Measuring Distribution and Mobility of Income and Wealth 9780226816043

A collection of twenty-three studies that explore the latest developments in the analysis of income and wealth distribut

190 108 14MB

English Pages 792 [736] Year 2022

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Contents
Prefatory Note
Acknowledgments
Introduction
I Income Inequality
1. In Search of the Roots of American Inequality Exceptionalism: An Analysis Based on Luxembourg Income Study (LIS) Data
2. Rising Between- Firm Inequality and Declining Labor Market Fluidity: Evidence of a Changing Job Ladder
3. United States Earnings Dynamics: Inequality, Mobility, and Volatility
4 Evidence from Unique Swiss Tax Data on the Composition and Joint Distribution of Income and Wealth
II Wealth Inequality
5 The Wealth of Generations, with Special Attention to the Millennials
6. Wealth Transfers and Net Wealth at Death: Evidence from the Italian Inheritance Tax Records, 1995– 2016
7. On the Distribution of Estates and the Distribution of Wealth: Evidence from the Dead
8. Structuring the Analysis of Wealth Inequality Using the Functions of Wealth: A Class- Based Approach
9 Social Security Wealth, Inequality, and Life- Cycle Saving
III Income and Wealth Mobility
10 Parental Education and the Rising Transmission of Income between Generations
11. Inequality of Opportunity for Income in Denmark and the United States: A Comparison Based on Administrative Data
12 Presence and Persistence of Poverty in US Tax Data
13 Intergenerational Home Ownership in France over the Twentieth Century
14 Inequality and Mobility over the Past Half- Century Using Income, Consumption, and Wealth
IV Mitigating Inequality
15. The Accuracy of Tax Imputations: Estimating Tax Liabilities and Credits Using Linked Survey and Administrative Data
16. Geographic Inequality in Social Provision: V ariation across the US States
17 Inequality and the Safety Net in American Cities throughout the Income Distribution, 1929– 1940
18 The EITC and Linking Data for Examining Multigenerational Effects
V Distributional National Accounts
19. Distributing Personal Income: Trends over Time
20 Developing Indicators of Inequality and Poverty Consistent with National Accounts
21. Distributional National Accounts: A Macro- Micro Approach to Inequality in Germany
22 The Distributional Financial Accounts of the United States
23 Using Tax Data to Better Capture Top Incomes in Official UK Income Inequality Statistics
Contributors
Author Index
Subject Index
Recommend Papers

Measuring Distribution and Mobility of Income and Wealth
 9780226816043

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Measuring Distribution and Mobility of Income and Wealth

Studies in Income and Wealth Volume 80

Measuring Distribution and Mobility of Income and Wealth

Edited by

Raj Chetty, John N. Friedman, Janet C. Gornick, Barry Johnson, and Arthur Kennickell

The University of Chicago Press Chicago and London

The University of Chicago Press, Chicago 60637 The University of Chicago Press, Ltd., London © 2022 by the National Bureau of Economic Research All rights reserved. No part of this book may be used or reproduced in any manner whatsoever without written permission, except in the case of brief quotations in critical articles and reviews. For more information, contact the University of Chicago Press, 1427 E. 60th St., Chicago, IL 60637. Published 2022 Printed in the United States of America 30 29 28 27 26 25 24 23 22 21

1 2 3 4 5

ISBN-13: 978-0-226-81603-6 (cloth) ISBN-13: 978-0-226-81604-3 (e-book) DOI: https://doi.org /10.7208/chicago/9780226816043.001.0001 Library of Congress Cataloging-in-Publication Data Names: Chetty, Raj, editor. | Friedman, John N., editor. | Gornick, Janet C., editor. | Johnson, Barry (Economist), editor. | Kennickell, Arthur B., editor. Title: Measuring distribution and mobility of income and wealth / [edited by] Raj Chetty, John N. Friedman, Janet C. Gornick, Barry Johnson, Arthur Kennickell. Other titles: Studies in income and wealth ; v. 80. Description: Chicago ; London : The University of Chicago Press, 2022. | Series: Studies in income and wealth; volume 80 | Revised versions of papers presented at the Conference on Research in income and Wealth titled “Measuring and understanding the distribution and intra/inter-generational mobility of income and wealth,” held in Bethesda, Maryland, on March 5– 6, 2020. | Includes bibliographical references and index. Identifiers: LCCN 2022026311 | ISBN 9780226816036 (cloth) | ISBN 9780226816043 (ebook) Subjects: LCSH: Income distribution— Congresses. | Wealth— Congresses. | BISAC: BUSINESS & ECONOMICS / Econometrics | BUSINESS & ECONOMICS / Economics / General | LCGFT: Conference papers and proceedings. Classification: LCC HC79.I5 M37 2022 | DDC 339.2— dc23/eng/ 20220607 LC record available at https://lccn.loc.gov/2022026311 ♾ This paper meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of Paper).

National Bureau of Economic Research Officers John Lipsky, Chair Peter Blair Henry, Vice-Chair James M. Poterba, President and Chief Executive Officer Robert Mednick, Treasurer

Kelly Horak, Controller and Assistant Corporate Secretary Alterra Milone, Corporate Secretary Denis Healy, Assistant Corporate Secretary

Directors at Large Susan M. Collins Kathleen B. Cooper Charles H. Dallara George C. Eads Jessica P. Einhorn Mohamed El-Erian Diana Farrell Helena Foulkes Jacob A. Frenkel

Robert S. Hamada Peter Blair Henry Karen N. Horn Lisa Jordan John Lipsky Laurence H. Meyer Karen Mills Michael H. Moskow Alicia H. Munnell

Robert T. Parry Douglas Peterson James M. Poterba John S. Reed Hal Varian Mark Weinberger Martin B. Zimmerman

Directors by University Appointment Timothy Bresnahan, Stanford Pierre-André Chiappori, Columbia Maureen Cropper, Maryland Alan V. Deardorff, Michigan Graham Elliott, California, San Diego Edward Foster, Minnesota Bruce Hansen, Wisconsin-Madison Benjamin Hermalin, California, Berkeley

Samuel Kortum, Yale George Mailath, Pennsylvania Joel Mokyr, Northwestern Richard L. Schmalensee, Massachusetts Institute of Technology Lars Stole, Chicago Ingo Walter, New York David B. Yoffie, Harvard

Directors by Appointment of Other Organizations Timothy Beatty, Agricultural and Applied Economics Association Martin J. Gruber, American Finance Association Philip Hoffman, Economic History Association Arthur Kennickell, American Statistical Association Robert Mednick, American Institute of Certified Public Accountants

Dana Peterson, The Conference Board Lynn Reaser, National Association for Business Economics Peter L. Rousseau, American Economic Association Gregor W. Smith, Canadian Economics Association William Spriggs, American Federation of Labor and Congress of Industrial Organizations

Directors Emeriti George Akerlof Peter C. Aldrich Elizabeth E. Bailey Jagdish Bhagwati John H. Biggs

Don R. Conlan Ray C. Fair Saul H. Hymans Marjorie B. McElroy Rudolph A. Oswald

Andrew Postlewaite John J. Siegfried Craig Swan Marina v. N. Whitman

Relation of the Directors to the Work and Publications of the NBER 1. The object of the NBER is to ascertain and present to the economics profession, and to the public more generally, important economic facts and their interpretation in a scientific manner without policy recommendations. The Board of Directors is charged with the responsibility of ensuring that the work of the NBER is carried on in strict conformity with this object. 2. The President shall establish an internal review process to ensure that book manuscripts proposed for publication DO NOT contain policy recommendations. This shall apply both to the proceedings of conferences and to manuscripts by a single author or by one or more co-authors but shall not apply to authors of comments at NBER conferences who are not NBER affiliates. 3. No book manuscript reporting research shall be published by the NBER until the President has sent to each member of the Board a notice that a manuscript is recommended for publication and that in the President’s opinion it is suitable for publication in accordance with the above principles of the NBER. Such notification will include a table of contents and an abstract or summary of the manuscript’s content, a list of contributors if applicable, and a response form for use by Directors who desire a copy of the manuscript for review. Each manuscript shall contain a summary drawing attention to the nature and treatment of the problem studied and the main conclusions reached. 4. No volume shall be published until forty-five days have elapsed from the above notification of intention to publish it. During this period a copy shall be sent to any Director requesting it, and if any Director objects to publication on the grounds that the manuscript contains policy recommendations, the objection will be presented to the author(s) or editor(s). In case of dispute, all members of the Board shall be notified, and the President shall appoint an ad hoc committee of the Board to decide the matter; thirty days additional shall be granted for this purpose. 5. The President shall present annually to the Board a report describing the internal manuscript review process, any objections made by Directors before publication or by anyone after publication, any disputes about such matters, and how they were handled. 6. Publications of the NBER issued for informational purposes concerning the work of the Bureau, or issued to inform the public of the activities at the Bureau, including but not limited to the NBER Digest and Reporter, shall be consistent with the object stated in paragraph 1. They shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The Executive Committee of the Board is charged with the review of all such publications from time to time. 7. NBER working papers and manuscripts distributed on the Bureau’s web site are not deemed to be publications for the purpose of this resolution, but they shall be consistent with the object stated in paragraph 1. Working papers shall contain a specific disclaimer noting that they have not passed through the review procedures required in this resolution. The NBER’s web site shall contain a similar disclaimer. The President shall establish an internal review process to ensure that the working papers and the web site do not contain policy recommendations, and shall report annually to the Board on this process and any concerns raised in connection with it. 8. Unless otherwise determined by the Board or exempted by the terms of paragraphs 6 and 7, a copy of this resolution shall be printed in each NBER publication as described in paragraph 2 above.

Contents

Prefatory Note Acknowledgments Introduction Raj Chetty, John N. Friedman, Janet C. Gornick, Barry Johnson, and Arthur Kennickell

xi xiii 1

I. Income Inequality 1. In Search of the Roots of American Inequality Exceptionalism: An Analysis Based on Luxembourg Income Study (LIS) Data Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson 2. Rising Between-Firm Inequality and Declining Labor Market Fluidity: Evidence of a Changing Job Ladder John Haltiwanger and James R. Spletzer 3. United States Earnings Dynamics: Inequality, Mobility, and Volatility Kevin L. McKinney, John M. Abowd, and John Sabelhaus 4. Evidence from Unique Swiss Tax Data on the Composition and Joint Distribution of Income and Wealth Isabel Z. Martínez

19

45

69

105

vii

viii

Contents

II. Wealth Inequality 5. The Wealth of Generations, with Special Attention to the Millennials William G. Gale, Hilary Gelfond, Jason J. Fichtner, and Benjamin H. Harris 6. Wealth Transfers and Net Wealth at Death: Evidence from the Italian Inheritance Tax Records, 1995– 2016 Paolo Acciari and Salvatore Morelli 7. On the Distribution of Estates and the Distribution of Wealth: Evidence from the Dead Yonatan Berman and Salvatore Morelli 8. Structuring the Analysis of Wealth Inequality Using the Functions of Wealth: A Class-Based Approach Pirmin Fessler and Martin Schürz 9. Social Security Wealth, Inequality, and Life-Cycle Saving John Sabelhaus and Alice Henriques Volz

145

175

205

221

249

III. Income and Wealth Mobility 10. Parental Education and the Rising Transmission of Income between Generations 289 Marie Connolly, Catherine Haeck, and Jean-William Laliberté 11. Inequality of Opportunity for Income in Denmark and the United States: A Comparison Based on Administrative Data Pablo A. Mitnik, Anne-Line Helsø, and Victoria L. Bryant 12. Presence and Persistence of Poverty in US Tax Data Jeff Larrimore, Jacob Mortenson, and David Splinter 13. Intergenerational Home Ownership in France over the Twentieth Century Bertrand Garbinti and Frédérique Savignac

317

383

411

Contents

14. Inequality and Mobility over the Past Half-Century Using Income, Consumption, and Wealth Jonathan D. Fisher and David S. Johnson

ix

437

IV. Mitigating Inequality 15. The Accuracy of Tax Imputations: Estimating Tax Liabilities and Credits Using Linked Survey and Administrative Data 459 Bruce D. Meyer, Derek Wu, Grace Finley, Patrick Langetieg, Carla Medalia, Mark Payne, and Alan Plumley 16. Geographic Inequality in Social Provision: V ariation across the US States Sarah K. Bruch, Janet C. Gornick, and Joseph van der Naald 17. Inequality and the Safety Net in American Cities throughout the Income Distribution, 1929– 1940 James Feigenbaum, Price Fishback, and Keoka Grayson 18. The EITC and Linking Data for Examining Multigenerational Effects Randall Akee, Maggie R. Jones, and Emilia Simeonova

499

529

569

V. Distributional National Accounts 19. Distributing Personal Income: Trends over Time Dennis Fixler, Marina Gindelsky, and David S. Johnson

589

20. Developing Indicators of Inequality and Poverty Consistent with National Accounts 605 Richard Tonkin, Sean White, Sofiya Stoyanova, Aly Youssef, Sunny Valentineo Sidhu, and Chris Payne 21. Distributional National Accounts: A Macro-Micro Approach to Inequality in Germany Stefan Bach, Charlotte Bartels, and Theresa Neef

625

x

Contents

22. The Distributional Financial Accounts of the United States Michael Batty, Jesse Bricker, Joseph Briggs, Sarah Friedman, Danielle Nemschoff, Eric Nielsen, Kamila Sommer, and Alice Henriques Volz 23. Using Tax Data to Better Capture Top Incomes in Official UK Income Inequality Statistics Dominic Webber, Richard Tonkin, and Martin Shine Contributors Author Index Subject Index

641

679

701 707 717

Prefatory Note

This volume contains revised versions of the papers presented at the Conference on Research in Income and Wealth titled “Measuring and Understanding the Distribution and Intra/Inter-Generational Mobility of Income and Wealth,” held in Bethesda, Maryland, on March 5– 6, 2020. We gratefully acknowledge the financial support for this conference provided by the Stone Center on Socio-Economic Inequality at the CUNY Graduate Center, the Stone Wealth and Income Inequality Project at Brown University, Statistics of Income (SOI)/Internal Revenue Service (SOI), Opportunity Insights, and the NBER. Support for the general activities of the Conference on Research in Income and Wealth is provided by the following agencies: Bureau of Economic Analysis, Bureau of Labor Statistics, the Census Bureau, the Board of Governors of the Federal Reserve System, SOI/IRS, and Statistics Canada. We thank Raj Chetty, John N. Friedman, Janet C. Gornick, Barry Johnson, and Arthur Kennickell, who served as conference organizers and as editors of the volume. Executive Committee, November 2020 Katharine G. Abraham (chair) John M. Abowd Susanto Basu Ernst R. Berndt Alberto Cavallo Carol A. Corrado Lucy Eldridge John C. Haltiwanger Ron S. Jarmin

J. Bradford Jensen Barry Johnson Greg Peterson Valerie A. Ramey Peter K. Schott Daniel E. Sichel Erich H. Strassner William Wascher

xi

Acknowledgments

The conference organizers thank those who contributed to making the conference successful, especially Brett Maranjian and the other staff members at the National Bureau of Economic Research (NBER) conference department. The authors of the chapters, the discussants, and the inquisitive audience played the most essential role. Discussants for the conference sessions were Pirmin Fessler, William Gale, David Johnson, Maggie R. Jones, Bruce D. Meyer, Frédérique Savingnac, and Alexander Yuskavage. Katharine Abraham, director of the Conference on Research in Income and Wealth (CRIW), provided helpful guidance. The organizers are also grateful to Helena Fitz-Patrick at NBER and to the staff at the University of Chicago Press (UCP) for guiding the volume through the publication process, and to anonymous referees from the NBER and the UCP. The conference would not have been possible without generous financial support from the Stone Center on Socio-Economic Inequality at the CUNY Graduate Center, the Stone Wealth and Income Inequality Project at Brown University, Statistics of Income/Internal Revenue Service, Opportunity Insights, and the NBER. The conference was part of a commemoration of the 100th anniversary of the NBER. The CRIW was launched as an NBER initiative in 1935. NBER president and CEO James Poterba delivered at lunchtime talk at the conference, explaining how income distribution and the share of national income received by labor and capital were among the key issues that led to the founding of the NBER.

xiii

Introduction Raj Chetty, John N. Friedman, Janet C. Gornick, Barry Johnson, and Arthur Kennickell

Economic research on the efficient allocation of resources has a very long history— for many it defines the core of the field. Increasingly over time, however, attention has also turned to inequality in the distribution of personal resources and outcomes, as well as to the related question of whether individuals are locked in their respective initial place in this distribution or whether there is the broadly shared possibility for mobility. Research has focused not only on measuring inequality and mobility but also on understanding its historical, economic, and social determinants, and on how policies might affect these distributions. In addition, it is now recognized with increased clarity that distributional differences may affect the transmission of macroeconomic shocks or responses to fiscal or monetary stimulus. In March 2020, the Conference on Research in Income and Wealth (CRIW) convened a meeting held in Bethesda, Maryland, to explore the latest developments in our understanding of issues related to income and Raj Chetty is the William A. Ackman Professor of Economics at Harvard University, director of Opportunity Insights, and a research associate and director of the Public Economics Program at the National Bureau of Economic Research. John N. Friedman is a professor of economics and international and political affairs at Brown University, and a research associate of the National Bureau of Economic Research. Janet C. Gornick is a professor of political science and sociology, director of the Stone Center on Socio-Economic Inequality, and holds the James M. and Cathleen D. Stone Distinguished Chair in Socio-Economic Inequality at the City University of New York. Barry Johnson is Deputy Chief Data and Analytics Officer and Director of the Statistics of Income Division at the Internal Revenue Service. Arthur Kennickell is a Stone Center Affiliated Scholar at the City University of New York, and a member of the board of directors of the National Bureau of Economic Research. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber.org/books-and-chapters/measuring -distribution-and-mobility-income-and-wealth /introduction.

1

2

Chetty, Friedman, Gornick, Johnson, and Kennickell

wealth distribution and mobility. This was the last NBER-affiliated meeting in 2020 to be held in-person, before COVID-19 concerns made virtualization of later meetings a necessity. Disruptions caused by quarantine shutdowns shortly after the conference prevented several of the conference authors from fully realizing their research goals by the publication date of this volume. The worldwide economic impact of the pandemic certainly makes the topics presented at this conference all the more relevant. The chapters included here highlight new findings, which push forward our knowledge in this area, but also bring new challenges to the fore that the next wave of scholars in this area must confront. A starting point for many of these chapters is an exploration of the difficulties that arise in the definition of income and wealth. Scholars often study these variables to stand as proxies for deeper aspects of inequality that are far more difficult to define and measure consistently. But however straightforward income and wealth may seem at first glance, they also entail many such problems. At the basic level of definition, there is a broad range of possibilities. In the case of income, should one include, for example, service flows from durables and owner-occupied housing or withdrawals from tax-deferred retirement accounts? Similarly, with wealth, should one include contingent assets, such as pension rights, and how should one treat incomeproducing assets in which there is no right to the underlying assets, such as some types of trust, or strongly illiquid assets? These questions require serious thought, especially as the appropriate definition may vary according to the particular intended analytical purpose. Whose income or wealth is often a critical question. Ownership within a household, or an extended family, is sometimes a fuzzy notion. Even when exact ownership can be determined, it may not be relevant— for example, in the case of jurisdictions with community property laws. Ownership rights through legal entities, whether businesses of some sort or trust arrangements, also may substantially veil some types of income or wealth. Even with clear definitions suitable to purpose, there remains the thorny question of how to measure income and wealth, and how to track changes in these variables over time to measure mobility. At least limited information on some measures of income is reasonably available, but wealth data at the individual or household level are much more limited in most countries. While many countries collect income data as part of administrative sources, including tax registers, and some collect partial elements of wealth, only a small number of countries collect broad wealth measures for the full population. While researchers have developed methods to impute wealth from capital income flows, these can be quite noisy. As a result, in many countries survey datasets continue to play a more important role in the measurement of wealth than income, alongside government and private sources. There is also information from sources such as financial institutions and invest-

Introduction

3

ment advisors, but at least for now, this information is not available for the full spectrum of the population and the data elements available are often fragmentary. Increasingly, momentum has been building to link multiple sources of survey, administrative, and other data in order to mitigate measurement problems in single sources or to provide more comprehensive data on income and wealth. While traditional research on income and wealth mobility uses data collected from surveys, recent research has highlighted the fragilities of this data source. Wealthy or high-income households are generally less likely to participate in surveys, and some evidence also suggests that poor households are also less likely to participate. Only in a small number of surveys, such as the US Survey of Consumer Finances, is it even feasible currently to detect and potentially address this deficiency directly. Reporting errors in surveys, driven perhaps by low financial literacy or privacy fears, add noise to the data. Moreover, surveys face two potentially fatal trends: declining response rates in many cases and escalating costs. In particular, the public’s declining willingness to participate complicates the use of survey data to study income or wealth mobility, since it is often difficult to follow individuals or households over successive rounds of a survey without serious attrition, which may bias the results. These pressures add to the incentives to merge and exploit multiple sources of data. More recently, the focus of the income and wealth inequality and mobility literature has turned to the use of administrative datasets. In principle, these sources eliminate some problems inherent in survey datasets— for instance, noisy individual recall and measurement error, or attrition over time— but they raise new issues as well. The contents of administrative datasets are defined by their administrative purposes. Importantly, variables are included or defined by the governing law or regulations, which may change over time. For example, individual income tax data are considered very important for the study of income, but laws and regulatory decisions may have great influence on what is reported, when it is reported, and how it is reported. What is reported may change over time, as a result of changes in the administrative needs. Such considerations may even affect incentives about who reports the information— whether a different person or a legal entity. Sometimes administrative data serve as a basis for projecting patterns for other variables or populations. As noted, under some modeling assumptions about rates of return and other factors, income tax data may be used to project patterns of wealth holding. Similarly, estate tax data have been used to project patterns of wealth holding among the full population; such projection requires assumptions about the “selection probability” appropriate to decedents, the stationarity of the underlying processes, and the parts of the population not covered by such taxation. However, for most countries, the absence in administrative data of full direct measures of all relevant economic and

4

Chetty, Friedman, Gornick, Johnson, and Kennickell

contextual variables for the full spectrum of the population indicates that for at least the intermediate term, both survey and administrative data in blended or other complementary form will be needed to further research. Finally, it is worth noting that privacy concerns often limit what information can be collected accurately or shared. As noted, privacy concerns may affect the incentives persons providing data face to respond at all and to answer faithfully. But data holders also face important privacy considerations. For example, government agencies typically cannot release personal data that can be identified with specific individuals. Agencies may address such constraints by limiting access or by using disclosure limitation techniques to reduce privacy risks by reducing the information content of the data. This volume contains revised versions of most of the papers presented at the conference. They cover an array of topics; some are primarily substantive, others focus more on advancements related to data, measurements, and methods. The 23 chapters are organized into five sections: income inequality, wealth inequality, income and wealth mobility, mitigating inequality, and distributional national accounts. Below, we provide an overview of these five sections and offer previews of the 23 chapters. I.

Income Inequality

For most households, income is the principal driver of consumption and wealth accumulation. Thus, changes in the distribution of income and the transitions of individual income over time have important implications for both short-term and long-term welfare. Income has components derived from labor supply, capital returns, and transfers. Differences in capital income explain much of the inequality observed at the very top of the income distribution. However, labor income is the largest component of personal income and its path over time is therefore a key determinant of inequality and welfare among working households. Unlike the straightforward humpshaped pattern of income in the simplest life-cycle models, labor income may have a variety of trajectories over time, depending on personal choices, labor market fluidity, returns to skills, and larger social forces. The four chapters in this first section address the observed patterns of income inequality and shifts in compensation and fluidity that drive or reinforce income inequality. Gornick, Milanovic, and Johnson (“In Search of the Roots of American Inequality Exceptionalism: An Analysis Based on Luxembourg Income Study (LIS) Data”) assess cross-national variation in households’ market income, focused on the question of what is driving the unusually high level of inequality observed in the US. Using micro data on labor income from 24 OECD countries, they disaggregate the working-age population into household types, defined by the number and gender of the household’s earners and the partnership and parenting status of its members. The authors find that

Introduction

5

the pattern for the US is explained more by relatively high inequality within groups rather than variation in mean income across groups. Haltiwanger and Spletzer (“Rising Between-Firm Inequality and Declining Labor Market Fluidity: Evidence of a Changing Job Ladder”) look at potential connections between the observed rise in earnings inequality and declining labor market fluidity, building on earlier evidence of a rise in between-firm inequality and other work on labor market fluidity. The authors bring data from the Longitudinal Employer Household Dynamics (LEHD) data and other contextual information to bear on the question of the extent to which the observed patterns reflect changes in hiring across industries with different earnings profiles. They find that such changes have made it more difficult for workers both to get on a career ladder and to proceed up the ladder. McKinney, Abowd, and Sabelhaus (“United States Earnings Dynamics: Inequality, Mobility, and Volatility”) look at earnings inequality and dynamics at the subnational level, focusing on for large metropolitan statistical areas (Detroit, Los Angeles, New York, and San Francisco), using data from 1998 to 2017 from the LEHD through the new Earnings and Mobility Statistics (EAMS) application developed by the US Census Bureau. They find an upward shift toward greater concentration among the top of the wage distribution, though with differing trends across these areas. Among other findings, they also report a marked decline in earnings mobility in Detroit and New York. The results in the chapter exemplify analysis that will be possible using the new EAMS web application. Martínez (“Evidence from Unique Swiss Tax Data on the Composition and Joint Distribution of Income and Wealth”) uses administrative data for eight Swiss cantons to examine the joint distribution and composition of income and wealth, revealing both substantial heterogeneity of composition across the distribution and a high correlation of income and wealth at the top. The author finds that age is a powerful determinant of wealth holdings, that gender shapes income more than it does wealth, and that an exceptionally low level of real estate wealth among the bottom 50 percent renders Switzerland distinct from other high-income countries. II.

Wealth Inequality

There appears to be a broad trend across many countries toward an increase in wealth inequality. Understanding the drivers and deeper patterns of inequality is often limited by the availability of data. In part to cope with measurement difficulties, wealth is often treated as a household-level phenomenon, thus obscuring other dimensions of inequality and consequent differences in bargaining power within households. Moreover, the very definition of wealth affects what can be said. While market-based contingent assets are usually included as a part of wealth, there is no definitive rule for

6

Chetty, Friedman, Gornick, Johnson, and Kennickell

how to include a value of contingent economic income-flow entitlements, such as forms of social benefits, pensions, or social security. In addition, it may be that focusing on accounting measures of the value of wealth overlooks the instrumentality of wealth in a social context. The five chapters in this second section address all these questions. Gale, Gelfond, Fichtner, and Harris (“The Wealth of Generations, with Special Attention to the Millennials”) use the US Survey of Consumer Finances for the years 1989– 2016 to investigate the demographic structure of the observed increase in the concentration of wealth over this period. Among other results, they find an upward shift in wealth for older age groups and a decline for the young. Acciari and Morelli (“Wealth Transfers and Net Wealth at Death: Evidence from the Italian Inheritance Tax Records, 1995– 2016”) use data from inheritance tax files to study the concentration of wealth in Italy. Inferring the wealth distribution from estate data requires a means of mapping the wealth of the dead to that of the living. As is usual with such data, they take the form of a “multiplier,” which is the inverse of the probability of death of the decedent. The authors document a substantial rise in the total value of inheritance and gifts as a share of national income, from 8.4 percent in 1995 to 15.1 percent in 2016. At the same time, there was a marked decline in tax revenues linked to these wealth transfers. Berman and Morelli (“On the Distribution of Estates and the Distribution of Wealth: Evidence from the Dead”) look more generally at what can be learned from estate tax data. In particular, they consider how sensitive wealth estimates by this method are to the multipliers typically used to extrapolate estate wealth to the general population. They conclude for the set of countries examined that wealth estimates are sufficiently insensitive to plausible variations in the multipliers that unadjusted estate tax data can give a good indication of wealth among the living. Fessler and Schürz (“Structuring the Analysis of Wealth Inequality Using the Functions of Wealth: A Class-Based Approach”) consider inequality from the perspective of a decomposition of the wealth distribution that relies on a categorization that focuses on the social implications of wealth. The categories are renters (who mainly hold wealth for “precautionary” reasons), homeowners who occupy the homes that they own, and homeowners who also own a business or real estate other than a home. Based on these measures, and analyses of US and European data, the authors propose new measures of inequality they believe are more directly linked to social dynamics and choices. Sabelhaus and Volz (“Social Security Wealth, Inequality, and Life-Cycle Saving”) consider the distributional implications of incorporating measures of net Social Security wealth as part of household net worth. Including such a measure adds substantially to the wealth of otherwise low-wealth house-

Introduction

7

holds. They conclude that including Social Security wealth in an overall wealth measure generally reduces estimated levels of wealth inequality but it does not reverse the upward trend in top wealth shares. III.

Income and Wealth Mobility

Research indicates that inequality has a strong element of persistence across generations. Understanding the intergenerational transmission of inequality requires sorting out what factors reflect innate characteristics, which are the result of effort, and which are the result of actions by others. Families with large material resources may pass assets directly to subsequent generations through gifts or bequests. Financial investment in the human capital of children is another way of transmitting advantage. Relative advantage for children later in life may also stem from the nature of their home life. For example, a stable home, well-educated parents, or simply a caring and engaged parent may provide the support with which a person may more easily develop to their potential. Discrimination of many sorts is also an important factor. The five chapters in this section provide new evidence on the intergenerational patterns of inequality and the mechanisms that sustain those patterns. Connolly, Haeck, and Laliberté (“Parental Education and the Rising Transmission of Income between Generations”) investigate the causal link between the education of parents and the future income of their children. Using linked Canadian census data and intergenerationally linked tax return data, they show that income mobility has declined, especially for children of mothers without a high-school diploma. They claim that encouraging higher educational attainment among the young has the effect of increasing their earning potential as well as the prospects of their children. Mitnik, Helsø, and Bryant (“Inequality of Opportunity for Income in Denmark and the United States: A Comparison Based on Administrative Data”) use administrative data for Denmark and the US on the 1972– 73 birth cohort to study inequality of long-run income. Taking care to apply a coherent and consistent analytical framework to each country, they are able to characterize inequality in the two countries and bound key estimates of the extent to which observed inequality is a function of people’s initial conditions over which they have no control. Larrimore, Mortenson, and Splinter (“Presence and Persistence of Poverty in US Tax Data”) use linked US tax return data from 2007 to 2018 to study incidence and persistence of poverty among households since the Great Recession. Over 40 percent of the households were in poverty in at least one of those years. Although there is considerable mobility in and out of poverty, there is also substantial persistence, with about a third of those in poverty in 2007 being so in at least half of the years studied. The authors

8

Chetty, Friedman, Gornick, Johnson, and Kennickell

also find important age effects, with older people showing lower rates of poverty but relatively greater persistence, and younger people experiencing the opposite. Garbinti and Savignac (“Intergenerational Home Ownership in France over the Twentieth Century”) consider the correlation of housing tenancy across parents and children, using data from the French Wealth Survey. Their analysis shows that the intergenerational correlation of home ownership is increasing, as children of homeowners have a stable probability of ownership while children of nonowners have a declining probability of ownership. Although receipt of an inheritance or intergenerational transfers tends to be associated a higher level of ownership in general, the effect of parental home ownership remains strong. The authors suggest that their results may be explained by intergenerational correlations in income or preferences. Fisher and Johnson (“Inequality and Mobility over the Past Half-Century Using Income, Consumption, and Wealth”) use consumption, income, and wealth data from the Panel Study of Income Dynamics (PSID) from 1968 to 2017 to construct a multidimensional portrait of the inequality and mobility of individuals and families. They find that, while resources are increasing overall, inequality is also increasing and intragenerational mobility is falling or flat. They conclude that their study provides further evidence for the existence of the Great Gatsby Curve— the negative correlation between inequality and mobility. IV.

Mitigating Inequality

Most high-income countries have some policies in place that mitigate extreme inequality by providing income support, housing, food, or other resources. Such support has its most direct effect near the time it is delivered, but it may also have lasting effects, by helping people to avoid sinking into a state harder to escape, by providing a more stable environment for the longrun development of children, or by triggering other persistent behavioral or psychological reactions. To design effective interventions in the face of the harsh budgetary constraints, it is important to understand the nature of interventions and their short- and long-term effects. The four chapters in this section address variations in intervention strategies across time and geography, and assess the effects of diverse policies for supplementing the income of low-wage workers and low-income households. Meyer, Wu, Finley, Langetieg, Medalia, Payne, and Plumley (“The Accuracy of Tax Imputations: Estimating Tax Liabilities and Credits Using Linked Survey and Administrative Data”) use a data set linking a wide variety of US administrative sources with the Current Population Survey to construct a comprehensive picture of the distributional effects of transfer and tax-credit policies. The data linkage is especially important for capturing income sources missed in surveys and for addressing measurement error in

Introduction

9

survey variables. The chapter provides improved measures for the US of net redistribution and poverty reduction. Bruch, Gornick, and van der Naald (“Geographic Inequality in Social Provision: Variation across the US States”) assess the role of state governments, in the United States, in the design and provision of social policies, directing attention to the consequences of decentralization. Using a unique cross-state, over-time policy dataset, they examine the magnitude of crossstate variation in benefit generosity and program inclusiveness. They find substantial cross-state inequality states in social provision and conclude that this constitutes a meaningful form of inequality: inequality in the treatment of similar needs and claims by people who happen to live in different states. Feigenbaum, Fishback, and Grayson (“Inequality and the Safety Net in American Cities throughout the Income Distribution, 1929– 1940”) look at the period after the Great Depression in the US to examine the effects of that economic collapse and the programs of the New Deal on income inequality. To do so, they piece together micro data collected in a large number of cities by the Civil Works Administration and the decennial census. They conclude that inequality increased broadly, but that the shift was most notable in cities where per capita income fell the most. Among other results, they find that some New Deal Programs had a mitigating effect on inequality. Akee, Jones, and Simeonova (“The EITC and Linking Data for Examining Multigenerational Effects”) link US demographic micro data with time series data derived from individual income tax returns to study the effects on intergenerational mobility of the Earned Income Tax Credit (EITC), a refundable credit available to low-income workers first enacted in 1975. Using information on dependents on tax returns of workers claiming the EITC, the authors track outcomes for children who were exposed to differing intensities or durations of the EITC. Their findings suggest significant and mostly positive effects of more generous EITC refunds on the next generation; those effects vary substantially depending on the child’s gender and their household type. V. Distributional National Accounts For researchers and policymakers trying to use micro data in conjunction with more frequently available aggregate data, differences in the alignment of totals in the two sources have long been an obstacle. Conceptual differences are an important explanation and they are often quite difficult to address. Errors may also play a role, as survey respondents may not answer accurately or nonrandom nonresponse may skew the observed population, and/or projections or other estimates used in construction of aggregates may be inadequate or erroneous. Nonetheless, the benefits of being able to pair such data, especially in considering macroeconomic policy— and, increasingly, for inequality studies— have driven researchers to design strategies for

10

Chetty, Friedman, Gornick, Johnson, and Kennickell

achieving sufficient comparability. The final section in this volume includes five chapters focused mainly on creating ways of placing surveys on a comparable basis with national accounting data. Fixler, Gindelsky, and Johnson (“Distributing Personal Income: Trends over Time”) use publicly available micro data to construct a time series of the distribution of income as defined in the National Income and Product Accounts. Focusing on the period 2007– 16, they consider trends in growth and in inequality over this especially volatile period, including the Great Recession. They find that inequality changed little during the 2007– 16 period, aside from a slight increase derived from growth in the top quintile; that there was substantial change in the composition of personal income during the study years, with compensation decreasing as a share of income and transfers increasing; and that both mean and median real income increased during the period, with gains in every income quintile. Tonkin, White, Stoyanova, Youssef, Sidhu, and Payne (“Developing Indicators of Inequality and Poverty Consistent with National Accounts”) address the differences between survey measurements and national accounting measures of income for the UK. They note the importance of conceptual and coverage differences, but identify underreporting among survey households at the top of the income distribution as the largest source of discrepancies. Taking into account both conceptual differences and underreporting, they propose a method for adjusting survey measures to develop plausible indicators of inequality, poverty, and shared prosperity based on and consistent with national accounts. They also introduce the possibility of using a microsimulation approach to update survey measurements to support more frequent monitoring of distributional trends, given the most recent aggregate data. Bach, Bartels, and Neef (“Distributional National Accounts: A MacroMicro Approach to Inequality in Germany”) pursue a strategy for creating distributional national accounts following, with necessary adaptations, the approach of the World Inequality Database. They combine survey data, tax-based data, and national accounts data for Germany to bridge gaps in any one source alone, in order to create a consistent time series of income data, together with a variety of distributional, geographic, and demographic indicators. Batty, Bricker, Briggs, Friedman, Nemschoff, Nielsen, Sommer, and Volz (“The Distributional Financial Accounts of the United States”) describe the development of system of quarterly distributional accounts for wealth, blending data from the Federal Reserve Board’s triennial Survey of Consumer Finances (SCF) and a version of the quarterly Financial Accounts of the United States (FAOTUS) that includes nonprofits in service to the household sector (NPISH). A particular advantage of the SCF in this context is that it provides an implied value of aggregate wealth that is generally

Introduction

11

close to the FAOTUS estimates, most likely because the SCF has unusually good effective coverage of the top of the wealth distribution. There are, however, many differences in the two data sources at a more disaggregated level. The authors address the range of differences, and even develop a means of distributing FAOTUS values for estimates that are not directly collected in the SCF, such as the value of assets underlying defined benefit pensions due to households. Given the fully reconciled survey data, the authors develop a system for incorporating more frequently observed information in order to update the distributional characteristics in the survey. The resulting linkage provides policymakers with a timely basis for judging the effects of macroeconomic changes on households at a more detailed level. Finally, Webber, Tonkin, and Shine (“Using Tax Data to Better Capture Top Incomes in Official UK Income Inequality Statistics”) address the problem of random and nonrandom effective undercoverage of the top of the income distribution in surveys, with data from a sample of administrative records for taxpayers in the UK. Such differences greatly complicate the ability to use survey data to integrate survey information with data from national accounting systems. The authors investigate two methods: one using the administrative data to directly replace survey data on top values of gross income with values from equivalent quantile groups and the other reweighting the survey data according to the population observed in groups in the administrative data. They find that the reweighting method is preferable and that its use is most compelling for the top few percent of the income distribution. Papers Presented but not Included in This Volume Three additional papers were presented at the conference, but for a variety of reasons were not included in this volume. Because these papers, like those included, were selected to represent an important topic in inequality research, for each of those papers we provide a short description and an external link to a subsequent version of the paper. Meriküll, Kukk, and Rõõm (“What Explains the Gender Gap in Wealth? Evidence from Administrative Data”) are able to look at the patterns of wealth holdings at the individual level, thus allowing insight into gender differences within and across households.1 Using Estonian administrative data together with the Household Finance and Consumption Survey for Estonia, the authors find a very substantial unconditional gender wealth gap in favor of men, though much of the gap is driven by the top of the wealth 1. A published version of the paper by Jaanika Meriküll, Merika Kukk, and Tairi Rõõm, “What Explains the Gender Gap in Wealth? Evidence from Administrative Data,” may be found in the Review of Economics of the Household 19 (2021): 501– 47; https://doi.org /10.1007 /s11150 -020 -09522-x.

