomized control trials can dramatically advance our understanding of the sensitivity of systems to induced environmental change (Cicchetti & Gunnar, 2008), as well as providing us with information to refine and expand our theoretical models. And by offering such interventions to children at differing periods of development, such research can also tell us about points in development in which the system is plastic and open to change from intervention. Notably, however, caution is in order: interventions may be limited in their contribution to the understanding of risk and resilience if the neural processes deriving from positive events are not merely the mirror image of those precipitated by negative events. Indeed, it is even possible that the windows of sensitivity to stress of physiological processes and mental health outcomes do not always coincide with windows of sensitivity to the positive aspects of intervention. In effect, it is not clear whether developmental periods are plastic or merely vulnerable. Or if the neural effects of positive contextual change (intervention) are the same as the effects of negative contextual change (stress). If they are not, then learning about periods of plasticity to intervention may not tell us all we need to know about periods of sensitivity to stress. They certainly will tell us about one side of plasticity (the sensitivity of neural systems to positive inputs) but may leave us still wondering about the other. That said, such research can still provide sorely needed information about how to best reduce the risk of psychopathology. In this paper, we draw from research across the disciplines of biology, medicine, neuroscience, and psychology to present a new diathesis–stress framework for the development of psychopathology across the life span. The field is at a pivotal moment—new work is coming on line rapidly, dramatically increasing our understanding of neural processes in ways we could not have imagined a decade ago. It is imperative that we use this new information to develop a cross-disciplinary life span perspective on the emergence of psychopathology—one that highlights our deepening understanding of the neural processes underlying the stress/diathesis process. And, perhaps most importantly, it is critical that this information be used for the design of interventions that can most effectively mitigate the negative effects of psychopathology. In doing so, we will fully realize the commitment of a developmental psychopathology perspective (Cichetti & Toth, 2009) in leveraging the interdisciplinary-informed study of normality and psychopathology across the life span to more fully inform effective intervention at scale for those who are most at risk.

613

REFERENCES Adamec, R. E., Blundell, J., & Burton, P. (2005). Neural circuit changes mediating lasting brain and behavioral responses to predator stress. Neuroscience & Behavioral Reviews, 29, 1225–1241. doi: 10.1016/j .neubiorev.2005.05.007 Ade, P. A. R., Aghanim, N., Armitage-Caplan, C., Arnaud, M., Ashdown, M., Atrio-Barandela, F., . . . & Davies, R. D. (2014). Planck 2013 results. XVI. Cosmological parameters. Astronomy & Astrophysics, 571, A16. doi: 10.1051/0004-6361/201321591 Admon, R., Lubin, G., Stern, O., Rosenberg, K., Sela, L., Ben-Ami, H., & Hendler, T. (2009). Human vulnerability to stress depends on amygdala’s predisposition and hippocampal plasticity. Proceedings of the National Academy of Sciences of the United States of America, 106(33), 14120–14125. doi: 10.1073/pnas.0903183106 Albeck, D., McKittrick, C., Blanchard, D. C., Blanchard, R., Nikulina, J., McEwen, B. S., & Sakai, R. (1997). Chronic social stress alters levels of corticotropin-releasing factor and arginine vasopressin mRNA in rat brain. Journal of Neuroscience, 17, 4895–4903. Alexander, G. E., Chen, K., Merkley, T. L., Reiman, E. M., Caselli, R. J., Aschenbrenner, M., . . . Teipel, S. J. (2006). Regional network of magnetic resonance imaging gray matter volume in healthy aging. Neuroreport, 17(10), 951. doi: 10.1097/01.wnr.0000220135.16844.b6 Alleva, E., & Santucci, D. (2001). Psychosocial vs. “physical” stress situations in rodents and humans: Role of neurotrophins. Physiology & Behavior, 73, 313–320. doi: 10.1016/S0031–9384(01)00498-X Allman, J. M., Tetreault, N. A., Hakeem, A. Y., Manaye, K. F., Semendeferi, K., Erwin, J. M., . . . Hof, P. R. (2011). The Von Economo neurons in the frontoinsular and anterior cingulate cortex. Annals of the New York Academy of Sciences, 1225(1), 59–71. doi: 10.1111/j.1749 –6632.2011.06011.x Allman, J. M., Watson, K. K., Tetreault, N. A., & Hakeem, A. Y. (2005). Intuition and autism: A possible role for Von Economo neurons. Trends in Cognitive Sciences, 9(8), 367–373. doi: 10.1016/j.tics .2005.06.008 Alvarez-Buylla, A., & Kriegstein, A. (2013). Neural stem cells among glia. In J. Rubenstein & P. Rakic (Eds.), Patterning and cell types specification in the developing CNS and PNS (pp. 685–705). New York, NY: Academic Press. Anan, F., Masaki, T., Shimomura, T., Fujiki, M., Umeno, Y., Eshima, N., . . . Yoshimatsu, H. (2010). Abdominal visceral fat accumulation is associated with hippocampus volume in non-dementia patients with type 2 diabetes mellitus. NeuroImage, 49(1), 57–62. doi: 10.1016/j .neuroimage.2009.08.021 Andersen, S. L., & Teicher, M. H. (2008). Stress, sensitive periods and maturational events in adolescent depression. Trends in Neuroscience, 31, 183–191. doi: 10.1016/j.tins.2008.01.004 Andersen, S. L., Tomada, A., Vincow, E. S., Valente, E., Polcari, A., & Teicher, M. H. (2008). Preliminary evidence for sensitive periods in the effect of childhood sexual abuse on regional brain development. Journal of Neuropsychiatry and Clinical Neuroscience, 20, 292–301. Anderson, A. K. (2005). Affective influences on the attentional dynamics supporting awareness. Journal of Experimental Psychology: General, 134, 258–281. doi: 10.1037/0096–3445.134.2.258 Anderson, J. R., Fincham, J. M., Qin, Y., & Stocco, A. (2008). A central circuit of the mind. Trends in Cognitive Sciences, 12(4), 136–143. doi: 10.1016/j.tics.2008.01.006 Andrews-Hanna, J. R., Smallwood, J., & Spreng, R. N. (in press). The default network and self-generated thought: Component processes, dynamic control, and clinical relevance. Annals of the New York Academy of Sciences.