12

Chetty, Friedman, Gornick, Johnson, and Kennickell

distribution. In general, men tend to have somewhat more diversified assets than women and men are more likely to own personal businesses, one of the sources of large wealth disparity in Estonia. Guyton, Langetieg, Reck, Risch, and Zucman investigate the question of tax evasion through a variety of mechanisms such as offshore accounts, shell companies and trusts, as well as through financial engineering and other means.2 The paper combines random audit data with new data on offshore bank accounts to estimate the size and distribution of individual income tax evasion in the US. They find that evasion through offshore financial institutions is highly concentrated at the very top of the income distribution. Thus, measures of income inequality based on typically observed sources are likely to be biased downward. Asher, Novosad, and Rafkin focus on educational mobility across generations, as a proxy for income mobility, which is substantially more difficult to observe clearly in many countries.3 They develop a methodology allowing for the use of coarsely binned education data, which they apply to the US and India to make estimates of educational mobility for subgroups. Directions for Additional Work Throughout the past decade, in the United States and abroad, there has been an explosion of interest in high and rising economic inequality. A broad national and international conversation has developed, one that has included academics, journalists, policymakers, political figures, NGOs, and general publics. The global financial crisis of 2007– 9, and the Occupy movements that unfolded shortly after, provided crucial sparks. Since then, this intensified interest has driven— and has been driven by— methodological advances, new research institutes, enlarged data options, expanded media coverage, and a mountain of scholarship. Inequality had, in fact, been studied in select corners of academia for decades— but the current level of interest is of a different order. Our hope is that this volume will make a notable contribution to this rapidly growing field. The 23 chapters in this volume have covered extensive ground— cross-cutting income and wealth, as well as poverty, inequality, and mobility. The studies included here address policy impacts, geographic variation, change over time, and a multitude of issues related to data, measures, and methods. Yet, 2. A revised version of the paper by John Guyton, Patrick Langetieg, Daniel Reck, Max Risch, and Gabriel Zucman, “Tax Evasion at the Top of the Income Distribution: Theory and Evidence,” is available as NBER Working Paper No. 28542, at http://nber.org /papers/w28542. 3. A paper by the authors, Sam Asher, Paul Novosad, and Charlie Rafkin, that addresses the methodology in this presentation, “Intergenerational Mobility in India: New Methods and Estimates across Time, Space, and Communities,” is available at http://paulnovosad.com /pdf /anr-india-mobility.pdf.

Introduction

13

as always, for each research question addressed here, countless more come to mind. In this brief section, we raise some potential areas for future work. We first turn our attention to possible directions for extending the substance covered. We envision future lines of work aimed at assessing the effects of structural changes, disaggregating national populations, and expanding country coverage with respect to both geography and economic development. We close with some remarks about future directions with regard to data, measures, and methods. Substantive Extensions An array of structural changes seems likely to be an important factor in shaping and sustaining patterns of inequality and mobility, and more work on them would be welcome. For several decades, the bargaining power of labor has declined in the US and elsewhere. In the US, labor union membership has decreased and, in real terms, the federal minimum wage peaked before 1970. At the same time, the composition of occupations shifted more in the direction of service jobs. Technology and offshoring eliminated many types of jobs while creating others. The industrial structure has shifted as well, including the emergence of some entirely new industries. In recent years, “gig” work has become more common, appearing similar in some ways to patterns of self-employment in less developed countries. The effects of all of these shifts call out for further research. Increasingly, scholars and practitioners engaged with economic inequalities have called for further disaggregation of populations. The United Nations Sustainable Development Goals (SDGs), adopted in 2015, emphasize moving “beyond averages.” SDG Goal 10, reducing inequality, calls for promoting inclusion “irrespective of age, sex, disability, race, ethnicity, origin, religion or economic or other status.” Other supranational organizations have followed suit. Several of the chapters in this volume include analyses of intergroup disparities— mostly comparing age groups or cohorts, and, in some cases, disaggregating by gender, family structure, or level of educational attainment. Much more work is needed to assess how earnings, income, and wealth— levels, trends, and mobility (both “intra and inter”)— vary across other crucial axes of disparity, including race, ethnicity, religion, citizenship, sexual orientation, disability status, and urbanicity. A large share of research on income and wealth inequality, and mobility, focuses on the US or on other high-income countries. Many cross-national studies— including those in this volume— include groups of relatively homogeneous countries. That homogeneity is understandable; cross-national variation is more easily interpreted when national/economic contexts are reasonably similar, and many sources of high-quality data are available only for one or more high-income countries. Research on economic inequality

14

Chetty, Friedman, Gornick, Johnson, and Kennickell

and intergenerational mobility in middle- and low-income countries has, of course, been carried out— much of it by development economists— and a growing literature assesses global inequality, where the unit of analysis is the whole world. Still, among inequality scholars, silos persist, with one set of scholars/institutions mainly addressing (essentially) the high-income Global North and another, the lower-income Global South. We urge scholars of poverty, income and wealth inequality, and mobility to bridge these geographic and economic divides to more fully assess the extent to which lessons learned in rich/northern countries do, and do not, apply in less affluent countries or regions— and vice versa. Data, Measures, Methods Several chapters in this volume highlight the value of linking various sources of data, especially administrative data, in order to increase the accuracy and power of analysis. More progress is needed in this area, especially in linking various government data sources, where agency-specific rules and differing views of their mandates may be inhibiting. Among the more promising signs of progress in this area, in the US, is work under the Foundations for Evidence-Based Policymaking Act and the Federal Data Strategy, aimed at making more federal data available for research purposes and exploring potential structural changes, such as a US National Secure Data Service, as envisioned by the Congressionally Chartered Commission on EvidenceBased Policymaking. Our hope is that inequality scholars, including the authors in this volume, will engage in efforts to create new sources of linked data, to raise the availability of these linked data, and to aim for widespread and equitable data access. Surveys, and the challenges that they face, demand continued attention. Survey data remain an important source for studying inequality but, as noted earlier, data providers face serious challenges related to cost and data quality. To support the collection and dissemination of survey data and to anticipate future difficulties, urgent attention should be given to developing linkages between survey data and other types of data, and to improving tools for measuring the impact of nonlinkages and incorrect linkages on inferences. In the short run, more linkage would facilitate new lines of research and would allow potential improvements in data editing and nonresponse adjustment. Over the longer run, linkage of survey data on wealth with contemporaneous income data would allow a more detailed evaluation of models used to project wealth information from income data and other sources. Linkage with multiple years of nonsurvey data might support simulation of wealth beyond the survey year, as well as research into other questions that require panel data to place the survey data in context. The US Survey of Consumer Finances, which employs data based on individuals’ income tax in its sample design, is a natural candidate for such work. Our overarching

Introduction

15

hope is that diverse scholars and practitioners will commit to supporting the production, improvement, expansion, and analysis of survey data in new and innovative ways. Despite the well-known flaws of survey data, research on inequality and mobility would suffer immeasurably if the volume, quality, and/or accessibility of survey data were to decline substantially.

I

Income Inequality

1

In Search of the Roots of American Inequality Exceptionalism An Analysis Based on Luxembourg Income Study (LIS) Data Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

1.1 1.1.1

Introduction Background

It has been known for at least two decades that disposable household income— income after accounting for transfers and taxes— is more unequally distributed in the United States than in comparable high-income economies (see, e.g., Brandolini and Smeeding 2006; Gornick and Jäntti 2013; OECD 2009, 2011; Piketty and Saez 2006). Broadly speaking, there are two possible underlying explanations. First, market income inequality (i.e., income before direct taxes and transfers are taken into account) may be similar in the US as elsewhere, but US taxes and transfers are less redistributive, either because the overall size of the welfare state is smaller or because the redistribution is less progressive. Second, market income inequality may itself be higher in the US than in many other countries, thus driving up the high level of inequality even after redistribution is taken into account. The first explanation has generally held sway because US market income inequality calculated across households— importantly, households of all ages— Janet C. Gornick is a professor of political science and sociology, director of the Stone Center on Socio-Economic Inequality, and the James M. and Cathleen D. Stone Distinguished Chair in Socio-Economic Inequality, at the Graduate Center at the City University of New York. Branko Milanovic is a senior scholar at the Stone Center on Socio-economic Inequality at the Graduate Center at the City University of New York. Nathaniel Johnson is a data scientist at Amenity Analytics. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber.org/books-and-chapters/measuring-distribution -and-mobility-income-and-wealth /search-roots-american-inequality-exceptionalism-analysis -based-luxembourg-income-study-lis-data.

19

20

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

is not especially exceptional, across the OECD countries, while disposable income inequality is substantially greater. Recent work, however, by Gornick and Milanovic (2015) shifts that conclusion about the market income inequality in the US, in comparative perspective. They begin with the insight that market income inequality, when calculated across households of all ages, may be depressed— especially relative to many European countries— because Americans tend to stay in the labor market until later in life compared with their counterparts elsewhere. Because the market income in pensioners’ households is often very small or zero, the existence of a developed system of social protection paradoxically exaggerates market income inequality (among older households) in other OECD countries and brings the overall market income inequality in line with that reported in the US. Thus, the comparatively high level of US market income inequality— net of older households— is obscured. Gornick and Milanovic’s main conclusion is that, for persons under 60 years of age, weaker US redistribution is not the main cause of greater inequality at the disposable income stage. The “problem” is that the distribution of “original” labor and capital incomes is substantially more unequal in the US than elsewhere, and government redistribution, at the average OECD level, does not compensate for the inequality generated in the market. Gornick and Milanovic’s (2015) analysis had precursors in the work of scholars of earnings distributions, who argued that weaker redistribution in the US could not alone explain the entire disposable income inequality gap between the US and the rest of the OECD countries. Mishel (2015), for example, argues that the underlying market income distribution, most importantly the earnings distribution, in the US, is highly unequal in crossnational terms. He and others point to, on the bottom end of the earnings distribution, the low US minimum wage and the high prevalence of low-paid jobs (Gautié and Schmitt 2009; Lucifora and Salverda 2009), and, on the upper end, the extremely high earnings of managers, doctors, lawyers, CEOs and the financial sector (Gabaix and Landler 2008). The exceptionally large gap between CEOs’ salaries in the US and in the rest of OECD countries is well documented (see Mishel and Davies 2015; Piketty 2014). Indeed, the findings in Gornick and Milanovic (2015) confirm that market income inequality is a major explanation for comparatively high levels of disposable income inequality in the US, among working-age households. 1.1.2

Objective

The objective of this chapter is to further investigate the nature of the high level of market income inequality found among US working-age households, compared to their counterparts in several other affluent countries. Because the major component of market income is labor income, we focus exclusively on it— disregarding income from capital, which is a relatively

In Search of the Roots of American Inequality Exceptionalism

21

minor component in the market income package of working-age households in these countries.1 Our main strategy is to disaggregate working-age households— in the US and in the comparison countries— into household subgroups. These subgroups are distinguished by the number and gender of earners in the household, and (subsequently) by the partnership and parenting status of the household. Clearly, a household’s labor income is shaped by the number of earners present. The logic of further disaggregating by gender, partnership, and parenting is rooted in the labor economics literature, which has long established that individuals’ earnings (gross and net of other workerand job-level characteristics) are affected by their gender and whether they have partners and/or children (for a review, see Blau and Winkler 2017). We assess inequality that exists both within and between various household types and we compare the results for the US with those in other OECD countries. Our objective is to establish whether the greater underlying US market income inequality is the result of (1) higher earnings inequality within each of the relevant groups, (2) an unusual composition (for example, a high share of groups where earnings inequality is high), or (3) large gaps between groups in mean earnings.2 A substantial prior literature on economic inequality in the US addresses the question of the levels, and/or drivers, of within-group versus betweengroup inequality. Much of this literature focuses on earnings, and most of it locates the question of within-versus-between in the context of change over time. Two decades ago, McCall (2000) observed that most research on (earnings) inequality in the US was concerned with growing gaps between groups— with workers differentiated by race, age, education, and income. She noted that, in fact, a large share of rising inequality had occurred within these groups. Her own study assessed variation in within-group inequality across 500 local labor markets. Western, Bloome, and Percheski (2008) assessed rising income inequality among US families, between 1975 and 2005. They concluded that most of the increase in family income inequality during that period was driven by rising within-group inequality; their disaggregation combined family type and educational attainment. Introducing his own study of the drivers of within-group inequality between 1970 and 2001, VanHeuvelen (2018, 1– 65) summarized the literature as follows: “An increasing number of studies have begun to note that within-group inequality— or the inequality that remains after accounting for average between-group pay differences . . . such as human capital, 1. Among the working-age population, and in the countries included here, income from labor accounts for 97 percent of total market income, on average. In no country is the labor income share of market income less than 93 percent. 2. In this chapter, we use the terms “labor income,” “earnings,” and “wages” interchangeably.

22

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

occupational characteristics, sex, socio-demographics, and household composition— is of growing importance for overall inequality.” While these and other earlier studies influenced our analytic strategy, our study is a departure. First, no earlier research disaggregates household types as we do. Our typology includes unusually finely drawn categories; our groups are defined by the number and gender of a household’s earners, further disaggregated by partnership and parenting status. Second, we depart from earlier research with respect to our income measure and unit of analysis. Most existing within-versus-between research assesses either earnings at the individual level, or posttax, posttransfer income at the household (or family) level. In contrast, we focus on earnings (what we call market income) at the household level. Our framework allows us to place our work in the large cross-national literature, much of it using the same data that we use, concerned with the extent to which inequality in disposable household income is driven by inequality in household-level market income. 1.1.3

Analytic Strategy

To carry out our analyses, we use microdata, drawn from household surveys, contained in the Luxembourg Income Study (LIS) Database Wave VIII, which is centered on the year 2010.3 We include 24 OECD countries:4 Australia, Canada, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Israel, Italy, Luxembourg, Netherlands, Norway, Poland, Russia, Slovakia, Slovenia, Spain, the UK, and the US.5 In all cases, but one, the data are from the year 2010; the exception is Hungary, for which we have 2009 data. Appendix table 1A.1 reports the list of countries and datasets used. Our analysis is conducted across households whose members are all below age 60 and which have at least one member reporting labor income. To assess labor income, we use LIS’s harmonized variable hil (that is, household income from labor). This variable includes: (1) cash wage and salary income, and the value of nonmonetary goods and services received as a substitute for cash; (2) monetary supplements to the basic wage and the value of nonmonetary supplements; (3) cash wage and salary income, and the value of nonmonetary goods and services, received by directors of their own enterprise; (4) monetary payments and the value of nonmonetary goods and services 3. This means that the datasets report income earned in the year 2010; the surveys may have been fielded in the subsequent year. 4. Russia is not officially an OECD member state, but a “roadmap to accession” has been approved. For convenience, when we use the term “OECD countries” in this chapter, we include Russia. 5. The LIS data are available from LIS, the cross-national data center in Luxembourg. Extensive documentation is available on the website: www.lisdatacenter.org (multiple countries; microdata runs carried out April 2017 to December 2019).

In Search of the Roots of American Inequality Exceptionalism

23

Fig. 1.1 Inequality of labor income across working-age households, in 24 OECD countries Note: Ginis based on equivalized labor income.

received from casual, irregular, or occasional dependent employment; and (5) profits/losses from self-employment activities. Because one of our motivating interests is the relationship, at the household-level, between earnings inequality and disposable income inequality, our unit of observation is not an individual worker (earner) but the household. Total household earnings are adjusted for household size, using the well-known “square-root adjustment.” In other words, total household earnings are divided by the square root of the number of household members.6 Thus, we arrive at a variable that measures potential individual welfare (assuming equal division of earnings within the households) derived from labor income. As our measure of inequality, we mainly use the Gini coefficient. The Gini is preferred largely because it enables us to easily relate our results about inequality within different demographic subgroups to the well-known Gini values of market and disposable income inequality seen in the US and elsewhere. In one part of our analysis, we use two Theil indices. 1.2

Labor Income Inequality across Various Household Types

In figure 1.1, we report inequality, across households, of labor incomes. The four countries with the most unequal earnings distributions (at the 6. This assumes economies of scale midway between perfect economies of scale (parameter = 0) and no economies of scale (parameter = 1).

24

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

household level) are Israel and three Anglophone countries; the US is ranked second highest. These labor income Ginis range from 0.277– 0.311 for the highly egalitarian Slovenia and Slovakia, respectively, to 0.436– 0.442 in the US and Israel. The median and mean labor income Gini is about 0.36. Thus, we establish immediately that labor income inequality in the US is, relative to other OECD countries, on the high end. What lies behind this comparatively high level of earnings inequality among US households? Our approach, as already mentioned, is to disaggregate working-age households into several demographic groups (defined below) and to assess labor income inequality within each of them. The Gini decomposition when the population is divided into different groups has three terms: a weighted sum of within-group inequalities (narrowly defined within-inequality), inequality that is the result of differences in mean incomes between the groups, and an overlap (residual) term that reflects the homogeneity of the underlying populations. To understand the meaning of the last, note that when incomes of the groups into which we have divided the population are so different that there is absolutely no overlap (e.g., all individuals from a mean-richer group have higher incomes than all individuals from a mean-poorer group), the overlap term becomes zero. It increases as there is more overlap between the incomes of individuals belonging to different groups. The overlap terms move together with the narrowly defined within-inequality, and we shall treat them together. We can write the Gini decomposition across recipients belonging to groups i (1, 2,. . . r) as (1.1)

1 G= μ

r

r

r

( yj i=1 j>i

yi ) pi p j +

pi siGi + L, i=1

where μ = overall mean income, yi = mean income of i-th group, pi = population share of i-th group, si = share of i-th group in total income, Gi = Gini of i-th group, and L = the overlap term. The first term in equation (1.1) is the between-group inequality; the second term, the narrowly defined withingroup inequality; the third, the overlap term. The second and third terms are in the further text considered as “within-group inequality.” We can now see that higher overall US labor income Gini (G) may be the result of greater group Ginis (Gi), or greater share (si ) of groups that have higher inequality of earnings, or finally, may be due to large mean income gaps between the groups (that is, to the between-component). 1.2.1

Disaggregating into Household Types, Based on the Number and Gender of Earners

In all countries, we first divide the population into six main groups, based on the number and the gender of the earners in these households: households that contain (1) one female earner, (2) one male earner, (3) one male and one female earner, (4) two female earners, (5) two male earners and,

In Search of the Roots of American Inequality Exceptionalism

25

Fig. 1.2 Typology of households based on number and gender of earners, further disaggregated by demographic groups based on partnership and parenting status Note: The six main types of households are indicated by numbers 1–6.

finally, (6) three or more earners. Later in the chapter, groups (1), (2), and (3) will be further subdivided into demographic groups, based on partnership and parenting status. Throughout this chapter, results are presented at the person level— albeit drawing on their household characteristics. When we refer to various household types, either their prevalence or their outcomes, we are reporting results about the persons who live in those household types. Figure 1.2 summarizes our typology of households. Earners are defined as people who report having received nonzero labor income during the survey’s reference period. Table 1.1 reports the composition of the working-age population, across the six household types, in these study countries.7 As can be expected, three household types dominate to the extent that they include more than 80 percent of all persons in all countries— except for Hungary, Ireland, and Russia.8 The three dominant groups are: the “traditional”9 two-earner households composed of one female and one male earner (with a cross-country average share of 46 percent), one-male-earner households with an average share of 21 percent, and households with three or more earners, with 16 percent. The other three groups are less prevalent, although households with only one female earner (cross-country average share of 12 percent) do play, as we shall see below, an important role. 7. It should be kept in mind that the typology presented in table 1.1 takes no account of partnership status. For example, in households with a one female earner, those female earners may or may not have partners. Later in the chapter, we integrate partnership and parenting status. 8. In all three countries, the reason is a relatively high presence of one-female-earner households (17– 18 percent). 9. When referring to two-earner households, we use the term “traditional” to denote that one of these earners is male is one is female (as opposed to two earners of the same gender).

26

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

Table 1.1

Country/group Australia Canada Czech Republic Denmark Estonia Finland France Germany Greece Hungary Iceland Ireland Israel Italy Luxembourg Netherlands Norway Poland Russia Slovakia Slovenia Spain United Kingdom United States Unweighted means

Composition of working-age population, across six main household types (where household types are based on the number and gender of earners) 1 female earner 1

1 male earner 2

1 male, 1 female earner 3

2 female earners 4

2 male earners 5

3+ earners 6

Sum of columns 2+3+6 7

9.2 9.5 8.6 11.8 16.0 12.0 14.7 14.0 8.2 17.6 10.1 18.2 10.7 10.1 10.7 9.3 12.0 14.0 16.9 8.3 9.3 10.8 13.2 14.8 12.1

21.6 14.7 23.3 13.4 20.3 15.5 19.7 19.7 30.9 24.7 11.1 23.6 24.1 34.0 25.0 15.6 15.0 28.7 17.3 14.4 15.8 25.7 21.2 22.1 20.7

39.7 43.5 47.7 47.8 47.4 53.1 55.8 48.6 48.6 39.6 45.3 41.1 40.8 44.8 51.5 51.7 48.3 42.3 39.6 43.4 50.6 46.6 46.6 42.2 46.1

2.3 2.5 1.5 2.1 2.7 1.4 1.1 1.0 0.9 1.6 2.1 2.2 1.9 0.8 0.7 1.3 1.4 1.5 2.9 1.4 1.4 1.5 1.8 2.3 1.7

3.6 3.2 2.2 2.1 1.3 0.7 1.4 1.7 2.3 0.7 1.0 3.9 3.1 4.0 2.3 2.2 1.5 3.3 2.6 1.9 1.9 2.9 2.2 3.0 2.3

22.7 25.5 16.8 22.1 12.3 17.2 6.8 15.0 7.3 9.1 30.4 11.0 19.2 6.3 9.7 18.8 20.2 10.2 20.7 30.5 21.1 10.0 14.7 15.3 16.4

83.9 83.8 87.7 83.3 79.9 85.8 82.3 83.3 86.8 73.4 86.8 75.7 84.1 85.1 86.2 86.1 83.5 81.2 77.6 88.3 87.4 82.3 82.5 79.6 83.2

In figure 1.3, we take a first look at US labor income inequality within each of these household types in comparative context. For each type, the figure indicates the position of the US Gini (in black) compared to the other 23 countries. For three household types (one-male-earner, one male and one female earner, and two male earners), the US has the most unequal distribution of all countries; for the other three household types, the US distribution is the second most unequal.10 In no case, is the US Gini even close to the median Gini for a given household type, much less lower than it. Therefore, breaking the overall labor earnings distribution into household 10. Note that the Ginis of these various household types differ substantially in countries considered here. Labor income inequality among “traditional” two-earner households is within a rather narrow range between 0.22 and 0.36 whereas, for example, one-female-earner and one-male-earner households display much greater ranges of inequality. However, this is not the topic with which we are concerned here. Our objective is to find the sources of differences between the US and comparable countries.

In Search of the Roots of American Inequality Exceptionalism

27

Fig. 1.3 Inequality in six main household types (where household types are based on the number and gender of earners) Notes: Each bar shows the Gini of a given group and country. The US Gini is black. Ginis are ordered from the highest to the lowest.

Fig. 1.4

Relative income of six main household types

Notes: Each bar shows mean income of a group compared to the mean income of the country. The US values are black. Values are ordered from the highest to the lowest.

types reinforces our previous finding: US labor income is very unequally distributed, not only in the aggregate, but within each household type. We need to also look at between-group inequality (that is, between the six household types). Consider now figure 1.4, which is constructed similarly to figure 1.3 but where we look at relative earning levels of household types. For

28

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

example, the one-female-earner households’ mean earnings11 are in relative terms the lowest in Israel (only 45 percent of the country mean) and the highest in Hungary (75 percent of the country mean). The US at 54 percent is somewhat below the median for OECD countries included here. A look at figure 1.4 shows that the US position, with the exception of the two-female-earner households (relatively low) and the one-male-earner households (relatively high) is not exceptional. In other words, when it comes to the relative earnings of various demographic groups, the US is far from being a cross-national outlier: groups’ relative earning levels track closely other high-income countries’ relative earnings levels. This, in turn, implies that the origin of high labor income inequality in the US is not to be found in unusually high earnings of some demographic groups and unusually low earnings of others, but in systematically high earnings inequalities within each individual household type. We confirm this conclusion by looking at the results of the decomposition exercise using equation (1.1). Each country’s overall inequality is broken into between- and within-inequalities (vis-à-vis the six groups). The US within-inequality Gini (shown in table 1.2, column 3) is 0.311. This means that if the mean earnings of the six household types were exactly equal, the overall labor income inequality would be 0.311, which is by far the highest value among the countries considered here. Canada and Luxembourg have the second highest within-inequality, with a Gini of 0.282, some 10 percent lower than the US. When we look at the between-inequality, however, the US is far from exceptional. Although the within-inequality of the US is 34 percent higher than the mean of the other 23 countries, the between-inequality is practically the same as the mean for other countries. Finally, we can assess this from another vantage point by using the Theil index instead of the Gini. The advantage of the Theil, in this particular case, is that it is exactly decomposable between different components. Table 1.3 reports the results of two Theil decompositions for the US case. The first column presents the Theil T— or the GE(1)— where the weights are income shares. The second column presents the Theil L, or the GE(0)— the mean log deviation— where the weights are population shares. When we assume that the US has both the same demographic structure and the same relative group incomes as the average of the other 23 OECD countries, the Theil index, in its two variants, is reduced by either 3 or 6 percent. The changes seem minimal and reinforce our view that the dominant factor explaining high market income inequality in the US is high inequality within each demographic group.12 11. Note that this is household-size-adjusted (equivalent) labor income. 12. The two Theil indexes, because of their different weighting structures, give different answers as to the relative importance of demographics versus relative group incomes. According to Theil L, US demographic structure (in the sense that it is different from the OECD average) contributes more to high US inequality. According to Theil T, the divergence of US relative group incomes from the OECD average pattern is more important.

Table 1.2

Australia Canada Czech Republic Denmark Estonia Finland France Germany Greece Hungary Iceland Ireland Israel Italy Luxembourg Netherlands Norway Poland Russia Slovakia Slovenia Spain United Kingdom United States Non-US mean US/non-US mean

Decomposition: Between-group and within-group components (for six household types) Overall labor Gini (1)

Between Gini component (2)

Within Gini component (3)

0.357 0.394 0.323 0.323 0.368 0.335 0.365 0.363 0.365 0.394 0.330 0.430 0.442 0.320 0.366 0.336 0.337 0.358 0.368 0.311 0.277 0.366 0.400 0.436 0.358 1.21

0.119 0.112 0.129 0.112 0.124 0.103 0.114 0.109 0.127 0.149 0.127 0.186 0.184 0.149 0.084 0.100 0.119 0.135 0.156 0.136 0.128 0.136 0.124 0.125 0.129 0.97

0.238 0.282 0.193 0.211 0.245 0.232 0.251 0.254 0.238 0.245 0.202 0.243 0.258 0.171 0.282 0.236 0.218 0.223 0.212 0.175 0.149 0.230 0.277 0.311 0.229 1.34

Note: Within-inequality includes the narrowly defined within-inequality and the overlap component; see equation (1.1). Table 1.3

Theil counterfactual: US inequality with OECD average demographic structure and relative mean group incomes

Actual US inequality US inequality if demographic structure were as OECD average (change) US inequality if relative group incomes were as OECD average (change) US inequality if both demographic structure and relative group incomes were as OECD average (change)

Theil T—GE(1)

Theil L—GE(0)

0.342 0.364 (+6%) 0.312 (−9%) 0.333 (−3%)

0.380 0.334 (−12%) 0.334 (+6%) 0.360 (−6%)

30

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

We have thus established that US labor income inequality is, together with Israel’s, the highest among all of the OECD countries included here and that the source of that inequality is not to be found in vastly different mean labor incomes across different household types, but in the consistently higher inequality with which labor incomes are distributed within each household type. We now continue by looking in greater detail into three prevalent household types: one-female-earner households, one-male-earner households, and two-earner “traditional” households (which contain one female and one male earner). 1.3

1.3.1

Earnings Inequality within One-Earner and “Traditional” Households: Further Disaggregation by Partnership and Parenting Status One-Female-Earner Households

We begin by looking at households that contain only one earner— one who is female. The prevalence of these households across the countries included here is very uneven: at the low end are Greece, Slovakia, and the Czech Republic where fewer than 9 percent of households contain only one earner, who is female. At the other end are Estonia and (as mentioned earlier) Hungary, Ireland, and Russia, which each contain more than 16 percent of households of this type. The US falls in the upper range, with the share of one-female-earner households being 15 percent. In our next analysis, we divide one-female-earner households into five demographic subgroups, corresponding to the households in which they live: couple-headed households with one or more children, couple-headed households without children, single-headed13 households with children, single-headed household without children, and others.14 As we did before for all households, here we look first at inequality levels within each household type and then at the relative incomes of each type. The most common type among one-female-earner households in the US, and across these 24 countries, is a household headed by a single woman with children. The next most prevalent types are couple-headed households with children (where, by definition, a female is the only earner) and single-female-headed households 13. We use the word “single” to mean, exclusively, a person who is not married/partnered. We do not use it to refer to the number of earners or persons in a household. 14. Throughout the chapter, households are defined as “coupled” if the head reports having a partner in the household and there are no other adults in the household. Households are further coded as having “children” if they contain children (under age 18) who are the children of the household head. Households are classified as “other” if the household—with or without children— contains adults who are not the head or the head’s partner (for example, the head’s parent or sibling, or a roommate).

In Search of the Roots of American Inequality Exceptionalism

31

Fig. 1.5. Inequality of five subgroups among one-female-earner households Notes: Each bar shows the Gini of a given group and country. The US Ginis are black. Ginis are ordered from the highest to the lowest.

without children. In the US, these three household types comprise over 80 percent of one-female-earner households. But is the distribution of labor income in such American households more unequal than in the other countries? Figure 1.5, with the same interpretation as figure 1.3, provides an answer. In all cases, US inequality is greater than the median inequality among 24 countries, and is always ranked either the fourth or the fifth from the top. Particularly interesting is the situation of single-headed one-female-earner households with children, where the US Gini is (a high) 0.48 while the mean Gini for this type of household, is 0.40. Very high inequality among single-headed one-female-earner households, both with and without children, in the US clearly implies that they are economically and socially diverse. We shall find similar high heterogeneity among single one-male-earner households without children. Next, we look at relative incomes (see figure 1.6). The situation here is familiar: US subgroup mean relative incomes are not dissimilar to the median relative incomes across the 24 countries. The differences are minimal (e.g., for a couple with a child, the average labor income is 41 percent of US overall mean vs. 45 percent across the 24 countries). The exception is the low income level of one-female-earner households with children (that is, single mothers): their relative income in the US is 40 percent of the overall mean while the countries’ average is 50 percent. An ethnic/racial component may be important here, as we find (not reported here) that these households, when headed by Hispanics and African Americans, have mean labor incomes that are only about 30 percent of the overall US mean.

32

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

Fig. 1.6

Relative income of five subgroups among one-female-earner households

Notes: Each bar shows mean income of a subgroup compared to the mean income of the country. The US values are black. Values are ordered from the highest to the lowest.

1.3.2

One-Male-Earner Households

We now move to one-male-earner households, where we keep the same household classification as for one-female-earner households. The prevalence of these households varies markedly across countries. At the low end, in Iceland, Denmark, Canada, and Slovakia, their share is less than 15 percent. But at the high end, Italy and Greece— with comparatively low levels of female employment— have more than 30 percent of one-male-earner households. The US result (22 percent) falls near the cross-national mean (21 percent).15 The results for inequality are familiar (see figure 1.7): US households have a much greater labor income inequality than the rest of the countries, and for two groups in particular (couple-headed households with and without children) US inequality is the highest of all. But it is among the highest in the other three types of one-male-earner households as well. Figure 1.8 shows the results for the relative income of single one-maleearner households. In three out of five types here, US relative mean income is around the cross-country median. The exceptions are one-male-earner households (couples with or without children) whose relative income is among the highest. These two groups are interesting because they display unusually high relative mean incomes with similarly unusually high inequality. 15. Note that the share of one-female-earner households across these OECD countries ranges from 8 to 18 percent. The share of one-male-earner households varies from 11 to 31 percent. The corresponding US values are 15 and 22 percent. Thus, neither US value is exceptional.

In Search of the Roots of American Inequality Exceptionalism

Fig. 1.7

33

Inequality of five subgroups among one-male-earner households

Notes: Each bar shows the Gini of a given group and country. The US Ginis are black. Ginis are ordered from the highest to the lowest.

Fig. 1.8

Relative income of five subgroups among one-male-earner households

Notes: Each bar shows mean income of a group compared to the mean income of the country. The US values are black. Values are ordered from the highest to the lowest.

1.3.3

“Traditional” Households

“Traditional” (one male earner and one female earner) households comprise the largest share of all households, from just under 40 percent in Australia, Hungary, and Russia to 56 percent in France. (The US with 42 percent is on the low side, modestly below the unweighted mean of 46 percent).

34

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

Fig. 1.9

Inequality of two subgroups of “traditional” households

Notes: Each bar shows the Gini of a given group and country. The US Ginis are black. Ginis are ordered from the highest to the lowest.

Here, we look at only two subgroups: “traditional” households with and without children. US inequality is again very high (see figure 1.9). US inequality is the highest of all countries, among these two-earner couples with children— with a Gini of 0.37 compared to the cross-country median Gini of just less than 0.30. US inequality is second highest, among two-earner couples without children. When it comes to relative incomes (see figure 1.10), US relative labor income for two-earner households with children is very close to the median for the 24 countries; it is higher than the cross-country median, however, for two-earner couples without children. 1.3.4

Regression Analysis

To tease out the specificity of US inequality, we estimate regressions where the Gini coefficient for each country/group is regressed on groups’ relative mean income (i.e., relative to the mean of that country) and dummy variables for the subgroups (N = 15) and countries (N = 24). The omitted household type is one-male-one-female-earner with children and the omitted country is Denmark (with low inequality). We use two specifications of the regression: an unweighted one, and a weighted regression where each group is weighted by its share in the population of a given country. The latter adjusts for variation in household compositions across countries. We are, of course, interested in the coefficient on the dummy variable for the US. The results are reported in table 1.4. Compared to the omitted country (Denmark), the coefficient on the US dummy is 0.069 in the unweighted formulation, and 0.101 in the weighted

In Search of the Roots of American Inequality Exceptionalism

Fig. 1.10

35

Relative income of two subgroups of “traditional” households

Notes: Each bar shows mean income of a group compared to the mean income of the country. The US values are black. Values are ordered from the highest to the lowest.

formulation. It is statistically significant at less than 0.01 in both cases. This means that, on average (whatever demographic group we take), US inequality is between 6.9 and 10.1 Gini points greater than Denmark’s. Perhaps more revealing is the fact that in both formulations, the US coefficient is the largest of all country dummies. The next largest positive coefficient in the unweighted formulation is Canada’s (5.4 Gini points more unequal than Denmark) and, in the weighted formulation, Israel’s (8.2 Gini points more unequal than Denmark). So, in terms of within-group inequalities, the US is, on average, more unequal than the second most unequal OECD country by between 1.5 and 1.9 Gini points. 1.3.5

Robustness of the Results

There are two possible limitations of our results that need to be addressed. The first refers to the composition of the population (i.e., shares of different demographic groups); the second to the year of study (2010) selected here. Consider group composition first. Earlier in this chapter, we noted that the higher overall labor income Gini in the US, compared to other relatively similar countries, could be the result of (1) greater group Ginis (the “within” component); (2) larger mean income gaps between the groups (the “between” component); and/or (3) greater shares of groups that have higher level of inequality. Throughout this chapter, we formally assessed the contributions of the first two of these three factors— the “within” and “between” components of inequality— but we did not present a detailed look at the third. The regres-

Table 1.4

US income inequality exceptionalism (dependent variable: Gini coefficient of household type/country) Coefficient ( p-value) * = significance < 0.05 ** = significance < 0.01 Unweighted regression

Variable

Population-share weighted regression

Relative group mean

−0.036 (0.20)

−0.003 (0.89)

Three or more earners

−0.028 (0.09)

−0.034** (0.00)

Two earners

Female Male

One female earner

Couple with children Couple without children Other Single with children Single without children

One male earner

Couple with children Couple without children Other Single with children Single without children

One male one female earner US dummy Adjusted R-squared (F) Number of observations

Couple without children

0.022 (0.23) 0.034* (0.04)

0.035* (0.02) 0.033** (0.01)

0.099** (0.00) 0.048* (0.03) 0.057* (0.03) 0.082** (0.00) 0.066** (0.00)

0.136** (0.00) 0.065** (0.00) 0.081** (0.00) 0.098** (0.00) 0.078** (0.00)

0.089** (0.00) 0.054** (0.00) 0.049* (0.05) 0.086** (0.00) 0.087** (0.00)

0.097** (0.00) 0.067** (0.00) 0.074** (0.00) 0.117** (0.00) 0.094** (0.00)

0.104 (0.54)

0.002 (0.80)

0.069** (0.00)

0.101** (0.00)

0.59 (12.3) 360

0.82 (38.9) 360

Note: The regression is based on 360 observations, i.e., 24 countries × 15 subgroups. The omitted household type is one male, one female earner with children, and the omitted country is Denmark. Coefficients on dummy variables for countries other than the US are not shown.

In Search of the Roots of American Inequality Exceptionalism Table 1.5

37

Population shares of household types

Type of household One female earner Couple with children Couple without children Other Single with children Single without children One male earner Couple with children Couple without children Other Single with children Single without children “Traditional” with children without children Two female earners Two male earners Three+ earners

Share in the US (percent) (1)

Average share in other 23 countries (percent) (2)

Difference between US share and average share in other countries (percentage points) (3) = (1) − (2)

2.3 1.0 1.8 6.5 3.2

3.5 0.9 0.6 4.1 3.2

−1.2 0.1 1.1 2.4 0.0

13.1 2.2 1.5 1.2 4.0

13.0 1.7 0.8 1.0 4.3

0.1 0.5 0.7 0.2 −0.3

29.6 12.6 2.3 3.0 15.3

32.9 13.4 1.7 2.3 16.4

−3.3 −0.8 0.6 0.7 −1.1

sion analysis, however, shows that the introduction of the demographic composition does not affect the results; it rather makes them stronger because the US dummy variable in the population-weighted regression is greater than in the unweighted formulation. So, if anything, the US has a “favorable” demographic composition. In table 1.5 we show the share of each subgroup in the US and the unweighted average shares of the same subgroups across the 23 comparator countries. The US shares diverge by more than 2 percentage points in only two cases. The first case is the one-male-one-female-earner couple with children: about 30 percent of the US population is living in such households versus 33 percent, on average, in the rest of these OECD countries. The second case is one-female-earner households where that earner is single with children; about 6.5 percent of the US population lives in that type of household but only 4 percent (on average) in the other OECD countries. (In common parlance, the US is slightly low on “traditional” households and slightly high vis-à-vis single mothers). In short and even leaving the regression results aside, we note that the US family composition is very similar to that of other countries. Thus, a unique compositional structure does not explain the high level of overall earnings inequality reported in the US. Second, is the “story” that we report here stable over time, or is there something unusual about the year that we chose (2010)?