614

Life Span Diathesis–Stress Model of Psychopathology

Anisman, H., Merali, Z., & Hayley, S. (2008). Neurotransmitter, peptide and cytokine processes in relation to depressive disorder: Comorbidity between depression and neurodegenerative disorders. Progress in Neurobiology, 85(1), 1–74. doi: 10.1016/j.pneurobio.2008.01.004

Barrett, L., & Satpute, A. B. (2013). Large-scale brain networks in affective and social neuroscience: towards an integrative functional architecture of the brain. Current Opinion in Neurobiology, 23(3), 361–372. doi: 10.1016/j.conb.2012.12.012

Ansell, E. B., Rando, K., Tuit, K., Guarnaccia, J., & Sinha, R. (2012). Cumulative adversity and smaller gray matter volume in medial prefrontal, anterior cingulate, and insula regions. Biological Psychiatry, 72(1), 57–64. doi: 10.1016/j.biopsych.2011.11.022

Bartels, A., & Zeki, S. (2004). The neural correlates of maternal and romantic love. NeuroImage, 21, 1155–1166. doi: 10.1016/j.neuroimage .2003.11.003

Araque, A., Parpura, V., Sanzgiri, R. P., & Haydon, P. G. (1999). Tripartite synapses: Glia, the unacknowledged partner. Trends in Neurosciences, 22(5), 208–215. doi: 10.1016/S0166–2236(98)01349–6

Bavelier, D., Levi, D. M., Li, R. W., Dan, Y., & Hensch, T. K. (2010). Removing brakes on adult brain plasticity: from molecular to behavioral interventions. Journal of Neuroscience, 30(45), 14964–14971. doi: 10.1523/JNEUROSCI.4812–10.2010

Armony, J. L., & LeDoux, J. (1997). How the brain processes emotion. In R. Yehuda & A. McFarlane (Eds.), Annals of the New York Academy of Sciences: Psychobiology and posttraumatic stress disorder (Vol. 821, pp. 259–271). New York, NY: Academy of Sciences.

Bechara, A., Damasio, A. R., Damasio, H., & Anderson, S. W. (1994). Insensitivity to future consequences following damage to human prefrontal cortex. Cognition, 50(1), 7–15. doi: 10.1016/0010 –0277(94)90018–3

Armstrong, E., Schleicher, A., Omran, H., Curtis, M., & Zilles, K. (1995). The ontogeny of human gyrification. Cerebral Cortex, 5(1), 56–63. doi: 10.1093/cercor/5.1.56

Bechara, A., Damasio, H., Damasio, A. R., & Lee, G. P. (1999). Different contributions of the human amygdala and ventromedial prefrontal cortex to decision-making. Journal of Neuroscience, 19(13), 5473–5481.

Arnsten, A. F. T. (2009). Stress signalling pathways that impair prefrontal cortex structure and function. Nature Reviews Neuroscience, 10(6), 410–422. doi: 10.1038/nrn2648 Atsak, P., Hauer, D., Campolongo, P., Schelling, G., McGaugh, J. L., & Roozendaal, B. (2012). Glucocorticoids interact with the hippocampal endocannabinoid system in impairing retrieval of contextual fear memory. Proceedings of the National Academy of Sciences of the United States of America, 109(9), 3504–3509. doi: 10.1073/pnas.1200742109 Attwell, D., & Laughlin, S. B. (2001). An energy budget for signaling in the grey matter of the brain. Journal of Cerebral Blood Flow & Metabolism, 21, 1133–1145. doi: 10.1097/00004647–200110000–00001 Averbeck, B. B., & Seo, M. (2008). The statistical neuroanatomy of frontal networks in the macaque. PLoS Computational Biology, 4(4), e1000050. doi: 10.1371/journal.pcbi.1000050 Bachis, A., Cruz, M. I., Nosheny, R. L., & Mocchetti, I. (2008). Chronic unpredictable stress promotes neuronal apoptosis in the cerebral cortex. Neuroscience Letters, 442(2), 104–108. doi: 10.1016/j.neulet .2008.06.081 Baer, R. A. (2003). Mindfulness training as a clinical intervention: A conceptual and empirical review. Clinical Psychology: Science and Practice, 10(2), 125–143. doi: 10.1093/clipsy.bpg015 Bal, N., Figueras, G., Vilaro, M. T., Sunol, C., & Artigas, F. (1997). Antidepressant drugs inhibit a glial 5-hydroxytryptamine transporter in rat brain. European Journal of Neuroscience, 9(8), 1728. doi: 10.1111 /j.1460–9568.1997.tb01530.x Banasr, M., Valentine, G. W., Li, X., Gourley, S. L., Taylor, J. R., & Duman, R. S. (2007). Chronic unpredictable stress decreases cell proliferation in the cerebral cortex of the adult rat. Biological Psychiatry, 62(5), 496–504. doi: 10.1016/j.biopsych.2007.02.006 Barch, D. M., & Dowd, E. C. (2010). Goal representations and motivational drive in schizophrenia: The role of prefrontal–striatal interactions. Schizophrenia Bulletin, 36(5), 919–934. doi: 10.1093/schbul /sbq068 Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51(6), 1173. doi: 10.1037/0022–3514.51.6.1173 Barres, B. A. (2008). The mystery and magic of glia: A perspective on their roles in health and disease. Neuron, 60(3), 430–440. doi: 10.1016/j.neuron.2008.10.013 Barrett, L., Mesquita, B., Ochsner, K. N., & Gross, J. J. (2007). The experience of emotion. Annual Review of Psychology, 58, 373–403. doi: 10.1146/annurev.psych.58.110405.085709