38

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

Table 1.6

US inequality rankings among 24 OECD countries (1 = highest; 24 = lowest)

Type of household One female earner Couple with children Couple without children Other Single with children Single without children One male earner Couple with children Couple without children Other Single with children Single without children “Traditional” with children without children Two female earners Two male earners Three+ earners Unweighted mean rank

Subgroup share (US 2010)

Ranks 1995

2010

Difference in US rank between two time points

2.3 1.0 1.8 6.5 3.2

2 6 2 3 1

5 4 6 3 3

−3 2 −4 0 −2

13.1 2.2 1.5 1.2 4.0

1 1 2 5 2

1 1 3 8 4

0 0 −1 −3 −2

29.6 12.6 2.3 3.0 15.3

1 2 3 1 2 2.3

1 2 2 1 2 3.1

0 0 1 0 0

Note: The five rows in bold account for more than 75 percent of persons in the US.

Table 1.6 provides a window onto the answer to that question. It shows, for each subgroup, how US inequality (captured by the Gini) is ranked with respect to the 24 countries in our study at two points in time, 1995 (Wave IV)16 and 2010 (Wave VIII). This is not, of course, a huge sweep of time but it is the longest interval for which we had data on all 24 countries; and 13 years (including the onset of the global financial crisis) is not a trivial passage of time. Consider the five most prevalent subgroups— shown in bold. These groups constitute over 75 percent of the US population. In each of these five subgroups, the US rank (within the 24 countries) is exactly the same at both time points. Across all 15 subgroups, the average change in rank, over this 13-year period, is 0.8, from 2.3 to 3.1— that is, less than one rank position. Thus, we conclude, our results are sustainable over time. The main year of this study— 2010— does not appear to be unique, as least not with the respect to the past two decades.

16. LIS’s Wave IV is centered on 1995, but the precise years vary between 1992 and 1997. The Wave IV US dataset is from 1997.

In Search of the Roots of American Inequality Exceptionalism

1.4

39

Conclusions and Directions for Future Research

We began by noting that prior literature establishes that the high level of inequality in US disposable household income, calculated across workingage households, is not only the product of modest redistribution in the US as compared with similar OECD countries; it is also the result of a comparatively high level of inequality in the underlying market income. Furthermore, the primary component of market income is income from labor. In this study, we have shown that equivalized labor income across households is indeed more unequally distributed in the US than in all (but one) of 24 OECD countries included. We were also interested in assessing whether labor income inequality is pervasive across household types and demographic subgroups or whether it may be due to either exceptionally high or exceptionally low average labor incomes received by some groups. We conclude that within-group inequality of labor incomes in the US is, in almost all cases, high by OECD standards. So, it is neither an unusual household composition nor unusually high mean labor incomes of some demographic groups that explains high US earnings inequality, but simply the fact that high and low labor incomes are widely spread across all of our household/demographic categories. We have seen that, in 2010, when we look at 15 (mutually exclusive) demographic groups, the inequality rankings of the US are consistently high. In 10 out of 15 groups (within-group) inequality places the US among the three most unequal countries out of 24; in three more cases, the US falls among the five most unequal countries. Our overall conclusion, clearly, is that the high level of market income inequality in the US, in cross-national perspective, is found across diverse subgroups. A detailed policy analysis is beyond the goals and scope of this chapter, but we offer a few final comments about research that would extend what we have reported here. The large cross-national literature on policies and institutions that shape economic inequality can be divided, in general, into two bodies of work: comparative research on the determinants of earnings inequality and comparative studies on income redistribution. The former literature mainly focuses on regulations and other tools that set floors under earnings (mainly minimum wages) and institutions that shape workers’ bargaining power (mainly unions); these public interventions are increasingly referred to as instruments of “predistribution” (Chwalisz and Diamond 2015; Hacker 2011).17 The latter literature focuses on the design, mix, and effectiveness of 17. Hacker (2011) is widely credited with coining the term “predistribution,” referring to institutions that prevent or reduce market-driven inequalities. The term is intentionally contrasted with redistribution, specifically with the classic redistributive instruments— transfers and taxes— that reshape inequalities produced by markets.

40

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

the classic tools of redistribution— public income transfers and households’ direct taxes (see Gornick and Smeeding 2018 for a review). The study that we report in this chapter turns our attention to the instruments of predistribution. Institutions that affect earnings inequality have received heightened attention in recent years among economists studying income inequality (see the influential policy proposals in Atkinson 2015; see also OECD 2008, 2011, 2015). They have also attracted attention among political scientists, sociologists, labor scholars, and legal analysts (e.g., Alderson and Nielsen 2002; Alexander, Haley-Lock, and Ruan 2015; Anker and Anker 2017; Golden and Wallerstein 2011; Kenworthy 2001). A substantial strand in that literature focuses on how, and the extent to which, policies and institutions vary across high-income countries (e.g., Blau and Kahn 2002; Salverda and Checchi 2015). As is well known by now, many studies have indicated that, in the US, minimum wages are low and unions are weak, relative to other high-income countries, especially among OECD countries (see OECD 2015). Furthermore, several studies have concluded that the low minimum wages and weak collective bargaining in the US do, in fact, explain a substantial share of the higher level of earnings inequality in the US (see Gornick and Smeeding 2018 for a review). Future work that builds on this chapter might address two issues/questions: (1) Public policies and institutions that shape earnings distributions— such as minimum wages, structures of collective bargaining, and other mechanisms for wage setting, including on the high end— are understood to affect the distribution of individuals’ earnings. Our work focuses on households’ earnings; those are clearly shaped by the earnings of individual household members but also by the ways in which households are formed, vis-à-vis combinations of earners. Little if any research assesses the extent to which these earnings-related policies and institutions shape households’ earnings— either directly or indirectly (by influencing household-level employment behavior or even household formation).18 (2) The extensive cross-national literature on the major tools of “predistribution” has not, thus far, focused on their varied effects across subgroups of workers, differentiated by “bundles” of characteristics— much less subgroups of households. Little if any research assesses whether (or how or why) earnings-related policies differentially affect households, when those

18. We are certainly not the first to note this lacuna in the inequality literature. Salverda and Checchi’s (2015, 1537) review of labor market institutions and wage dispersion begins by observing that there are two massive literatures— one on wage dispersion and one on income inequality— and that “the two strands of study are . . . miles apart.” Salverda and Checchi lament that split because income from labor is the largest component of working-age households’ income. They attribute the lack of integration of the two literatures to the complexity of their interaction.

In Search of the Roots of American Inequality Exceptionalism

41

households are distinguished by their earners’ gender, partnership, and/or parenting status. A rich and growing supply of institutional databases in combination with high-quality microdata (such as the LIS data)– available both across countries and over time— offers the basis for future studies that might tackle these questions.

Appendix Table 1A.1

Australia Canada Czech Republic Denmark Estonia Finland France Germany Greece Hungary Iceland Ireland Israel Italy Luxembourg Netherlands Norway Poland Russia Slovakia Slovenia Spain United Kingdom United States

LIS datasets used Name of survey

Year

Household Expenditure Survey (HES) and Survey of Income and Housing (SIH) Survey of Labour and Income Dynamics (SLID) Survey of Income and Living Conditions (EU-SILC) Statistics Denmark: Law Model Estonian Social Survey (ESS); Survey on Income and Living Conditions (EU-SILC) Survey of Income and Living Conditions (EU-SILC), formerly known as Income Distribution Survey (IDS) Family Budget Survey (BdF) German Social Economic Panel Study (GSOEP) Survey of Income and Living Conditions (EU- SILC), 2011 Household Monitor Survey (HES) Survey of Income and Living Conditions (EU-SILC) Survey of Income and Living Conditions (EU-SILC) Household Expenditure Survey (HES) Survey of Household Income and Wealth (SHIW) Panel socio-économique “Liewen zu Letzebuerg” (PSELL III); Survey of Income and Living Conditions (EU-SILC) Survey of Income and Living Conditions (EU-SILC) Household Income Statistics (formerly based on the Income Distribution Survey) Household Budget Survey Russia Longitudinal Monitoring Survey, Higher School of Economics (RLMS-HSE) Survey of Income and Living Conditions (EU SILC), 2011 Household Budget Survey Encuesta de Condiciones de Vida (ECV); Survey of Income and Living Condition (EU-SILC), 2010 Family Resources Survey (FRS) Current Population Survey (CPS) Annual Social and Economic Supplement (ASEC)

2010 2010 2010 2010 2010 2010 2010 2010 2010 2009 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010

42

Janet C. Gornick, Branko Milanovic, and Nathaniel Johnson

References Alderson, Arthur S., and Francois Nielsen. 2002. “Globalization and the Great U-Turn: Income Inequality Trends in 16 OECD Countries.” American Journal of Sociology 107 (5): 1244– 99. Alexander, Charlotte, Anna Haley-Lock, and Nantiya Ruan. 2015. “Stabilizing Low-Wage Work.” Harvard Civil Rights-Civil Liberties Law Review 50: 1– 48. Anker, Richard, and Martha Anker. 2017. Living Wages Around the World: Manual for Measurement. Northampton, MA: Edward Elgar. Atkinson, Anthony B. 2015. Inequality: What Can Be Done? Cambridge, MA: Harvard University Press. Blau, Francine D., and Lawrence M. Kahn. 2002. At Home and Abroad: U.S. Labor Market Performance in International Perspective. New York: Russell Sage Foundation. Blau, Francine D., and Anne E. Winkler. 2017. The Economics of Women, Men, and Work, 8th ed. New York: Oxford University Press. Brandolini, Andrea, and Timothy M. Smeeding. 2006. “Patterns of Economic Inequality in Western Democracies: Some Facts on Levels and Trends.” Political Science and Politics 39 (1): 21– 26. Chwalisz, Claudia, and Patrick Diamond, eds. 2015. The Predistribution Agenda: Tackling Inequality and Supporting Sustainable Growth. London: IB Taurus. Gabaix, Xavier, and Augustin Landler. 2008. “Why Has CEO Pay Increased so Much?” Quarterly Journal of Economics 123 (1): 49– 100. Gautié, Jérôme, and John Schmitt, eds. 2009. Low-Wage Work in the Wealthy World. New York: Russell Sage Foundation. Golden, Miriam, and Michael Wallerstein. 2011. “Domestic and International Causes for the Rise of Pay Inequality in OECD Nations between 1980 and 2000. In Comparing European Workers Part A: Experiences and Inequalities, edited by David Brady, 209– 49. Bingley: Emerald. Gornick, Janet C., and Markus Jäntti, eds. 2013. Income Inequality Economic Disparities and the Middle Class in Affluent Countries. Stanford, CA: Stanford University Press. Gornick, Janet C., and Branko Milanovic. 2015. “Income Inequality in the United States in Cross-National Perspective: Redistribution Revisited.” LIS Center Research Brief 1/2015. New York: Stone Center on Socio-Economic Inequality. https://stonecenter.gc.cuny.edu/research/income-inequality-in-the-united-states -in-cross-national-perspective-redistribution-revisited/. Gornick, Janet C., and Timothy M. Smeeding. 2018. “Redistributional Policy in Rich Countries: Institutions and Impacts in Nonelderly Households.” Annual Review of Sociology 44: 441– 68. Hacker, Jacob. 2011. “The Institutional Foundations of Middle-Class Democracy.” Policy Network. 6 (5): 33– 37. Kenworthy, Lane. 2001. “Wage-Setting Measures: A Survey and Assessment.” World Politics 54: 57– 98. Lemieux, Thomas. 2006. “Increasing Residual Wage Inequality: Composition Effects, Noisy Data, or Rising Demand for Skill?” American Economic Review 96 (3): 461– 98. Lucifora, Claudio, and Wiemer Salverda W. 2009. “Low Pay.” In The Oxford Handbook of Economic Inequality, edited by Brian Nolan, Wiemer Salverda, and Timothy M. Smeeding, 257– 83. Oxford: Oxford University Press.

In Search of the Roots of American Inequality Exceptionalism

43

Luxembourg Income Study (LIS) Database. http://www.lisdatacenter.org. Luxembourg: Luxembourg Income Study Center. McCall, Leslie. 2000. “Explaining Levels of Within-Group Wage Inequality in U.S. Labor Markets.” Demography 37 (4): 415– 30. Mishel, Lawrence. 2015. “Causes of Wage Stagnation.” Economic Policy Institute, January 6. https://www.epi.org /publication /causes-of-wage-stagnation/. Mishel, Lawrence, and Alyssa Davis. 2015. “Top CEOs Make 300 Times More than Typical Workers.” Economic Policy Institute Issue Brief No. 399. http://www.epi .org /files/2015/top -ceos-make-300 -times-more-than-typical-workers.pdf. Organization for Economic Cooperation and Development (OECD). 2008. Growing Unequal: Income Distribution and Poverty in OECD Countries. Paris: OECD. Organization for Economic Cooperation and Development (OECD). 2011. Divided We Stand: Why Inequality Keeps Rising. Paris: OECD. Organization for Economic Cooperation and Development (OECD). 2015. Minimum Wages After the Crisis: Making Them Pay. Paris: OECD Piketty, Thomas. 2014. Capital in the Twenty-First Century. Translated by Arthur Goldhammer. Cambridge, MA: Belknap Press of Harvard University Press. Piketty, Thomas, and Emmanuel Saez. 2006. “The Evolution of Top Incomes: A Historical and International Perspective.” American Economic Review 96 (2): 200– 205. Salverda, Wiemer, and Daniele Checchi. 2015. “Labour-Market Institutions and the Dispersion of Wage Earnings.” In Handbook of Income Distribution, edited by Anthony B. Atkinson and François Bourguignon, 2:1535– 727. Amsterdam: Elsevier. VanHeuveline, Tom. 2018. “Recovering the Missing Middle: A Mesocomparative Analysis of Within-Group Inequality, 1970– 2011.” American Journal of Sociology 123 (4): 1064– 116. Western, Bruce, Deirdre Bloome, and Christine Percheski. 2008. “Inequality among American Families with Children, 1975 to 2005.” American Sociological Review 73 (6): 903– 20.

2

Rising Between-Firm Inequality and Declining Labor Market Fluidity Evidence of a Changing Job Ladder John Haltiwanger and James R. Spletzer

2.1

Introduction

A large literature has documented the growth of real earnings dispersion in the US economy since the late 1970s, often referred to as increasing earnings inequality. During this same time, labor market fluidity in the US has declined as evidenced by a decline in the overall pace of hires and separations (see Davis, Faberman and Haltiwanger 2012; Davis and Haltiwanger 2014; Hyatt and Spletzer 2013; and Molloy et al. 2016). The decline in the hiring rate includes both a decline in the pace of employer-to-employer flows as well as hires from nonemployment. In this chapter, we explore potential connections between the rise in earnings inequality and declining labor market fluidity. Our analysis of these issues uses matched employer-employee data from the LEHD program at Census to conduct a series of empirical exercises that help understand the connections from the findings from the distinct literatures on inequality and labor market fluidity. We use this data infrastructure John Haltiwanger is a professor of economics at the University of Maryland, and a research associate of the National Bureau of Economic Research. James R. Spletzer is principal economist of the Center for Economic Studies at the U.S. Census Bureau. We thank Keith Bailey, Henry Hyatt, Ron Jarmin, Bruce Meyer, and participants at the CRIW-NBER conference on “Measuring and Understanding the Distribution and Intra/ Inter-Generational Mobility of Income and Wealth,” Bethesda, Maryland, March 5– 6, 2020, for helpful comments. Any opinions and conclusions expressed in this chapter are those of the authors and do not necessarily represent the views of the US Census Bureau. All results have been reviewed to ensure that no confidential information is disclosed (CBDRB-FY20CED006– 0010). For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber.org /books-and -chapters /measuring-distribution -and-mobility-income -and-wealth /rising-between -firm -inequality-and-declining-labor-market-fluidity-evidence-changing-job -ladder.

45

46

John Haltiwanger and James R. Spletzer

to show increasing inequality in the upper tail of the earnings distribution during the last two decades (1998– 2018). Using the same data infrastructure, we illustrate key components of the observed declining fluidity, focusing on the decomposition of workers into four hires types: stayers, job switchers within the same industry, job switchers across industries, and hires from nonemployment. We find that the share of stayers has been increasing as a fraction of employment while the share of hires has declined, with especially large declines of hires from nonemployment. Our empirical analysis also builds on the recent literature that shows substantial firm and industry dimensions to increasing inequality. Recent findings emphasize that much of the rise in earnings inequality in the US over the last few decades is accounted for by rising between-firm inequality (see Barth et al. 2016; Song et al. 2019). Our recent work (Haltiwanger and Spletzer 2020) shows that this rising between-firm inequality is dominated by rising industry inequality. For our sample and definition of firms, we replicate that finding in our analysis. The dominant role of rising between-firm and between-industry inequality provides a potential connection to the changing patterns of fluidity via a changing job ladder. There is much evidence that individuals tend to start their careers at lower earnings (lower rungs of the job ladder) and move up over the course of their careers. Topel and Ward (1992) found that a large fraction of earnings increases for young workers is accounted for by job switches rather than within-firm increases in earnings. A core prediction of job ladder models (see, e.g., Burdett and Mortensen 1998; Moscarini and Postel-Vinay 2013) is that high-wage firms should have more of their hires via job switchers while low-wage firms should have more of their hires via nonemployment. Recent evidence provides empirical support for this prediction. Haltiwanger et al. (2018) show that high-wage firms have a large share of hires from other firms while low-wage firms have large share of hires from nonemployment. These patterns hold for job switches both within and between industries.1 Our findings in this chapter along with those in the recent literature support the hypothesis that there has been a change in the job ladder. Rising between firm inequality suggests that the rungs of the job ladder have become further apart. Declining fluidity suggests that it has become more difficult to get on the ladder and the pace of climbing the ladder has slowed. The current work explores this hypothesis of a changing job ladder on a number of dimensions. In turn, we assess the contribution of the changing job ladder for understanding the increase in earnings inequality. We exploit the dominant role of industry effects to investigate the con1. Haltiwanger et al. (2018) include both within- and between-industry job switchers in their analysis. Haltiwanger, Hyatt, and McEntarfer (2016) provide evidence that there is a betweenindustry job ladder.

Rising Between-Firm Inequality and Declining Labor Market Fluidity

47

nection between changing interindustry earnings differentials and changes in the job ladder. Using detailed industry-level data, we find that industries with a high share of hires from job switchers, and especially from job switchers between industries, have significantly higher earnings. Relatedly we find that industries with a high share of hires from nonemployment have significantly lower earnings. These patterns also hold for earnings of different hires types: stayers, job switchers, and hires from nonemployment. These patterns also hold whether or not we control for the demographic composition of workers (e.g., worker age, education, and gender) and firms (i.e., firm size and firm age) in the industry. These results are consistent with the empirical job ladder evidence above and are also consistent with the theoretical predictions of job ladder models cited above. Not only do industries with a larger share of hires from job switchers have especially high wages but the earnings differential for such industries has been rising during the past two decades. The differentials for both hires from the same industry and hires from other industries have been increasing. Likewise, the industries with a larger share of hires from nonemployment have increasingly lower earnings differentials over the past two decades. Using simple accounting decompositions, we find that changing differentials by hires types in combination with the changing distribution of hires types accounts for about 30 percent of rising interindustry earnings differentials. This finding is without any controls. Using only firm and worker demographic controls, we can account for about 60 percent of the rising interindustry earnings differentials. In specifications including both hires types and firm and worker controls, we can account for about 80 percent of rising interindustry earnings differentials. The latter differs from the implied 90 percent (adding the separate 30 + 60 contributions) given covariance effects in the accounting decompositions. We also investigate the role of composition effects resulting from declining fluidity. We find that using either individual-level or detailed industry-level data, there is rising inequality within each of the hires types: stayers, job switchers within industries, job switchers between industries, and hires from nonemployment. This finding highlights that composition changes in hires types from declining fluidity does not help account for rising inequality. If anything, this composition effect works in the wrong direction, since the variance of earnings of stayers is the lowest and the variance of earnings for hires from nonemployment is the highest among the four groups. The chapter proceeds as follows. Section 2.2 describes the data infrastructure. Section 2.3 shows that rising overall earnings inequality is dominated by rising between-firm inequality and in turn by rising betweenindustry inequality. Section 2.4 explores the patterns of declining fluidity through the lens of the four hires types we use in our subsequent analysis: stayers, job switchers within industries, job switchers between industries, and hires from nonemployment. Section 2.5 analyzes the variance of earnings for

48

John Haltiwanger and James R. Spletzer

each of the four hires types. Section 2.6 investigates the connection between rising interindustry earnings differentials and earnings differentials by hires types along with controlling for and exploring the contribution of changing firm and worker demographic effects. Section 2.7 provides concluding remarks. We view our results as exploratory, bringing together two distinct literatures. We focus on a range of open questions in our concluding remarks. 2.2

Data Infrastructure

All of our analysis is based on data from the Longitudinal EmployerHousehold Dynamics (LEHD). The LEHD is a longitudinally linked employer-employee dataset created by the US Census Bureau as part of the Local Employment Dynamics federal-state partnership. The data are derived from state-submitted unemployment insurance (UI) wage records and the Quarterly Census of Employment and Wages (QCEW) data. Every quarter, employers who are subject to state UI laws— approximately 98 percent of all private-sector employers, plus state and local governments— are required to submit to the states information on their workers (the wage records, which lists the quarterly earnings of every individual in the firm) and their workplaces (the QCEW, which provides information on the industry and location of each establishment). The wage records and the QCEW data submitted by the states to the US Census Bureau are enhanced with Census and survey microdata in order to incorporate information about worker demographics (age, gender, and education) and the firm (firm age and firm size). A job in the LEHD is defined as the presence of an individual-employer match, and earnings are defined as the amount earned from that job during the quarter. We use full-quarter (FQ) jobs in our analysis, where an FQ job is defined as a contemporaneous employer-employee match that also exists in the previous quarter and in the following quarter. The underlying assumption is that individuals in FQ jobs are working all 13 weeks of the quarter, which avoids the issue of not knowing the number the weeks worked during the quarter for individuals who start a job or end a job during that quarter. Restricting to FQ jobs is similar in spirit to the full-time or full-year restriction used when analyzing inequality with household survey data. We impose two recodes on the LEHD earnings data. First, to minimize the effect of outliers and smooth the first two moments of the earnings time series, we topcode earnings at the 99.5th percentile of the state-year-quarter distribution. Second, all of our analysis uses the natural log of real quarterly earnings, where nominal values are converted to real using the 2018Q1 Consumer Price Index for Urban Consumers Research Series (CPI-U-RS) deflator. Because states have joined the LEHD program at different times and have provided various amounts of historical data on joining the LEHD program, the length of the time series of LEHD data varies by state. We use data from the 20 states that have data available from 1996Q4 through 2018Q2, which

Rising Between-Firm Inequality and Declining Labor Market Fluidity

49

gives us full-quarter data from 1997Q1 to 2018Q1.2 We restrict the LEHD data to the private sector. In order to focus on long-run trends and avoid issues of seasonality, we use data from the first quarter of the year.3 The primary definition of firms we use in our analysis is business units defined by the state UI number, referred to by users of the LEHD data as the state employer identification number (SEIN). This definition of firms is narrower than the enterprise definition used in Haltiwanger and Spletzer (2020) and the definition based on the federal Employer Identification Number (EIN), as used by Song et al. (2019). We explore the sensitivity of analysis to using the SEIN vs. EIN vs. Census enterprise firm (Census firm IDs) below. The SEIN has the advantage that is includes more geographic variation, which is relevant for declining labor market fluidity since part of the latter is declining geographic mobility (see, e.g., Molloy et al. 2016). Key statistics from our annual data are given in figure A.1 of the online appendix (http://www.nber.org /data-appendix /c14447/appendix.pdf). To summarize, in 2018Q1, there are over 50 million FQ jobs and approximately 3.2 million SEIN firms in our 20-state LEHD data. The variance of FQ LEHD earnings is increasing between 1998 and 2108. This rising variance, often referred to as “increasing earnings inequality,” is the focus of our analysis in this chapter. In figure A.2 of the online appendix, we present percentiles of the LEHD full-quarter earnings distribution from 1996 to 2018, as well as published percentiles of full-time wage and salary earnings from the Current Population Survey (CPS). The time series of the LEHD and CPS percentiles, indexed to 100 in 1996, are similar. 2.3

Rising Earnings Inequality: The Dominant Role of Between-Firm and Between-Industry Effects

We focus on the variance as the measure of the dispersion of LEHD fullquarter earnings. This focus facilitates the decomposition of the variance of individual earnings into within-firm and between-firm components: (2.1a)

Var(Wif ) = Var(Wif

Wf ) + Var(Wf ),

2. These 20 states are: California, Colorado, Connecticut, Hawaii, Idaho, Illinois, Kansas, Louisiana, Maryland, Minnesota, Missouri, Montana, North Carolina, New Jersey, New Mexico, Oregon, Rhode Island, Texas, Washington, and Wyoming. These 20 states account for roughly 46 percent of national employment. The time series of employment from these 20 states closely tracks the national time series of total private sector employment published by the QCEW program at the Bureau of Labor Statistics (BLS). 3. The key findings from our variance decomposition are not sensitive to whether we use full-quarter earnings from the first, second, third, or fourth quarter of the year, nor are they sensitive to whether we sum the LEHD quarterly earnings into an annual measure of earnings with a minimum earnings threshold. Annual earnings are used by Song et al. (2019) using SSA data, as well as by Abowd, McKinney, and Zhao (2018), using LEHD data. The key findings do change dramatically when no minimum earnings threshold is applied to annual earnings data, most likely due to a decline in short-duration jobs and thus a compositional change in the lower part of the earnings distribution— see Hyatt and Spletzer (2017) for further elaboration on this point.

50

John Haltiwanger and James R. Spletzer

where i refers to the individual and f refers to the firm. The first term on the right side of the equation is the variance within firms, and the second term is the variance between firms. Furthermore, letting k refer to industries, we can further write this variance decomposition as (2.1b)

Var(Wifk) = Var(Wifk Wfk) + Var(Wfk Wk) + Var(Wk).

The middle term on the right side of the equation is the between-firm withinindustry variance, and the third term is the variance between industries. Calculating this variance decomposition in each year, and letting Δ denote changes across time, we have (2.1c)

Var(Wifk) = Var(Wifk Wfk) + Var(Wfk Wk) + Var(Wk).

The increase in the variance of individual level wages can be decomposed into a change within firms (the first term on the right-hand side of equation (2.1c)), the change between firms within industries (the second term), and the change between industries (the third term). The variance decompositions with the LEHD full-quarter earnings data are presented in figure 2.1. The top line is the variance of individual earnings, which is the same as in online appendix figure A.1. This variance increases from 1.109 in 1998 to 1.291 in 2018. The within-firm variance in figure 2.1 is roughly constant across time (rising slightly from 0.566 in 1998 to 0.575 in 2018). The between-firm variance in figure 2.1, from equation (2.1a), rises from 0.543 in 1998 to 0.716 in 2018. These statistics tell us that 95.1 percent of total variance growth from 1998 to 2018 is between firms, with only 4.9 percent of the variance growth within firms. This finding that most variance growth is between firms rather than within firms is consistent with much of the recent literature (Barth et al. 2016; Haltiwanger and Spletzer 2020; Handwerker and Spletzer 2016; Song et al. 2019), as well as a much earlier literature (Davis and Haltiwanger 1991; Dunne et al. 2004). The rising between-firm variance can further be decomposed into withinindustry and between-industry components. Using four-digit North American Industrial Classification System (NAICS) industries, the between-firm within-industry variance rises from 0.272 in 1998 to 0.337 in 2018, and the between-industry variance rises from 0.271 in 1998 to 0.379 in 2018. These statistics show that 62.4 percent of the large increase in between-firm variance is between industries, and 37.6 percent is within industries. This finding that a substantial amount of variance growth is between industries is the focus of recent work by Haltiwanger and Spletzer (2020), and it plays an important role in the methodology we use later in this chapter. As we emphasize in that companion paper, this finding of a dominant role for industry effects challenges conventional wisdom from the recent literature. We argue that this reflects limitations in industry codes in the prior literature that we overcome with high-quality industry codes on business-level data at BLS and Census. Our approach and methodology build on the finding in the

Rising Between-Firm Inequality and Declining Labor Market Fluidity

Fig. 2.1

51

Variance decomposition

companion paper of a dominant role for industry effects in rising betweenfirm inequality. We contribute to that finding here by extending this result for a longer sample period and using the SEIN as the definition of the firm. We conclude this section with two sensitivity analyses. Table 2.1 presents the basic variance decomposition (from the equations above) using different levels of NAICS industry detail. To read this table, begin with the column titled “4-digit naics.” The first panel presents the 2018 decomposition of earnings discussed above, and the second panel presents the 1998– 2018 decomposition of variance growth. The key panel is the fourth panel, where we present the decomposition of variance growth in percentage terms. Staying with the 4-digit naics column, we see that 59.3 percent of total variance growth is between industries, which translates into 62.4 percent of the between-firm variance growth being between industries. How does this 62.4 percent statistic vary with the level of industry detail? There are 23 two-digit industries, and 30.6 percent of between firm variance growth is between these 23 industries.4 The amount of between firm variance growth between industries rises with the level of industry detail, to 53.8 percent of variance growth between the 91 three-digit industries and 62.4 percent between the 304 four-digit industries. Additional industry detail shows that 65.3 percent of between-firm variance growth is between the 682 five-digit industries, and 66.5 percent is between the 1,034 six-digit industries. Our second sensitivity analysis is to examine how changing the definition 4. Our reference to two-digit industries refers to the first two digits of the six-digit NAICS code. This is slightly different from NAICS sectors, in which 31– 33 are aggregated into Manufacturing, 44– 45 are aggregated into Retail Trade, and 48– 49 are aggregated into Transportation and Warehousing.

52

John Haltiwanger and James R. Spletzer

Table 2.1

Variance decomposition

2018 levels Variance LN($) Within firms Between firms Within industry Between industry 1998–2018 growth Variance LN($) Within firms Between firms Within industry Between industry 2018 levels Variance LN($) Within firms Between firms Within industry Between industry Between firms Within industry Between industry 1998–2018 growth Variance LN($) Within firms Between firms Within industry Between industry Between firms Within industry Between industry Number of industries

2-digit NAICS

3-digit NAICS

4-digit NAICS

5-digit NAICS

6-digit NAICS

1.291 0.575 0.716 0.474 0.242

1.291 0.575 0.716 0.387 0.329

1.291 0.575 0.716 0.337 0.379

1.291 0.575 0.716 0.316 0.400

1.291 0.575 0.716 0.306 0.410

0.182 0.009 0.173 0.120 0.053

0.182 0.009 0.173 0.080 0.093

0.182 0.009 0.173 0.065 0.108

0.182 0.009 0.173 0.060 0.113

0.182 0.009 0.173 0.058 0.115

100.0% 44.5% 55.5% 36.7% 18.7% 100.0% 66.2% 33.8%

100.0% 44.5% 55.5% 30.0% 25.5% 100.0% 54.1% 45.9%

100.0% 44.5% 55.5% 26.1% 29.4% 100.0% 47.1% 52.9%

100.0% 44.5% 55.5% 24.5% 31.0% 100.0% 44.1% 55.9%

100.0% 44.5% 55.5% 23.7% 31.8% 100.0% 42.7% 57.3%

100.0% 4.9% 95.1% 65.9% 29.1% 100.0% 69.4% 30.6%

100.0% 4.9% 95.1% 44.0% 51.1% 100.0% 46.2% 53.8%

100.0% 4.9% 95.1% 35.7% 59.3% 100.0% 37.6% 62.4%

100.0% 4.9% 95.1% 33.0% 62.1% 100.0% 34.7% 65.3%

100.0% 4.9% 95.1% 31.9% 63.2% 100.0% 33.5% 66.5%

23

91

304

682

1034

of the firm affects our results. In almost all of this chapter, we use the SEIN as the definition of the firm. The SEIN is the UI number that represents the firm within the state. We have two other firm identifiers in the LEHD data— the EIN and the enterprise-level firm ID. The latter encompasses all activity under common operational control. Both the EIN and the enterprise firm ID are national whereas the SEIN is state specific. We present results in the online appendix (http://www.nber.org/data-appendix/c14447/appendix.pdf, see table A.1); they show that our finding that more than half of variance growth is between four-digit NAICS industries is unaffected by the definition of the firm.

Rising Between-Firm Inequality and Declining Labor Market Fluidity

2.4

53

Declining Labor Market Fluidity

Many studies have found a decline in indicators of labor market fluidity (see, for example, Davis et al. 2007; Davis, Faberman and Haltiwanger 2012; Davis and Haltiwanger 2014; Hyatt and Spletzer 2013; Molloy et al. 2016). Such indicators include a decline in the pace of worker reallocation (hires + separations), job reallocation (job creation + destruction), and employer-toemployer flows. These findings on declining labor market fluidity are drawn from studies that use administrative data such as the LEHD and the LBD, business survey data such as the Job Openings and Labor Turnover Survey (JOLTS), and individual survey data such as the CPS. The LEHD data are the most comprehensive, in that the decline in fluidity can be analyzed by characteristics of the firms as well as characteristics of the workers. In addition, the LEHD data permit decomposing hires (and separations) into employer-to-employer flows and hires from nonemployment. In this chapter, we are interested in the potential connection between rising earnings variance and declining labor market fluidity. We start with the simple observation that persons employed today were either in the same firm last year (stayers) or not in the firm last year (hires): (2.2a)

Total Employment = Stayers + Hires.

A hire can be either a person working in a different firm last year (employerto-employer hire) or a person who was not employed last year (hire from nonemployment): (2.2b)

Total Employment = Stayers + Employer-to-Employer Hires + Hires NonEmp.

Persons hired from a different firm could be persons hired from a firm in the same industry (E2E Same Ind) or persons hired from a different industry (E2E Diff Ind): (2.2c)

Total Employment = Stayers + E2E Same Ind + E2E Diff Ind + Hires NonEmp.

Equation (2.2b) identifies the four “hires type” groups we use in our subsequent analysis. Some details are required to implement this decomposition in practice. Our measurement approach is designed to yield a decomposition of FQ jobs in Q1 of each year given our focus on earnings of FQ jobs in Q1 of each year. Stayers are thus jobs where the individual holds a FQ job at the same firm in Q1 of adjacent years. Job Switchers are those that switch firms while holding FQ jobs in Q1 of adjacent years. “Hires from Nonemp” are residual reflecting hires from non-FQ employment in the year before a FQ Q1 job in the current year. These definitions are distinct from related

54

John Haltiwanger and James R. Spletzer

Fig. 2.2

Labor market fluidity

measures in the literature as we discuss in more detail below. It is also worth noting that our dataset is jobs rather than persons, so accounting for multiple jobholding is a slight complication.5 Figure 2.2 presents our measures of hires types as percentages of total full-quarter employment. Figure 2.2a shows that the percentage of fullquarter jobs that are stayers increased from 63.0 percent in 1998 to 68.5 per5. Persons holding one FQ job last year and more than one FQ job this year (1:N) are coded as follows: if last year’s job is also held this year, then that job is a stayer and the other “N−1” jobs this year are classified as hires from nonemployment. Persons holding more than one FQ job last year but only one FQ job this year (N:1) are classified based on whether this year’s job could be found last year (stayers) or if the current year’s job is new (E2E Same Ind or E2E Diff Ind). Persons holding two FQ jobs this year and two FQ jobs last year are classified by looking for the same job across years (stayers) or whether the current year’s jobs are new (E2E same ind or E2E diff ind). A very small number of persons with N1 FQ jobs last year and N2 full quarters jobs this year, where N1 > 2, N2 > 2, and N1 > 2 and/or N2 > 2, are deleted from the data.

Rising Between-Firm Inequality and Declining Labor Market Fluidity

55

cent in 2018. Expressed in terms of hires rather than stayers, our data shows evidence of declining labor market fluidity— the percentage of full-quarter jobs that are hires fell from 37.0 percent in 1998 to 31.5 percent in 2018. Figure 2.2b shows the decomposition of total hires into employer-toemployer flows and hires from nonemployment. Employer-to-employer hires only slightly declined from 10.0 percent in 1998 to 9.1 percent in 2018, whereas hires from nonemployment fell from 27.0 percent to 22.4 percent. Figure 2.2c shows the decomposition of employer-to-employer hires based on whether the hire was from the same four-digit NAICS industry or a different four-digit NAICS industry. Hires from the same industry are relatively small without much movement over time, whereas hires from a different industry are cyclical with a slight downward trend during our time period. Figure 2.2d shows the four key labor market flows that we will use in the following analysis. Our measures of labor market fluidity are, as noted, based on the status of employment for workers in the first quarter across years. These measures are related to but distinct from the published quarterly measures from the LEHD Quarterly Workforce Indicator (QWI) and Job-to-Job (J2J) programs (see https:// lehd.ces.census.gov/data/). In figure A.3 of the online appendix (http://www.nber.org /data-appendix /c14447/appendix.pdf) we provide comparisons of our measures with the published QWI and J2J series from LEHD.6 As described in the appendix, our takeaway is that our annual measures are capturing the well-known findings of a declining pace of hires with an especially large decline in hires from nonemployment. As will become clear, these measures not only are highly correlated with related published measures of fluidity but also are closely connected to interindustry earnings differentials both in the cross section and over time. 2.5

Earnings Dispersion by Hires Types

Figure 2.3 presents mean earnings for the various types of stayers and hires (hires from nonemployment, E2E hires from the same industry, and E2E hires from a different industry). The dotted black line in figure 2.3 is mean earnings of all FQ jobs, which is the same as in online appendix figure A.1. The data in figure 2.3 are broadly consistent with a job ladder. Mean earnings of stayers are the highest, and mean earnings of hires from nonemployment are the lowest. Mean earnings of persons hired from a different firm in the same industry are somewhat higher than mean earnings of persons hired from a different firm in a different industry. The variance of earnings for each of the classifications of hires and stayers are presented in figure 2.4. Figure 2.4a shows the total variance, figure 2.4b 6. We intentionally use the term employer-to-employer flows in this chapter (and shorthand E2E) to avoid confusion with the published job-to-job flows (J2J) series from LEHD.

Fig. 2.3

Mean full-quarter earnings by type of annual flow

Fig. 2.4

Variance of full-quarter earnings by type of annual flow

Rising Between-Firm Inequality and Declining Labor Market Fluidity

57

shows the between-industry variance, and figure 2.4c shows the withinindustry variance. In all panels of figure 2.4, the dotted black line is the variance of all FQ jobs. There are two striking results in figure 2.4. First, the variance of earnings is increasing over time for stayers and for each type of hire. This pattern of within hires type increase in earnings dispersion holds at the individual level overall, between industry, and within industry. Second, the variance of earnings of hires from nonemployment is greater than the variance of stayers. This is consistent with the predictions of the Burdett and Mortensen (1998) model of a job ladder, since transitions from nonemployment include all rungs of the job ladder while employer-to-employer flows include only rungs of the ladder above the current position of the ladder for workers. This pattern may also reflect the role of ex ante heterogeneity of workers. For example, heterogeneous individuals transit from nonemployment to substantially different starting earnings (e.g., high school versus college graduates transiting from nonemployment to employment). These findings from figure 2.4 imply that compositional changes in hires types cannot account for rising earnings inequality. First, the rise in earnings inequality is pervasive within each hires type. Second, declining fluidity implies that, over time, there is a larger share of stayers (low variance) and a smaller share of hires from nonemployment (high variance), and the resulting composition effects act to dampen the overall increase in variance. Put differently, there is even more rising inequality to account for after considering such composition effects. 2.6 2.6.1

The Contribution of Earnings Differentials by Hires Types Accounting Decomposition Methodology

Since the rising interindustry earnings differentials are within hires types groups, in this section we explore the potential connection between rising interindustry earnings differentials and the job ladder within groups. We use simple accounting decompositions for this purpose and focus our attention on rising between-industry earnings inequality. The focus on rising between-industry dispersion is motivated by our findings above that the vast majority of rising overall inequality is due to between-firm effects and in turn most of the latter is due to between-industry effects. Using the rising interindustry earnings differentials has numerous advantages since it permits a transparent mapping between the characteristics of the industry in terms of its position on the job ladder while also permitting controlling for firm and worker demographics of the industry. The simple regression and associated accounting decompositions we use in this section are intended to be exploratory and descriptive. Such regressions and decompositions don’t identify causal channels for rising interindustry differentials but help provide

58

John Haltiwanger and James R. Spletzer

guidance about the nature of the connection between rising inequality and the changing job ladder. We start by exploring the relationship between FQ industry earnings Wktj for hires type j and industry-level measures of the share of workers in the four hires types (Hkt ) as well as industry-level measures of firm and worker demographics (Dkt ).7 We estimate the following two specifications: (2.3a)

Wktj = Hkt

j

+ Dkt

j

+

j kt

(2.3b)

Wktj = Hkt

j t

+ Dkt

j t

+

j kt

.

Specification (2.3a) is a pooled specification with time invariant coefficients, and specification (2.3b) permits the coefficients to vary over time. Observe that we permit the shares of all hires types to impact the earnings of each hires type (more generally, the right-hand side variables are the same for each type j but the coefficients vary by j). Specification (3b) can be rewritten as (2.3c)

Wktj = Hktʹ

j

+ Dʹkt

j

+ Hktʹ (

j t

j

) + Dktʹ (

j t

j

)+

j kt

.

Following Juhn, Murphy, and Pierce (1993) (hereafter JMP), Davis and Haltiwanger (1991), and Dunne et al. (2004), the changes in dispersion (either the variance or other moments) can be decomposed into quantity (Hkt and Dkt ) effects for average prices ( j , j ), price effects ( tj and tj ), and the residual. We do not pursue the full distribution accounting insights from this approach but focus on the decomposition of variance.8 The estimation and decomposition is on an employment-weighted basis to be consistent with the variance trends reported in figure 2.4. 2.6.2

Regressions and Decompositions

We present estimates of regression equation for (2.3a) for each of the hires type groups and for overall earnings in the industry. The explanatory variables include the hires types shares (with stayers as the omitted group) and the firm and worker demographic variables. Worker characteristics (age, gender, and education) are meant to capture differences in the mix of workers across industries, and firm characteristics (firm age and firm size) capture differences in firm observables across industries.9 The industry-level employ7. By design the right-hand side variables are the same for each of the specifications by hire type. For example, each regression in table 2.3 includes the percentage of females in the industry as an explanatory variable, and each regression includes the share of hires from nonemployment in the industry as an explanatory variable. The right-hand side variables represent characteristics of the industry. 8. There are some limitations of the JMP decomposition methodology as highlighted by DiNardo, Fortin, and Lemieux (1996) and Fortin, Lemieux, and Firpo (2010). These limitations primarily apply to the full distribution accounting (e.g., decomposing the 90– 50 vs. the 50– 10) which we do not pursue. 9. To be precise, we create industry-year means of worker age, gender, education, firm age, and firm size, and then take the natural log of the industry-year means for worker age, education, firm age, and firm size. Worker and firm demographics are deviations from pooled means.

Rising Between-Firm Inequality and Declining Labor Market Fluidity Table 2.2

Intercept

59

Regressions and decompositions using industry-by-year earnings by hires type Earnings all jobs

Earnings same firm stayers

Earnings hires same industry

Earnings hires different industry

Earnings hires nonemployment

9.566

9.624

9.806

9.044

9.028

Hire same industry Hire diff industry Hire non-FQ-emp

4.861 5.030 −4.165

4.070 4.959 −3.679

2.281 7.004 −5.046

5.066 6.783 −3.167

7.360 5.570 −3.943

LN(worker age) female LN(education) LN(firm age) LN(firm size)

0.913 −1.068 5.583 −0.263 0.054

0.628 −1.038 5.681 −0.197 0.045

−0.110 −0.990 5.121 −0.248 0.027

1.428 −1.048 5.431 −0.205 0.038

1.441 −1.176 5.371 −0.449 0.082

0.839

0.835

0.819

0.830

0.750

Variance growth Predicted X(t) ∗ β Predicted X(t) ∗ β(t) Residual Total

−0.122 0.085 0.023 0.108

−0.097 0.092 0.024 0.116

−0.120 0.043 0.014 0.057

−0.105 0.043 0.011 0.054

−0.133 0.092 0.027 0.119

% contribution Changing X Changing β Residual

−113.0 191.7 21.3

−83.6 162.9 20.7

−210.5 286.0 24.6

−194.4 274.1 20.4

R-squared

−111.8 189.1 22.7

Notes: Dependent variable is LN real full-quarter earnings of the hires type listed at the top of the row. N = 6384 industry year observations. Weighted regressions, where weight is number of industry-year full-quarter jobs for the hire type. Worker and firm demographic variables are deviations from pooled means. All regression coefficients have an estimated t-statistic greater than 2.

ment weights in each regression reflect the share of the hires type of the dependent variable for that industry relative to the economywide total. This implies that the mean of the dependent variable is the earnings for that hires type in the overall economy, and the variances of the dependent variable replicate the between-industry variances in the top right panel of figure 2.4. Table 2.2 presents estimates from these specifications. We report the time invariant pooled estimated coefficients from equation (2.3a). In the botWe acknowledge that the education variable in the LEHD is mostly imputed— Vilhuber (2018) reports that 92 percent of Protected Identification Keys (PIKs) have an education impute. Earnings is one of the variables used to impute education, which limits the value added of this variable in accounting for rising variance of earnings. Formally, this implies we are controlling for the covariance between education and earnings in our analysis. We include this variable in the main specification since our focus is on the hires type variables and we seek to understand the impact of those variables even after controlling for a rich set of firm and worker controls. In unreported results, we find that many of the basic patterns reported in this section are robust to the exclusion of this variable, and if anything, the relative effect of the changing job ladder contribution (i.e., the hires types) is even larger without including education.

60

John Haltiwanger and James R. Spletzer

tom of table 2.2, we report the variance decompositions that are based on equation (3c). All of the specifications include controls for firm and worker demographics in an industry. These demographic variables have the expected effects (for all hires types): industries with older workers have higher earnings, industries with more females have lower earnings, and industries with higher educated workers have higher earnings. On the firm side, industries with larger firms and younger firms have higher earnings.10 We find broadly similar patterns for the relationship between the shares of hires types in the industry and earnings for each hires type. Industries with a higher share of employer-to-employer flows (especially from job switchers between industries) have higher earnings for stayers, job switchers within industries, job switchers from other industries, and hires from nonemployment (these represent the pooled time invariant s in equation (2.3a)).11 We also find that industries with a higher share of hires from nonemployment have lower earnings for stayers, job switchers from the same industry, job switchers from different industries, and hires from nonemployment. While there are some quantitative differences across hires types, our conclusion is that the hires shares in an industry have basically similar effects on the earnings of each hires type. The finding that the factors influencing earnings of each hires type at the industry level are quite similar is interesting in its own right. These patterns are consistent with our interpretation of a job ladder with earnings for all hires types being higher in industries with a high share of hires from employer-to-employer flows and lower in industries with a high share of hires from nonemployment. It is striking, for example, that earnings for stayers are higher in industries with a larger share of hires from employerto-employer flows, and similarly, earnings for stayers are lower in industries with a larger share of hires from nonemployment. This is consistent with top-of-the-job-ladder industries paying higher wages for all workers. But it may also reflect the type of competitive pressures discussed in Faberman and Justiniano (2015), wherein a higher pace of employer-to-employer flows puts upward pressure on wage growth within an industry. Given that the patterns are so similar for each of the hires type groups considered separately, it is not surprising that the first column of table 2.2 shows that overall earnings for an industry is higher with a larger share of employer-to-employer flows and lower for an industry with a higher share 10. The finding that earnings are higher at younger firms might seem surprising but in table 2.3 this is the marginal effect of firm age controlling for a rich set of other factors. We find that without the hires types controls that the marginal effect of firm age is positive. The relationship between earnings and firm age is not our focus but it is interesting that this effect flips sign once we control for hires types. 11. Given that we include an exhaustive set of hires types with the omitted group being stayers, the estimated effect of an increase in hires of a specific type can be interpreted as an increase in the share of hires from that type (since this estimated effect holds the hires of other types constant).

Rising Between-Firm Inequality and Declining Labor Market Fluidity

61

of hires from nonemployment. We exploit that finding below to dig into the findings in more detail. The lower panel of table 2.2 shows the results of JMP style decompositions. The results of these accounting decompositions are quite similar for each of the hires type groups and overall industry earnings. We find that taking into account both the changing distribution of characteristics including hires types and firm and worker demographics (the Xs) and the changing earnings differentials from these characteristics (the βs accounts for about 80 percent of the rising variance in interindustry earnings differentials.12 Overwhelming the positive contribution derives from the changing βs while the changing distribution of characteristics is a drag on rising interindustry earnings differentials. To dig into the patterns in table 2.2 in more detail, we focus on the results of the first column, using overall industry earnings (mean ln real earnings) as the dependent variable.13 Table 2.3 and figure 2.5 present additional results for this specification. Summary statistics in table 2.3 provide more information about the changing distribution of characteristics. Declining fluidity is evident in the second column with declining means of hires shares of employer-to-employer flows and from nonemployment. For the firm and worker demographics there is an increase over time in the age of workers and age of businesses as well as an increase in the average firm size. Of greater relevance for changing inequality is the fourth column showing changing dispersion in the characteristics. There is compression of dispersion in hires rates across industries accounted for mostly by compression of dispersion in hires from nonemployment and job switchers across industries. Thus, not only is there a decline in the average pace of fluidity but there is also declining less dispersion across industries. There is also a large decline in dispersion in education and firm size across industries. These patterns help explain the findings in table 2.2 about the negative contribution of the changing distribution of characteristics in the decompositions. Specifications 1a, 1b, and 1c in table 2.3 present estimates of equation (2.3a) with time invariant coefficients and only the hires types as explanatory variables. The specification in column 1a shows that industries with more hires have lower earnings, but as seen in column 1b, industries with more employer-to-employer hires have higher earnings and industries with more hires from nonemployment have lower earnings. Column 1c shows that industries with more job switchers from other industries have especially 12. We use changing βs as a label for the combined contribution of changes in δs and χs and changing Xs as a label for the combined contribution of changing Hkt s and Dkts. In table 2.3, we provide guidance of the marginal contribution of the hires type variables in terms of both changing differentials and changing characteristics. Even there we use the same type of placeholder labeling. 13. In unreported results we have found the patterns we discuss from table 2.3 and figure 2.5 are broadly similar for all hires types.

−21.3 63.0 58.3

% contribution Changing X Changing β Residual

−93.5 125.0 68.5

−0.101 0.034 0.074 0.108

0.615

No No

−6.875

8.951

9.869

(1b)

−74.1 104.6 69.4

−0.080 0.033 0.075 0.108

0.633

No No

5.134 10.80 −6.751

9.824

(1c)

−43.5 102.8 40.7

−0.047 0.064 0.044 0.108

0.746

No No

2.172 −1.272 7.658 0.039 0.046

9.004

(2)

−113.0 191.7 21.3

−0.122 0.085 0.023 0.108

0.839

No No

0.913 −1.068 5.583 −0.263 0.054

4.861 5.030 −4.165

9.566

(3)

−96.3 181.5 14.8

−0.104 0.092 0.016 0.108

0.903

Yes Yes

0.510 −0.851 6.555 −0.244 0.062

1.761 3.653 −3.398

9.459

(4)

Notes: Dependent variable is LN real full-quarter earnings. Mean of the dependent variable is 9.004 (standard deviation = 0.576). N = 6384 industry-year observations. Weighted regressions, where weight is number of industry-year full-quarter jobs. Worker and firm demographic variables are deviations from pooled means. All regression coefficients have an estimated t-statistic greater than 2.

−0.023 0.045 0.063 0.108

Variance growth Predicted X(t) ∗ β Predicted X(t) ∗ β(t) Residual Total

0.418

R-squared

0.009 −0.004 −0.014 0.049 −0.054 No No

0.093 0.207 0.045 0.290 1.590

Year dummies 2-digit industry

0.079 0.018 −0.006 0.515 0.394

0.000 0.000 0.000 0.000 0.000

−0.009 −0.002 0.002 −0.004 −0.011

LN(worker age) female LN(education) LN(firm age) LN(firm size)

0.102 0.027 0.017 0.022 0.083

−3.632

−0.055 −0.009 −0.001 −0.008 −0.046

(1a)

0.325 0.087 0.026 0.061 0.238

ΔStd. dev.

Hires Hires E2E Hire same industry Hire different industry Hire non-FQ employment

Std. dev. 10.18

ΔMean

Intercept

Mean

Table 2.3 Regressions and decompositions using industry-by-year earnings

Rising Between-Firm Inequality and Declining Labor Market Fluidity

Fig. 2.5

63

Year-specific coefficient estimates from earnings regressions

high earnings. Industries with a larger share of hires from nonemployment have lower earnings. Specification 2 of table 2.3 shows the results from only using the firm and worker demographic controls. Specification 3 repeats the results from table 2.2 for overall earnings. We also consider a specification in 4 which includes year effects and two-digit industry dummies (we could not estimate the yearspecific regressions if we included four-digit industry dummies). The basic patterns are robust to the inclusion of these additional controls. Figure 2.5 presents the estimated year-specific coefficients from specification 3 of table 2.3— these are the coefficient estimates tj and tj from equation (2.3b). Figure 2.5a shows the coefficients of the hires type variables. The coefficients on both of the employer-to-employer hires variables, hires from the same industry and hires from a different industry, are positive and increasing over time. On the other hand, the year-specific coefficients for hires from nonemployment are negative declining over time, from −3.7 in 1998 to −5.0 in 2018. Figure 2.5b presents the estimated year-specific coefficients for the worker

64

John Haltiwanger and James R. Spletzer

and firm demographic variables. The education coefficient is on the right axis, and all other coefficients are measured on the left axis. The education coefficients are increasing over time, from 3.8 in 1998 to 7.9 in 2018. The other worker and firm demographic coefficients are not changing much over time. The coefficients on worker age increase from 1.004 in 1998 to 1.172 in 2018 (the coefficient on worker age spikes in 2011 for reasons we do not fully understand), and the coefficients on female gradually decline from −0.887 in 1998 to −1.220 in 2018. The coefficients on firm age and firm size are essentially invariant over time. The lower half of table 2.3 presents the results from the JMP variance decompositions. We are particularly interested in quantifying the marginal contribution of the hires type variables. We find that without firm and worker demographic controls (specification 1c), the combined contribution of changing distribution of hires types along with the changing pattern of earnings differentials by hires types accounts for 30 percent of the rising dispersion in interindustry earnings differentials. The analogous contribution of combined characteristics and changing prices for firm and worker demographics (specification 2) accounts for as much as 60 percent of rising dispersion in interindustry earnings differentials. Together, hires types and firm and worker demographics account for about 80 percent of rising interindustry earnings differentials. The latter differs from the “implied” 90 percent from adding up the separate contributions and reflects covariance effects in the accounting decompositions. Overall, then, we find that the marginal contribution of the hires type variables in accounting for rising between-industry inequality is about 20 percent (with firm and worker demographic controls) to 30 percent (without firm and worker demographic controls). As noted above, this positive contribution is overwhelmingly coming through the changing “prices”— the δts of equation (2.3b). We interpret the regression results and variance decompositions through the lens of a changing job ladder over time. Consistently tables 2.2 and 2.3 and figure 2.5 show that industries with a larger share of hires from nonemployment are low-earnings industries. In addition, figure 2.5 shows that the negative earnings differential associated with these bottom-of-the-ladder industries is growing in magnitude over time. In contrast, figure 2.5 shows that the top-of-the-ladder industries have a growing positive differential.14 2.7

Concluding Remarks

Rising earnings inequality in the last few decades is dominated by rising between-firm inequality. In turn, rising between-firm inequality is domi14. The online appendix (http://www.nber.org /data-appendix /c14447/appendix.pdf) section D includes supplementary analysis of selected industries. We show that industries such as software publishers are at the top of the ladder in terms of average and growing earnings differentials. In contrast, industries such as grocery stores are at the bottom of the ladder in terms of average and decreasing relative earnings differentials.

Rising Between-Firm Inequality and Declining Labor Market Fluidity

65

nated by rising interindustry earnings differentials. Over this same period, there has been declining labor market fluidity. The pace of hires and separations has slowed. Viewed from the perspective of hires, there has been an especially large decline in the pace of hires from nonemployment. We present evidence that these patterns are connected through the lens of a changing job ladder. Stated simply, our results suggest it has become more difficult to get on the job ladder, as evidenced by the declining hires from nonemployment. Moreover, the rungs of the job ladder have moved further apart as evidenced by the year-specific coefficients on both of the employer-to-employer hires variables, which are increasing over time, as well as by the year-specific coefficients for hires from nonemployment, which are declining over time. The widening of the rungs of the ladder is also evident in the rising between-firm and between-industry differentials. In combination, our results suggest there has been an increase in inequality accompanied by a decline in an important form of economic mobility— that is, it has become more difficult to get on and climb the job ladder. We view our results as exploratory, with many open questions. We have focused on rising interindustry earnings differentials since rising betweenindustry dispersion accounts for much of the rising between-firm dispersion in earnings. The finding of rising interindustry earnings differentials is important since it implies that the structural change underlying rising earnings inequality is working through mechanisms that change the structure of industries. This points toward looking more intensively at changes in technology, globalization, and market structure that vary across industries. Identifying these industry-specific driving forces should be a high priority for future research. There is also rising between-firm dispersion within industries that deserves further attention. In principle, the approach we have taken here can be used at the firm level for exploring within-industry rising between-firm dispersion. In companion research (Haltiwanger and Spletzer 2020), we have found that the rising interindustry earnings differentials are almost completely accounted for by occupation effects. The latter reflect differences across industries in the changing mix of occupations as well as changing differentials for occupations that vary widely across industries. These findings are consistent with the findings of Acemoglu and Autor (2011) and related literature highlighting the increasingly important role of changing tasks and changing returns for tasks. Our contribution in this companion research is to show that that the changing role of occupations is working primarily through rising interindustry earnings differentials. An open question is how to relate this occupation/task-based perspective with the findings in this chapter. The job ladder is changing over time and we find this is closely connected to rising interindustry earnings differentials. Getting on the job ladder has become more difficult and the earnings differential for starting at the bottom of the ladder has declined. Presumably, our findings on the changing job ladder can be related to the changing relative

66

John Haltiwanger and James R. Spletzer

demand for occupations and tasks. Understanding this connection should be an important area for future research.

References Abowd, John M., Kevin L. McKinney, and Nellie L. Zhao. 2018. “Earnings Inequality and Mobility Trends in the United States: Nationally Representative Estimates from Longitudinally Linked Employer-Employee Data.” Journal of Labor Economics 36 (51): S183– S300. Acemoglu, Daron, and David H. Autor. 2011. “Skills, Tasks and Technologies: Implications for Employment and Earnings.” In Handbook of Labor Economics, vol. 4, edited by Orley Ashenfelter and David Card, 1043– 171. Amsterdam: Elsevier-North Holland. Barth, Erling, Alex Bryson, James C, Davis, and Richard Freeman. 2016. “It’s Where You Work: Increases in the Dispersion of Earnings across Establishments and Individuals in the United States.” Journal of Labor Economics 34 (2, pt. 2): S67– S97. Burdett, Kenneth, and Dale T. Mortensen. 1998. “Wage Differentials, Employer Size, and Unemployment.” International Economic Review 39 (2): 257– 73. Davis, Steven J., R. Jason Faberman, and John C. Haltiwanger. 2012. “Labor Market Flows in the Cross Section and Over Time.” Journal of Monetary Economics 59 (1): 1– 18. Davis, Steve J., and John Haltiwanger. 1991. “Wage Dispersion between and within U.S. Manufacturing Plants, 1963– 86.” Brookings Papers on Economic Activity: Microeconomics (Spring): 115– 200. Davis, Steven J., and John Haltiwanger. 2014. “Labor Market Fluidity and Economic Performance.” NBER Working Paper No. 20479. Cambridge, MA: National Bureau of Economic Research. (Published in the 2015 Federal Reserve Bank of Kansas City Jackson Hole Conference Symposium.) Davis, Steven J., John Haltiwanger, Ron Jarmin, and Javier Miranda. 2007. “Volatility and Dispersion in Business Growth Rates: Publicly Traded versus Privately Held Firms.” NBER Macroeconomics Annual 2006 21: 107– 80. DiNardo, John, Nicole M. Fortin, and Thomas Lemieux. 1996. “Labor Market Institutions and the Distribution of Wages, 1973– 1992: A Semiparametric Approach.” Econometrica 64 (5): 1001– 44. Dunne, Timothy, Lucia Foster, John Haltiwanger, and Kenneth R. Troske. 2004. “Wage and Productivity Dispersion in United States Manufacturing: The Role of Computer Investment.” Journal of Labor Economics 22 (2): 397– 429. Faberman, R. Jason, and Alejandro Justiniano. 2015. “Job Switching and Wage Growth.” Chicago Fed Letter 337. Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2010. “Decomposition Methods in Economics.” NBER Working Paper No. 16045. Cambridge, MA: National Bureau of Economic Research. Haltiwanger, John C., Henry R. Hyatt, Lisa B. Kahn, and Erika McEntarfer. 2018. “Cyclical Job Ladders by Firm Size and Firm Wage.” American Economic Journal: Macroeconomics 10 (2): 52– 85. Haltiwanger, John, Henry R. Hyatt, and Erika McEntarfer. 2016. “Do Workers Move Up the Firm Productivity Job Ladder?” Unpublished paper. Haltiwanger, John, Henry R. Hyatt, and Erika McEntarfer. 2018. “Who Moves Up the Job Ladder?” Journal of Labor Economics 36 (S1): S301– S336.

Rising Between-Firm Inequality and Declining Labor Market Fluidity

67

Haltiwanger, John, and James Spletzer. 2020. “Between Firm Changes in Earnings Inequality: The Dominant Role of Industry Effects.” NBER Working Paper No. 26786. Cambridge, MA: National Bureau of Economic Research. Handwerker, Elizabeth Weber, and James R. Spletzer. 2016. “The Role of Establishments and the Concentration of Occupations in Wage Inequality.” Research in Labor Economics 43: 167– 93. Hyatt, Henry R., and James R. Spletzer. 2013. “The Recent Decline in Employment Dynamics.” IZA Journal of Labor Economics 2 (3): 1– 21. Hyatt, Henry R., and James R. Spletzer. 2017. “The Recent Decline in Single Quarter Jobs.” Labour Economics 46: 166– 76. Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. “Wage Inequality and the Rise in Returns to Skill.” Journal of Political Economy 101 (3): 410– 42. Molloy, Raven, Riccardo Trezzi, Christopher L. Smith, and Abigail Wozniak. 2016. “Understanding Declining Fluidity in the U.S. Labor Market.” Brookings Papers on Economic Activity (Spring): 183– 259. Moscarini, Giuseppe, and Fabien Postel-Vinay. 2013. “Stochastic Search Equilibrium.” Review of Economic Studies 80 (4): 1545– 81. Song, Jae, David J. Price, Fatih Guvenen, Nicholas Bloom, and Till von Wachter. 2019. “Firming up Inequality.” Quarterly Journal of Economics 134 (1): 1– 50. Topel, Robert H., and Michael P. Ward. 1992. “Job Mobility and the Careers of Young Men.” Quarterly Journal of Economics 107 (2): 439– 79. Vilhuber. Lars. 2018. “LEHD Infrastructure S2014 Files in the FSRDC.” Center for Economic Studies Working Paper No. CES-18– 27R. Washington, DC: Center for Economic Studies. https://www2.census.gov/ces/wp/2018/CES -WP-18-27R .pdf.

3

United States Earnings Dynamics Inequality, Mobility, and Volatility Kevin L. McKinney, John M. Abowd, and John Sabelhaus

3.1

Introduction

Using data from the US Census Bureau’s Longitudinal EmployerHousehold Dynamics (LEHD) infrastructure files, we study changes over time and across subnational populations in the distribution of real labor earnings and earnings dynamics. At the national level, LEHD administrative data has been used to show earnings inequality is increasing, while worker mobility is declining (Abowd, McKinney, and Zhao 2018; hereafter AMZ). In addition, overall earnings volatility is declining in administrative data (Bloom et al. 2017; Sabelhaus and Song 2010), but earnings volatility of workers with weak labor force attachment is increasing (McKinney and Abowd 2019). Although these national-level trends are well established, relatively little is known about earnings inequality, mobility, and volatility at subnational geographies. This chapter is a first step in that direction, using LEHD data to study earnings distributions and earnings dynamics Kevin L. McKinney is an economist at the US Census Bureau. John M. Abowd is the Chief Scientist and Associate Director for Research and Methodology at the US Census Bureau, on leave from Cornell University John Sabelhaus is a nonresident senior fellow at the Brookings Institution and adjunct research professor at the Survey Research Center of the University of Michigan. Any opinions and conclusions expressed in this chapter are those of the authors and do not represent the views of the US Census Bureau or other sponsors. All results have been reviewed to ensure that no confidential information is disclosed (DRB clearance number CBDRB-FY20CED006– 0013). This research uses data from the US Census Bureau’s Longitudinal EmployerHousehold Dynamics Program, which was partially supported by NSF grants SES-9978093, SES-0339191, and ITR-0427889; National Institute on Aging grant AG018854; and grants from the Alfred P. Sloan Foundation. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber .org /books-and-chapters/measuring-distribution-and-mobility-income -and-wealth /united -states-earnings-dynamics-inequality-mobility-and-volatility.

69

70

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.1 Output growth, unemployment, employment growth, and earnings growth by MSA, 2001 to 2018

across four large metropolitan statistical areas (MSAs) over the period 1998 through 2017. The results exemplify the sorts of analyses that will be possible with a new data exploration tool— the Earnings and Mobility Statistics (EAMS) web application— currently under development at the US Census Bureau. Disaggregating earnings distributions and earnings dynamics by geography is motivated in large part by observed differences in economic and labor market conditions across local areas. Figure 3.1 shows a wide range of outcomes for real GDP, unemployment, employment, and real annual earnings during our study period across the four MSAs (Detroit, Los Angeles, New York, and San Francisco) we consider in this chapter. All four MSAs show the negative effects of the Great Recession and subsequent slow recovery, but the size of the shocks and postrecession trajectories

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.1

71

(cont.)

differ substantially. For example, Detroit experienced larger labor market and output shocks than the other three areas, from which they have been slower to recover, while San Francisco experienced less of a shock, followed by a much stronger recovery in employment and earnings. There are also clear differences in the prerecession economic conditions across MSAs, with Detroit experiencing notably high unemployment rates and slow output and earnings growth in the period 2001 through 2007, relative to the other areas and the overall national average. The differences in output, employment, and earnings across MSAs can be cautiously interpreted in terms of the same economic and demographic factors generally put forth as explaining rising earnings inequality and wage polarization. For example, Detroit and San Francisco are thought to be representative of two distinct types of local economies. Detroit is gener-

72

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

ally characterized as manufacturing-oriented, and thus more exposed to the direct effects of import penetration and automation. The persistent decline in manufacturing employment and consequent increase in the relative supply of lesser-skilled employment has arguably combined with skill-biased technical change to limit earnings growth. San Francisco is generally characterized as emblematic of a local economy dominated by booming high-tech industries, and thus much less exposed to those same forces. What is not clear is whether there are differences in labor market outcomes between Detroit and San Francisco for otherwise similar workers. For example, earnings and employment outcomes for high school– educated males at the national level are deteriorating generally. Is this because workers in that group are more concentrated in areas such as Detroit where they are much worse off? Is it possible the same demographic group in San Francisco is only slightly worse off or even experiencing earnings growth more in line with the rest of the population? Although the four overall measures of economic outcomes in figure 3.1 are suggestive of underlying factors driving earnings inequality, mobility, and volatility, the measures are incomplete. For example, starting in 2012 real average earnings in San Francisco grew faster than the rest of the country generally— and Detroit in particular— but that could be due to very rapid growth at the top of the earnings distribution. Alternatively, is upward mobility more prevalent throughout the entire earnings distribution, meaning a rising local area tide is lifting all boats? Overall differences in employment and output growth across MSAs lead to another set of questions about the role of entry and exit into the paid labor force. Detroit saw a huge drop in employment during the Great Recession relative to the other MSAs and the national average, but since 2012 it has seen similar employment growth rates. How much of the differences in levels is due to (presumably low or negative) population growth and how much is due to persistently lower labor force participation? Questions about what is driving the overall labor force and earnings outcomes in figure 3.1 at the local level can be answered with the LEHD data using an empirical approach recently developed and implemented at the national level by AMZ. The LEHD data begin with the universe of jobs, and AMZ show that limiting the universe to observations with valid social security numbers (SSNs) effectively transforms the LEHD data from a “found” to a “designed” frame. AMZ show that the designed LEHD frame tracks the trends (if not the levels) in the data sets commonly used to study earnings inequality, such as the Current Population Survey (CPS) and American Community Survey (ACS). In addition, the scale, scope, and longitudinal structure of the LEHD data make it possible to study earnings dynamics in ways that are not possible with the CPS or ACS. For example, the patterns of earnings volatility in the LEHD data reported by McKinney and Abowd

US Earnings Dynamics: Inequality, Mobility, and Volatility

73

(2019) are shown to track the volatility patterns based on Social Security Administration earnings data in Bloom et al. (2017). The fixed real earnings “bin” is the key methodological building block in the AMZ empirical approach to studying earnings inequality, mobility, and volatility, and we take the same approach here. Most other analyses of earnings inequality are based on relative distributions, for example, considering the average earnings within a given distributional fractile, or the ratio of (say) the 90th to the 10th percentile cutoff. That approach is useful for describing trends in earnings levels within a given population but it is less useful for studying earnings dynamics or comparing outcomes across subpopulations. Percentile cutoffs can be problematic because they vary over time and across subpopulations in ways that may be correlated with the phenomenon being studied. For example, a drop in employment among previously low earners will shift all percentile cutoffs up and make it appear (erroneously) as though earnings have become more equal, when in fact the previously low earners are now much worse off. Establishing a fixed overall earnings distribution based on all time periods and subpopulations makes it possible to observe and evaluate where in the earnings distribution there are differences across subpopulations and at different points in time. Does San Francisco have higher mean earnings growth than Detroit because workers are generally shifting to the right across all or most fixed earnings cells or is it the case that earnings in San Francisco are just becoming more skewed, meaning the binned employment distributions are stable but earnings within the top earnings cell are increasing? Fixing the reference earnings distribution also makes it possible to disaggregate the source of the change across distributional fractiles. Is the flow between unemployment/nonparticipation and various earnings fractiles the same across MSAs or (for example) is someone who loses a job in Detroit more likely to remain out of the labor force? Also, are the positive and negative flows somehow different, meaning (for example) Detroit sees much more earnings-reducing job destruction than other MSAs? The LEHD data enable drilling down into the published MSA-level GDP, unemployment, employment, and earnings statistics to provide some preliminary answers to these overarching questions. We present standard measures of earnings inequality, such as the Gini coefficient, but the mixed signals (inverted U-shape between 1998 and 2017 but generally little changed on net over the entire period in all four MSAs) could reflect offsetting movements in different parts of the earnings distributions. Therefore, we also look at pairwise discretized earnings densities within and across MSAs and find both common cyclical components and divergent longer run trends. Consistent with the overall macro charts (figure 3.1) all four MSAs experienced large employment and output shocks in the Great Recession, and that is reflected in earnings distributions (for those who are employed) that are

74

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

essentially unchanged between 2007 and 2011. Earnings distributions are shifting steadily to the right in the prerecession period in all four local areas, though to different degrees. In the postrecession period, only San Francisco has seen anything like a resumption of prerecession widespread earnings growth across the entire earnings distribution. Conventional inequality measures and univariate earnings distributions capture the earnings only of the employed; hence, those statistics fail to capture the distributional impacts of cyclical downturns associated with increased transitions to unemployment. The LEHD data permit the analysis of earnings mobility as we can track workers as they move in and out of employment covered by unemployment insurance (UI). We find both trend differences and common cycles in the entry and exit rates across our four MSAs. The most obvious commonality is in the cyclical entry to and exit from UI-covered employment, as exits from the UI-covered employment sector surged in 2008 and 2009, while rates of entry to covered UI employment fell. Rates of entry (which include reentry of those who moved to inactivity in 2008 and 2009) rose only slowly thereafter, consistent with a slow decline in unemployment and the prolonged declines in measured labor force participation in the wake of the Great Recession. On net, by the end of the study period in 2017, the number of workers entering and exiting paid employment had generally converged back to the 1998– 99 levels in most of the MSAs we study here, except in Detroit, where inflows and outflows were each about 20 percent below the base period. Earnings mobility and earnings volatility are complementary ways to characterize longitudinal earnings dynamics of the continuously employed. In the fixed real earnings bin methodology, mobility is the movement between earnings bins measured over some time period. We disaggregate workers into mobility types in a given year using distinct mobility paths, such as the transition from earnings bin 1 to earnings bin 2, earnings bin 1 to earnings bin 3, and so on. This mobility path approach makes it possible to address, for example, how the longer-term earnings of workers who experienced a negative earnings shock in a given year compared to workers who were in the same base period real earnings bin but did not experience the shock. The different mobility paths are also key to understanding declining earnings volatility for all four MSAs. Some mobility paths are associated with substantial volatility as they involve economically meaningful earnings changes (say, bin 5 to bin 1, or vice versa) but in fact, overall volatility is dominated by the effects of large percentage movements in relatively low earnings. Workers who remained in the lowest real earnings bin (below $18,000 annually) in two adjacent periods account for roughly 25 percent of overall earnings volatility over the study period. Our MSA-level observations about earnings inequality, mobility, and volatility complement the growing literature on how substantial geographic differences in economic outcomes in the US have important implications

US Earnings Dynamics: Inequality, Mobility, and Volatility

75

for labor market and macroeconomic policies. Abel and Dietz (2019) look at earnings distributions across select MSAs (including San Francisco and Detroit) using Census and ACS data, and find that earnings growth in San Francisco exceeded earnings growth in Detroit at every percentile of the earnings distribution over the period 1980 to 2015. Our findings are consistent with the Abel and Dietz paper in focusing attention on the role of better overall local labor market conditions and/or agglomeration, as opposed to fundamentals such as schooling or other human capital considerations. Other subnational labor market research has focused attention on international trade, housing, and even monetary policy, with an emphasis on how some initial shock or policy innovation generates spillovers that dominate local labor market outcomes. For example, one well-known paper considers how increased international trade differentially impacted local economies. Autor, Dorn, and Hanson (2013) use local labor market data to show substantial negative impacts from rising import penetration in areas where production was more concentrated in import-sensitive industries. More importantly, they show that there are substantial adjustment costs and second-round employment effects associated with import-related job destruction, and that fully considering those costs might substantially change one’s views about the gains from trade and the overall value of cheap imports. Housing policy also became a prominent policy topic in the Great Recession, especially given substantial differences in outcomes across subnational areas, and again the implications for local labor markets are key. Mian, Rao, and Sufi (2013) focus on the role of the housing boom and bust in determining regional labor market outcomes through both collateral and wealth channels. The key insight is that— and this is independent of what caused the housing boom and bust in the first place— carefully tracking outcomes in tradable and nontradable consumer goods across regions shows how a wealth shock can have disproportionate negative effects on a local economy. The extent to which the shock is distributed to other local labor markets depends on the extent to which local production is tradable. For example, a worker employed in the restaurant sector in a local area where tradable production declines is likely to be severely impacted, as the workers in the tradable sector cut back on their restaurant spending. Monetary policy has also been shown to have important differential geographic impacts, depending on local economic conditions. Beraja et al. (2018) show that the effects of expansionary monetary policy in the wake of the financial crisis varied by regions because of differences in loan-to-value ratios and other initial conditions. Similarly, Beraja, Hurst, and Ospina (2019) use regional data on employment and wages to separate the effects of shocks (aggregate demand and labor force participation) from the effects of wage stickiness in the Great Recession and find support for the idea that Phillips Curve principles may be operative regionally but that the relation-

76

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

ship between labor market tightness and wage growth is not observed at the national level because of vast differences by geography. These sorts of findings are consistent with what we see in the MSA-level LEHD earnings inequality, mobility, and volatility. It is likely that the different parts of the US have simultaneously experienced very different trend and cyclical phenomena, and thus different fiscal (and even monetary) policies across regions may be warranted. Indeed, Austin, Glaeser, and Summers (2018) characterize these issues in terms of “place-based” policies, arguing, for example, that policies focused on nonemployment are likely to have more bang for the buck in areas with high (and perhaps rising) rates of nonemployment. In addition to directly contributing to the literature on regional economic differences, the other important contribution of this paper is to lay the foundation for a new data dissemination application under development at the US Census Bureau. The EAMS data extraction tool will complement several other tools made available to Census Bureau data users in recent years. These other tools include the Quarterly Workforce Indicators (QWI), Job to Job (J2J) Employment Flows, LEHD Origination Destination Employment Statistics (LODES), and most recently, the Post-Secondary Employment Outcomes (PSEO).1 As in those other applications, users will be able to disaggregate labor market outcomes by a number of characteristics and display the results in many possible ways. Although our focus in this chapter is on subnational geography, we are investigating the feasibility of including demographic and firm characteristics from the LEHD infrastructure in the EAMS web application. This implies, for example, that users could see labor force entry/exit or movement across earnings bins disaggregated by age and gender. The remainder of the chapter proceeds as follows. In section 3.2 we describe the LEHD infrastructure, focusing on the particular criteria used to decide which LEHD records are included in the EAMS database generally, and the four MSAs here in particular. Section 3.3 turns to measures of inequality, including both conventional summary statistics, such as the Gini coefficient and top earnings shares, and much more detailed perspectives from (for example) discretized univariate earnings distributions. Section 3.4 focuses on earnings mobility, including average earnings dynamics among continuously employed workers based on their mobility paths across earnings bins, as well as movements into and out of paid employment. Section 3.5 builds on the mobility analysis and shows how earnings volatility varies across and along various earnings mobility paths and how the volatility of earnings along any given mobility path contributes to overall earnings volatility. Section 3.6 concludes. 1. See Abowd et al. (2009) for a discussion of the QWI; Hyatt et al. (2014) for a discussion of J2J; and Foote, Machanavajjhala, and McKinney (2019) for a discussion of PSEO.