Belsky, J. (1997). Variation in susceptibility in environmental influences: An evolutionary argument. Psychological Inquiry, 8(3), 182–186. doi: 10.1207/s15327965pli0803_3 Belsky, J., Bakermans-Kranenberg, M. J., & van IJzendoorn, M. H. (2007). For better and for worse: Differential susceptibility to environmental influence. Current Directions in Psychological Science 16, 300–304. Beltramo, R., D’Urso, G., Dal Maschio, M., Farisello, P., Bovetti, S., Clovis, Y., . . . Fellin, T. (2013). Layer-specific excitatory circuits differentially control recurrent network dynamics in the neocortex. Nature Neuroscience, 16(2), 227–234. doi: 10.1038/nn.3306 Ben-Ari, Y., Gaiarsa, J., Tyzio, R., & Khazipov, R. (2007). GABA: a pioneer transmitter that excites immature neurons and generates primitive oscillations. Physiological Reviews, 87(4), 1215–1284. doi: 10.1152/physrev.00017.2006 Bennett, M. R. (2011). The prefrontal–limbic network in depression: A core pathology of synapse regression. Progress in Neurobiology, 93(4), 457–467. doi: 10.1016/j.pneurobio.2011.01.001 Berkes, P., Orba’n, G., Lengyel, M., & Fiser, J. (2011). Spontaneous cortical activity reveals hallmarks of an optimal internal model of the environment. Science, 331, 83–87. doi: 10.1126/science.1195870 Berman, M. G., Peltier, S., Nee, D. E., Kross, E., Deldin, P. J., & Jonides, J. (2011). Depression, rumination and the default network. Social Cognitive and Affective Neuroscience, 6(5), 548–555. doi: 10.1093/scan/nsq080 Bernard, C. (1878). Leçons sur les phénomènes de la vie communs aux animaux et aux végétaux [Lectures on the phenomena of life common to animals and plants]. Paris, France: Ballieres. Bertoni-Freddari, C., Fattoretti, P., Delfino, A., Solazzi, M., Giorgetti, B., Ulrich, J., & Meirer-Ruge, W. (2002). Deafferentative synaptopathology in physiological aging and Alzheimer’s disease. Annals of the New York Academy of Sciences, 977(1), 322–326. doi: 10.1111/j.1749–6632.2002.tb04833.x Beumer, W., Gibney, S. M., Drexhage, R. C., Pont-Lezica, L., Doorduin, J., Klein, H. C., . . . Versnel, M. A. (2012). The immune theory of psychiatric diseases: A key role for activated microglia and circulating monocytes. Journal of Leukocyte Biology, 92(5), 959–975. doi: 10.1189/jlb.0212100 Binder, J. R., Frost, J. A., Hammeke, T. A., Bellgowan, P. S., Rao, S. M., & Cox, R. W. (1999). Conceptual processing during the conscious resting state: A functional MRI study. Journal of Cognitive Neuroscience, 11, 80–95. doi: 10.1162/089892999563265

References

615

Blair, K. S., Vythilingam, M., Crowe, S. L., McCaffrey, D. E., Ng, P., Wu, C. C., . . . Charney, D. S. (2013). Cognitive control of attention is differentially affected in trauma-exposed individuals with and without post-traumatic stress disorder. Psychological Medicine, 43, 85–95. doi: 10.1017/S0033291712000840

Brooks-Gunn, J., Duncan, G. J., & Aber, J. L. (1997). Neighborhood poverty: Context and consequences for children (Vol. 1). New York, NY: Russell Sage Foundation Publications.

Bloom, D., & Michalopoulos, C. (2001). How welfare and work policies affect employment and income: A synthesis of research. New York, NY: MDRC.

Brown, M. (1813). Sketch of the medical topography of the country that is watered by the upper streams of the Mohawk River (Letter 6). Annals of Medicine, Natural History, Agriculture and the Arts, 4, 170.

Bluhm, R. L, Miller, J., Lanius, R. A., Osuch, E. A., Boksman, K., Neufeld, R. W.J., . . . Williamson, P. (2007). Spontaneous lowfrequency fluctuations in the BOLD signal in schizophrenic patients: Anomalies in the default network. Schizophrenia Bulletin, 33(4), 1004–1012. doi: 10.1093/schbul/sbm052

Bruce, M. L. (1998). Divorce and psychopathology. In B. P. Dohrenwend (Ed.), Adversity, stress, and psychopathology (pp. 219–232). New York, NY: Oxford University Press.

Bonetti, C., & Surace, E. M. (2010). Mouse embryonic retina delivers information controlling cortical neurogenesis. PLOS ONE, 5(12), e15211. doi: 10.1371/journal.pone.0015211 Bora, E., Fornito, A., Yücel, M., & Pantelis, C. (2010). Voxelwise metaanalysis of gray matter abnormalities in bipolar disorder. Biological Psychiatry, 67(11), 1097–1105. doi: 10.1016/j.biopsych.2010 .01.020 Borsook, D., Edwards, R., Elman, I., Becerra, L., & Levine, J. (2013). Pain and analgesia: The value of salience circuits. Progress in Neurobiology, 104, 93–105. doi: 10.1016/j.pneurobio.2013.02.003 Bouret, S., & Sara, S. J. (2005). Network reset: A simplified overarching theory of locus coeruleus noradrenaline function. Trends in Neurosciences, 28(11), 574–582. doi: 10.1016/j.tins.2005.09.002 Boyce, W. T., & Ellis, B. J. (2005). Biological sensitivity to context: An evolutionary-developmental theory of the origins and functions of stress reactivity. Development and Psychopathology, 17, 271–301. doi: 10.1017/S0954579405050145

Brown, G. W., & Harris, T. O. (1978). Social origins of depression: A study of depressive disorders in women. New York, NY: Free Press.

Buchel, C., Morris, J., Dolan, R. J., & Friston, K. J. (1998). Brain systems mediating aversive conditioning: An event-related fMRI study. Neuron, 20, 947–957. doi: 10.1016/S0896–6273(00)80476–6 Buckner, R. L., Andrews-Hanna, J. R., & Schacter, D. L. (2008). The brain’s default network: Anatomy, function, and relevance to disease. Annals of the New York Academy of Science, 1124, 1–38. doi: 10.1196/annals.1440.011 Buckner, R. L., Krienen, F. M., & Yeo, B. T.T. (2013). Opportunities and limitations of intrinsic functional connectivity MRI. Nature Neuroscience, 16, 832–837. doi:10.1038/nn.3423 Buckwalter, J. A., Schumann, C. M., & Van Hoesen, G. W. (2008). Evidence for direct projections from the basal nucleus of the amygdala to retrosplenial cortex in the Macaque monkey. Experimental Brain Research, 186(1), 47–57. doi: 10.1007/s00221–007–1203-x Burchfield, S., Woods, S., & Elich, M. (1980). Pituitary adreno-cortical response to chronic intermittent stressors. Physiology & Behavior, 24, 297–302.

Brainard, M. S., & Knudsen, E. I. (1998). Sensitive periods for visual calibration of the auditory space map in the barn owl optic tectum. Journal of Neuroscience, 18, 3929–3942.

Burghy, C. A., Stodola, D. E., Ruttle, P. L., Molloy, E. K., Armstrong, J. M., Oler, J. A., . . . Essex, M. J. (2012). Developmental pathways to amygdala-prefrontal function and internalizing symptoms in adolescence. Nature Neuroscience, 15, 1736–1741. doi: 10.1038/nn .3257

Bredy, T. W., Grant, R., Champagne, D. L., & Meaney, M. J. (2003). Maternal care influences hippocampal neuron survival in the rat. European Journal of Neuroscience, 18(10), 2903–2909. doi: 10.1111/j.1460–9568.2003.02965.x

Burke, C. A. (2010). Mindfulness-based approaches with children and adolescents: A preliminary review of current research in an emergent field. Journal of Child and Family Studies, 19(2), 133–144. doi: 10.1007/s10826–009–9282-x

Bredy, T. W., Zhang, T. Y., Grant, R. J., Diorio, J., & Meaney, M. J. (2004). Peripubertal environmental enrichment reverses the effects of maternal care on hippocampal development and glutamate receptor subunit expression. European Journal of Neuroscience, 20, 1355–1362. doi: 10.1111/j.1460–9568.2004.03599.x

Burri, A., Maercker, A., Krammer, S., & Simmen-Janevska, K. (2013). Childhood trauma and PTSD symptoms increase the risk of cognitive impairment in a sample of former indentured child laborers in old age. PLOS ONE, 8(2), e57826. doi: 10.1371/journal.pone.0057826

Bremner, J. D., Randall, P., Scott, T., Bronen, R., Southwick, S., Seibyl, J., & Innis, R. (1995). MRI-based measurement of hippocampal volume in patients with posttraumatic stress disorder. American Journal of Psychiatry, 152, 973–981. Bremner, J. D., & Vermetten, E. (2001). Stress and development: Behavioral and biological consequences. Development and Psychopathology, 13, 473–490. Brody, G. H., Yu, T., Chen, Y., Kogan, S. M., Evans, G. W., Beach, S.R.H., . . . Gibbons, F. X. (2013). Cumulative socioeconomic status risk, allostatic load, and adjustment: A prospective latent profile analysis with contextual and genetic protective factors. Developmental Psychology, 49(5), 913–927. doi: 10.1037/a0028847 Bronfenbrenner, U., & Morris, P. A. (1998). The ecology of developmental processes. In R. M. Lerner (Ed.), Handbook of child psychology: Theoretical models of human development (5th ed., Vol. 1, pp. 993–1028). New York, NY: Wiley. Bronfenbrenner, U., & Morris, P. A. (2006). The Bioecological Model of Human Development. In R. M. Lerner (Ed.), Handbook of child psychology (Vol. 1). Hoboken, NJ: John Wiley & Sons.

Bush, G., Luu, P., & Posner, M. I. (2000). Cognitive and emotional influences in anterior cingulate cortex. Trends in Cognitive Science, 4(6), 215–222. doi: 10.1016/S1364–6613(00)01483–2 Butterworth, P., Cherbuin, N., Sachdev, P., & Anstey, K. J. (2012). The association between financial hardship and amygdala and hippocampal volumes: Results from the PATH through life project. Social Cognitive and Affective Neuroscience, 7(5), 548–556. doi: 10.1093/scan /nsr027 Cacioppo, J. T., & Berntson, G. G. (1992). Social psychological contributions to the decade of the brain: Doctrine of multilevel analysis. American Psychologist, 47(8), 1019. doi: 10.1037/0003–066X.47.8.1019 Cacioppo, J. T., Berntson, G. G., Sheridan, J. F., & McClintock, M. K. (2000). Multilevel integrative analyses of human behavior: Social neuroscience and the complementing nature of social and biological approaches. Psychological Bulletin, 126(6), 829. doi: 10.1037/0033 –2909.126.6.829 Cameron, H. A., Woolley, C. S., McEwen, B. S., & Gould, E. (1993). Differentiation of newly born neurons and glia in the dentate gyrus of the adult rat. Neuroscience, 56(2), 337–344. doi: 10.1016/0306 –4522(93)90335-D

616

Life Span Diathesis–Stress Model of Psychopathology

Campbell, I. G., & Feinberg, I. (2009). Longitudinal trajectories of non-rapid eye movement delta and theta EEG as indicators of adolescent brain maturation. Proceedings of the National Academy of Sciences of the United States of America, 106(13), 5177–5180. Campolongo, P., Roozendaal, B., Trezza, V., Hauer, D., Schelling, G., McGaugh, J. L., & Cuomo, V. (2009). Endocannabinoids in the rat basolateral amygdala enhance memory consolidation and enable glucocorticoid modulation of memory. Proceedings of the National Academy of Sciences of the United States of America, 106(12), 4888–4893. doi: 10.1073/pnas.0900835106 Cannon, W. B. (1932). The wisdom of the body. New York, NY: Norton. Cao, X., Li, L., Wang, Q., Wu, Q., Hu, H., Zhang, M., . . . Xiong, W. (2013). Astrocyte-derived ATP modulates depressive-like behaviors. Nature Medicine. doi: 10.1038/nm.3162 Carrion, V. G., Weems, C. F., Richert, K., Hoffman, B. C., & Reiss, A. L. (2010). Decreased prefrontal cortical volume associated with increased bedtime cortisol in traumatized youth. Biological Psychiatry, 68(5), 491–493. doi: 10.1016/j.biopsych.2010.05.010 Carstensen, L. L., Gross, J. J., & Fung, H. H. (1997). The social context of emotional experience. In K. Warner Schaie & M. Powell Laughton (Eds.), Annual review of gerontology and geriatrics (Vol. 17, pp. 325–352). New York, NY: Springer. Carter, A. R., Shulman, G. L., & Corbetta, M. (2012). Why use a connectivity-based approach to study stroke and recovery of function? NeuroImage, 62(4), 2271–2280. doi: 10.1016/j.neuroimage .2012.02.070 Casey, B. J., Getz, S., & Galvan, A. (2008). The adolescent brain. Developmental Review, 28(1), 62–77. doi: 10.1016/j.dr.2007.08.003 Casey, B. J., Jones, R. M., & Hare, T. A. (2008). The adolescent brain. Annals of the New York Academy of Sciences, 1124(1), 111–126. doi: 10.1196/annals.1440.010 Catalano, S. M., & Shatz, C. J. (1998). Activity-dependent cortical target selection by thalamic axons. Science, 281(5376), 559–562. doi: 10.1126/science.281.5376.559 Caviness, V. S., & Takahashi, T. (1995). Proliferative events in the cerebral ventricular zone. Brain Development 21, 289–295. doi: 10.1016/0387–7604(95)00029-B Cerqueira, J. J., Mailliet, F., Almeida, O. F., Jay, T. M., & Sousa, N. (2007). The prefrontal cortex as a key target of the maladaptive response to stress. Journal of Neuroscience, 27, 2781–2787. doi: 10.1523/JNEUROSCI.4372–06.2007 Cerqueira, J. J., Pego, J. M., Taipa, R., Bessa, J. M., Almeida, O. F.X., & Sousa, N. (2005). Morphological correlates of corticosteroid-induced changes in prefrontal cortex-dependent behaviors. Journal of Neuroscience, 25, 7792–7800. doi: 10.1523/JNEUROSCI.1598–05.2005 Champagne, F. A. (2010). Early adversity and developmental outcomes: Interaction between genetics, epigenetics, and social experiences across the lifespan. Perspectives on Psychological Sciences, 5, 564–574. Changeaux, J., & Danchin, A. (1976). Selective stabilization of developing synapses as a mechanism for specification of neural networks. Nature, 264, 705–712. doi: 10.1038/264705a0 Chapillion, P., Patin, V., Roy, V., Vincent, A., & Caston, J. (2002). Effects of pre- and postnatal stimulation on developmental, emotional, and cognitive aspects in rodents. A review. Developmental Psychobiology, 41, 373–387. doi: 10.1002/dev.10066 Chen, Z. J., He, Y., Rosa-Neto, P., Gong, G., & Evans, A. C. (2011). Age-related alterations in the modular organization of structural cortical network by using cortical thickness from MRI. NeuroImage, 56(1), 235–245. doi: 10.1016/j.neuroimage.2011.01.010 Chick, C. F., & Reyna, V. F. (2012). A fuzzy trace theory of adolescent risk taking: Beyond self-control and sensation seeking. In V. F. Reyna,