US Earnings Dynamics: Inequality, Mobility, and Volatility

3.2

77

Data and Methods

The empirical work in this chapter uses job-level earnings information from the LEHD infrastructure files, developed and maintained by the US Census Bureau.2 In the LEHD data infrastructure, a “job” is the statutory employment of a worker by a statutory employer as defined by the UI system in a given state. Mandated reporting of UI-covered wage and salary payments between one statutory employer and one statutory employee is governed by the state’s UI system. Reporting covers private employers and state and local government. There are no self-employment earnings unless the proprietor draws a salary, which is indistinguishable from other employees in this case. The LEHD program is based on a voluntary federal-state partnership. When a state becomes a member of the partnership, current as well as all available historical data for that state are ingested into the LEHD internal database. By 2004, LEHD data represent the complete universe of statutory jobs covered by the UI system in the US. However, studying joblevel inequality— the task for which having a complete job frame is well suited— as a proxy for person-level inequality may be misleading due to the time-varying many-to-one assignment of jobs to workers. Therefore, we use all jobs to construct person-level annual real earnings (2017 Consumer Price Index for All Urban Workers) analysis files covering the period 1998– 2017.3 It is preferable to have both a person frame that covers a known population of interest and to have a relatively high level of confidence that the persons in that population use a consistent person identifier across all jobs. To that end we use the US Census Bureau’s edited version of the Social Security Administration’s master SSN database (the Numident) to create a set of “eligible” workers each year, removing annual earnings records for ineligible workers. The first condition is that an eligible worker must have an SSN that appears on the Numident. Second, each year an “eligible” worker must meet an additional set of conditions: age is between 18 and 70 (inclusive), is not reported dead, and has an active SSN. If the worker has reported earnings in a given year, the worker must also not have more than 12 reported employers during the year, otherwise we assume the SSN is being used by multiple persons and the annual earnings report is discarded. The overarching data selection and processing decisions here largely mirror AMZ, and the reader is referred to that paper for additional details. 2. See Abowd et al. (2009) for a detailed summary of the construction of the LEHD infrastructure. 3. Although our sample begins prior to the complete data period, none of the missing data states are highly connected to the four MSAs (Detroit [DT], Los Angeles [LA], New York [NY], and San Francisco [SF]) we study in this chapter.

78

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

However, there are a number of additional decisions and assumptions associated with analyzing subnational populations. Because the LEHD data use a job-level frame, locating a worker within a given subnational area (one of the four MSAs versus somewhere else in the country) involves mapping each job to an employer location. This is straightforward for single-establishment employers, where we use the location of the single establishment. Geolocating the job is more difficult for multiestablishment employers because the earnings data are reported at the employer level, not the establishment level in LEHD data. A statistical model is used to impute the location of each job in a multiestablishment employer to one of the physical locations of its establishments. For multiestablishment employers, we use the results of these imputation models to assign each worker to one of the firm’s establishments. Workers with multiple employers in a given year may also have work locations in more than one subnational area and, in that case, we assign the individual’s work location to the establishment at the employer with whom the worker had the highest earnings (dominant) in that year. Assigning subnational geography is also complicated when an individual is inactive. For example, we might observe a worker in paid employment in Detroit in a given year, but the same worker may no longer have positive reported earnings in the subsequent year. Although we have the complete set of statutory UI employment records for that individual, we do not know if the worker has entered self-employment, is inactive in Detroit, or inactive in some other subnational area. If and when the worker reappears with positive earnings in a subsequent year, we do not assume a location. Instead, the location of the worker is determined by the location of the dominant employer in the adjacent year. For example, if a worker reappears in Detroit after a year or more of inactivity, then the worker is a new entrant to the Detroit labor market, whether they actually left Detroit or remained in the MSA during the period with no reported UI earnings. For workers with a continuous work history, the location of the dominant employer allows us to observe both within and across MSA earnings mobility. Privacy is a substantial concern in studies involving disaggregated LEHD data or other large-scale administrative data sources. In this study, we avoid disclosure risk by limiting ourselves to four very large MSAs and report statistics for very large cells (annual earnings data with wide earnings bins). In the production version of EAMS, where the analysis cell counts and sums are likely to be much smaller, the approach will be to build on existing Census Bureau privacy protection methods and use noise infusion to mitigate the risks of unauthorized disclosure. For an overview of one approach to noise infusion, see Abowd and McKinney (2016). Also, Foote, Machanavajjhala, and McKinney (2019) discuss how to use differentially private noise infusion to estimate earnings distributions and quantiles for the Census PSEO public-use data dissemination tool.

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.2

3.3

79

Distributions of employment and earnings, all years (1998 to 2017)

Inequality

Our analysis of earnings inequality and earnings dynamics begins with the overall distributions of employment and earnings across five broad real earnings bins for the US and four large MSAs (DT, LA, NY, SF) over the entire 1998– 2017 study period (figures 3.2a and 3.2b). The five real earnings bins are $1– 18,000, $18,000– 54,000, $54,000– 96,000, $96,000– 132,000, and greater than $132,000. For the US as a whole, almost 75 percent of the person-year employment observations (figure 3.2a) are in the bottom two bins, a bit over 15 percent are in the third bin, and just under 5 percent of employment is in each of the two top earnings bins. Total earnings (figure 3.2b) skew very differently than employment, with only about 35 percent of total earnings in the first two bins, a bit over 25 percent in the third bin, and over 35 percent in the top two earnings bins combined. While most workers (almost 75 percent) are in the bottom two earnings bins, the 25 percent of workers in the top three earnings bins are responsible for about 65 percent

80

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

of total earnings. Perhaps even more striking is the just over one-third of person-year employment in the less than $18,000 real earnings bin accounted for only a bit over 5 percent of total earnings. The distributions of employment and total earnings within the four MSAs are broadly similar, but a closer look provides the first indication of how inequality differs at the subnational level. Relative to the US totals, all four MSAs have more person-year employment (figure 3.2a) in the higher earnings bins, consistent with higher earnings in larger MSAs generally. The differences at the very top are most prominent in New York and San Francisco, with Los Angeles not far behind. Detroit has a larger fraction of person-year employment than the US in the top three earnings bins, but the employment is more concentrated in the $54,000– 96,000 and $96,000– 132,000 bins. The same relative patterns are even more pronounced in the total earnings distributions (figure 3.2b). For example, the $132,000 and higher earnings bin accounted for over 40 percent of total in earnings in New York and San Francisco, but only 25 percent for the US as a whole. The Kullback-Leibler (K-L) statistic is a useful summary measure of how each of the MSA-level employment and earnings distributions diverge from the overall US distributions. The K-L statistics for employment (figure 3.3a) and real earnings (figure 3.3b) indicate substantial differences in both levels and trends across the four MSAs. In general, the employment and earnings distributions in Los Angeles and Detroit are most similar to the entire country, and the divergence between the MSA-level and national distributions is not changing substantially over time. The employment distributions in New York and San Francisco are generally more divergent from the national distribution, and the divergence in San Francisco increased dramatically after the Great Recession. The total earnings K-L statistics in New York and San Francisco are generally above the employment K-L statistics and trending up throughout the study period, indicating that in addition to New York and San Francisco having more workers in the higher real earnings bins, average real earnings in the top earnings bins are also higher, and the differences in average real earnings at the top are increasing over time.4 Although the measures of inequality shown here (and in the online appendix, http://www.nber.org/data-appendix/c14448/appendix.pdf) are informative, the LEHD data are rich enough to answer questions about differences between specific points in the real earnings distributions, whether between MSAs in a given year, or for the same MSA over time. For the next set of results, we discretize the real total earnings distribution into 25 earnings bins. The first 20 real earnings bins have a width of $6,000, the next four bins 4. The online appendix (http://www.nber.org /data-appendix /c14448/appendix.pdf) shows additional inequality measures for the same time period and MSAs, including mean real earnings, Gini coefficients, and ratios of top-to bottom earnings shares. In general, the additional measures are in line with the increasing inequality captured by the K-L statistics.

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.3

81

Kullback-Leibler measures of distributional divergence

have a width of $12,000, and the final bin captures yearly real earnings above $168,000. In figure 3.4 we plot pairwise densities for all four MSAs in 1998 and 2017 and for various years for each MSA separately in figures 3.5– 3.8. Before discussing the results, we remind the reader that figures 3.5– 3.8 are total earnings densities. For each bin, rather than sum the number of workers, we sum the earnings for all workers with real annual earnings greater than the minimum bin real earnings value and less than or equal to the top earnings value. Traditional earnings densities are often characterized as lognormal in shape, and the results for the discretized total earnings densities are roughly consistent with a mixture of a log-normal or a log-normal– like distribution with fatter tails (e.g., log-Student-t). The starting point for the density analysis is a comparison of all four

82

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.4

Total earnings densities, 1998 and 2017

MSAs, in the first and last years of our study period (figure 3.4).5 Figure 3.4a shows Detroit and Los Angeles in 1998. Los Angeles had more lowerpaying (less than $50,000) and more higher-paying (above $168,000) jobs than Detroit, indicated by the thin line above the thick line. Earnings in Detroit were more concentrated in the middle of the earnings distribution (between $50,000 and $100,000).6 Detroit, in 1998, had substantially more earnings equality than Los Angeles, because of the concentration of middleearnings jobs. San Francisco (figure 3.4b) was also a relatively equal MSA in 1998 (the Gini was well below New York and Los Angeles) for the same reason— a large fraction of earnings in the $50,000– 100,000 range. The four earnings distributions all shifted to the right between 1998 and 2017, though to very different degrees. Comparing Detroit and Los Angeles 5. Although the right tail in the total earnings density graphs ends at $300,000, all earnings values above $168,000 are included when calculating the density. 6. The differences in the bottom, middle, and top of the earnings distributions are consistent with summary statistics like the Gini coefficient (see online appendix figure A1, http://www .nber.org /data-appendix /c14448/appendix.pdf).

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.4

83

(cont.)

(figure 3.4c), the rightward shift in Los Angeles is more pronounced, with earnings in the middle of the distribution reallocated to the long right tail. In contrast, the changes in Detroit were relatively modest. The earnings distribution shifts in New York and San Francisco were more dramatic, with a substantial reduction of total earnings in the $50,000– 100,000 range and a corresponding greatly increased long right tail.7 Fluctuations over time in the summary statistics like the Gini coefficient and top-to-bottom share ratios over the study period indicate the rate of change in the shift to the right throughout the study period is not constant. Indeed, this is borne out by comparing discretized densities for each of the four MSAs in 1998, 2007, 2011, and 2017. Figures 3.5a, 3.6a, 3.7a, and 3.8a isolate the prerecession years (1998 and 2007); figures 3.5b, 3.6b, 3.7b, and 3.8b focus on the early years of the Great Recession (2007 and 2011); figures 3.5c, 3.6c, 3.7c, and 3.8c look at changes in the latter stages of the recovery 7. Again, these shifts can be tied to the summary statistics discussed above. See, in particular, the ratio of top to bottom earnings shares in online appendix figures A4 and A5.

84

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.5

Total earnings densities, Detroit, various years

(2011 and 2017); and figures 3.5d, 3.6d, 3.7d, and 3.8d show the change for the entire period (1998 and 2017). Detroit (figure 3.5) is clearly an outlier among the four MSAs, with relatively little change in the total earnings distribution over the period. There is a modest rightward shift in the middle of the distribution between 1998 and 2007 (figure 3.5a), subsequently reversed by a leftward shift during the recession years (figure 3.5b). There is very little change in the Detroit earnings distribution during the postrecession period 2011– 17 (figure 3.5c), and hence little overall change during the entire study period (figure 3.5d). The patterns of shifting earnings distributions in the other three MSAs during the study subperiods all tell a similar story, though with different magnitudes. In Los Angeles (figure 3.6), New York (figure 3.7), and San Francisco (figure 3.8), earnings were shifting to the right, and especially into the long right tail, between 1998 and 2007. During the Great Recession, earnings distributions essentially locked down, as in Detroit. The stability in the total earnings distributions is the result of lost jobs and labor force exits in the bottom half of the MSA earnings distributions, offset by a lack of growth in earnings at the top. Excluding workers who exited the labor

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.5

85

(cont.)

force after 2007 for economic reasons provides a distorted view of inequality. Earnings certainly became more unequal during the Great Recession. Limiting the population to workers with observed earnings obscures that fact.8 The postrecession differences in earnings density shifts across MSAs are also notable and help clarify some of the earlier summary inequality statistics. Los Angeles (figure 3.6c) and New York (figure 3.7c) are to a large extent similar to Detroit for the 2011– 17 period, with only a modest additional rightward shift in the earnings distributions. However, San Francisco (figure 3.8c) is a clear outlier, with a dramatic rightward shift in the earnings distribution. This is consistent with the dramatic rise in average earnings in San Francisco relative to the other MSAs after 2011 (see online appendix 8. Workers with zero earnings receive zero weight in a total earnings distribution. However, we discuss flows of workers into zero reported earnings status in the next section, and AMZ discuss these workers in even more detail. See AMZ, table 5, for a detailed accounting of the national net flows of eligible workers into no reported earnings status. For example, between 2007 and 2011 approximately 11 million eligible workers moved into no reported earnings status. AMZ also present parametric measures of earnings inequality that specifically take into account eligible workers with no reported UI earnings.

86

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.6

Total earnings densities, Los Angeles, various years

figure A1, http://www.nber.org /data-appendix /c14448/appendix.pdf), and the jump in the K-L divergence (figure 3.3). Indeed, the continued rightward shift in the San Francisco earnings distribution suggests very different labor market dynamics were in play across the entire distribution. 3.4

Mobility

Snapshots of earnings distributions and summary inequality statistics across years are a useful way to describe a given local economy at a point in time, but the static pictures tell us little about individual earnings dynamics. One recurring example from the previous section— the finding that earnings inequality seemed to fall or was stagnant during the Great Recession— is an artifact of earnings distributions and summary statistics excluding those who exited the labor market. Of course, the workers who suffered the biggest earnings losses during the Great Recession are excluded from measures such as the Gini, top shares, and earnings densities. As a result, those earnings losses are not captured in the traditional comparative snapshot approach.

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.6

87

(cont.)

The solution is to shift the perspective from static to dynamic, and to focus on employment and earnings mobility. Shifting to a dynamic perspective involves comprehensively tracking workers across earnings bins and nonemployment status. All workers in the mobility samples in this section meet the eligibility criteria described in section 3.2 in all of the periods considered for the given statistic. Thus, for example, a worker must be eligible in both year t and t + 1 and have positive earnings in at least one of the two years; workers who are not active both years are not included. We allocate workers within each MSA across the five real earnings bins used in the first figures in the previous section (figures 3.2a and 3.2b), along with eligible but inactive workers who are active in the adjacent year, and eligible active workers who transition to or from a different MSA.9 Thus, there are seven distinct possible bins for an eligible

9. The five real earnings bins are $1– 18,000, $18,000– 54,000, $54,000– 96,000, $96,000– 132,000, and greater than $132,000.

Fig. 3.7

Total earnings densities, New York, various years

Fig. 3.8

Total earnings densities, San Francisco, various years

90

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

worker in a given year: one of the five earnings bins, not active, and active in a different MSA.10 It is useful to begin with a high-level view of two-year mobility across the four MSAs in the base period, 1998. Eligible workers in 1998 experienced one of seven broadly defined earnings transitions. An individual could have stayed in the same earnings bin in 1999 (S), moved up to a higher earnings bin (U), moved down to a lower earnings bin (D), exited to inactivity (X), entered from inactivity (E), left the reference MSA for employment elsewhere (L), or moved into the MSA from elsewhere (M). At this very high level of aggregation, there is a great deal of commonality across MSAs in terms of mobility. In particular, about half of all workers in the four MSAs were in the same earnings bin in both 1998 and 1999. Flows in and out of activity within the given MSA were generally on the order of 5 percent of workers, and gross migration (inflows and outflows) were generally balanced, each between 5 and 10 percent of the population. Most workers who were continuously employed in an MSA between 1998 and 1999 but changed earnings bins experienced upward mobility (roughly 10– 13 percent) as opposed to downward mobility (roughly 6– 8 percent). Although the transition rates between 1998 and 1999 seem fairly homogeneous across MSAs, transition patterns evolved somewhat differently after 1999. To show this, we plot transitions for each year-pair 1999/2000, 2001/2002, . . . , 2016/2017, relative to the base 1998/1999 transitions (figures 3.9, 3.10, 3.11, and 3.12). For each MSA, we show whether the worker stayed in the same bin, moved up one or more bins, or moved down one or more bins earnings transitions (figures 3.9a, 3.10a, 3.11a, and 3.12a); entrants and exits from inactivity (figures 3.9b, 3.10b, 3.11b, and 3.12b); and leavers and movers to the reference MSA(figures 3.9b, 3.10b, 3.11b, and 3.12b). In each MSA/figure, a value of 1 for a given year-pair indicates that the number of workers experiencing that transition is identical to the number of workers who experienced that transition in 1998/99. Values above 1 indicate more workers experiencing the transition (relative to 1998/99) in the given year, and vice versa. There are both trend differences and common cycles in the relative mobility rates across MSAs. The most obvious commonality is in the entry and exit between paid employment and inactivity (b in each of the figures). Rates of exit to inactivity from paid employment were higher over most of the prerecession period and surged in 2008 and 2009 at the start of the Great Recession, while rates of entry from inactivity to paid employment fell. In addition, except for San Francisco there was no increase in the rate of return to paid employment from inactivity after 2010. Rather, rates of 10. In principle, it may eventually be possible to distinguish inactive workers who remained in an MSA from inactive workers who subsequently moved using other LEHD data. See the discussion in section 3.2.

Fig. 3.9

Mobility relative to base year, Detroit

Note: Base year 1998 shares are in parentheses.

Fig. 3.10

Mobility relative to base year, Los Angeles

Note: Base year 1998 shares are in parentheses.

Fig. 3.11

Mobility relative to base year, New York

Note: Base year 1998 shares are in parentheses.

Fig. 3.12

Mobility relative to base year, San Francisco

Note: Base year 1998 shares are in parentheses.

US Earnings Dynamics: Inequality, Mobility, and Volatility

95

entry (including reentry) rose quickly until 2010 and then stagnated or fell, consistent with a slow decline in unemployment and the prolonged declines in measured labor force participation in the wake of the Great Recession. On net, by the end of the study period, the number of workers entering and exiting paid employment had generally converged back to the 1998/99 levels, except in Detroit, where inflows and outflows were each about 20 percent below the base period. What happened to earnings for those who remained employed in each year/pair combination? It is important to keep in mind that the reference point for mobility among the continuously employed is the 1998/99 yearpair, and upward mobility in Detroit was relatively strong in that period (figure 3.9a) and thus all of the subsequent years have noticeably lower upward mobility. The more salient observation about upward mobility in Detroit is that the relative number of workers experiencing upward mobility in Detroit, fell from 2000 forward, compared to 1998, but the absolute number remained constant, as evidenced by the flat line. Conversely, the number of workers experiencing downward mobility was higher than in the base period between 2000 and into the Great Recession, but has since remained lower. Relative patterns of upward and downward mobility across the other MSAs differ to some extent, although there are similarities. For example, there are temporary offsetting movements in upward and downward mobility during the Great Recession, while unlike in Detroit the (relative) number of workers remaining in the same real earnings bin climbed steadily over the 20-year study period. While job destruction and the increased level of inactivity associated with recessions (deservedly) get most of the attention in the macro-labor literature, the cyclical decrease in upward mobility is also an important feature, because those who remained in paid employment were much less likely to see large earnings increases. And, although the fraction of continuously employed remaining in a given earnings bin from one year to the next is obviously dependent on the earnings bin specification, the general upward trend in “same bin” transitions and the lower level (relative) of up and down earnings mobility across MSAs is consistent with decreased wage dynamism. The final transition needed to complete our mobility taxonomy is leavers and movers for each MSA. Again, the fact that these are relative transitions should be kept in mind when evaluating Detroit over time in comparison to the other three MSAs. In the 1998– 99 reference period, Detroit experienced fairly high rates of leaving (to another MSA) and moving in (from another MSA). Somewhat counterintuitively, workers moving in also outpaced workers moving out in Detroit during the base year-pair, so the fact that both leaving and moving are lower after 2000 is less of a mystery than a first impression suggests. In general, geographic mobility is cyclical across MSAs, as rates of both leaving and moving declined during the Great Recession

96

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.13

Earnings dynamics between 2008 and 2012, San Francisco

in all four areas. The sense in which Detroit stands out is that geographic mobility did not increase after the Great Recession ended, as it did in the other three MSAs. Classifying mobility using the seven broad categories is a good starting point, but it is possible to drill down even further and investigate how, for example, mobility varies by where the worker started in terms of earnings bin, inactivity, or working in a different MSA. One can study how earnings dynamics (as measured by average earnings) interact with the starting point and mobility path. Mobility is more than a two-period concept, and it is also useful to investigate how multiperiod mobility differs from single-period mobility along a given dynamic path. Is there evidence of mean reversion or reinforcing positive or negative earnings shocks along a given path? These are the sorts of detailed questions which the new Census Bureau EAMS web application is being designed to answer, and we conclude this section with an example of how one might deconstruct earnings dynamics in a given MSA for a given time period. Our specific example is the San Francisco MSA for the years 2008 through 2012 (figure 3.13). Each subfigure (figure 3.13a– g) has two components, a

US Earnings Dynamics: Inequality, Mobility, and Volatility

Fig. 3.13

97

(cont.)

pie chart showing the fraction of workers along a given mobility path, and a line chart showing the average earnings of workers on that mobility path in each of the five years. The seven subfigures comprehensively capture the workers in the seven status bins as of 2008, where 0 represents inactivity (no earnings in 2008); bins 1, 2, 3, 4, and 5 represent the five fixed real earnings bins (less than $18,000 through $132,000 or more); and 6 represents workers

98

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.13

(cont.)

outside the reference MSA. A given transition path within any given subfigure is represented using bin numbers pairs, so “01” indicates the worker was in bin 0 (inactive) in 2008, and bin 1 (positive earnings, less than $18,000) in 2009. We will refer to that as the “01” mobility path. One gets a sense of how complex nonparametric analysis of earnings dynamics quickly becomes by first noting we need seven subfigures, each with two separate charts, simply to describe earnings paths and average earnings along those paths for one MSA in one base year. The first subfigure shows outcomes for workers who were in the inactive group (bin 0) in 2008. The pie chart shows that the most likely path for such workers who entered paid employment was by far entry into the lowest earnings bin— the “01” path. The second most likely path was 02, and the third was 03. Only a very small fraction of workers (too small to display) transitioned from inactivity to earnings bins 4 and 5 between 2008 and 2009. Conditional on entering a given earnings bin, the trajectory of average workers entering from inactivity were all positive. The immediate fanning out of average earnings is determined by the bin into which the worker entered, with the 01 group earning about $10,000 in 2009, the 02 group about

US Earnings Dynamics: Inequality, Mobility, and Volatility

99

$30,000 in 2009, and the 03 group about $70,000 in 2009. All three groups experienced continued average earnings growth between 2009 and 2012, though in relative terms the most substantial growth was for the 01 group, who saw their average earnings more than double during the four years after they entered from inactivity. Workers in the 03 group still saw substantial real gains, with average earnings approaching $100,000 by 2012. The earnings dynamics of workers who started the 2008 to 2012 period with positive earnings in 2008 confirm the findings on earnings stability noted earlier in this section, and the findings on similar trajectories in years three and beyond just noted for the inactive in 2008. The pie charts in figure 3.13b– f show that the majority of workers who had positive earnings in the San Francisco MSA in 2008 remained in the same earnings bin in 2009. Low earners were more likely to transition to inactivity in 2009, as indicated by the slices of the respective pie charts associated with the 10, 20, 30, 40, and 50 earnings paths. However, the line charts show that, conditional on experiencing a transition to inactivity, average earnings bounced back quickly for those workers after 2009. Average earnings for those experiencing inactivity in 2009 moved back into line with the levels and trajectories of average earnings for workers who remained in paid employment during 2009. One particularly interesting subset of earnings paths involves those who leave the reference MSA— in this case, San Francisco— in 2008, and immediately find paid employment in another MSA. These transitions are captured in the 16 path (figure 3.13b), 26 (figure 3.13c), 36 (figure 3.13d), 46 (figure 3.13e), and 56 path (figure 3.13f). These MSA leavers account for nearly a fifth of bin 1 earners in 2008, and about 10 percent of workers in bins 2– 5. In every case, the average earnings of MSA leavers track the average earnings of those who remain in their same earnings bin between 2008 and 2009. Average earnings are rising over time for workers in bin 1 who left San Francisco, and generally flat for workers in bins 2– 5 who left San Francisco, but in all cases they move in the same direction as those who stayed in the earnings bin (and likely the same job) but did not leave San Francisco. The final subfigure (figure 3.13g) captures movers to San Francisco in 2009. For these workers, we observe earnings in some other MSA in 2008, and thus we can bin their earnings as of 2008. As indicated by the pie chart, the majority of movers to San Francisco in 2009 were in the two lowest earnings bins, with about 40 percent in bin 1, and another 30 percent in bin 2. Thus, at least during the depths of the Great Recession, moving to San Francisco was not dominated by high-earning workers. In addition, the basically flat average earnings trajectories across origination earnings bins suggest that, again— at least during this time period— moving to San Francisco was not associated with observable upward changes in earnings trajectories. One has to look at the subset of high earnings— the 56-transition group— to see any positive earnings gains, and that is only in 2012.

100

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Fig. 3.14

3.5

One-year arc-percent change in real earnings

Volatility

Upward earnings mobility— as defined in the previous section— is an unambiguously desirable economic outcome. The more workers who move up the job ladder to higher-paying jobs from one year to the next, the better. Earnings volatility is a bit more nuanced, however. While it is desirable that workers should not be subject to increased uncertainty about their real annual earnings, measured overall earnings volatility will also decrease when upward mobility decreases. Thus, it is important to measure overall volatility, but then disaggregate that overall volatility using the same fixed earnings bins approach we have used to study inequality and mobility to get a sense of where in the mobility distribution measured volatility is most prominent. There are various ways to measure overall earnings volatility in a given year, and here we focus on the standard deviation of the one-year arc-percent change between the current and the subsequent year (figure 3.14). At the overall MSA level, there is a clear downward trend in earnings volatility over the study period, which is consistent with a continuation of the trends found in earlier studies (Bloom et al. 2017; McKinney and Abowd, 2019). As expected, the Great Recession is associated with a cyclical uptick in volatility, especially in Detroit, but the downward trend resumes after the recession in all four local economies. By 2016– 17, measured overall volatility is noticeably lower than in 1998– 99, especially in Detroit and New York. A decline in measured earnings volatility is a normatively good thing if it is associated with particular earnings trajectories. For example, if all workers are on a general upward earnings trend, then a decline in measured volatility around that trend is good news, because workers are achieving the same longrun earnings outcomes with less uncertainty. However, measured volatility can also decline because of a trend decline in upward mobility. Although overall measured earnings volatility increased during the Great Recession, earnings volatility moved in different directions at different points in the

US Earnings Dynamics: Inequality, Mobility, and Volatility

101

Fig. 3.15 Absolute value of the arc-percent change (bars) and the share of MSA year sum of squares (line)

earnings distribution (Bloom et al. 2017). Workers in the bottom half of the earnings distribution saw a spike in volatility associated with job loss, while workers in the top half of the distribution saw a decrease in volatility because real salary increases were very limited during the recession.11 It is unclear whether the continued decline in measured overall earnings volatility after the recession is being driven by a reversal of the volatility for low-earning workers or by a continued decline in upward mobility. Our particular measure of earnings volatility— the standard deviation of the one-year arc-percent change— is disproportionately influenced by large percentage changes in very low earnings. For example, workers in our real earnings bin 1 have total annual earnings of less than $18,000. A worker who moves from (say) $5,000 in one year to $15,000 in the next year contributes an arc-percent change of 1 or 100 percent to the overall average, even though that change is much less economically significant relative to another worker moving from $50,000 to $150,000, which contributes the same 100 percent to the overall average. One solution to this problem is to limit the sample to workers with earnings above a preset threshold but, as suggested by the mobility analysis above, this sort of sample exclusion reduces the impact of labor force inactivity on actual earnings volatility.12 In addition, these thresholds are generally set so low (say, part-time at the minimum wage or the Social Security qualifying threshold) that some relatively small changes in dollar earnings are still large in percentage terms. There are various ways to sort out the impact of volatility in different parts of the earnings distribution (figure 3.15), and the approach we take here is to 11. In datasets with only annual earnings, such as the one used by Bloom et al. (2017), job loss generally shows up as reduced earnings for workers who remain “employed” because they have positive earnings for some period during the year. 12. Although the arc-percent change measure does allow for transitions to or from zero earnings, we do not include these transitions in the volatility results presented in this chapter.

102

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

tie that decomposition back to our mobility analysis in the previous section. The vertical bars show the average of the absolute value of each worker’s arc percentage change in each one-year mobility path, again denoted 11, 12, 13, and so on, to refer to the origination and destination bin. The average absolute arc percentage changes are then ranked from highest volatility transitions (earnings bin 5 to earnings bin 1, earnings bin 1 to earnings bin 5, etc.) to lowest volatility transitions (those who remained in earnings bin 4). The rank-order of the bars captures the two distinct determinants of measured volatility, as movements across bins far apart (15 or 51) suggests a large absolute earnings change, and if one of those bins is a low earnings bin, the dollar change is magnified because the base for the arc-percent change is lower. The thick line overlaid on the bars in figure 3.15 shows how much the variability along each of the different mobility paths contributes to overall measured volatility. For example, although the 15 and 51 paths for earnings mobility exhibit extreme volatility (as indicated by the height of the bars), there are so few workers on those paths that the impact on overall volatility is negligible (the thick line is close to zero). The largest single contributor to overall volatility is the 11-path for earnings mobility, which has workers with real earnings between $1 and $18,000 in both years of the pairwise arcpercent change. Measured volatility along the 11 path is about one-third that of the 15 or 51 path, but there are so many workers on the 11 path that they account for almost one-third of overall volatility during the study period. Again, this reinforces the observations above that volatility is a highly nonlinear concept, and specific trends in overall measures (as in figure 3.14) should be interpreted with caution. 3.6

Conclusion

The primary goal of this chapter is to demonstrate the substantial heterogeneity across subnational areas of the US. For the four large MSAs we analyze, there are clear national trends represented in each of the local areas, the most prominent of which is the increase in the share of earnings accruing to workers at the top of the earnings distribution in 2017 compared with 1998. However, the magnitude of these trends varies across MSAs, with New York and San Francisco showing relatively large increases and Los Angeles somewhere in the middle relative to Detroit, whose total real earnings distribution is relatively stable over the period. A second goal is to show the important role of earnings mobility. Large changes in earnings typically occur either though job change or internal promotion. Our measure captures both and provides a comprehensive view of the change in the earnings distributions. One potentially concerning trend is the decrease in the ratio of the sum of workers moving up to a higher earnings bin or down to a lower earnings bin relative to the number of workers

US Earnings Dynamics: Inequality, Mobility, and Volatility

103

staying in the same bin. The reduced worker earnings mobility observed over the analysis period potentially has long-term productivity implications if workers choose to stay in jobs with a relatively poor match rather than move to a better match either at the same or a new firm. The reduction in earnings mobility is especially strong in both Detroit and New York, a result worthy of further investigation. When estimating earnings distributions and earnings mobility, we take a nonparametric approach to estimation. This approach allows us to show detailed local area information in a flexible way, although at the cost of a large number of estimated parameters. The traditional venue of the academic research paper is not ideal for displaying our results, which is why we are developing an interactive dissemination application at the US Census Bureau. We hope the reader is able to see in this chapter a glimpse of our ultimate goal, which is to allow for the interactive display of detailed earnings and mobility statistics for MSAs across the US.

References Abel, Jaison R., and Richard Deitz. 2019. “Why Are Some Places so Much More Unequal than Others?” Economic Policy Review 25 (1): 58– 75. Abowd, John M., and Kevin L. McKinney. 2016. “Noise Infusion as a Confidentiality Protection Measure for Graph-based Statistics.” Statistical Journal of the International Association for Official Statistics 32: 127– 32. Abowd, John M., Kevin L. McKinney, and Ian M. Schmutte. 2019. “Modeling Endogenous Mobility in Earnings Determination.” Journal of Business and Economic Statistics 37 (3): 405– 18. Abowd, John M., Kevin L. McKinney, and Nellie Zhao. 2018. “Earnings Inequality and Mobility Trends in the United States: Nationally Representative Estimates from Longitudinally Linked Employer-Employee Data.” Journal of Labor Economics 36 (S1): 183– 300. Abowd, John M., Bryce Stephens, Lars Vilhuber, Fredrik Andersson, Kevin L. McKinney, Marc Roemer, and Simon Woodcock. 2009. “The LEHD Infrastructure Files and the Creation of the Quarterly Workforce Indicators.” In Producer Dynamics: New Evidence from Micro Data, edited by Timothy Dunne, J. Bradford Jensen and Mark J. Roberts, 149– 230. Chicago: University of Chicago Press. Austin, Benjamin, Edward Glaeser, and Lawrence Summers. 2018. “Jobs for the Heartland: Place-Based Policies in 21st-Century America.” Brookings Papers on Economic Activity (Spring): 151– 255. Autor, David H., David Dorn, and Gordon H. Hanson. 2013. “The China Syndrome: Local Labor Market Effects of Import Competition in the United States.” American Economic Review 103 (6): 2121– 68. Beraja, Martin, Andreas Fuster, Erik Hurst, and Joe Vavra. 2018. “Regional Heterogeneity and the Refinancing Channel of Monetary Policy.” Quarterly Journal of Economics 134 (1): 109– 83. Beraja, Martin, Erik Hurst, and Juan Ospina. 2019. “The Aggregate Implications of Regional Business Cycles.” Econometrica 87 (6): 1789– 833.

104

Kevin L. McKinney, John M. Abowd, and John Sabelhaus

Bloom, Nicholas, Fatih Guvenen, Luigi Pistaferri, John Sabelhaus, Sergio Salgadok, and Jae Song. 2017. “The Great Micro Moderation.” Working paper (April). Bureau of Economic Analysis. 2015. “GDP by Metropolitan Area Methodology.” https://www.bea.gov/sites/default/files/methodologies/GDPMetro2015.pdf. Bureau of Economic Analysis. 2019. “GDP by County, Metro, and Other Areas.” https://www.bea.gov/data/gdp/gdp -county-metro-and-other-areas. Congressional Budget Office (CBO). 2019. “The Distribution of Household Income, 2016.” Research Report. Washington, DC: Congressional Budget Office. https:// www.cbo .gov /system /files /2019 -07/55413 -CBO -distribution -of -household -income-2016.pdf. Foote, Andrew, Ashwin Machanavajjhala, and Kevin McKinney. 2019. “Releasing Earnings Distributions Using Differential Privacy: Disclosure Avoidance System for Post-Secondary Employment Outcomes (PSEO).” Journal of Privacy and Confidentiality 9 (2). https://journalprivacyconfidentiality.org /index.php/jpc/article /view/722. Haltiwanger, John C., Henry R. Hyatt, Lisa B. Kahn, and Erika McEntarfer. 2018. “Cyclical Job Ladders by Firm Size and Firm Wage.” American Economic Journal: Macroeconomics 10 (2): 52– 85. Haltiwanger, John C., Henry Hyatt, and Erika McEntarfer. 2018. “Who Moves up the Job Ladder?” Journal of Labor Economics 36 (S1): S301– S336. Hyatt, Henry, Erika McEntarfer, Kevin Mckinney, Stephen Tibbets, and Doug Walton. 2014. “Job-to-Job (J2J) Flows: New Labor Market Statistics from Linked Employer-Employee Data.” JSM Proceedings: Business and Economics Statistics Section, 98– 110. Alexandria, VA: American Statistical Association. McEntarfer, Erika, John Haltiwanger, Melissa Bjelland, and Bruce Fallick. 2011. “Employer-to-Employer Flows in the United States: Estimates Using Linked Employer-Employee Data.” Journal of Business and Economic Statistics 29 (4): 493– 505. McKinney, Kevin L., and John M. Abowd. 2019. “Male Earnings Volatility in LEHD before, during, and after the Great Recession.” Paper presented at the 2019 AEA Meetings in Atlanta, GA. Mian, Atif, Kamalesh Rao, and Amir Sufi. 2013. “Household Balance Sheets, Consumption, and the Economic Slump.” Quarterly Journal of Economics 128 (4): 1687– 726. Piketty, Thomas, Emmanuel Saez, and Gabriel Zucman. 2018 “Distributional National Accounts: Methods and Estimates for the United States.” Quarterly Journal of Economics 133 (2): 553– 609. Sabelhaus, John, and Jae Song. 2010. “The Great Moderation in Micro Labor Earnings.” Journal of Monetary Economics 57 (4): 391– 403.

4

Evidence from Unique Swiss Tax Data on the Composition and Joint Distribution of Income and Wealth Isabel Z. Martínez

4.1

Introduction

Recent research on inequality has shifted its focus from income to wealth. With rising top income inequality, it comes as no surprise that wealth, which is already distributed more unequally than income, has become more concentrated too— especially at the top. Little is known, however, about the joint distribution of income and wealth at the individual level. Are those at the top of the income distribution also among the wealthiest or are these different groups? Furthermore, detailed evidence on the individual demographics as well as on the composition of income and wealth along their respective distributions is still limited. In this chapter, I make several contributions to the growing literature on wealth inequality. The first contribution is a new, unique dataset, which I construct out of individual income and wealth tax data obtained from eight Swiss cantons. I harmonize these very detail-rich datasets and pool the data for 2010, the year covered in all cantonal datasets obtained. This is the first time, to my knowledge, that individual income and wealth tax Isabel Z. Martínez is a senior researcher at the KOF Swiss Economic Institute at ETH Zürich. I thank Reto Föllmi, Andreas Peichl, Frank Pisch; participants at the CRIW-NBER conference on “Measuring and Understanding the Distribution and Intra/Inter-Generational Mobility of Income and Wealth,” Bethesda, Maryland, March 5– 6, 2020; and seminar participants at LMU Munich for helpful comments. Larissa Luchsinger and especially Oliver Hümbelin provided excellent research assistance. We all greatly appreciate financial support through SNSF Grant 176458 “The Influence of Taxation on Wealth and Income Inequality.” For acknowledgments, sources of research support, and disclosure of the author’s material financial relationships, if any, please see https://www.nber.org /books-and-chapters/measuring-distribution -and-mobility-income -and-wealth /evidence -unique -swiss-tax-data-composition-and-joint -distribution-income-and-wealth.

105

106

Isabel Z. Martínez

data from different cantons have been combined into one large harmonized dataset, representing about half the Swiss population of taxpayers. I show that this pooled dataset is representative of Switzerland as a whole along many dimensions, including the top of the income and wealth distributions and demographic characteristics. Using these data allows me to study the distribution of income and wealth in more depth than has been possible in Switzerland so far. First, I show how the high concentration of wealth in Switzerland documented previously in Föllmi and Martínez (2017) plays out at lower points in the distribution: those in the bottom 30 percent have virtually zero or even negative net wealth. Due to the high incidence of debt in the bottom quintile of the distribution, the share of the bottom half of the population in total net wealth is negative, the bottom 60 percent own 1 percent of total net wealth. The Gini index for the net wealth distribution amounts to 0.80— almost double the Gini index for gross income, which is 0.41. Second, I show individuals’ characteristics within the different income and wealth percentile groups. This was not possible in prior research on top income and wealth shares in Switzerland, as it was based on aggregate tax statistics, where individual information is lost (Föllmi and Martínez 2017; Frey, Gorgas, and Schaltegger 2016; Schaltegger and Gorgas 2011). Other research that relied on detailed cantonal tax data did not cover more than one canton (examples include Martínez 2022; Brülhart et al. 2021; Gallusser and Krapf 2019; Moser 2019). This analysis reveals several important features. (1) Retirees are strongly overrepresented in the top decile of the wealth distribution. They make up more than half of the wealth-rich individuals, but only 23 percent of the population in my sample. (2) In the income distribution, retirees are overrepresented in the second quintile and among the top 1 percent. (3) Single women— many of whom are retirees— are less likely to be in the top decile of the income distribution than single men. However, single women are more likely to be found in the top decile of the total net wealth distribution than single men. Given the vast evidence on the gender wealth gap (e.g., Neelakantan and Chang 2010; Schneebaum et al. 2018; Sierminska, Brandolini, and Smeeding 2010), this finding is especially surprising. A possible explanation is that single women are much more likely to be retired than single men and hence belong to a population which is a priori more likely to be wealthy. In addition, since women have longer liveexpectancy than men, single retired women are more likely to be widows which means they may have inherited wealth from their late husbands. Taken together, these features all point to the strong life-cycle patterns in wealth accumulation. Indeed, the pronounced age-wealth gradient is an important feature throughout the analysis of this chapter. Third, I carve out the composition of income and wealth along their respective distributions and for different subgroups. I find that financial assets, including personal accounts, are the most important wealth compo-

Unique Swiss Tax Data

107

nent for households below the median of the wealth distribution. Real estate wealth is held only by those in the upper half of the wealth distribution, even when looking at subgroups such as retirees. This is in line with the low home ownership rate of less than 40 percent— and very different from other countries, such as the US or Spain, where real estate is much more widespread. In Spain, for example, real estate amounts to 90 percent of total net wealth for individuals around the median (P40– P60) (Martínez-Toledano 2020). While there are overall important differences in the composition along the distribution, age is the most influential factor for differences in the composition of income and wealth. Retirees along the whole wealth distribution have lower debt levels, very low shares in business and other movable assets, and real estate is distributed more evenly among retirees than among nonretirees— although it remains limited to those in the upper half of the wealth distribution. In contrast, gender differences in the composition of wealth are small. Similarly, the composition of income reveals substantial heterogeneity along the distribution and between retirees and nonretirees. For the latter, labor income tends to be the most important income source, especially for the bottom 99 percent. Even those in the top 0.01 percent of the income distribution draw on average 35 percent of their income from labor, the remaining 65 percent is different forms of capital income. In addition, income composition varies by gender. Women draw a lower share of their income from labor, hence they rely more heavily than men on transfers or— at the very top— on capital incomes. Finally, I shed light on the joint distribution of income and wealth. The overall correlation between someone’s income and wealth rank is 0.32. However, this number masks substantial heterogeneity, as there is a strong tail dependence between the two distributions. Those who already are very rich therefore also derive the largest incomes. This is especially pronounced at the very top: 78 percent of those in the top 0.01 percent of the gross income distribution are in the top 0.1 percent of the net wealth distribution. At the same time, a considerable share of individuals across all income ranks are in the bottom quintile of the wealth distribution, that is, they have very low or even negative net wealth— even if they are in the top 10 percent of the income distribution. In contrast, those belonging to the top 10 percent of the wealth distribution have a low likelihood of having low incomes. Lowincome wealth millionaires are therefore very rare, while about one out of six top earners can be considered wealth-poor. Overall, it is relatively unlikely to be in a higher wealth group compared to one’s income group. These findings can have important implications for life-cycle models as well as for optimal tax theory. If joint inequality of income and wealth is even larger than income or wealth inequality taken alone, optimal redistributive taxation may, for example, be more progressive. Similarly, understanding the composition of income and wealth is important to draw conclusions on the

108

Isabel Z. Martínez

incidence and the distributional effects of differential taxation of, such as labor and capital income, or financial assets and real estate. Understanding the joint distribution of income and wealth is further relevant for research on regional tax competition. Mobility of high earners in response to income taxes has been shown to be quite large (see Kleven et al. 2020 for an overview). Nevertheless, jurisdictions engaging in such tax competition for top earners may not break even in terms of income tax revenue (Agrawal and Foremny 2019; Agrawal, Foremny, and Martínez-Toledano 2020; Martínez 2022). However, if those top earners also increase the wealth and inheritance tax base, foregone income tax revenue may be compensated by revenue from taxes on wealth. Finally, my descriptive findings also have implications for macroeconomic policies: if low-income earners also have low wealth, it is harder for them to cope with shocks. Given that in many datasets it is possible to observe income but not wealth in detail, these are valuable insights for future research. The remainder of this chapter is organized as follows. Section 4.2 embeds the chapter in the previous literature. Section 4.3 describes the dataset I compiled for this project. In sections 4.4 and 4.5 I present the results on the composition of wealth and income, respectively, followed by results on the joint distribution of income and wealth (section 4.6). Section 4.7 concludes. 4.2

Previous Research

There is a rapidly growing literature on wealth inequality, including, for example, Piketty, Yang, and Zucman (2019), Kopczuk and Saez (2004), and Saez and Zucman (2016). Data constraints are often a limiting factor for this research, as data on wealth is much less readily available than data on income. To estimate the wealth distribution, researchers have relied on surveys, bequest tax data, capital income tax data, or wealth tax data— although the latter is available in only a small number of countries: while 12 countries had net wealth taxes in 1990, there were only four OECD countries that still levied recurrent taxes on individuals’ net wealth in 2017 (see OECD 2018 for an overview on wealth taxation). Since wealth is much more concentrated than income, many papers have put special focus on the evolution of top wealth shares. For Switzerland, Dell, Piketty, and Saez (2007) and Föllmi and Martínez (2017) have documented the evolution of top wealth shares over the past century until 2010. Based on aggregate wealth tax statistics, this research shows that wealth is highly concentrated in Switzerland, where the top 1 percent holds around 40 percent of total wealth. Föllmi and Martínez (2017) find that correcting for nontaxable pension wealth of the active population reduces this share, but that since the mid-1990s there is nevertheless an upward trend in top wealth shares. While empirical research on income but also wealth inequality has made much progress, especially over the past two decades, research on the joint dis-

Unique Swiss Tax Data

109

tribution of income and wealth remains scattered— mainly due to the lack of high-quality individual data covering both individual income and wealth distributions. Aiming at better measurement of “economic position” or “economic well-being,” Wolfson (1979) made adjustments to the Canadian income distribution by (1) accounting for family size, (2) including imputed rent, and (3) including the annuity equivalent of net worth. More recent contributions include Aaberge, Atkinson, and Königs (2018), Chauvel et al. (2019), Jäntti, Sierminska, and Smeeding (2008), Kuhn, Schularick, and Steins (2020), Peichl and Pestel (2013), and Sierminska, Brandolini, and Smeeding (2007). Most papers, including this one, rely on nonparametric measures of the joint distribution. A notable exception is that by Jäntti, Sierminska, and Van Kerm (2015), which presents a new, parametric approach based on copula functions. The difficulty in this approach lies in accommodating the extensive mass at income and especially wealth zero, as the copula is not uniquely defined across mass points.1 All these previous papers base their analysis on surveys. Besides typically not covering the upper tail of the distributions very well, survey data excludes people living in institutions. This is especially problematic when studying the distribution of wealth, which is more concentrated among the elderly, who in turn are more likely to live in nursing homes and similar institutions. The recent paper by Gallusser and Krapf (2019) is the only other paper that I am aware of that studies the joint distribution of income and wealth based on administrative tax records. Being based on cantonal tax data, it is also the study most similar to mine. Nevertheless, our papers differ in several aspects. First, I combine data from several cantons to cover more than 50 percent of the population in Switzerland, while Gallusser and Krapf (2019) use data from the canton of Lucerne only. Second, their focus lies on new inequality measures combining annuitized wealth and annual labor income flows, while I present evidence on the association between income and wealth along several dimensions. Similar to their findings, I find a very strong tail dependence— especially at the top— and I further show that the strong tail dependence is driven by the top 1 percent within the top 1 percent. 4.3 4.3.1

A New Income and Wealth Tax Dataset for Switzerland Cantonal Tax Data

Switzerland is a federal country with 26 states, called cantons. The federal government levies an annual personal and corporate income tax. On top of 1. Some recent papers go even further and include consumption inequality as a third dimension (e.g., Fisher et al. 2022; Linder and Schürz 2020; Ruiz 2011). While such a multidimensional approach is appropriate to measure well-being in an encompassing manner, the goal of the present chapter is to gain a deeper understanding of the complex relationship between income and wealth.

110

Isabel Z. Martínez

this tax, each canton levies income as well as wealth taxes on an annual basis for both individuals and corporations. The wealth and income tax bases are very broad and include all income earned and wealth held in and outside Switzerland. As a rather unique feature, Swiss tax data therefore contains detailed information on income and wealth for the whole population, including the upper tail of the distribution. What are not taxed, and therefore not recorded separately, are realized capital gains on personal assets.2 Due to their large tax autonomy and in order to reduce administrative burdens, cantons collect the direct federal taxes on behalf of the federal government such that taxpayers file only one tax return each year. All personal taxes are residence based. This institutional setting has important implications for the availability of tax data. Cantons enjoy large tax autonomy and are the owners of the data collected. They forward only a limited set of income variables to the Federal Tax Administration, including taxable and net income after itemized deductions. Income is therefore aggregated and the information on the different income sources (e.g., employment, self-employment, capital income, pensions, etc.) is lost. Most importantly, because the federal government does not levy a wealth tax, it has no individual-level information on wealth in its tax data. Cantons share only aggregate wealth statistics with the Federal Tax Administration. Hence while tax data available from the Federal Tax Administration, which cover the full population living in Switzerland, have been used in previous research on income and wealth inequality in Switzerland (including work on top income and wealth shares by Dell, Piketty, and Saez 2007, and Föllmi and Martínez 2017), they do not allow us to uncover the composition nor the joint distribution of income and wealth. I obtained anonymized cantonal tax data based on taxpayer’s tax returns from the following eight out of 26 cantons:3 Aargau (AG), Bern (BE), BaselStadt (BS), Jura (JU), Luzern (LU), St. Gallen (SG), Obwalden (OW), and Zurich (ZH). Figure 4.1 shows the regional data coverage. I am able the cover most of the German-speaking areas and some French-speaking parts but unfortunately miss the Italian-speaking south of the country. These cantons cover 53 percent of the universe of regular taxpayers who had filed a tax return in 2010 according to federal income tax statistics. Since my dataset 2. Capital gains incurred on personal assets are taxed indirectly through the wealth tax. Since the wealth tax is based on assets’ worth on December 31, capital gains— especially from financial assets— are therefore taxed even when not realized. Realized capital gains on business assets are taxed under corporate taxation. 3. To obtain the cantonal tax data, requests have to be made at each canton on a project-byproject basis. The application process as well as costs for data access vary widely across cantons, and ultimately not all cantons are willing to provide tax data for research purposes. For this project, data access was granted within the SNSF Grant 176458, “The Influence of Taxation on Wealth and Income Inequality.” To facilitate the data application process and reduce costs, some of the data used here were approved as part of an earlier SNFS Grant, “Inequality in Income and Wealth in Switzerland from 1970 to 2010,” and kindly made available for this research project. See http://inequalities.ch/ for details on that earlier project.

Unique Swiss Tax Data

Fig. 4.1

111

Cantons and tax units covered in the data

Notes: The map shows the cantons for which data are available, along with the share of tax units covered by each canton. The shares of tax units are based on the number of regular taxpayers in federal income tax statistics of 2010. In contrast with regular population statistics, this metric takes into account that some groups, especially foreigners without permanent residence and employees at international organizations, do not file a tax return. Together, the data cover 53 percent of all taxpayers in Switzerland and roughly three-quarters of the population in German-speaking areas. French-speaking parts in my data include the whole canton of Jura (JU) and the western part of the canton of Bern (BE).

is based on filed tax returns, this is the relevant comparison. Note that true nonfiling is not an issue: filing is mandatory for all Swiss citizens and permanent residents. In case of nonfiling, the tax administration will file a tax return on the taxpayer’s behalf in an unfavorable way (e.g., overestimating their income and disregarding deductions) and add a fine to the tax bill. These nonfilers’ imputed tax returns are then included in the statistics. Due to the financial penalties involved (which increase with each year of nonfiling), their share is however extremely low. Nonpermanent residents and employees at international organizations as well as diplomats do not usually file a tax return and are hence excluded. Individuals employed at the many international organizations located in Switzerland as well as diplomats are tax-exempt. Nonpermanent residents are taxed at the source without filing a tax return— unless their annual income exceeds 120,000 CHF, in which case they can opt to file a tax return and are part of my data. I combine these cantonal datasets into one, large, harmonized dataset. There are some important limitations. First, in the canton of Zurich the data does not contain the full population of taxpayers There, detailed cantonal tax data is available for only 45 out of 161 municipalities. These 45 municipalities include the large city of Zurich and cover roughly 60 percent of all taxpayers in the canton. According to the tax administration, this sample is representative of the canton as a whole. Unfortunately, no sampling weights were provided. Out of this sample, I obtained a 50 percent random sample

112

Isabel Z. Martínez

of all the taxpayers belonging to the bottom 95 percent of the gross income or net wealth distribution, and a 100 percent sample of those belonging to the top 5 percent of the income and/or wealth distribution. I use sampling weights to take this into account. Second, each of the cantonal datasets covers different time periods, including the years 2000– 2016. For all cantons except Zurich, the year 2010 is in the data. I merge cantonal data from 2010 to obtain a cross-section dataset covering the eight cantons described above. For the canton of Zurich, where data is available only in intervals of three years, I use data from 2011. I refer to this cross-sectional dataset as pooled tax data. Dynamic analyses, however, are still only possible using data for single cantons. Third, some of the variables on income, wealth, and deductions differ in their level of detail across cantons. While the tax base is the same across cantons (defined in the 1990 Federal Tax Harmonization Act), the individual tax data differs across cantons due to differences in how tax returns are structured and what is recorded in the main taxpayer file. In each canton, I have access only to data that are recorded in the main tax file, and some cantons did not include all the variables due to privacy concerns. To ensure comparability, I harmonize the data across cantons. Sections 4.3.2 and 4.3.3 describe the variables I use in detail. Fourth, in Switzerland married couples have to file jointly. Therefore, a tax file might represent one or two adults. I individualize the data, so every observation represents a single person. This leaves me with a total of 2,755,938 observations in 2010. While some income components could be attributed exactly to one of the spouses, this is not possible in every canton nor for every income component. Wealth components are always reported for the tax unit as a whole and cannot be attributed to one of the spouses. For married couples I therefore split all income and wealth components equally between spouses. Such equal division of resources is appropriate to depict the distribution— assuming that married couples share income and wealth even if they do not contribute to the same extent. Since this assumption is likely to be violated in reality, my analysis will slightly underestimate true individual income and wealth inequality. 4.3.2

Income Measures

I use gross income net of all mandatory contributions but not net of taxes. I differentiate between income from labor, capital, and transfers, and further break these components down into subcomponents. For some income categories, only net income is available, namely income from real estate. Below I explain all income components used in the analysis in detail. • Labor income is the sum of income from employment and selfemployment: – Income from employment. Tax filers declare gross income from employment net of the following contributions withheld by the

Unique Swiss Tax Data

113

employer at the source: social security, disability, military, maternity, and unemployment insurance contributions as well as occupational pension contributions. Annual gross income is reported in a legal form issued by the employer which needs to be enclosed with the tax return. – Income from self-employment includes profits from nonincorporated businesses, namely, sole proprietorship, partnerships, and limited partnerships. While legally mandatory only for businesses with turnover above 500,000 CHF, in practice also most small businesses conduct orderly—that is, double-entry—accounting. Even if a business keeps only simple accounting, expenditures need to be proven and in direct relation to the business. Losses can be carried forward seven years. Self-employment income is subject to social security, disability, military and maternity insurance contributions. Self-employed further have the option to voluntarily join an occupational pension fund. Contributions are deductible and wealth held in these funds is tax-free. To maintain equal treatment, all of the above require the self-employed to pay both the employee’s and the employer’s part of the contribution. As income from self-employment is commonly considered as mixed income (see, e.g., Martínez-Toledano 2020), I follow the literature and allocate 70 percent of these profits to income from self-employment and 30 percent to capital income. • Total capital income includes all incomes from capital and real estate: – Capital income encompasses income from financial assets, namely interests and dividends, income from undistributed inheritances (Erbengemeinschaften), plus 30 percent of income from selfemployment. – Real estate income consists of income from renting out real estate and imputed rent of home owners. Imputed rent is part of the income tax base of homeowners in Switzerland and is reported under real estate income in the tax return. Only in the tax data from ZH, OW, SG, and AG are imputed rents listed separately, allowing me to distinguish between net rental real estate income and imputed rents. Their amount is defined by the tax laws and specified by cantonal authorities. All real estate income is reported net of maintenance costs, which are tax deductible. • Total transfer income is the sum from all transfers and pensions: – Transfers contain benefits from unemployment, accident, disability, and military insurances, as well as from child, family, maternity, and sickness allowances. They further include private transfers from other households, especially alimonies from ex-partners for the spouse and minor children. Means-tested benefits are excluded, as they are not taxable and hence are not declared. Since means-tested benefits depend on a variety of factors, including the household composition, living and

114

Isabel Z. Martínez

health conditions, and are determined on a case-by-case basis, I cannot impute these benefits with the data at hand. I therefore underestimate true income for low-income individuals. – Pensions. This component summarizes all incomes stemming from pensions. It is available for all cantons except AG. In all other cantons, it can be further broken down into social security pensions and occupational and private pensions. Social security pensions. Pensions from the public pension system, the first pillar in the Swiss pension system. All labor income is subject to contributions. Nonworking individuals pay contributions based on their wealth. Everyone is covered and pensions are capped. Occupational and private pensions. Pensions from the second and third pillars of the Swiss pension system. The occupational pension system has some similarities with the US 401(k)s; the main differences are that (1) contributions are mandatory for employment income above CHF 23,940 (in 2010), (2) contribution rates are age-dependent, and (3) rates are set by the government. Private pensions pensions (the third pillar) stem from life insurances and private (usually tax-exempt) retirement saving accounts. • Other income includes all other incomes which do not belong to any of the categories above. In particular, this category contains lump-sum settlements for recurrent benefits and, at least in some cantons, cash payouts at retirement from the second and third pillars of the pension system. • Gross income is the sum of all the income components listed above. I use the term gross as it is income before taxes and before any tax-related deductions, even though some components, such as real estate income, are net of expenses and deductions. 4.3.3

Wealth Measures

As far as possible, I base my analysis on total net wealth. The data allow me to distinguish between financial assets, movable business assets, movable personal assets, real estate, and debt. Since wealth on retirement accounts from the mandatory occupational pension system (second pillar) and the voluntary tax-exempt saving scheme (pillar 3a) are not subject to taxation until they are either cashed in or transformed into a pension at retirement, I have to exclude these assets from the analysis. However, since voluntary contributions toward the second and third pillars are deductible from annual income up to thresholds fixed by the federal government, I see who makes such contributions. Financial assets include securities, credit balances, cash on bank deposits, gold and other precious metals as well as the value of life insurance poli-

Unique Swiss Tax Data

115

cies. Excluded are personal retirement savings accounts from the third pillar of the pension system, as they are tax-exempt. These savings can only be accessed after retirement or to purchase a private home. The same is true for wealth held within the occupational pension system (second pillar). Business assets are movable assets held within nonincorporated businesses (sole proprietorship, partnerships, and limited partnerships). They include all movable business assets such as inventories, livestock, vehicles, machinery/furniture, equipment, and so on. Also, financial assets held within a business are included here. As opposed to movable personal assets, movable business assets are not valued at market but at book value. Therefore, assets are discounted annually for depreciation such that they tend to be valued below market value. Furthermore, business assets are reported net of business debts in the main tax file. This implies that movable business assets are undervalued compared to personal assets. Real estate held within nonincorporated businesses is reported together with privately held real estate. Real estate wealth includes all real estate, including private homes, secondary homes, land and property held within a business. In the majority of cantonal tax data it is not possible to further distinguish between these categories. Reported values are gross values excluding any debt. However, real estate is deliberately undervalued in Swiss tax data to avoid an excessive tax load on homeowners. Assessment methods vary by canton (except for real estate used for agriculture or forestry, which is valued uniformly in the whole country). Because individuals might own real estate in different cantons, tax authorities use so-called repartition values (Repartitionswerte) to rescale real estate valued by another canton. I use these values, published by the Federal Tax Administration, to adjust for different valuation practices across cantons. In addition, assessments happen only approximately every decade. To account for developments in real estate prices over time, I further adjust real estate prices since the last valuation year using regional house price indices. These are collected by the real estate firm Wuest + Partner and published online by the Swiss National Bank. Unfortunately, this second adjustment is only possible in cantons that assess all properties in a given year. This is the case in BE, ZH, OW, BS, AG and JU. In LU and SG, real estate valuations are done on a rolling basis, where every year about 10 percent of all properties are reevaluated. Therefore, the development of real estate prices should at least partly be captured. For ZH, AG, and SG I further have access to real estate information that enables me to distinguish between owner-occupied houses like main residences and vacation homes from other real estate– like properties for rent, business properties, and land.

116

Isabel Z. Martínez

Movable assets. This component includes motor vehicles, shares in undistributed inheritances, shares in nonlisted companies and other assets, such as jewelry and art, that do not fall into any other category. These assets are valued according to their insurance value, if possible. Nonlisted companies are valued at book value. Valuation of cars takes into account depreciation at rates stipulated by the tax authority. Debt. In principle all types of debt are tax deductible. Debt therefore includes mortgages, but also personal loans, consumer credits, and other verifiable outstanding financial liability, including taxes owed. Unfortunately, it is not possible to distinguish the different types of debt in the data: in all the cantons, only the sum of all debt is recorded in the main tax file. It is therefore not possible to define the different asset categories net of debt. The descriptive analysis suggests that by far the largest component of debt is mortgages. Gross wealth is the sum of financial assets, business assets, other movable assets and real estate wealth. Strictly speaking, this is not a true gross value, since business assets are net of debt. Because it is not possible to attribute reported debts to any corresponding gross asset in the data at hand, for some analyses I have to revert this admittedly imprecise definition of gross wealth. Net wealth. Total net wealth is built by subtracting total debt from gross wealth. Since debt can be deducted at market value, but real estate tends to be undervalued (even though I try to account for this as well as possible), I may underestimate net wealth, especially for real estate owners. 4.3.4

Demographics

For all cantons, the data include the following demographics: marital status, number of dependent children, gender of the main taxpayer (in the case of married couples), age of the main taxpayer (although in BS only age categories are available). Other characteristics are available only for some cantons, namely age of the second taxpayer (ZH, BE, SG, JU) and gender of the second taxpayer (SG, BE, JU). In all other cantons, I define the gender of the second person in a married couple as the opposite of that of the main taxpayer. Because I cannot identify same-sex couples, I will assign a wrong gender to the second person in a same-sex couple. I estimate that in the cantons, where I have to impute gender in this way, 0.43 percent of couples were of the same gender in 2010. The measurement error is therefore small. I further define dummy variables based on income streams to indicate whether someone is an employee, self-employed, or a retiree. Someone is an employee or self-employed, respectively, depending on what was their main source of labor income. The retiree dummy variable takes on the value of 1 if someone draws a social security pension and is allowed to retire according to their age. For men, early retirement is possible at age 63 and for women at age 62. I introduce the age cut because social security pensions include

Unique Swiss Tax Data

117

disability pensions for nonretired individuals. In the cantons ZH, LU, OW, BS, and AG, where I do not know the age of the spouse, I impute their age based on the age structure of couples where the main taxpayer is between 55 and 80 years old in the canton of Bern. On average, the spouse is three years younger. This allows me to define retirement for individuals in all cantons except AG, where I lack information on pension income. An alternative approach is to define retirement according to legal retirement (65 years for men, 64 years for women). With this approach, however, the missing age of the spouse is more problematic: there are either more missing observations, since age of the spouse is not available in LU, OW, BS, and AG, or I have to base the definition solely on the imputed age of these individuals. Nevertheless, most results are very similar for both definitions of retirement. Finally, I create an indicator for home ownership. This variable takes on the value of 1 if someone has an imputed rent in their income tax base. This variable is defined only for the cantons of ZH, OW, BS, SG and AG. 4.3.5

Summary Statistics

Most of the analyses in this chapter are carried out by percentile groups. Below P90, these groups correspond to deciles, but then I use smaller fractions of the population within the top 10 percent. Earlier findings on top income groups suggest that there are considerable differences between the rich and the super-rich (e.g., Atkinson and Piketty 2007, 2010; Föllmi and Martínez 2017). I classify individuals into percentile groups based on the total gross income and the total net wealth distributions, respectively. Even subgroups— such as men and women or retirees and workers— are grouped into percentiles based on the total distribution and not based on the distribution within their respective subgroup. This allows for direct comparisons across groups, as the income and wealth thresholds remain unchanged. Table 4.1 shows the income and wealth group thresholds, medians, and averages for each group. The reported income and wealth shares correspond well with the results in Föllmi and Martínez (2017) for Switzerland, suggesting that the sample is representative of Switzerland as a whole. Differences between the net and gross wealth distributions are largest at the tails of the distributions but are small overall. All inequality measures show that wealth is considerably more unequally distributed than income. The Gini index of the net wealth distribution reaches 0.8, almost twice as much as the Gini for gross income. While the bottom 80 percent earn just short of 55 percent of all income, the bottom 80 percent of the wealth distribution own less than 13 percent of total net wealth. Taken together, the bottom 50 percent of the wealth distribution have negative net worth, corresponding to 1.3 percent of total net wealth. Wealth is therefore heavily concentrated at the top. Tables 4.2 and 4.3 further show population averages by gross income and net wealth percentile groups, respectively. Comparing the share of single

27 35 42 49 56 64 75 97 126 252 348 821 3566

0.41 6.54 2.06 2.26 0.31

13 31 39 46 52 60 69 85 109 163 292 485 1338 21935

Mean

2,755,938

Threshold 15 31 39 46 52 60 69 84 108 153 289 448 1109 5979

Median

Gross income (1,000 CHF)

4.4 5.2 6.5 7.7 8.8 10.0 11.6 14.2 9.2 11.0 2.5 3.3 2.0 3.7

Share %

Income and wealth percentiles, 2010

0 5 17 38 73 131 231 454 776 2427 4030 12798 64728

Threshold

2,755,938

0.80 1045275 9.31 24.29 0.00

−35 2 11 27 54 99 176 324 587 1251 3060 6557 23436 300250

Mean −1 2 10 26 53 98 174 315 574 1124 2983 5838 19531 109234

Median

Net wealth (1,000 CHF)

−3.0 0.1 0.5 1.1 2.3 4.2 7.5 13.8 12.5 21.4 6.5 11.2 9.0 12.8

Share % 4 13 35 88 191 286 403 658 1042 3120 5032 15274 69597

Threshold

2,755,938

0.76 1357388 8.03 43.58 0.00

0 8 23 57 138 239 340 508 813 1636 3891 7980 26731 310831

Mean

0 8 22 55 137 239 338 498 796 1471 3810 7193 22325 115143

Median

Gross wealth (1,000 CHF)

0.0 0.2 0.7 1.7 4.0 6.9 9.9 14.8 11.8 19.0 5.7 9.3 7.0 9.0

Share %

Notes: The table contains the thresholds, mean, and median wealth and income, as well as income and wealth shares within each percentile group of the gross income, net wealth, and gross wealth distributions, respectively. Overall inequality measures and total number of observations are reported at the bottom of the table. Statistics are based on individual data, where wealth and income are split equally among married adults. Pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively. See online appendix figure A1 (http://www.nber.org /data-appendix /c14452 /appendix.pdf) for a graphical representation of the percentile thresholds.

N (weighted)

P0–P20 P20–P30 P30–P40 P40–P50 P50–P60 P60–P70 P70–P80 P80–P90 P90–P95 P95–P99 P99–P99.5 P99.5–P99.9 P99.9–P99.99 P99.99–P100 Inequality measures Gini P90/P10 P90/P50 P75/P25 P10/P50

Percentile group

Table 4.1

Unique Swiss Tax Data

119

women and single men across both distributions shows that women are less likely than men to be found at the top of the income distribution, but more likely than men to be found at the top of the wealth distribution. This suggests that, in Switzerland, single women— many of whom are retired— are likely to be found at the lower end of the income distribution and / or at the upper end of the wealth distribution. Twenty-three percent of all individuals in my data are retirees (i.e., drawing a pension and 62 years and older as described above). This corresponds well with total population statistics, according to which 21 percent of the adult population was aged 65 and older in 2010. Retirees are overrepresented at the bottom and at the very top of the income distribution. Within the wealth distribution, retirees are clearly concentrated at the top. They are more than twice as likely to be found within the top 10 percent of the wealth distribution than within the population as a whole. Figure 4.2 further shows the composition of gender and retirees along both distributions. Especially women belonging to the top 10 percent of the wealth distribution are very likely to be retired. But also for men, the share of retirees increases further up in the wealth distribution. This is not true for income, where especially the distribution of retired women is bimodal: they are most likely to be found at the bottom or the very top of the distribution. Looking at the probability of retirees to work (tables 4.2 and 4.3) reveals that retirees are more likely to continue working the higher up they are in either of the two distributions: while retirees who also earn some kind of labor income make up 4.2 percent of the total population, they represent almost one-fifth of those in the top 0.01 percent. This suggests that those who were doing well in life before retiring are also those most likely to continue working after retirement. At the same time, even though those at the bottom of the income and wealth distributions would benefit most from continuing to work from a resource perspective, they are least likely to do so. Likely explanations are worse health, lower education, and less attachment to the labor market before reaching retirement. Unfortunately, I lack the information on these characteristics. The data further show that individuals within the top 40 percent of the income distribution are more likely than the average to contribute to taxexempt, private, retirement accounts (third pillar). From those in the top 10 percent, excluding the top 1 percent, more than 65 percent contribute to these schemes. The picture is similar for the more regulated contributions to occupation pension schemes (second pillar). Here it is the top 1 percent who benefit most from such contributions. Contributions toward retirement accounts are spread more evenly across the wealth distribution. Especially contributions toward the third pillar are most likely in the range of P40– P95. This suggests that individuals who are building up wealth put part of it aside for retirement. Finally, I look at the distribution of home ownership. An estimated

Retirees (%)

Self-employed (%)

Employees (%)

With children (%)

Single male (%)

Single female (%)

Married (%)

Net wealth (1,000 CHF)

Table 4.2

N

N

N

N

N

N

N

N

53 (0.484) 531,711 31.3 (0.064) 531,711 36.6 (0.066) 529,787 32.4 (0.064) 529,787 24.0 (0.059) 531,711 44.5 (0.071) 495,508 3.8 (0.035) 304,922 23.4 (0.060) 501,584

P0–P20

117 (0.485) 268,600 57.0 (0.095) 268,600 28.7 (0.087) 267,224 14.6 (0.068) 267,224 33.2 (0.091) 268,600 40.8 (0.098) 253,519 5.4 (0.054) 173,356 45.6 (0.102) 239,988

P20–P30

130 (0.604) 268,935 63.8 (0.093) 268,935 24.0 (0.083) 267,509 12.4 (0.064) 267,509 39.8 (0.094) 268,935 52.6 (0.099) 253,956 5.4 (0.054) 172,473 36.3 (0.099) 237,548

P30–P40

126 (0.529) 268,548 60.9 (0.094) 268,548 25.2 (0.084) 267,322 14.1 (0.067) 267,322 42.7 (0.095) 268,548 65.5 (0.094) 254,697 4.7 (0.051) 173,595 26.4 (0.090) 238,695

P40–P50 126 (0.566) 269,233 57.6 (0.095) 269,233 24.0 (0.083) 268,223 18.5 (0.075) 268,223 43.9 (0.096) 269,233 74.1 (0.086) 256,738 4.2 (0.047) 177,304 19.6 (0.081) 240,599

P50–P60 141 (0.683) 272,126 55.7 (0.095) 272,126 23.0 (0.081) 271,074 21.6 (0.079) 271,074 44.9 (0.095) 272,126 78.3 (0.081) 259,840 4.0 (0.046) 181,093 16.2 (0.075) 243,595

P60–P70 173 (0.793) 278,849 53.8 (0.094) 278,849 23.3 (0.080) 277,795 23.1 (0.080) 277,795 47.0 (0.094) 278,849 80.3 (0.077) 266,386 4.2 (0.046) 190,197 14.5 (0.071) 250,066

P70–P80

Population averages by gross income percentile group, 2010

242 (1.042) 288,994 53.2 (0.093) 288,994 22.2 (0.078) 287,724 24.8 (0.080) 287,724 50.6 (0.093) 288,994 81.4 (0.074) 274,872 4.7 (0.047) 202,038 13.5 (0.067) 259,295

P80–P90 359 (2.148) 155,723 54.0 (0.126) 155,723 19.0 (0.100) 154,623 27.3 (0.113) 154,623 56.2 (0.126) 155,723 80.7 (0.103) 147,246 6.2 (0.071) 113,738 13.2 (0.090) 140,686

P90–P95 672 (6.454) 135,034 58.6 (0.134) 135,034 14.6 (0.097) 133,578 27.3 (0.122) 133,578 64.4 (0.130) 135,034 75.4 (0.121) 126,820 9.4 (0.090) 104,547 14.6 (0.101) 122,620

P95–P99 1500 (37.50) 16,553 60.0 (0.381) 16,553 12.4 (0.258) 16,361 28.0 (0.351) 16,361 69.9 (0.356) 16,553 66.6 (0.378) 15,587 14.1 (0.303) 13,253 18.0 (0.313) 15,073

P99–P99.5 3480 (116.30) 12,296 60.6 (0.441) 12,296 12.8 (0.304) 12,144 27.1 (0.403) 12,144 68.7 (0.418) 12,296 62.3 (0.449) 11,653 14.6 (0.354) 9,984 23.6 (0.400) 11,246

P99.5–P99.9 13,475 (504.7) 2,374 63.6 (0.99) 2,374 13.4 (0.704) 2,336 23.6 (0.879) 2,336 64.2 (0.984) 2,374 61.5 (1.03) 2,236 12.6 (0.751) 1,960 33.0 (1.02) 2,121

P99.9–P99.99

231,281 (47730) 246 65.0 (3.05) 246 15.6 (2.35) 238 20.6 (2.63) 238 63.4 (3.08) 246 52.3 (3.38) 220 8.5 (2.03) 189 39.4 (3.29) 221

P99.99–P100

225 (4.512) 2,769,222 52.3 (0.030) 2,769,222 25.8 (0.026) 2,755,938 22.2 (0.025) 2,755,938 41.6 (0.030) 2,769,222 64.5 (0.030) 2,619,278 5.0 (0.016) 1,818,649 23.0 (0.027) 2,503,337

Total

N

N

N

N

N

2.2 (0.030) 237,774 3.1 (0.025) 495,508 1.7 (0.019) 495,508 7.1 (0.057) 201,598 3.8 (0.043) 201,598

5.3 (0.062) 131,703 8.5 (0.055) 253,519 1.7 (0.026) 253,519 18.7 (0.117) 110,118 13.1 (0.102) 110,118

5.4 (0.063) 127,993 15.2 (0.071) 253,956 1.7 (0.025) 253,956 23.5 (0.129) 108,026 14.3 (0.107) 108,026

4.6 (0.059) 127,577 22.8 (0.083) 254,697 1.7 (0.026) 254,697 26.2 (0.135) 106,149 12.4 (0.101) 106,149

3.9 (0.054) 129,223 30.4 (0.091) 256,738 1.8 (0.026) 256,738 28.4 (0.139) 105,751 10.6 (0.095) 105,751

3.8 (0.053) 130,648 39.4 (0.096) 259,840 2.2 (0.029) 259,840 32.4 (0.143) 107,563 9.9 (0.091) 107,563

3.8 (0.052) 137,302 48.2 (0.097) 266,386 3.0 (0.033) 266,386 36.8 (0.142) 115,045 9.7 (0.087) 115,045

4.2 (0.052) 147,002 58.1 (0.094) 274,872 5.1 (0.042) 274,872 43.2 (0.139) 127,755 9.7 (0.083) 127,755

4.7 (0.072) 86,327 65.1 (0.124) 147,246 8.8 (0.074) 147,246 50.1 (0.177) 79,924 9.9 (0.106) 79,924

5.5 (0.078) 85,379 66.1 (0.133) 126,820 14.7 (0.100) 126,820 56.8 (0.172) 83,057 10.2 (0.105) 83,057

7.3 (0.243) 11,538 61.2 (0.390) 15,587 20.8 (0.325) 15,587 61.1 (0.459) 11,289 11.8 (0.303) 11,289

9.8 (0.317) 8,798 53.8 (0.462) 11,653 20.5 (0.374) 11,653 66.7 (0.512) 8,485 16.4 (0.402) 8,485

16.3 (0.883) 1,753 40.9 (1.04) 2,236 18.9 (0.828) 2,236 69.5 (1.14) 1,638 21.1 (1.01) 1,638

18.5 (3.00) 168 16.8 (2.53) 220 7.3 (1.75) 220 71.7 (3.24) 194 23.2 (3.04) 194

4.2 (0.017) 1,363,185 30.5 (0.028) 2,619,278 3.5 (0.011) 2,619,278 30.1 (0.043) 1,166,592 9.9 (0.028) 1,166,592

Notes: The table contains population averages within each gross income percentile group. Standard errors are reported in parentheses. Statistics are based on individual data, where wealth and income are split equally among married adults. Individuals are classified as married, single female, or single male. The dummy for children is defined irrespective of civil status. Employees are individuals whose income from employment is larger than income from self-employment. Self-employed are individuals whose income from self-employment is larger than income from employment. Retirees are defined as those drawing social security pensions and being above age 63 (m) or 62 (f), respectively. This variable is not defined in the canton of AG. Working retirees are defined as retirees (according to the aforementioned definition) who have individual labor income larger than zero. Hence their share is defined with respect to the total population, not with respect to the population of retirees. This variable is not defined in the cantons BE, BS, and AG. The two variables on savings toward pillar 3 (private pensions schemes) and pillar 2 (occupational pension schemes) indicate whether individuals claimed such deductions. “Homeowners” is a dummy variable which takes on the value of 1 if someone has an imputed rent in their income tax base. “Retired owners” indicates the share of the total population who are at the same time homeowners and retirees according to the definition above. These last two variables are defined only for the cantons of ZH, OW, BS, SG and AG. Pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

Retired owners (%)

Homeowners (%)

Savings pillar 2 (%)

Savings pillar 3 (%)

Working retirees (%)

Retiree (%)

Self-employed2a (%)

Employed 1 (%)

with children (%)

Single male (%)

Single female (%)

Married (%)

Gross income (1,000 CHF)

Table 4.3

N

N

N

N

N

N

N

N

50 (0.087) 579,804 48.1 (0.066) 579,804 24.1 (0.056) 575,608 28.2 (0.059) 575,608 48.4 (0.066) 579,804 68.0 (0.064) 539,961 4.7 (0.034) 387,453 7.6 (0.036) 526,410