S. B. Chapman, M. R. Dougherty, & J. Confrey (Eds.) The adolescent brain. Learning, reasoning, and decision making (pp. 379–428). Washington, DC: American Psychological Association. Choi, J., & Brown, T. H. (2003). Central amygdala lesions block ultrasonic vocalization and freezing as conditional but not unconditional responses. Journal of Neuroscience, 23(25), 8713–8721. Chrousos, G., & Gold, P. (1992). The concepts of stress and stress system disorders: Overview of physical and behavioral homeostasis. Journal of the American Medical Association, 267, 1244–1252. doi: 10.1001/jama.1992.03480090092034 Chugani, H. T. (1998). A critical period of brain development: Studies of cerebral glucose utilization with PET. Preventive Medicine, 27(2), 184–188. doi: 10.1006/pmed.1998.0274 Cicchetti, D. (1993). Developmental psychopathology: Reactions, reflections, projections. Developmental Review, 13(4), 471–502. doi: 10.1006 /drev.1993.1021 Cicchetti, D., & Gunnar, M. R. (2008). Integrating biological measures into the design and evaluation of preventive interventions. Development and Psychopathology, 20(3), 737. doi: 10.1017 /S0954579408000357 Cicchetti, D., & Posner, M. I. (2005). Cognitive and affective neuroscience and developmental psychopathology. Development and Psychopathology, 17(3), 569–575. doi: 10.1017/S0954579405050273 Cicchetti, D., Rogosch, F. A., Gunnar, M. R., & Toth, S. L. (2010). The differential impacts of early abuse on internalizing problems and diurnal cortisol activity in school-aged children. Child Development, 25, 252–269. doi: 10.1111/j.1467–8624.2009.01393.x Cicchetti, D., Rogosch, F. A., Toth, S. L., & Sturge-Apple, M. L. (2011). Normalizing the development of cortisol regulation in maltreated infants through preventive interventions. Development and Psychopathology, 23(3), 789–800. doi: 10.1017/S0954579411000307 Cicchetti, D., & Toth, S. L. (2009). The past achievements and future promises of developmental psychopathology: The coming of age of a discipline. Journal of Child Psychology and Psychiatry, 50, 16–25. doi: 10.1111/j.1469–7610.2008.01979.x. Cicchetti, D., & Valentino, K. (2006). An ecological-transactional perspective on child maltreatment: Failure of the average expectable environment and its influence upon child development. In D. Cicchetti & J. Cohen (Eds.), Developmental psychopathology (2nd ed., Vol. 3, pp. 129–201). New York, NY: Wiley. Cisler, J. M., Elton, A., Kennedy, A. P., Young, J., Smitherman, S., Andrew James, G., & Kilts, C. D. (2013). Altered functional connectivity of the insular cortex across prefrontal networks in cocaine addiction. Psychiatry Research: Neuroimaging. 213(1), 39–46.] doi: 10.1016/j.pscychresns.2013.02.007 Cisler, J. M., James, G. A., Tripathi, S., Mletzko, T., Heim, C., Hu, X. P., . . . Kilts, C. D. (2012). Differential functional connectivity within an emotion regulation neural network among individuals resilient and susceptible to the depressogenic effects of early life stress. Psychological Medicine, 43(3), 1–12. doi: 10.1017/S0033291712001390 Clark, C. M., Kessler, R. C., Buchsbaum, M. S., Margolin, R. A., & Holcomb, H. H. (1984). Correlational methods for determining regional coupling of cerebral glucose metabolism: A pilot study. Biological Psychiatry, 19, 663–678. Cohen, S., & Williamson, G. M. (1991). Stress and infectious disease in humans. Psychological Bulletin, 109(1), 5. doi: 10.1037/0033–2909 .109.1.5 Colcombe, S. J., Kramer, A. F., Erickson, K. I., Scalf, P., McAuley, E., Cohen, N. J., & al, et. (2003). Cardiovascular fitness, cortical plasticity, and aging. Proceedings of the National Academy of Sciences of the United States of America, 101, 3316–3321. doi: 10.1073/pnas .0400266101