P0–P20

35 (0.048) 260,487 36.4 (0.094) 260,487 34.1 (0.093) 259,403 29.8 (0.090) 259,403 38.0 (0.095) 260,487 74.3 (0.089) 241,883 2.3 (0.037) 165,636 9.8 (0.061) 240,574

P20–P30

44 (0.052) 271,787 36.7 (0.092) 271,787 34.3 (0.091) 270,851 29.2 (0.087) 270,851 38.2 (0.093) 271,787 78.6 (0.081) 256,780 2.9 (0.040) 177,761 11.0 (0.062) 252,283

P30–P40

51 (0.061) 273,621 43.6 (0.095) 273,621 31.4 (0.089) 272,655 25.2 (0.083) 272,655 40.1 (0.094) 273,621 74.5 (0.086) 259,265 3.5 (0.043) 181,393 16.3 (0.074) 250,814

P40–P50 57 (0.060) 273,941 54.2 (0.095) 273,941 25.5 (0.083) 272,945 20.5 (0.077) 272,945 42.4 (0.094) 273,941 71.3 (0.089) 260,570 4.2 (0.047) 182,149 20.2 (0.081) 246,387

P50–P60 61 (0.069) 273,879 62.4 (0.093) 273,879 21.0 (0.078) 272,807 16.7 (0.071) 272,807 42.9 (0.095) 273,879 65.7 (0.093) 262,266 5.3 (0.053) 178,286 25.9 (0.089) 243,968

P60–P70 65 (0.080) 272,732 67.2 (0.090) 272,732 18.7 (0.075) 271,658 14.3 (0.067) 271,658 40.4 (0.094) 272,732 57.9 (0.097) 261,832 6.3 (0.058) 173,323 33.7 (0.096) 241,540

P70–P80

Population averages by net wealth percentile group, 2010

69 (0.101) 274,128 66.3 (0.090) 274,128 20.0 (0.076) 273,008 13.9 (0.066) 273,008 35.7 (0.091) 274,128 48.7 (0.097) 263,105 7.3 (0.062) 175,325 44.1 (0.101) 242,026

P80–P90 80 (0.205) 141,781 62.4 (0.129) 141,781 22.6 (0.111) 140,967 15.2 (0.096) 140,967 36.8 (0.128) 141,781 41.5 (0.134) 135,305 8.1 (0.089) 94,037 52.6 (0.140) 126,356

P90–P95 110 (0.665) 119,378 56.0 (0.144) 119,378 27.0 (0.129) 118,533 17.3 (0.110) 118,533 44.1 (0.144) 119,378 38.2 (0.145) 112,758 8.5 (0.096) 83,625 58.0 (0.150) 107,925

P95–P99 190 (2.12) 14,577 52.8 (0.414) 14,577 29.7 (0.380) 14,477 17.8 (0.318) 14,477 48.8 (0.414) 14,577 38.7 (0.418) 13,562 8.8 (0.276) 10,501 59.3 (0.428) 13,186

P99–P99.5 332 (4.38) 10,676 54.1 (0.482) 10,676 27.8 (0.435) 10,622 18.3 (0.375) 10,622 48.1 (0.484) 10,676 43.7 (0.501) 9,808 7.9 (0.313) 7,458 55.3 (0.505) 9,680

P99.5–P99.9 1,074 (64.6) 2,197 60.5 (1.04) 2,197 23.3 (0.907) 2,172 16.7 (0.800) 2,172 51.0 (1.07) 2,197 48.3 (1.12) 1,998 8.5 (0.703) 1,571 51.3 (1.12) 1,976

P99.9–P99.99

17,386 (4837) 234 55.6 (3.26) 234 29.3 (2.99) 232 15.5 (2.38) 232 48.7 (3.27) 234 47.0 (3.68) 185 6.1 (2.10) 131 53.3 (3.43) 212

P99.99–P100

61 (0.425) 2,769,222 52.3 (0.030) 2,769,222 25.8 (0.026) 2,755,938 22.2 (0.025) 2,755,938 41.6 (0.030) 2,769,222 64.5 (0.030) 2,619,278 5.0 (0.016) 1,818,649 23.0 (0.027) 2,503,337

Total

N

N

N

N

N

1.8 (0.024) 303,879 22.7 (0.057) 539,961 2.0 (0.019) 539,961 28.3 (0.088) 261,739 3.3 (0.035) 261,739

1.2 (0.031) 123,167 12.4 (0.067) 241,883 1.4 (0.024) 241,883 3.8 (0.058) 108,454 0.4 (0.019) 108,454

1.5 (0.033) 134,720 22.2 (0.082) 256,780 1.6 (0.025) 256,780 8.1 (0.081) 112,647 1.0 (0.030) 112,647

2.1 (0.039) 138,482 30.9 (0.091) 259,265 2.1 (0.028) 259,265 14.3 (0.103) 114,600 2.1 (0.042) 114,600

3.2 (0.048) 137,421 39.5 (0.096) 260,570 2.6 (0.031) 260,570 22.7 (0.125) 112,767 4.0 (0.058) 112,767

4.6 (0.058) 131,501 43.1 (0.097) 262,266 3.8 (0.037) 262,266 32.6 (0.142) 108,348 8.0 (0.082) 108,348

6.2 (0.068) 125,758 42.0 (0.097) 261,832 5.0 (0.043) 261,832 43.8 (0.153) 104,901 15.1 (0.111) 104,901

8.2 (0.078) 122,513 37.4 (0.094) 263,105 6.2 (0.047) 263,105 53.2 (0.153) 105,855 24.0 (0.131) 105,855

10.2 (0.118) 65,466 33.1 (0.128) 135,305 7.3 (0.071) 135,305 60.5 (0.199) 60,182 33.3 (0.192) 60,182

12.3 (0.130) 63,649 29.6 (0.136) 112,758 8.8 (0.084) 112,758 61.9 (0.197) 61,101 37.5 (0.196) 61,101

13.6 (0.367) 8,704 26.4 (0.378) 13,562 9.0 (0.246) 13,562 64.1 (0.525) 8,359 38.2 (0.531) 8,359

14.5 (0.441) 6,387 26.1 (0.443) 9,808 9.6 (0.297) 9,808 66.0 (0.605) 6,136 35.5 (0.611) 6,136

17.0 (1.00) 1,420 22.2 (0.930) 1,998 9.2 (0.647) 1,998 67.9 (1.28) 1,336 32.7 (1.28) 1,336

19.5 (3.66) 118 9.2 (2.13) 185 5.9 (1.74) 185 67.1 (3.65) 167 32.3 (3.63) 167

4.2 (0.017) 1,363,185 30.5 (0.028) 2,619,278 3.5 (0.011) 2,619,278 30.1 (0.043) 1,166,592 9.9 (0.028) 1,166,592

Notes: The table contains population averages within each net wealth percentile group. Standard errors are reported in parentheses. Statistics are based on individual data, where wealth and income are split equally among married adults. Individuals are classified as married, single female, or single male. The dummy for children is defined irrespective of civil status. Employees are individuals whose income from employment is larger than income from self-employment. Self-employed are individuals whose income from self-employment is larger than income from employment. Retirees are defined as those drawing social security pensions and being above age 63 (m) or 62 (w), respectively. This variable is not defined in the canton of AG. Working retirees are defined as retirees (according to the aforementioned definition) who have individual labor income larger than zero. Hence their share is defined with respect to the total population, not with respect to the population of retirees. This variable is not defined in the cantons BE, BS, and AG. The two variables on savings toward pillar 3 (private pensions schemes) and pillar 2 (occupational pension schemes) indicate whether individuals claimed such deductions. “Homeowners” is a dummy variable which takes on the value of 1 if someone has an imputed rent in their income tax base. “Retired owners” indicates the share of the total population who are at the same time homeowners and retirees according to the definition above. These last two variables are defined only for the cantons of ZH, OW, BS, SG and AG. Pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

Retired owners (%)

Homeowners (%)

Savings pillar 2 (%)

Savings pillar 3 (%)

Working retirees (%)

Fig. 4.2 Gender composition over the wealth (top) and income (bottom) distributions, 2010 Notes: This figure shows the share of working-age and retired men and women, respectively, along the wealth and income distributions. Retirees are defined as those who draw social security pensions and are allowed to a retire according to their age (early retirement is possible at age 63 for men and at age 62 for women). In the cantons ZH, LU, OW, BS, and AG, where I do not know the ages of the spouses, I impute their ages based on the age structure of couples where the main taxpayer is between 55 and 80 years old in the canton of Bern. This allows me to define retirement for individuals in all cantons except AG, where I lack information on pension income. To enhance visibility in the upper part of the wealth distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. Both panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively. Wealth and income are split equally among married adults.

Unique Swiss Tax Data

125

30 percent of individuals live in their own house or apartment. This is lower than the national home ownership rate of 36 percent in 2010, which can be explained by the fact that I can measure home ownership only in some cantons, including Basel City and Zurich. These are mainly urban areas, where home ownership is considerably lower than in the countryside (Basel City: 15 percent; Zurich: 29 percent). An important factor in understanding the distributions of income and wealth is age. Figure 4.3 shows the age composition along both distributions. Age is monotonically increasing along the wealth distribution, such that wealth is getting older in the aggregate. This finding coincides with findings for the US in Saez and Zucman (2016). The picture is less clear for income. At the bottom one can find both the elderly and the young, while the middleaged cohorts dominate in the range of P50– P99. Interestingly, at the very top of the income distribution the share of people beyond retirement age increases again. This corroborates the finding that the elderly are likely to be overrepresented in both tails of the income distribution. 4.4

The Composition of Wealth along the Distribution

Figure 4.4 shows the composition of total gross wealth along its distribution. The two major private wealth components are financial assets and real estate. For all wealth groups, financial assets are the most important wealth component, making up 30– 90 percent of total gross wealth. Especially in the bottom half of the distribution, financial assets make up the largest asset type, followed by movable assets including cars. Since unincorporated business assets are only available net of debt, their share is very low and almost negligible compared to the US (Saez and Zucman 2016) or Spain (Martínez-Toledano 2020).4 When it comes to real estate, the distribution is very concentrated in the upper middle class. Those below the median of the gross wealth distribution own hardly any real estate, a finding that corresponds well with the low rate of home ownership in Switzerland and in my data. Results are qualitatively similar when looking at shares along the net wealth distribution, reported in online appendix figure A2 (http://www.nber.org /data-appendix /c14452 /appendix.pdf): those below the median of the net wealth distribution on average hold less than 20 percent of their wealth in real estate. Even if real estate were still undervalued by 20 percent throughout my data, despite attempts to correct for undervaluation, the picture would hardly change. While I cannot compute assets net of debt, figure 4.4 reveals that debt 4. In addition, the Swiss legal system favors limited liability companies (LLC) over partnerships and sole proprietorship. Founding a limited liability company requires founding capital in the amount of merely 20,000 CHF. An LLC is much easier to set up than a corporation, provides more flexibility and capital requirements are much lower than for setting up a corporation. Yet an LLC offers more protection than a partnerships and sole proprietorship.

126

Fig. 4.3 2010

Isabel Z. Martínez

Age composition over the wealth (top) and income (bottom) distributions,

Notes: This figure shows the share of individuals in each wealth (4.3a) and income group (4.3b), respectively. I drop spouses from the cantons ZH, LU, OW, BS, and AG, as I do not know the ages of the spouses in these cantons. Spouses represent about 25 percent of individuals in each canton. To enhance visibility in the upper part of the wealth distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. Both panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively. Wealth and income are split equally among married adults.

strongly mirrors real estate shares in gross wealth. This is true overall and by age group (see online appendix figure A3). Indeed, roughly 90 percent of all private debt is mortgage debt according to national account statistics.5 Real estate therefore plays a much less important role in the portfolios of most Swiss taxpayers compared to other countries. A strikingly different case is Spain, a country with a home ownership rate of approximately 82 per5. Because interest payments on debt can be deducted from taxable income and taxable wealth is always net of debt, it is common for homeowners to never fully pay off their mortgage.

Unique Swiss Tax Data

Fig. 4.4

127

Wealth composition along the distribution in Switzerland, 2010

Notes: This figure shows shares of wealth components in total gross wealth along the gross wealth distribution. Since debts cannot be linked directly to a single wealth component, debt is displayed as negative share in total gross wealth. A significant number of individuals in the lower part of the wealth distribution have no or hardly any assets, but they have debts, resulting in extremely large debt shares. Individual’s debt shares were therefore truncated at 500 percent. Wealth is split equally among married adults. Figure 4.4a uses data on 2.77 million individuals from all eight available cantons. Figure 4.4b uses data on 1.44 million individuals from three large cantons (ZH, SG, AG), which allows further decomposition of real estate wealth into owner-used and other real estate. See section 4.3.3 for details on the wealth components. To enhance visibility in the upper part of the wealth distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. Both panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

cent compared to 36 percent in Switzerland.6 Martínez-Toledano (2020) shows that in Spain real estate amounts to 90 percent of total net wealth for individuals around the median (P40– P60).7 In three cantons I can further decompose real estate into owner occupied and other real estate. Figure 4.4b shows that other real estate— which includes properties for rent, business properties, and land, but excludes personally used real estate like a main residence or vacation home— is significant only above the fourth quintile, especially for those belonging to the P95– P99.9 wealth groups. Real estate is a viable option for relatively safe 6. Switzerland has the lowest home-ownership rate across all Europe, while Spain has the highest in Western Europe. Source: Federal Statistical Office (FSO), https:// www .bwo.admin .ch /dam /bwo /de /dokumente /01 _Wohnungsmarkt /16 _Zahlen _und _Fakten /163 _Wohneigentumsquote /wohneigentumsquotenschweiz -eu2008 .pdf .download .pdf /wohneigentumsquotenschweiz-eu2008.pdf. 7. Martínez-Toledano (2020) uses net values while I have to rely on gross values.

128

Isabel Z. Martínez

investments for the wealthy, who may further use investments in real estate for diversification of their portfolio. The share of real estate investments has likely increased since 2010, due to the low and even negative interest rates prevailing in Switzerland since December 2014. As I cannot attribute debt to the different asset categories, I plot debt as share of total gross wealth. For the P0– P20 group, debt shares become very large due to very small denominators.8 I truncate individual debt shares at 500 percent to keep the graphs readable. Apart from the bottom, debt shares are largest at P60– P70. I find that debt is highly correlated with home ownership, implying that household debt it is strongly driven by mortgages. Figure 4.5 shows how the wealth composition changes when looking at subgroups by employment status and gender. Percentile thresholds are held constant across graphs: they correspond to the percentiles of the total gross wealth distribution and I hold them constant over the whole analysis. This allows for direct comparisons across groups. The biggest difference arises between retirees and nonretirees, shown in figures 4.5a and 4.5b. At each gross wealth level, retirees have lower debt levels than nonretirees. This reflects the life-cycle pattern of real estate acquisition in younger years, and the reduction of mortgage debt as individuals age. I find that retirees’ debt levels keep falling even after retirement, suggesting that the elderly keep saving during their 60s and 70s. At the same time, real estate is more evenly spread across the distribution in the case of retirees. Since the analysis is based on cross-section data, I cannot distinguish age from cohort effects here. Finally, business assets as well as movable assets like cars, which are concentrated at the bottom of the distribution, are less important among the retired population. Differences between single men and women (not shown) can be attributed to the different likelihood of being retired for men and women: 33 percent of single women are retirees compared to only 17 percent of single men. That is why single women have on average less debt, less business assets, and less movable assets but more real estate than men. Taking into account retirement status, gender differences are small (figure 4.5c– f): women still tend to have slightly lower debt and less business assets (net of debt) than men. At the very top, working women tend to have more movable assets and retired women have more rented out real estate than men— at the expense of financial assets. 4.5

The Composition of Income along the Distribution

The composition of total gross income varies considerably along the distribution as shown in figure 4.6.9 For the bottom 30 percent of the income 8. I attribute those roughly 10 percent of individuals with zero gross wealth one Swiss Franc of wealth to compute the debt shares, rather than dropping them. 9. Some individuals have negative incomes. For this part of the analysis, I drop observations with negative incomes, as it is not possible to represent these appropriately as shares. This leads

Unique Swiss Tax Data

Fig. 4.5 2010

129

Wealth composition by employment status and gender in Switzerland,

Notes: This figure shows shares of wealth components in total gross wealth along the gross wealth distribution. Individual’s debt shares were truncated at 500 percent. Wealth is split equally among married adults. In each panel, individuals are ranked according to the percentiles of the total gross wealth distribution for the entire population, hence wealth levels across panels are identical. Figure 4.5a shows the composition of gross wealth for nonretirees. The composition of wealth for retirees is displayed in figure 4.5b. Retirees are defined as those who draw social security pensions and are allowed to retire according to their age (early retirement is possible at age 63 for men and at age 62 for women). In the cantons ZH, LU, OW, BS, and AG, where I do not know the ages of the spouses, I impute their ages based on the age structure of couples where the main taxpayer is between 55 and 80 years old in the canton of Bern. This allows me to define retirement for individuals in all cantons except AG, where I lack information on pension income. Figures 4.5c–f further split the population by gender. For the main taxpayer, gender is reported in the individual tax data. In case of married individuals and in cantons where the gender of the spouse is not recorded, I assume the spouse is of opposite gender from the main taxpayer. See section 4.3.3 for details on the wealth components. To enhance visibility in the upper part of the distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. All panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively. Since not all cantonal data contain all variables, some panels rely on data from fewer cantons as indicated.

distribution, labor income is equally important as transfers. Together, they make up 90 percent of total gross income. Moving up the distribution, the importance of labor income increases at the expense of transfers. The remaining 10 percent are capital incomes, mostly imputed rents (see figure 4.6b). Within the top 10 percent, however, the composition changes considerably: not only transfer income, also the share of labor income declines to a loss of 7.7 percent of observations. I recompute the income percentiles, yet they remain unchanged below P90 and only change slightly above.

130

Fig. 4.5

Isabel Z. Martínez

(cont.)

strongly. For those in the top 0.01 percent of the income distribution, labor income represents about one-fifth of their total income. Three-fifths can be attributed to capital income (including income from real estate, see figure 4.6b), and almost one-fifth are other incomes— typically one-time capital payments, including capital gains (or losses) from business liquidation in the event of definitive cessation of self-employment. Hence these incomes distinguish the richest 1 percent within the top 1 percent considerably from the rest. Figure 4.6b further shows the subcomponents of each income component. Social security pensions are the most important type of transfer for

Unique Swiss Tax Data

Fig. 4.6

131

Income distribution and its components in Swiss Cantons, 2010

Notes: This figure shows shares of income components in total gross income along the gross income distribution. Income is split equally among married adults. Observations with negative incomes are dropped from this figure. Figure 4.6a uses 2.55 million individual observations from all eight available cantons. Figure 4.6b contains detailed information on income from pensions and real estate income for 1.30 million individuals. This information is not available for BE, BS, and AG. See section 4.3.2 for details on the income components. To enhance visibility in the upper part of the income distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. Both panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

low-income individuals, indicating the high share of retirees at the bottom of the income distribution. For most individuals, capital income consists of imputed rents from home ownership. Rental income from real estate is only relevant for those within the top 10 percent of the income distribution. Also, the share of income from self-employment is largest within the top 10 percent. For the middle and upper parts of the distribution in the range of P40– P99, income from employment makes up more than 50 percent of all income. When adding income from self-employment, even the bottom 90 percent within the top 1 percent (P99– P99.9) earn more than half their income through work. As before, I split the sample into different subgroups holding the percentile thresholds constant (figure 4.7). Again, the most important differences arise between nonretired individuals and retirees (figure 4.7a– b). Unsurprisingly, labor income is the most important source of income for the workingage population. On average, up to 90 percent of individuals’ gross income comes from work. Yet again, the share of income from work declines further up in the distribution, especially from the top 5 percent and beyond. There,

132

Fig. 4.7 2010

Isabel Z. Martínez

Income composition by employment status and gender in Switzerland,

Notes: This figure shows shares of income components in total gross income along the gross income distribution. Income is split equally among married adults. In each panel, individuals are ranked according to the percentiles of the total gross income distribution for the entire population, hence income levels across panels are identical. Figure 4.7a shows the composition of gross income for nonretirees. The composition of income for retirees is displayed in figure 4.7b. Retirees are defined as those who draw social security pensions and are allowed to retire according to their age (early retirement is possible at age 63 for men and at age 62 for women). In the cantons ZH, LU, OW, BS, and AG, where I do not know the ages of the spouses, I impute their ages based on the age structure of couples where the main taxpayer is between 55 and 80 years old in the canton of Bern. This allows me to define retirement for individuals in all cantons except AG, where I lack information on pension income. Figures 4.7c–f further split the population by gender. For the main taxpayer, gender is reported in the individual tax data. In case of married individuals and in cantons where gender of the spouse is not recorded, I assume the spouse is of opposite gender from the main taxpayer. See section 4.3.2 for details on the income components. To enhance visibility in the upper part of the distribution, percentile steps for the top 10 percent are displayed in smaller increments and the lowest 20 percent are summarized together. All panels use pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively. Since not all cantonal data contain all variables, some panels rely on data from fewer cantons as indicated.

incomes from capital, real estate, and other sources take over. Note, however, that on average about 35 percent of the income of individuals belonging to the top 0.01 percent still consists of labor income, mostly from employment. At the bottom end of the distribution, transfers including unemployment benefits and family allowances contribute to the income mix. For nonretirees, social security pensions refer to disability benefits, which are transformed into a social security pension at retirement. For retirees, pensions replace labor income by and large. Moving up the income distribution, however, income from capital and real estate becomes relatively more important than for the working-age population. Together,

Unique Swiss Tax Data

Fig. 4.7

133

(cont.)

these two sources make up almost 90 percent of the income going to retirees belonging to the top 0.01 percent. Interestingly, the share of income from employment, while small overall, increases as one moves up the distribution. In Switzerland it is therefore not those retirees who have low (pension) incomes who have the highest likelihood of continuing to work beyond retirement, but those who were successful in the labor market or their own business, respectively, and who can derive large incomes (in line with results in table 4.2). Online appendix figure A4 (http://www.nber.org /data-appendix /c14452/appendix.pdf) further shows the relative importance of transfers and labor income over the

134

Isabel Z. Martínez

distribution at different points over the life cycle, breaking up the income composition by age group. Besides labor income, very wealthy retirees can draw large incomes from capital and real estate. The latter stems mainly from renting out properties and not from imputed rent. Since the share of real estate wealth does not differ that much between retirees and the working-age population, this suggests that the elderly derive larger incomes from their real estate than younger generations, and again that the correlation between the income and wealth distribution varies by age group. There are some noteworthy differences by gender within the working-age and the retiree populations, respectively (figures 4.7c– d and 4.7e– f). First, nonretired women tend to draw a lower share of their income from labor than men, especially at the very top. Second, nonretired women in the middle of the distribution draw larger incomes from pensions (e.g., widhowhood or disability pensions) and especially from transfers (e.g., alimony, maternity or unemployment benefits) than their male counterparts. Third, among retirees, men are much more likely to earn income from work, especially those within the top 10 percent. Retired men belonging to the top 0.01 percent of the income distribution on average earn 20 percent of their total income from labor, a share that drops to almost zero for retired women in the same income class. Gender differences are very similar or even more pronounced when looking at singles only, whose individual income is not affected by splitting income equally among partners. Taken together, these findings reflect gender differences in the Swiss labor market. Although the labor force participation of women is high in international comparison, 44 percent of women work part-time. In 2010, the hours-adjusted wage gap was 15.6 percent. Only 62 percent of this difference could be explained by observables like education, industry, or job characteristics10—an indicator that gender discrimination against women is present in the Swiss labor market. Since these gender differences were also present (and even more pronounced) in the past, they translate into gender differences among retirees, too. 4.6

Joint Distribution of Income and Wealth

Next, I turn to the joint distribution of income and wealth. Figure 4.8 shows the joint distribution matrix of income and wealth. Figure 4.8a looks at how each income group on the y axis is distributed over the net wealth distribution. There is a clear tail dependence between the two distributions, especially at the very top: 78 percent of those in the top 0.01 percent of the income distribution belong to the top 0.1 percent of the wealth distribution. 10. Source: FSO, https://www.bfs.admin.ch /bfs/de/home/statistiken /arbeit-erwerb/loehne -erwerbseinkommen-arbeitskosten /lohnniveau-schweiz /lohnunterschied.html.

Unique Swiss Tax Data

Fig. 4.8

135

Joint distribution matrix by income and wealth group

Notes: This figure shows the joint distribution of individuals across the gross income and net wealth distributions. Figure 4.8a shows how income groups are distributed over the wealth distribution. For each income group on the y axis, the matrix shows the share of individuals from that group in each wealth group (relative row frequencies). Figure 4.8b shows how wealth groups are distributed over the income distribution. For each wealth group on the y axis, the matrix shows the share of individuals from that group in each income group (relative row frequencies). In both figures, the shares in each row sum to 100 percent (note that columns do not add to 100 percent). Analysis based on individual data, where wealth and income are split equally among married adults. Pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

Interestingly, within all income groups there is a substantial part belonging to the bottom of the wealth distribution, where average and median net wealth are negative. Even among those belonging to the top 0.01 percent of the income distribution, 5 percent fall into this lowest category of wealth. The fact that more mass lies below the main diagonal suggests that an individual is relatively unlikely to be in a higher wealth group relative to their

136

Isabel Z. Martínez

income group. At the same time, a relatively large mass is concentrated near though below the main diagonal. Therefore, many individuals are likely to be near to their position in the income distribution within the wealth distribution. Given that in many data we observe income but not wealth, these findings are helpful when assumptions need to be made about individual’s position in the wealth distribution— without necessarily knowing that position nor the corresponding wealth levels. Figure 4.8b takes on the opposite perspective and looks at how each wealth group on the y axis is distributed over the gross income distribution. Again, the tail dependency is clearly visible: those at the top of the wealth distribution are very likely to be at the top of the income distribution and those with hardly any wealth are concentrated at the bottom of the income distribution. However, note that very few of those belonging to the top 10 percent of the wealth distribution have low incomes. This implies that only few people among the wealth-rich are income-poor— while the opposite is not true, as shown in figure 4.8a. Taking a closer look at the top of both distributions, figure 4.9 shows how the P99– P99.9 and the top 0.01 percent of the wealth distribution are distributed over the income distribution and vice versa. The graph reveals how those belonging to the P99– P99.99 group— i.e., the bottom 99 percent within the top 1 percent— and the top 0.01 percent are substantially different: those belonging to the P99– P99.99 of the wealth distribution can be found over the whole income distribution, even though their share is highest among the top 10 percent of income earners, especially in the P95– P99 group. The top 0.01 percent of the wealthiest, in contrast, can only be found among the top 5 percent of income earners, and are mainly part of the top 0.1 percent of earners. The picture is similar when flipping the axes: the top 0.01 percent of income earners are almost exclusively found in the top 1 percent of wealth holders. The correlation between income and wealth therefore increases as one moves up along the tails. Gender differences are small, which is partly mechanical, as I have to split wealth and income equally between spouses. Looking at singles only, however, reveals that compared to women, men tend to be higher up in the income distribution given their wealth rank. In the bottom nine deciles of the income distribution, women tend to be distributed more evenly across the wealth distribution than men and they are less likely to find themselves in the bottom quintile of the wealth distribution (i.e., the first column of figure 4.8a). Within the top 10 percent and especially within the top 1 percent, however, the association between income and wealth is even stronger for women than for men. Finally, the strong age gradient in wealth also affects the joint distribution. The rank correlation remains relatively low for individuals below retirement age and almost doubles for individuals beyond age 65. The reason is because at all income levels retirees are on average considerably higher up

Fig. 4.9 Joint distribution of individuals in top 1 percent and top 0.01 percent wealth and income groups Notes: This figure shows the distribution of the top 1 percent and top 0.01 percent, respectively, of the gross income (net wealth) distribution over the net wealth (gross income) distribution. Figure 4.9a shows where those belonging to the P99–P99.99 of the wealth distribution are located in the income distribution (results are very similar for the P99–P99.9 fractile and the top 1 percent). Figure 4.9b shows where those belonging to the P99–P99.99 of the income distribution stand in the wealth distribution (results are very similar for the P99–P99.9 fractile and the top 1 percent). Figures 4.9c and 4.9d show the same relationships for the top 0.01 percent, i.e., the top 1 percent within the richest 1 percent. Analysis based on individual data, where wealth and income are split equally among married adults. Pooled tax data including the cantons BE, LU, OW, AG, SG, BS, and JU in the year 2010, and ZH in 2011, respectively.

138

Isabel Z. Martínez

Fig. 4.9

(cont.)

in the wealth distribution than nonretirees. The only exceptions are retirees in the bottom decile of the income distribution: they are just as likely to be wealth-poor as their nonretired counterparts, which in turn retains the tail dependency at the lower end of the distribution. As this is a static analysis for the year 2010, it is unfortunately not possible to disentangle age from cohort effects. It is quite possible that for younger cohorts the relationship between their income and wealth ranks will be different when they reach the same age range. 4.7

Conclusion

In this chapter, I construct a new dataset out of cantonal income and wealth tax data, which covers about half of the Swiss population of tax-

Unique Swiss Tax Data

139

payers and is representative of Switzerland as a whole. This unique dataset allows me to obtain detailed evidence on the composition of wealth and income in Switzerland and on the joint distribution of income and wealth. The results can be summarized as follows. First, I find that age has a very strong influence on the distribution of wealth among individuals. The older individuals are, the more likely they are to be wealthy. While this is to be expected intuitively, I show how pronounced this effect is. What is surprising is that this age effect continues well beyond retirement age. This strong age-wealth gradient in turn dominates many findings in this chapter: retirees are strongly overrepresented in the top decile of the wealth distribution, they have low debt and are more likely to hold real estate than younger cohorts. Second, gender differences (taking age into account) are stronger in the distribution and composition of income than wealth. Women draw a lower share of their income from labor, hence they rely more heavily on transfers or— at the very top— on capital incomes than men. This reflects past and present gender differences in the Swiss labor market. Despite the high female labor force participation of 75 percent (in 2010; 80 percent in 2019), 60 percent of women work part-time (especially mothers). The wealth composition, in contrast, bears only very small gender differences once age is taken into account. However, this finding masks wealth differences in tax-exempt retirement accounts, which compound labor income differences between men and women. Third, the distribution of real estate wealth along the wealth distribution sets Switzerland apart from other economies. On average, those in the bottom 50 percent of the distribution hardly hold any real estate wealth. This finding— which corresponds well with official statistics on home ownership, according to which less than 40 percent of the population are homeowners— has potential implications for the optimal design of wealth and property taxation. While there are in principle several different forms of wealth to tax, a typical distinction is made between (owner-used) real estate and other wealth when it comes to taxation. Finally, I shed light on the joint distribution of income and wealth. The joint distribution of income and wealth reveals a strong tail dependence, especially at the top. As wealth is more concentrated than income, an individual is relatively unlikely to be in a higher wealth group relative to their income group. Through the strong age-wealth gradient, age also affects the joint distribution: at almost all income levels, retirees are in substantially higher wealth percentiles. Overall, I find that while a substantial share of top earners have very low wealth, those belonging to the top of the wealth distribution are very unlikely to have low incomes. Low-income wealth millionaires are therefore a rare phenomenon. The new dataset and results presented here form part of the ongoing research project “The Influence of Taxation on Wealth and Income Inequality” (SNFS Grant 176458). This study shows the potential of this rich data-

140

Isabel Z. Martínez

set and sheds light on income and wealth in Switzerland. The goal of future research is to exploit the richness of this data to understand drivers of the observed patterns. Upon data availability, future analyses will also look at changes over time.

References Aaberge, Rolf, Anthony B. Atkinson, and Sebastian Königs. 2018. “From Classes to Copulas: Wages, Capital, and Top Incomes.” Journal of Economic Inequality 16: 295– 320. Agrawal, David R., and Dirk Foremny. 2019. “Relocation of the Rich: Migration in Response to Top Tax Rate Changes from Spanish Reforms.” Review of Economics and Statistics 101 (2): 214– 32. Agrawal, David R., Dirk Foremny, and Clara Martínez-Toledano. 2020. “Wealth Tax Mobility and Tax Coordination.” http://dx.doi.org/10.2139/ssrn.3676031. Atkinson, Anthony B., and Thomas Piketty. 2007. Top Incomes over the Twentieth Century: A Contrast between European and English-Speaking Countries. Oxford: Oxford University Press. Atkinson, Anthony B., and Thomas Piketty. 2010. Top Incomes: A Global Perspective. Oxford: Oxford University Press. Brülhart, Marius, Jonathan Gruber, Matthias Krapf, and Kurt Schmidheiny. 2019. “Behavioral Responses to Wealth Taxes: Evidence from Switzerland.” CESifo Working Paper No. 7908. http://dx.doi.org/10.2139/ssrn.3477721. Chauvel, Louis, Anne Hartung, Eyal Bar-Haim, and Philippe Van Kerm. 2019. “Income and Wealth above the Median: New Measurements and Results for Europe and the United States.” In What Drives Inequality?, edited by Koen Decanq and Philippe Van Kerm, 89– 104. Bingley: Emerald. Dell, Fabian, Thomas Piketty, and Emmanuel Saez. 2007. “Income and Wealth Concentration in Switzerland over the Twentieth Century,” In Top Incomes Over the Twentieth Century: A Contrast between Continental European and EnglishSpeaking Countries, edited by Anthony B. Atkinson and Thomas Piketty, 472– 500. Oxford: Oxford University Press. Fisher, Jonathan D., David S. Johnson, Timothy M. Smeeding, and Jeffrey P. Thompson. 2022. “Inequality in 3-D: Income, Consumption, and Wealth.” Review of Income and Wealth 68 (1): 16–42. Föllmi, Reto, and Isabel Z. Martínez. 2017. “Volatile Top Income Shares in Switzerland? Reassessing the Evolution between 1981 and 2010.” Review of Economics and Statistics 99 (5): 793– 809. Frey, Christian, Christoph Gorgas, and Christoph A. Schaltegger. 2016. “The Long Run Effects of Taxes and Tax Competition on Top Income Shares: An Empirical Investigation.” Review of Income and Wealth 63 (4): 792– 820. https://doi.org /10 .1111/roiw.12228. Gallusser, David, and Matthias Krapf. 2019. “Joint Income-Wealth Inequality: An Application Using Administrative Tax Data.” CESifo Working Paper Series 7876. Munich: CESifo. Jäntti, Markus, Eva Sierminska, and Timothy M. Smeeding. 2008. “The Joint Distribution of Household Income and Wealth: Evidence from the Luxembourg Wealth

Unique Swiss Tax Data

141

Study.” OECD Social, Employment and Migration Working Papers No. 65. Paris: OECD. Jäntti, Markus, Eva Sierminska, and Philippe Van Kerm. 2015. “Modelling the Joint Distribution of Income and Wealth.” IZA Discussion Paper No. 9190, July. Bonn: Institute of Labor Economics. Kleven, Henrik, Camille Landais, Mathilde Muñoz, and Stefanie Stantcheva. 2020. “Taxation and Migration: Evidence and Policy Implications.” Journal of Economic Perspectives 34 (2): 119– 42. Kopczuk, Wojciech, and Emmanuel Saez. 2004. “Top Wealth Shares in the United States, 1916– 2000: Evidence from Estate Tax Returns.” National Tax Journal 57 (2): 445– 88. Kuhn, Moritz, Moritz Schularick, and Ulrike I. Steins. 2020. “Income and Wealth Inequality in America, 1949– 2016.” Journal of Political Economy  128 (9): 3469– 519. Lindner, Peter, and Martin Schürz. 2020. “The Joint Distribution of Wealth, Income and Consumption in Austria: A Cautionary Note on Heterogeneity.” Monetary Policy and the Economy 19 (4): 57– 76. Martínez, Isabel. 2017. “Beggar-Thy-Neighbour Tax Cuts: Mobility after a Local Income and Wealth Tax Reform in Switzerland.” LISER Working Paper No. 2017– 08. Luxembourg: Institute of Socio-Economic Research. https://dx.doi .org /10.2139/ssrn.2979275. Martínez-Toledano, Clara. 2020. “House Price Cycles, Wealth Inequality and Portfolio Reshuffling.” WID.world Working Paper No. 2020/02. https://wid.world /document/house-price-cycles-wealth-inequality-and-portfolio-reshuffling-wid -world-working-paper-2020 -02/. Moser, Peter. 2019. “Vermögensentwicklung und -mobilität. Eine Panelanalyse von Steuerdaten des Kantons Zürich 2006– 2015.” Statistik.info 02/2019, Statistisches Amt des Kantons Zürich. Neelakantan, Urvi, and Yunhee Chang. 2010. “Gender Differences in Wealth at Retirement.” American Economic Review 100 (2): 362– 67. Organization for Economic Cooperation and Development (OECD). 2018. “The Role and Design of Net Wealth Taxes in the OECD.” OECD Tax Policy Studies No. 26. Paris: OECD. Peichl, Andreas, and Nico Pestel. 2013. “Multidimensional Affluence: Theory and Applications to Germany and the US.” Applied Economics 45 (32): 4591– 601. Piketty, Thomas, Li Yang, and Gabriel Zucman. 2019. “Capital Accumulation, Private Property, and Rising Inequality in China, 1978– 2015.” American Economic Review 109 (7): 2469– 96. Ruiz, Nicolas. 2011. “Measuring the Joint Distribution of Household’s Income, Consumption and Wealth Using Nested Atkinson Measures.” OECD Statistics Working Paper No. 2011/05. Paris: OECD. Saez, Emmanuel, and Gabriel Zucman. 2016. “Wealth Inequality in the United States since 1913: Evidence from Capitalized Income Tax Data.” Quarterly Journal of Economics 131 (2): 519– 78. Schaltegger, Christoph A., and Christoph Gorgas. 2011. “The Evolution of Top Incomes in Switzerland over the 20th Century.” Swiss Journal of Economics and Statistics 147 (4): 479– 519. Schneebaum, Alyssa, Miriam Rehm, Katharina Mader, and Katarina Hollan. 2018. “The Gender Wealth Gap across European Countries.” Review of Income and Wealth 64 (2): 295– 331. Sierminska, Eva M., Andrea Brandolini, and Timothy M. Smeeding. 2007. “CrossNational Comparison of Income and Wealth Status in Retirement: First Results

142

Isabel Z. Martínez

from the Luxembourg Wealth Study (LWS).” Working Paper No. 2007– 3. Boston: Center for Retirement Research at Boston College. Sierminska, Eva M., Joachim R. Frick, and Markus M. Grabka. 2010. “Examining the Gender Wealth Gap.” Oxford Economic Papers 62 (4): 669– 90. Wolfson, Michael C. 1979. “Wealth and the Distribution of Income, Canada 1969– 70.” Review of Income and Wealth 25 (2): 129– 40.