References Cole, S. W., Hawkley, L. C., Arevalo, J. M., Sung, C. Y., Rose, R. M., & Cacioppo, J. T. (2007). Social regulation of gene expression in human leukocytes. Genome Biology, 8(9), R189. doi: 10.1186/gb -2007–8–9-r189 Convit, A., Wolf, O. T., de Leon, M. J., Patalinjug, M., Kandil, E., Caraos, C., . . . Cancro, R. (2001). Volumetric analysis of the pre-frontal regions: Findings in aging and schizophrenia. Psychiatry Research: Neuroimaging, 107(2), 61–73. doi: 10.1016/S0925–4927(01)00097-X Convit, A., Wolf, O. T., Tarshish, C., & de Leon, M. J. (2002). Reduced glucose tolerance is associated with poor memory performance and hippocampal atrophy among normal elderly. Proceedings of the National Academy of Sciences of the United States of America, 100(4), 2019–2022. doi: 10.1073/pnas.0336073100 Copeland, W. E., Keeler, G., Angold, A., & Costello, E. J. (2007). Traumatic events and posttraumatic stress in childhood. Archives of General Psychiatry, 64(5), 577. doi: 10.1001/archpsyc.64.5.577 Coplan, J. D., Smith, E. L.P., Altemus, M., Scharf, B. A., Owens, Nemeroff, C. B., . . . Rosenblum, L. A. (2001). Variable foraging demand rearing: Sustained elevations in cisternal cerebrospinal fluid corticotropin-releasing factor concentrations in adult primates. Biological Psychiatry, 50(3), 200–204. doi: 10.1016/S0006 –3223(01)01175–1 Corbetta, M., Patel, G., & Shulman, G. L. (2008). The reorienting system of the human brain: From environment to theory of mind. Neuron, 58(3), 306–324. doi: 10.1016/j.neuron.2008.04.017 Corbetta, M., & Schulman, G. L. (2002). Control of goal-directed and stimulus-driven attention to the brain. Nature Reviews of Neuroscience, 3, 201–215. doi: 10.1038/nrn755 Cottrell, E. C., & Seckl, J. R. (2009). Prenatal stress, glucocorticoids and the programming of adult disease. Frontiers in Behavioral Neuroscience, 3. doi: 10.3389/neuro.08.019.2009 Courchesne, E., Chisum, H., & Townsend, J. (1994). Neural activity-dependent brain changes in development: Implications for psychopathology. Development and Psychopathology, 6, 697–722. doi: 10.1017/S0954579400004740 Craig, A. D. (2003). Pain mechanisms: Labeled lines verses convergence in central processing. Annual Review of Neuroscience, 21(1), 1–30. doi: 10.1146/annurev.neuro.26.041002.131022 Craig, A. D. (2008). Interoception and emotion: A neuroanatomical perspective. In M. Lewis, J. Haviland-Jones, & L. Barrett (Eds.), Handbook of emotions (pp. 272–287). New York, NY: Guildford Press. Critchley, H. D. (2004). The human cortex responds to an introceptive challenge. Proceedings of the National Academy of Sciences of the United States of America, 101, 6333–6334. doi: 10.1073/pnas .0401510101 Critchley, H. D., Mathias, C. J., Josephs, O., O’Doherty, J., Zanini, S., Dewar, B. K., . . . Dolan, R. J. (2003). Human cingulate cortex and autonomic control: Converging neuroimaging and clinical evidence. Brain, 126, 2139–2152. doi: 10.1093/brain/awg216 Critchley, H. D., Weins, S., Rotshtein, P., Ohman, A., & Dolan, R. J. (2004). Neural systems supporting interoceptive awareness. Natural Neuroscience, 7, 189–195. doi: 10.1038/nn1176 Crofts, J. J., Higham, D. J., Bosnell, R., Jbabdi, S., Matthews, P. M., Behrens, T. E.J., & Johansen-Berg, H. (2011). Network analysis detects changes in the contralesional hemisphere following stroke. NeuroImage, 54(1), 161–169. doi: 10.1016/j.neuroimage.2010.08.032 Cui, H, Sakamoto, H, Higashi, S, & Kawata, M. (2008). Effects of single-prolonged stress on neurons and their afferent inputs in the amygdala. Neuroscience, 152(3), 703–712. doi: 10.1016/j.neuroscience .2007.12.028 Damasio, A. R. (1999). The feeling of what happens: Body and emotion in the making of consciousness. New York NY: Harcourt Brace.

617

Damasio, A. R. (2003). Feelings of emotion and the self. Annals of the New York Academy of Sciences, 1001(1), 253–261. doi: 10.1196/annals .1279.014 Damasio, A. R., & Carvalho, G. B. (2013). The nature of feelings: Evolutionary and neurobiological origins. Nature Reviews Neuroscience, 14(2), 143–152. doi: 10.1038/nrn3403 Damon, W., Lerner, R. M., & Eisenberg, L. (2006). Handbook of child psychology: Theoretical models of human development (6th ed., Vol. 1). New York, NY: John Wiley & Sons, Inc. Davidson, R. J., Maxwell, J. S., & Shackman, A. J. (2004). Privileged statues of emotion in the brain. Proceedings of the National Academy of Sciences of the United States of America, 101, 11915–11916. doi: 10.1073/pnas.0404264101 Davies, A. M. (2013). Regulation of neuronal survival by neurotrophins in the developing peripheral nervous system. In J. Rubenstein & P. Rakic (Eds.), Patterning and cell types specification in the developing CNS and PNS (pp. 303–311). San Diego, CA: Academic Press. Davis, M., Walker, D., & Lee, Y. (1997). Roles of the amygdala and bed nucleus of the stria terminalis in fear and anxiety measured with the acoustic startle reflex: Possible relevance to PTSD. In R. Yehuda & A. McFarlane (Eds.), Annals of the New York Academy of Sciences: Psychobiology of posttraumatic stress disorder (Vol. 821, pp. 305–331). New York, NY: New York Academy of Sciences. De Bellis, M. D., Baum, A., Birmaher, B., Keshavan, M., & Ryan, N. D. (1999). A. E. Bennet Research Award. Developmental traumatology. Part I: Biological stress systems. Biological Psychiatry, 45(10), 1259–1270. doi: 10.1111/j.1749–6632.1997.tb48289.x De Bellis, M. D., Keshavan, M. S., Shifflett, H., Iyengar, S., Beers, S. R., Hall, J., & Moritz, G. (2002). Brain structures in pediatric maltreatment-related posttraumatic stress disorder: A sociodemographically matched study. Biological Psychiatry, 52(11), 1066–1078. doi: 10.1016/S0006–3223(02)01459–2 De Bellis, M. D., & Kuchibhatla, M. (2006). Cerebellar volumes in pediatric maltreatment-related posttraumatic stress disorder. Biological Psychiatry, 60(7), 697–703. doi: 10.1016/j.biopsych.2006.04.035 De Kloet, E. R., Rosenfeld, P., Van Eekelen, A. M., Sutanto, W., & Levine, S. (1988). Stress, glucocorticoids, and development. Progress in Brain Research, 73, 101–120. De Lima, A. D., Gieseler, A., & Voigt, T. (2009). Relationship between GABAergic interneurons migration and early neocortical network activity. Developmental Neurobiology, 69(2–3), 105–123. doi: 10.1002/dneu.20696 Dedovic, K., Duchesne, A., Andrews, J., Engert, V., & Pruessner, J. C. (2009). The brain and the stress axis: The neural correlates of cortisol regulation in response to stress. NeuroImage, 47(3), 864–871. doi: 10.1016/j.neuroimage.2009.05.074 Delprato, D. J., & McGlynn, F. D. (1985). Innovations in behavioral medicine. Progress in Behavior Modification, 20(1), 67–122. Demetriou, A., Christou, C., & Spanoudis, G. (2002). The development of mental processing: Efficiency, working memory, and thinking. Monographs of the Society for Research in Child Development, 67, 1–156. Denburg, N. L., Buchanan, T. W., Tranel, D., & Adolphs, R. (2003). Evidence for preserved emotional memory in normal older persons. Emotion, 3(3), 239. doi: 10.1037/1528–3542.3.3.239 Depue, R. A., & Morrone-Strupinsky, J. V. (2005). A neurobehavioral model of affiliative bonding: Implications for conceptualizing a human trait of affiliation. Behavioral and Brain Sciences, 28(3), 313–350. doi: 10.1017/S0140525X05000063 Di Cristo, G., Chattopadhyaya, B., Kuhlman, S. J., Fu, Y., Bélanger, M., Wu, C. Z., . . . Huang, Z. J. (2007). Activity-dependent PSA expression regulates inhibitory maturation and onset of critical period plasticity. Nature Neuroscience, 10(12), 1569–1577. doi: 10.1038/nn2008