II

Wealth Inequality

5

The Wealth of Generations, with Special Attention to the Millennials William G. Gale, Hilary Gelfond, Jason J. Fichtner, and Benjamin H. Harris

5.1

Introduction

The patterns and determinants of household wealth accumulation have long been of interest to economists, with seminal contributions dating back at least to Ando and Modigliani (1963), Friedman (1957), and Modigliani and Brumberg ([1954] 2005). Recent work by Piketty (2014) and Saez and Zucman (2019) has sparked a new generation of research interest in this topic. Wealth accumulation is of interest for several reasons. At the household level, wealth provides a source of future consumption, as well as insurance William G. Gale is the Arjay and Frances Fearing Miller Chair in Federal Economic Policy in the Economic Studies Program at the Brookings Institution. Hilary Gelfond is an analyst in the Quantitative Economics and Statistics (QUEST) practice of EY. Jason J. Fichtner is Vice President and Chief Economist of the Bipartisan Policy Center, and was a senior lecturer at the Johns Hopkins University School of Advanced International Studies when this chapter was written. Benjamin H. Harris is a visiting associate professor and Executive Director of the Kellogg Public-Private Interface at the Kellogg School of Management, Northwestern University. This chapter is an updated version of a paper prepared for the CRIW-NBER conference on “Measuring and Understanding the Distribution and Intra/Inter-Generational Mobility of Income and Wealth,” Bethesda, Maryland, March 5– 6, 2020. The authors thank Alec Camhi, Grace Enda, Claire Haldeman, Victoria Johnson, Aaron Krupkin and Lucie Parker for expert assistance; Karen Dynan, Douglas Holtz-Eakin, Alicia Munnell, and conference participants for helpful comments; and John Sabelhaus for sharing data. Part of the work on this project was funded by the US 2050 project, supported by the Peter G. Peterson Foundation and the Ford Foundation. The statements made and views expressed are solely the responsibility of the authors. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber.org /books-and -chapters /measuring-distribution -and -mobility-income -and -wealth /wealth -generations -special-attention-millennials.

145

146

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

against adverse economic shocks. At the aggregate level, wealth finances domestic and foreign investment, affects current consumption spending, and influences the efficacy of monetary and fiscal interventions. More broadly, as discussed further below, the sheer magnitude of changes in aggregate household wealth relative to GDP in recent decades merits attention. Documenting and determining the causes of changes in the level and distribution of household wealth and its components across generations and over time is an extraordinarily ambitious goal. This chapter takes several initial steps in that general direction, building on Gale, Gelfond, and Fichtner (2019), Gale and Harris (2020), and Gale and Pence (2006), and using data from the 1989 to 2016 waves of the Federal Reserve Board’s Survey of Consumer Finances (SCF). We obtain several key results. First, while the Great Recession in 2007– 9 reduced wealth in all age groups, the broader long-term trend has been that the wealth of older age groups has increased while the wealth of successive cross-sections of younger age groups has fallen. A significant share of these changes, in both directions, can be explained by the evolution of household demographic and economic characteristics. Second, we show that the millennial generation— people who were born between 1981 and 1996 and hence were between the ages of 20 and 35 in 2016— had less median and mean wealth in 2016 than any similarly aged cohort between 1989 and 2007. Predicting future wealth accumulation patterns is difficult, but we note that the millennials have certain advantages over previous generations in terms of wealth accumulation. They are the most educated generation in history and generally have higher earnings than their predecessors. Because of the evolution of the pension system toward defined contribution (DC) plans, millennials may well work longer than any previous generation, giving them additional years to save. And millennials may well end up inheriting more than any prior generation. Millennials also face numerous disadvantages. Their careers had rocky starts because of the financial crisis and Great Recession. They will be employed in contingent workforce jobs (which are more uncertain and have weaker benefits than traditional jobs) to a greater extent than previous generations. They are marrying, buying homes, and having children later. Longer lifespans mean that they have to accumulate more wealth, all else equal, to maintain preretirement living standards in retirement. Because their parents are living longer than previous generations did, millennials will also receive inheritances later in life. They will face increased burdens from any eventual resolution of the government’s long-term fiscal shortfalls in general, and the financial imbalances in Social Security and Medicare in particular. They face an economic future with projections of lower rates of return and economic growth than in the past. Third, we highlight the role that minorities will play in determining wealth prospects for millennials. Minorities constitute a substantially larger share

The Wealth of Generations: Millennials

147

of the millennial population than they do in any previous generation. Using cross-section and pooled regressions, we show that minority status is negatively associated with net worth, controlling for other household characteristics. The difference in wealth between Black and white households appears to be growing over time, controlling for other factors. One overarching caveat to all of the results and analysis is that the chapter applies to the period before the COVID-19 pandemic, an enormous shock that will clearly have significant impacts on wealth accumulation patterns for a wide range of birth cohorts. The chapter is thus best interpreted as addressing generational wealth patterns through 2016 and providing a pre-COVID benchmark against which future studies can be compared. The rest of the chapter is organized as follows. Section 5.2 describes the SCF data. Section 5.3 analyzes wealth accumulation over time for a wide range of birth cohorts. Section 5.4 discusses the status of the millennials as of 2016 and the advantages and disadvantages they face relative to prior generations. Section 5.5 addresses issues related to minorities and wealth accumulation. Section 5.6 concludes. 5.2

Survey of Consumer Finances

The SCF is a triennial household survey that is generally considered to provide the most reliable and complete survey-based measures of household wealth (or net worth, terms we use interchangeably below).1 The surveys covering the period 1989 to 2016 follow a generally consistent methodology. Raw sample sizes vary from about 3,100 to about 6,200 in surveys during that period. The survey includes information on characteristics of household demographics, income, assets, debts, and others. To capture how assets and debts are held broadly in the population, about two-thirds of the unweighted sample are drawn from a stratified, nationally representative random sample. The remainder of the sample is randomly selected from statistical records derived from tax returns, using a stratification technique that oversamples households likely to have substantial wealth. This sample design allows for more efficient and less biased estimates of wealth than are generally feasible through simpler designs. In particular, oversampling the wealthy is an important component of the survey, because wealth is so highly concentrated. All of the data presented in this chapter represent weighted statistics, using the sample weights provided by the SCF, which correct for selection probabilities and nonresponse. The SCF uses a multiple imputation procedure to fill in missing data. Five implicates form an approximate distribution of the missing data, creating 1. The SCF is conducted by the nonpartisan and objective research organization NORC at the University of Chicago on behalf of the Federal Reserve Board and with the cooperation of the Department of the Treasury.

148

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

a sample that is five times larger than the actual sample. For descriptive statistics, we use all five implicates by dividing the sample weights by five. In our regressions, we use the first implicate only. 5.3 5.3.1

Wealth across Generations Framework

We analyze the influence of changes in demographic characteristics on wealth accumulation across cohorts by utilizing basic median and ordinary least squares (OLS) regressions, in the absence and presence of demographic variables. We run median (least absolute deviation, or LAD) regressions and OLS regressions, pooling data from the 1989 and 2016 SCFs. We break the data into four age-category subsets, one for 25-to-34-year-olds, one for 35-to-44-year-olds, one for 45-to-54-year-olds, and one for 55-to-64-year-olds. For each household i in each age category k, we specify wealth as a function of a constant and a survey year indicator variable: (5.1)

wk1 = αk1 + βk1(year = 2016)i + εk1i .

In this model, the coefficient βk1 captures the change in median or mean wealth between the 1989 and 2016 samples of each age category. In a second basic regression specification, we add a vector of demographic indicators, denoted by X. This demographic specification is described in detail below: (5.2)

wk2 = αk2 + βk2(year = 2016)i + γk2Xi + εk2i .

If demographic changes explain most of the difference in wealth between 1989 and 2016 for age category k, βk2 should be close to zero, and the coefficients for the variables in the demographic vector should be statistically significant. For example, if βk1 = $100,000 and βk2 = $20,000 (and both are estimated precisely), we would say that demographics variables explained 80 percent of the rise in wealth accumulation. This approach could be expanded to account for the range of possible outcomes that exist to statistical imprecision, but in this chapter we take a “first cut” at looking at how demographic factors matter. Notably, the specifications above assume that the relationship between wealth and demographic characteristics is the same in both years (other than a shift in the intercept).2 5.3.2

Specification of Demographic Characteristics

The survey respondent and the household head are not necessarily the same person in the SCF. The SCF designates the household head to be the 2. Gale and Pence (2006) implement a similar approach, as well as a Blinder-Oaxaca decomposition (Blinder 1973; Oaxaca 1973) and counterfactual distributions based on DiNardo, Fortin, and Lemieux (1996) and Machado and Mata (2005). These different approaches generated similar conclusions.

The Wealth of Generations: Millennials

149

male in a mixed-gender relationship and the older partner in a same-gender relationship, while the respondent is (supposed to be) the person most familiar with the family’s finances. As a result, demographic characteristics do not always map neatly onto households, our unit of observation.3 We employ data on race, marital status, sex, educational attainment, and income category.4 Race applies to the respondent and is reported as indicator variables for non-Hispanic white, Black, Hispanic, or other (including those of Asian and Native descent). Marital status reflects whether the household consists of a single financially independent adult or two interdependent adults. Two financially interdependent unmarried people living together are considered a married couple. We control for the sex of the household head. We control for the educational attainment of the household head using indicators for less than a high school diploma, high school diploma, some college, bachelor’s degree, and graduate degree. We control for household income using indicators for income categories, with cut-offs at $20,000, $50,000, $100,000, and $200,000. 5.3.3

Specification of the Dependent Variable

The SCF covers all household assets and liabilities, with two notable exceptions. First, the survey excludes households in the Forbes 400, who would be easily identifiable in the data. Second, because the SCF defines net worth (assets minus debt) as resources that a household may access and control immediately, the survey does not report defined benefit (DB) pension wealth— the present value of future income (minus future contributions) that households expect to receive from DB pension plans. To present a more complete analysis of household wealth, we add to the SCF definition of net worth a measure of the present value of DB wealth, following Sabelhaus and Volz (2019). Our resulting wealth definition, like the SCF’s, does not include future Social Security or Medicare benefits (or taxes), which often comprise a significant share of households’ resources in retirement.5 We employ two different wealth specifications: one that uses the level of wealth and one that uses the inverse hyperbolic sine transformation of wealth. The results derived from the level- of-wealth analysis describe absolute changes in wealth over the period, while the results from the inverse hyperbolic sine specification describe proportional changes in 3. Lindamood, Hanna, and Bi (2007). 4. In the dataset, these variables correspond to categorical variables “RACE,” “MARRIAGE,” “HHSEX,” an “EDCL” categorical variable adjusted with information from “EDUC” to provide more granularity with less than high school and graduate school specifications, and a household income category variable generated from “INCOME,” all as defined in https:// www.federalreserve.gov/publications/files/scf17.pdf. See also https://www.federalreserve.gov /econres/files/Networth%20Flowchart.pdf. 5. Social Security provides about 90 percent or more of the income for one-third of retirees and 50 percent or more of the income for two-thirds of retirees (Social Security Administration 2019).

150

Fig. 5.1

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

Median age-wealth profiles by constant-age group (scaled, 1989 = 100)

Source: Federal Reserve Board of Governors Survey of Consumer Finances (1989–2016).

wealth over the period. We use this transformation, instead of the more traditional logarithmic transformation, because it approximates the logarithm while remaining defined for the nonpositive values common in wealth data. More formally, if θ is a scaling parameter and w is a measure of wealth, the inverse hyperbolic sine of wealth can be written as θ–1 sin h–1(θw) = θ–1 ln[θw + θ2w2 + 1)1/2]. This symmetric function is linear around the origin but approximates the logarithm for larger values of wealth. To see this, note that if w is large, ln[ w + ( 2 w2 + 1)1/2 ] ln2 + ln w, which is simply a vertical displacement of the logarithm. Following previous research, we set θ = 0.0001.6 When multiplied by this scaling parameter, coefficients and standard errors from an inverse hyperbolic sine specification, like logarithmic coefficients and standard errors, can be interpreted as the percentage change in wealth implied by a change in a particular demographic characteristic, assuming that wealth values are sufficiently large.7 5.3.4

Results

Figure 5.1 shows median age-wealth profiles for constant-age groups across birth cohorts. The data are scaled so that each generation’s 1989 value is set to 100. The graph demonstrates two points. First, the Great Recession in 2007– 9 significantly reduced household wealth in all age groups. Sec6. Burbidge, Magee, and Robb (1988) find the optimal value of θ to be 0.0000872 (within rounding distance of our choice). Pence (2002) finds that 0.0001 is the optimal value of θ, a value also used by Kennickell and Sundén (1997). See Gale and Pence (2006) for author’s prior work conforming to this methodology. 7. See Pence (2006) for further explanation of the logarithmic approximation, and Burbidge, Magee, and Robb (1988) for more information about the inverse hyperbolic sine transformation itself.

The Wealth of Generations: Millennials

Fig. 5.2

151

Average age-wealth profiles by constant-age group (scaled, 1989 = 100)

Source: Federal Reserve Board of Governors Survey of Consumer Finances (1989–2016). Table 5.1

Constant Year 2016 Observations

Pooled (least absolute deviations), 1989–2016 25–34 (1)

35–44 (2)

45–54 (3)

55–64 (4)

28,344*** (4,933) −3,504 (5,971)

165,024*** (17,388) −77,124*** (19,188)

267,826*** (27,166) −96,826*** (31,389)

303,986*** (31,336) 14,777 (39,818)

1,250

1,712

1,848

2,015

Note: Robust standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

ond, younger age groups have been doing worse than older age groups. For example, in 2016, all groups aged 55 and older had more median wealth than their 1989 counterparts. Households aged 25– 34 in 2016— roughly the millennial generation— held about 12 percent less wealth than did households who were the same age in 1989. Figure 5.2 shows scaled mean age-wealth profiles. Due to the growth of income and wealth at the top of the distribution, the mean increases exceed the median increases, but they follow the same general pattern, with wealth rising more slowly for younger age groups than for older age groups. Appendix figures 5A.1 and 5A.2 report median and mean net worth levels by age and year. Table 5.1 reports results from median regressions. The first specification follows equation (5.1), explaining the level of household wealth as a function of only a constant and an indicator of whether the observation occurred in 2016. The 2016 indicator shows that median wealth was substantially lower in 2016 than in 1989 for households aged 35– 44 and 45– 54— by about

152

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

$77,000 and $97,000, respectively. Median wealth for 25-to-34-year-olds and 55-to-64-year-olds was not significantly different in the two sample years.8 The effects shown in table 5.1 could be due to changes in the general economic environment and/or to changes in specific household characteristics. To isolate the impact of these two groups of determinants, the second specification includes several household-level demographic variables, following equation (5.2). We then (implicitly) assign the residual impact to the general economic environment. The results show that changes in household characteristics reduced wealth for households aged 25– 54 in 2016 relative to 1989 but raised wealth for households aged 55– 64 over the same period. For example, for households aged 35– 44, the coefficient on the 2016 indicator was about −$54,000, compared to about −$97,000 in the first specification. This implies that changes in household characteristics explain about 45 percent of the decline in wealth for this age group over time. Likewise, for households aged 35– 44, the 2016 effect reduced wealth by $77,000 when demographic variables were excluded but by only $15,000 when demographic variables were included. Thus, more than 80 percent of the decline in wealth for that group can be explained by demographic factors. In contrast, for households aged 55– 64, the coefficient on the 2016 indicator is about −$29,000, which is lower (algebraically) than the coefficient in the first equation— which is about $15,000 but not significantly different from zero. The coefficients on the demographic variables (not shown) are consistent with much prior work. Households that are Black or Hispanic have lower wealth than other households, even after controlling for observables. Households where the head has more formal education and/or higher income accumulate more wealth. To some extent, married households have more wealth and female-headed households often have lower wealth. Table 5.2 repeats the exercise using mean (OLS) regressions. The first specification shows that average wealth rises substantially in the 45– 54 and 55– 64 age groups. Coupled with the changes in median wealth shown in table 5.1, these figures suggest a substantial widening of the distribution of wealth in those age groups over time. The second specification shows, again, that changes in household demographic variables served to raise wealth substantially in the 55– 64 age group. More than two-thirds of the increase in wealth in that age group documented in the first specification can be explained by changes in demographic characteristics in the second specification. Including demographic variables 8. Because we are analyzing changes among age groups within years, our sample sizes are fairly limited, ranging from 452 to 1,446 people per age group per year. Notably, 2016 sample sizes for each age group are about two to three times the size of the sample sizes of the same age groups in 1989. This will manifest in standard errors and significance levels, which are reported in our results.

The Wealth of Generations: Millennials Table 5.2

Constant Year 2016 Observations R-squared

153

Pooled (least squares), 1989–2016 25–34 (1)

35–44 (2)

45–54 (3)

55–64 (4)

137,580*** (19,883) −29,057 (21,870)

324,112*** (21,669) 51,286 (34,665)

642,123*** (43,958) 239,161*** (72,015)

720,288*** (50,043) 678,944*** (85,037)

1,250 0.001

1,712 0.000

1,848 0.001

2,015 0.003

Notes: Robust standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

reduced the 2016-year effect on wealth from $678,944 to $228,623 in the age cohort approaching retirement. The impact of the individual demographic variables is qualitatively similar to those found in the median regressions in table 5.1— minorities and female-headed households have less wealth; households where the head is married or has more formal education, or where income is higher, tend to have higher wealth. (Appendix tables 5A.1 and 5A.2 provide regression results using the hyperbolic sine of wealth as the dependent variable and generate broadly similar conclusions.) 5.4

Millennials

The millennial generation includes individuals born between 1981 and 1996.9 Between the Great Recession and the COVID pandemic (the latter of which is not covered in the data presented here), millennials have already experienced two major economic disruptions during their adulthood. Substantial economic inequality has been an enduring fixture of millennials’ adulthood. While every generation faces its own unique opportunities and challenges, many people feel that the obstacles facing the millennial generation are especially acute.10 5.4.1

Current Status

Focusing first on their current status (that is, as of 2016), today’s young adults have accumulated less wealth than most previous generations at the same age. Figure 5.3 shows tabulations from each wave of the SCF from 1989 to 2016. In the 2016 survey year, millennials were between the ages of 20 and 35. We examine net worth accumulation among 20-to-35-year-olds 9. Dimock (2018). Various definitions of the millennial generation include those born between the early 1980s and the early 2000s. The Census Bureau (unofficially) defines millennials as the cohort born between 1982 and 2000 (US Census Bureau 2015). 10. Pew Research Center (2012).

154

Fig. 5.3

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

Median net worth among young households, 1989–2016

Source: Federal Reserve Board of Governors Survey of Consumer Finances (1989–2016).

in each of the SCF years (with all wealth data reported in 2016 dollars). Because wealth accumulation patterns may not be particularly informative for people who are still in college, we also examine wealth patterns among 25-to-35-year-olds in each year. The figure shows that, using either age-group comparison, median wealth among millennials in 2016 was lower than among similarly aged cohorts in any year from 1989 to 2007. As noted above, the Great Recession in 2007– 9 significantly reduced household wealth, which has been slowly recovering since then. Median wealth among millennials was about 25 percent lower in 2016 than among similarly aged households in 2007. The percentage declines in mean wealth are even larger.11 Focusing on retirement wealth, figure 5.4 shows that, relative to similarly aged people, millennials have about the same coverage rate for DB pensions and DC plans from 2004 on, but lower DC coverage than the 1995– 2001 cohort and sharply lower DB coverage relative to that in the late 1980s. Median DC balances among account holders has fallen since 2007.12 All of the results above likely overstate the relative wealth position of millennials because of the interaction of three factors. First, the SCF does not survey dependent members of households, including millennials who live 11. Dettling, Hsu, and Llanes (2018) provide further detail on wealth accumulation trends between 2007 and 2016. Considering components of wealth, millennials had more debt than similarly aged people in 1989 but have about the same level as the 2001 cohort. The latter result may be surprising, given the well-publicized growth of student loans, but millennials have less credit card and other debt than prior generations (Looney and Yannelis 2018). 12. Dettling and Hsu (2014) examine retirement saving trends for people aged 18– 31 in the successive SCFs. They find that millennials in 2013 were just as likely to have a DC retirement account as similarly aged people in 2001. Millennials had higher median balances, conditional on ownership (by about $2,000), but they had lower participation in defined benefit plans.

The Wealth of Generations: Millennials

155

Fig. 5.4 Defined benefit and defined contribution plan ownership among 25–34 cohort, 1989–2016 Note: Retirement account assets include the value of IRAs, Keoghs, thrift-type accounts, and future and current account–type pensions. Source: Federal Reserve Board of Governors Survey of Consumer Finances (1989–2016).

with their parents. Second, the share of millennials living with their parents is higher than the share of similarly aged people in prior generations. Among 25-to-34-year-olds, 16 percent lived with their parents in 2016, compared to 11 percent in 1990 and 10 percent in 2000.13 Third, some formal evidence (as well as casual observations) suggests that young adults who are living at their parents’ home are doing less well economically than other young adults. Among those aged 25– 34 and living at home in 2016, 26 percent were neither employed nor attending school.14 Less-educated people are less likely to live independently and those with higher wages are more likely to do so. Between 1989 and 2016, the distribution of wealth widened significantly. For example, for 25-to-35-year-olds, average net worth in the bottom 25 percent of the distribution fell from about −$1,200 in 1989 to −$5,000 in 2007 and to −$20,000 in 2016. Over the same period, average wealth in the top 10 percent of the distribution skyrocketed, rising from $1.9 million in 1989 to $3.3 million in 2007 to $4.8 million in 2016. These patterns are consistent with the heterogeneity in preparation for retirement seen for other generations and noted above. 5.4.2

Future Status

The millennials also face a distinctive set of issues and circumstances that will affect their ability to save for retirement, including both advantages 13. Fry (2017). In our analysis of SCF data, we use 20-to-35-year-olds in 2016 to represent millennials (or 25-to-35-year-olds, given that wealth accumulation of households where the head is less than 25-year-olds may be difficult to model). Many other analyses, focusing on different issues or time frames, use a more standard age classification of 25-to-34-year-olds. 14. Fry (2017), and St. Clair (2016).

156

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

and disadvantages compared to prior generations. The disruption to health, careers, and the economy due to the COVID-19 pandemic looms large in this regard but is not discussed further below as the relative effects on different generations is extremely hard to assess at this point. 5.4.2.1

Advantages

5.4.2.1.1 Education Millennials start out with the advantage of having the greatest amount of formal education of any generation in history. Over 60 percent of adult millennials have attended at least some college, compared to 46 percent of the baby boomer generation when they were the same age.15 Rising educational attainment among women drives this difference.16 As a result of increased educational attainment and other factors, median wages for employed women are generally higher for millennials than for earlier generations, controlling for age. Among men, the wages of employed millennials typically do not surpass wages of older generations until millennial workers reach their mid-30s.17 Having more education will make it easier to save for retirement. First, the higher wages that come with higher education will give households more opportunities to save. Second, people with more education tend to save more of their income, controlling for income.18 Third, people with higher education levels tend to have later retirement ages since they tend to have less physically demanding jobs, are healthier, and receive fringe benefits in addition to wages that may incentivize them to stay in the labor force.19 The overarching societal trend toward white- collar work may further increase average retirement ages for similar reasons. Good health status is also highly correlated with decisions to work longer.20 Working longer, of course, makes it easier to finance adequate retirement saving. On the other hand, higher education and income may make adequate saving more difficult to achieve in some ways. For example, Social Security benefits are progressive, replacing a smaller amount of average lifetime earnings as average lifetime earnings rise. And those who are better educated, and in better health, tend to live longer, meaning that they have a longer retirement period to finance, holding retirement age constant. 5.4.2.1.2 Longer careers due to change in type of retirement plan Since the 1980s, the share of people participating in DB plans has decreased while participation in DC plans has increased (figure 5.4). At the 15. Council of Economic Advisers (2014). 16. Johnson et al. (2017). 17. Percheski (2019). 18. Dynan, Skinner, and Zeldes (2004). 19. Burtless (2013). 20. Munnell (2015).

The Wealth of Generations: Millennials

157

same time, the overall share of the workforce participating in any plan has remained relatively constant.21 This trend is expected to continue, implying that today’s younger workers will have higher DC coverage than previous generations and lower DB coverage. This will likely lead to later retirement for millennials relative to previous generations, which would make it easier to accumulate funds necessary to finance retirement. Evidence suggests that DB plans often encourage comparatively early retirement through built-in incentives that maximize pension wealth at relatively early retirement ages.22 Since the 1980s the average retirement age has risen after decades of decline, consistent with the decline of DB plans and rise of DC plans. As with better education, however, the shift to DC is not an unambiguous gain for retirement saving adequacy. Greater DC coverage and less DB coverage shifts much of the planning burden and investment risk from the employer to the employee, as discussed further below. 5.4.2.1.3 Health insurance Millennials have higher rates of health insurance than prior generations, due largely to the Affordable Care Act. Among 19-to-25-year-olds in 2014, about 79 percent had coverage under a health insurance policy, 13.2 percentage points higher than earlier generations at that age.23 5.4.2.2

Disadvantages

Despite having some advantages relative to previous generations, the millennials face a variety of obstacles and concerns that increase their chances— in absolute terms and relative to previous generations— of saving too little. 5.4.2.2.1 Early-career labor market The early-career labor market experienced by many of the millennials has been dominated by the Great Recession and the tepid pace of recovery for several years after. The growth path of GDP has never recovered to the fullemployment trend that existed before the Great Recession.24 The weak job market and low overall labor force participation that existed at the beginning of their careers has probably adversely affected millennials’ career earnings paths. Research shows that entering the labor force during an economic downturn depresses long-run earnings.25 Evidence from the Great Depression further reveals that those who experience poor macroeconomic trends 21. Gale and John (2017). 22. Kotlikoff and Wise (1984); Stock and Wise (1990). 23. Council of Economic Advisers (2014). 24. CBO (2018). 25. Kahn (2010); Welch (1979).

158

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

while they are young are less likely to take on significant financial risk, invest in the stock market, or own bonds.26 5.4.2.2.2 The rise of contingent jobs The evolution of the labor force toward contingent jobs will also complicate retirement saving for millennials.27 In traditional employer-employee relationships, workers earn a salary or wage and receive fringe benefits, potentially including employer contributions to retirement plans. Contingent workers, on the other hand, work on an ad hoc basis and are paid based on the service or good they provide. They may or may not work full-time. Examples include Uber drivers, consultants, and contractors. Using a broad definition, there could be almost 20 million contingent workers in the United States.28 Among full-time workers, these individuals have median weekly earnings about 30 percent lower than traditional workers and face a variety of barriers to retirement saving. Conventional retirement savings mechanisms, such as payroll deductions and employer matching contributions, are not readily available. As a result, they are half as likely to have access to a work-provided retirement plan.29 While non-employer-based retirement options such as individual retirement accounts (IRAs) are available to this group, only a small percentage participate.30 Although a recent survey suggests that contingent work is not rising as fast as some had thought, it is nevertheless the case that millennials face higher probabilities of doing contingent work than previous generations.31 5.4.2.2.3 The added risks and responsibility of defined contribution plans As noted above, participants in DC plans tend to work longer than participants in DB plans; other things equal, longer working careers should improve the adequacy of retirement saving. But other structural features of DC plans may lead to lower retirement incomes. First, to establish a DC plan, employees must make significantly more decisions regarding contribution levels, asset allocations, and asset drawdown. This freedom may actually serve to undermine retirement security if retirees make poor financial decisions.32 Automatic mechanisms that govern enrollment, escalation of contributions, investment allocation, and rollovers can mitigate these problems. Second, workers bear all the investment risk in most DC plans, which can undermine retirement security if savers’ retirement portfolios underperform. 26. Malmendier and Nagel (2011). 27. Gale, Holmes, and John (2018); Harris and Krueger (2015); Katz and Krueger (2016). 28. Gale, Holmes, and John (2018). 29. Gale, Holmes, and John (2018). 30. GAO (2015). 31. US Bureau of Labor Statistics (2018). 32. Poterba (2014).

The Wealth of Generations: Millennials

Fig. 5.5

159

Share of young households owning a home, 1989–2016

Source: Federal Reserve Board of Governors Survey of Consumer Finances (1989–2016).

Fig. 5.6

Median age of first marriage, 1970–2016

Source: US Census Bureau (2017).

5.4.2.2.4 Delayed life decisions Compared to previous generations, millennials are more likely to delay home ownership, marriage, and childbearing. Young adults currently have the lowest home ownership rate of any similarly aged generation since at least 1989 (figure 5.5). The average age of first marriage has increased from age 21 for women (24 for men) in 1975 to age 27 (29) in 2016 (figure 5.6).33 The age at which parents have their first child has increased over time as well, 33. US Census Bureau (2017).

160

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

from 22 in 1975 to 26 in 2014.34 The changes in these behavioral patterns are so large that there is debate over whether to label “emerging adulthood” as a new stage of life between childhood and adulthood.35 These trends, in turn, may delay the onset of retirement saving if people feel the need to “get settled” by purchasing a house and raising children before beginning to think about saving for retirement. 5.4.2.2.5 Increasing lifespan Just as delayed life choices may postpone substantial retirement saving, increasing lifespans make it harder to maintain standards of living in retirement, other things equal. Over the past five decades, the average life expectancy at birth has increased from 67 to 76 for males and from 73 to 81 for females. These increases are not borne equally, however. Those at the top of the income distribution have received almost all of the increase, while life expectancy for those at the bottom has remained constant or has possibly even declined slightly.36 If households live longer and plan to maintain their preretirement standard of living in retirement, they will either need to work longer or save more. 5.4.2.2.6 Long-term federal fiscal imbalances The federal government faces a long-term debt problem that will create pressure to cut spending and raise taxes. Even before the COVID-19 pandemic, the federal government faced a long-term fiscal shortfall that will require spending cuts or tax increases at some point.37 Low interest rates, discussed below, will make this problem less severe, but the COVID-19 pandemic, and the associated policies and economic downturn, made the long-term fiscal situation significantly worse.38 The longer policymakers wait to institute fiscal adjustments, the larger the adjustments will have to be in each given year, and the greater sacrifices millennials will have to make. Changes to Social Security, Medicare, and taxes will be particularly relevant to addressing the long-term fiscal imbalance. If such fiscal adjustments are made over the next few decades, it seems highly likely that millennials, who will then be in their prime earning years, will bear a significant amount of the burden. 5.4.2.2.7 Low rates of return Real interest rates fell steadily from the mid-1990s, though they have risen slightly in the last few years. Many reasons have been put forward for the 34. Matthews and Hamilton (2016); US Census Bureau (2017). 35. Vespa (2017). 36. National Academies of Sciences, Engineering, and Medicine (2015). 37. Auerbach, Gale, and Krupkin (2018), and Gale (2019). 38. To be clear, we are not criticizing the size of the relief and stimulus packages. They were a necessary response to COVID-19.

The Wealth of Generations: Millennials

161

decline, and most of them suggest the low rates will persist somewhat.39 One explanation, popularized by former treasury secretary Larry Summers, involves a lack of aggregate demand due to the Great Recession and secular stagnation postrecession.40 Other explanations include a worldwide savings glut and a flight to safety.41 To the extent that such trends continue or do not reverse, and display themselves in lower overall asset returns, it will prove harder for millennials to accumulate sufficient retirement wealth.42 With a given pattern of retirement contributions over time, a lower rate of return will result in a smaller accumulation of balances during the accumulation phase of retirement saving. With a given balance at the point of retirement, lower interest rates will result in smaller feasible payouts— for example, through an annuity— during the withdrawal phase of retirement saving. 5.4.2.2.8 Slower wage growth Earnings trajectories— or “age-earnings profiles”— have been flattening over time. A college-educated worker turning 25 in 1940 could expect annual earnings to be 4.0 times as high by their 55th birthday compared to their earnings in 1940. For college-educated workers turning age 25 in 1980, this ratio had fallen to just 2.6. For workers with a high school diploma only, the same ratio fell from 3.6 for the 1940 cohort to 1.5 for the 1980 cohort.43 This decline in wage growth over time will reduce future income for millennials and make it harder for them to accumulate wealth over the life cycle. 5.5

Wealth Accumulation and Racial and Ethnic Minorities

Millennials are more racially and ethnically diverse than prior generations: for example, 44 percent of millennials identify as a minority (a race or ethnicity other than non-Hispanic white), compared to 25 percent of people aged 21– 36 in 1985.44 As a result of this increased diversity, the United States will be a “majority-minority” country by 2050, where minority is defined as any race other than non-Hispanic white.45 The projected growth of the minority population will present new challenges and opportunities for wealth accumulation. A substantial literature suggests that minorities are at a disadvantage with regard to wealth accumulation compared to their nonminority counterparts.46 39. Elmendorf and Sheiner (2016). 40. Summers (2016). 41. Bernanke (2007, 2015). 42. Fichtner and Seligman (2017), and Mitchell, Clark, and Maurer (2018). 43. Kong, Ravikumar, and Vandenbroucke (2018). 44. Fry, Igielnik, and Patten (2018). 45. US Census Bureau (2018). 46. See Hasler, Lusardi, and Oggero (2018), and Rhee (2013).

162

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

We provide further evidence on these issues below. We estimate regressions of the form (5.3)

w3 = α3 + β3Xi + g3Ri + ε3i ,

where w3 is a measure of wealth, X is a vector of characteristics for each household i— including education, marital status, sex (for singles), income and age categories— and R is a series of racial/ethnic indicators (Black, Hispanic, and nonwhite other, with white as the omitted category).47 Using the same methodology elaborated upon above, we employ cross-section data from each of the 10 survey years of the SCF (triennially from 1989 to 2016) and estimate least squares (LS) and median (LAD) regressions, each with robust estimation techniques. Thompson and Suarez (2015) examine similar issues and provide wealth decompositions using the 1989– 2013 SCFs.48 In the text, we present regressions using the level of wealth as the dependent variable. We emphasize that the coefficient on race shows differences in wealth accumulation after controlling for various factors but should not be interpreted as an estimate of the impact of racial discrimination. The LS regressions in table 5.3 show that Black households tend to have lower net worth than white households, controlling for other factors.49 In the 2016 SCF, controlling for other factors, Black households had on average $124,000 less net worth than white households. This difference may have increased over time. The Black-white differences in wealth in the 1989, 1992, and 1995 cross-sections are smaller than the 2016 difference, with p-values ranging from 0.035 to 0.105. This finding should be qualified carefully. Certainly, reductions in Blackwhite differences over time in educational attainment and in wages should serve to reduce Black-white wealth differences. Our results address a different point. We show that— controlling for any changes in education, wages, and other household characteristics— the difference in wealth between Black and white people may well have increased over time. Additional results presented below support this conclusion. Households where the head is of Hispanic origin do not generally have statistically significantly different net worth from whites, controlling for other factors. Other nonwhite individuals, on average, had significantly lower net worth than whites in three of the SCF years (2004, 2010, and 2016). This difference appears to have increased over time, with the coefficients on the 47. Regressions using a variable called “normal income” instead of income yielded similar results. 48. Other studies of racial wealth gaps include Altonji and Doraszelski (2005), Barsky et al. (2002), Masterson, Zacharias, and Wolff (2009), Pew Research Center (2011), Scholz and Levine (2003), and Shapiro, Meschede, and Osoro (2013). 49. This finding is consistent with Emmons and Ricketts (2017), who show that differences in observable variables cannot fully explain minorities’ wealth accumulation relative to whites.

0.028 3,143 0.114

p-value N R-squared

0.018 3,906 0.084

−40,433*** (13,804) 0.035 23,809 (24,879) 0.676 21,532 (49,493)

1992 (2)

0.045 4,299 0.088

−56,846*** (18,309) 0.105 −34,760 (28,864) 0.102 −20,803 (50,361)

1995 (3)

0.175 4,305 0.094

−66,146*** (19,132) 0.166 29,801 (26,233) 0.780 −66,383 (91,136)

1998 (4)

0.577 4,442 0.131

−96,659*** (31,543) 0.573 −5,895 (38,400) 0.366 −160,583 (134,387)

2001 (5)

0.848 4,519 0.102

−97,979*** (27,485) 0.572 −25,377 (40,387) 0.217 −229,736*** (86,197)

2004 (6)

0.491 4,417 0.104

−78,393*** (29,495) 0.335 70,958* (42,880) 0.618 −147,826 (115,956)

2007 (7)

0.866 6,482 0.099

−141,470*** (26,428) 0.704 −58,587** (29,682) 0.034 −279,171*** (87,955)

2010 (8)

0.175 6,015 0.093

−83,542*** (23,857) 0.358 1,298 (28,004) 0.377 −69,021 (88,076)

2013 (9)

n/a 6,248 0.071

−124,142*** (37,193) n/a 42,575 (37,425) n/a −255,881** (105,779)

2016 (10)

Notes: Control variables: education level category (less than high school diploma, high school diploma, some college, bachelor’s degree, graduate degree); marital status; sex of single heads of household; income category ($0–19,999, $20,000–49,999, $50,000–99,999, $100,000–199,999, $200,000+); age category (0–24, 25–34, 35–44, 45–54, 65–74, 75+). Robust standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Source: Board of Governors of the Federal Reserve (2017). Coefficients reported in 2016 values.

p-value Hispanic

p-value Nonwhite other

1989 (1)

Net worth regressions (least squares)

−42,109* (23,206) 0.061 −6,333 (26,919) 0.289 48,796 (89,674)

Black

Table 5.3

164

W. G. Gale, H. Gelfond, J. J. Fichtner, and B. H. Harris

Table 5.4

Black p-value Hispanic p-value Non-white other p-value N R-squared

Pooled net worth regressions (least squares) 1989–2016 (1)

1989–2007 (2)

2010–2016 (3)

−90,972*** (8,626) n/a −7,845 (11,002) n/a −112,807*** (29,078)

−71,877*** (9,386) 0.012 3,575 (13,499) 0.303 −70,197** (33,546)

−117,903*** (17,292) n/a −7,399 (18,635) n/a −199,954*** (54,600)

n/a 47,776 0.091

0.045 29,031 0.102

n/a 18,745 0.082

Notes: Control variables: same as in tables 5.3 and 5.5, plus year effects. Robust standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1. Source: Board of Governors of the Federal Reserve (2017). Coefficients reported in 2016 values.

1989, 1992, and 1995 regressions significantly smaller (in absolute value) than the coefficient in the 2016 regression. In all of these regressions, however, the relevant sample sizes are fairly small, so precise estimation is difficult. To help address the concern with sample size, we also pool the data across survey years, adding a control for the survey year, with results shown in table 5.4. As in the cross-section results, Black households have lower net worth, controlling for other factors. The coefficients are larger in absolute value in the 2010– 16 specification than in the 1989– 2007 specification, confirming the finding above about widening Black-white wealth differences, controlling for other factors. Households where the head is of Hispanic origin do not have a significantly different net worth in any of the specifications. Those who do not identify as white, Black, or Hispanic have significantly lower net worth than whites in each specification and the difference has grown over time. In cross-section LAD results reported in table 5.5, the typical Black household had $43,262 less in wealth than the typical white household, controlling for other factors. The difference in wealth has increased over time, again conditional on observable factors. The LAD regressions also show that the typical household with a head of Hispanic origin has less net worth than white households in several of the survey years, particularly in 2013 and 2016. The results suggest that wealth differences between whites and Hispanics may be increasing over time. Results are mixed for other nonwhite individuals, but the sample size for this group is relatively small, so precise estimation is difficult.

0.786 3,143 0.180

−29,966*** (4,534) 0.019 −25,656*** (9,338) 0.665 −12,710 (9,025)

1989 (1)

0.611 3,906 0.184

−20,971*** (3,181)