618

Life Span Diathesis–Stress Model of Psychopathology

Diorio, J., & Meaney, M. J. (2007). Maternal programming of defensive responses through sustained effects on gene expression. Journal of Psychiatry Neuroscience, 32, 275–284. Doetsch, F., Caillé, I., Lim, D. A., García-Verdugo, J. M., & Alvarez-Buylla, A. (1999). Subventricular zone astrocytes are neural stem cells in the adult mammalian brain. Cell, 97(6), 703–716. doi: 10.1016/S0092–8674(00)80783–7 Dohrenwend, B. S., Ashkenasy, A. R., Krasnoff, L., & Dohrenwend, B. P. (1978). Exemplification of a method for scaling life events: The PERI Life Events Scale. Journal of Health and Social Behavior, 19, 205–229. Dolcos, F., LaBar, K. S., & Cabeza, R. (2004). Dissociable effects of arousal and valence on prefrontal activity indexing emotional evolution and subsequent memory: An event-related fMRI study. NeuroImage, 23, 64–74. doi: 10.1016/j.neuroimage.2004.05.015 Dolcos, F., & McCarthy, G. (2006). Brain systems mediating cognitive interference by emotional distraction. Journal of Neuroscience, 26(7), 2072–2079. doi: 10.1523/JNEUROSCI.5042–05.2006 Dosenbach, N. U.F., Fair, D. A., Cohen, A. L., Schlagger, B. L., & Petersen, S. E. (2008). A dual-networks architecture of top-down control. Trends in Cognitive Science, 12, 99–105. doi: 10.1016/j.tics .2008.01.001 Dosenbach, N. U.F., Fair, D. A., Miezin, F. M., Cohen, A. L., Wenger, K. K., Dosenbach, R. A., . . . Petersen, S. E. (2007). Distinct brain networks for adaptive and stable task control in humans. Proceedings of the National Academy of Sciences of the United States of America, 104, 11073–11078. doi: 10.1073/pnas.0704320104 Dosenbach, N. U.F., Nardos, B., Cohen, A. L., Fair, D. A., Power, J. D., Church, J. A., . . . Lessov-Schlaggar, C. N. (2010). Prediction of individual brain maturity using fMRI. Science, 329(5997), 1358–1361. doi: 10.1126/science.1194144 Dosenbach, N. U.F., Visscher, K. M., Palmer, E. D., Miezin, F. M., Wenger, K. K., Kang, H. C., . . . Petersen, S. E. (2006). A core system for the implementation of task sets. Neuron, 50, 799–812. doi: 10.1016/j.neuron.2006.04.031 Downar, J., Crawley, A. P., Mikulis, D. J., & Davis, K. D. (2002). A cortical network sensitive to stimulus salience in a neutral behavioral context across multiple sensory modalities. Journal of Neurophysiology, 87, 615–620. doi: 10.1152/jn.00636.2001. Downar, J., Mikulis, D. J., & Davis, K. D. (2003). Neural correlates of the prolonged salience of painful stimulation. NeuroImage, 20, 1540–1551. doi: 10.1016/S1053–8119(03)00407–5 Dozier, M., Peloso, E., Lewis, E., Laurenceau, J. P., & Levine, S. (2008). Effects of an attachment-based intervention on the cortisol production of infants and toddlers in foster care. Development and Psychopathology, 20(3), 845–859. doi: 10.1017/S0954579408000400 Driscoll, I., & Sutherland, R. J. (2005). The aging hippocampus: Navigating between rat and human experiments. Reviews in the Neurosciences, 16(2), 87–122. Duric, V., Banasr, M., Licznerski, P., Schmidt, H. D., Stockmeier, C. A., Simen, A. A., . . . Duman, R. S. (2010). A negative regulator of MAP kinase causes depressive behavior. Nature Medicine, 16(11), 1328–1332. doi: 10.1038/nm.2219 Durlak, J. A., Weissberg, R. P., Dymnicki, A. B., Taylor, R. D., & Schellinger, K. B. (2011). The impact of enhancing students’ social and emotional learning: A meta-analysis of school-based universal interventions. Child Development, 82(1), 405–432. doi: 10.1111/j.1467–8624.2010.01564.x Earls, F., & Buka, S. (2000). Measurement of community characteristics. Handbook of Early Childhood Intervention, 2, 309–324. Edgar, N., & Sibille, E. (2012). A putative functional role for oligodendrocytes in mood regulation. Translational Psychiatry, 2(5), e109. doi: 10.1038/tp. 2012.34

Eisenberger, N. I., & Lieberman, M. D. (2004). Why rejection hurts: A common neural alarm system for physical and social pain. Trends in Cognitive Sciences, 8(7), 294–300. doi: 10.1016/j.tics.2004.05.010 Eisenberger, N. I., Lieberman, M. D., & Williams, K. D. (2003). Does rejection hurt? An fMRI study of social exclusion. Science, 302, 290–292. doi: 10.1126/science.1089134 Eisenberger, N. I., Taylor, S. E., Gable, S. L., Hilmert, C. J., & Lieberman, M. D. (2007). Neural pathways link social support to attenuated neuroendocrine stress responses. NeuroImage, 35(4), 1601–1612. doi: 10.1016/j.neuroimage.2007.01.038 Engel, G. (1977). The need for a new medical model: A challenge for biomedicine. Science, 196(4286), 129–136. doi: 10.1126/science .847460 Etkin, A., Prater, K. E., Schatzberg, A. F., Menon, V., & Greicius, M. D. (2009). Disrupted amygdalar subregion functional connectivity and evidence of a compensatory network in generalized anxiety disorder. Archives of General Psychiatry, 66(12), 1361. doi: 10.1001/archgenpsychiatry.2009.104 Etkin, A., & Wager, T. D. (2007). Functional neuroimaging of anxiety: A meta-analysis of emotional processing in PTSD, social anxiety disorder, and specific phobia. American Journal of Psychiatry, 164(10), 1476. doi: 10.1176/appi.ajp. 2007.07030504 Evans, G. W., Kim, P., Ting, A. H., Tesher, H. B., & Shannis, D. (2007). Cumulative risk, maternal responsiveness, and allostatic load among young adolescents. Developmental Psychology, 43(2), 341–351. doi: 10.1037/0012–1649.43.2.341 Evrard, H. C., Forro, T., & Logothetis, N. K. (2012). Von Economo neurons in the anterior insula of the macaque monkey. Neuron, 74(3), 482–489. doi: 10.1016/j.neuron.2012.03.003 Ewers, M., Sperling, R. A., Klunk, W. E., Weiner, M. W., & Hampel, H. (2011). Neuroimaging markers for the prediction and early diagnosis of Alzheimer’s disease dementia. Trends in Neurosciences, 34(8), 430–442. doi: 10.1016/j.tins.2011.05.005 Fagiolini, M., & Hensch, T. K. (2000). Inhibitory threshold for critical-period activation in primary visual cortex. Nature, 404(6774), 183–186. doi: 10.1038/35004582 Fair, D. A., Cohen, A. L., Dosenbach, N. U.F., Church, J. A., Miezin, F. M., Barch, D. M., . . . Schlaggar, B. L. (2008). The maturing architecture of the brain’s default network. Proceedings of the National Academy of Sciences of the United States of America, 105(10), 4028–4032. doi: 10.1073/pnas.0800376105 Fair, D. A., Dosenbach, N. U.F., Church, J. A., Cohen, A. L., Brahmbhatt, S., Miezin, F. M., . . . Schlagger, B. L. (2007). Development of distinct control networks through segregation and integration. Proceedings of the National Academy of Sciences of the United States of America, 104, 13507–13512. doi: 10.1073/pnas.0705843104 Farb, N. A.S., Segal, Z. V., & Anderson, A. K. (2013). Attentional modulation of primary interoceptive and exteroceptive cortices. Cerebral Cortex, 23(1), 114–126. doi: 10.1093/cercor/bhr385 Feinberg, I., & Campbell, I. G. (2009). Sleep EEG changes during adolescence: An index of fundamental brain reorganization. Brain and Cognition, 72(1), 56–65. doi: 10.1016/j.bandc.2009.09.008 Feldman, S., Conforti, N., Itzik, A., & Weidenfeld, J. (1995). The role of limbic structures in the modulation of ACTH responses following adrenalectomya. Annals of the New York Academy of Sciences, 771(1), 73–81. doi: 10.1111/j.1749–6632.1995.tb44671.x Fellin, T., Halassa, M. M., Terunuma, M., Succol, F., Takano, H., Frank, M., . . . Haydon, P. G. (2009). Endogenous nonneuronal modulators of synaptic transmission control cortical slow oscillations in vivo. Proceedings of the National Academy of Sciences of the United States of America, 106(35), 15037–15042. doi: 10.1073/pnas.0906419106

References Finn, D. P. (2010). Endocannabinoid-mediated modulation of stress responses: Physiological and pathophysiological significance. Immunobiology, 215(8), 629–646. doi: 10.1016/j.imbio.2009.05.011 Fischer, H., Sandblom, J., Gavazzeni, J., Fransson, P., Wright, C. I., & Bäckman, L. (2005). Age-differential patterns of brain activation during perception of angry faces. Neuroscience Letters, 386(2), 99–104. doi: 10.1016/j.neulet.2005.06.002 Fisher, P. A., Stoolmiller, M., Gunnar, M. R., & Burraston, B. O. (2007). Effects of a therapeutic intervention for foster preschoolers on diurnal cortisol activity. Psychoneuroendocrinology, 32(8), 892–905. doi: 10.1016/j.psyneuen.2007.06.008 Fiszbein, A., Schady, N. R., & Ferreira, F. H.G. (2009). Conditional cash transfers: Reducing present and future poverty. Washington, DC: World Bank Publications. Florian, C., Vecsey, C. G., Halassa, M. M., Haydon, P. G., & Abel, T. (2011). Astrocyte-derived adenosine and A1 receptor activity contribute to sleep loss-induced deficits in hippocampal synaptic plasticity and memory in mice. Journal of Neuroscience, 31(19), 6956–6962. doi: 10.1523/JNEUROSCI.5761–10.2011 Foa, E., & Kozak, M. (1986). Emotional processing of fear: Exposure to corrective information. Psychological Bulletin, 99, 20–35. doi: 10.1037/0033–2909.99.1.20 Fogel, S., Martin, N., Lafortune, M., Barakat, M., Debas, K., Laventure, S., . . . Carrier, J. (2012). NREM sleep and brain plasticity in aging. Frontiers in Neurology, 3, 176. Ford, D. H., & Lerner, R. M. (1992). Developmental systems theory: An integrative approach. Thousand Oaks, CA: Sage Publications, Inc. Fox, M. D., & Raichle, M. E. (2007). Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nature Reviews Neuroscience, 8(9), 700–711. doi: 10.1038 /nrn2201 Fox, M. D., Snyder, A. Z., Vincent, J. L., Corbetta, M., Van Essen, D. C., & Raichle, M. E. (2005). The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proceedings of the National Academy of Sciences of the United States of America, 102(27), 9673–9678. doi: 10.1073/pnas.0504136102 Francis, D., Diorio, J., Liu, D., & Meaney, M. J. (1999). Nongenomic transmission across generations of maternal behavior and stress responses in the rat. Science, 32, 892–905. doi: 10.1126/science.286.5442.1155 Frank, M. G. (2010). The functions of sleep. In J. W. Winkelman & D. T. Plante (Eds.), Foundations of psychiatric sleep medicine (pp. 59–78). Cambridge, UK: Cambridge University Press. Frank, M. G. (2012). Glial modulations of sleep and electroencephalographic rhythms. In M. G. Frank (Ed.), Sleep and brain activity (pp. 47–58). Oxford, UK: Academic Press. Frank, M. G. (2013). Why I am not SHY: A reply to Tononi and Cirelli. Neural Plasticity, 2013. Article ID 394946, 3 pages,] doi: 10.1155/2013/394946 Frankenhaeuser, M. (1980). Psychobiological aspects of life stress. In S. Levine & J. Ursin (Eds.), Coping and health (pp. 203–223). New York, NY: Plenum Press. Freund, T. F., Katona, I., & Piomelli, D.