Determinants and Economic Consequences of Youth Unemployment at the Beginning of the 21st Century: Themenheft Jahrbücher für Nationalökonomie und Statistik 4+5/2015 9783110510805, 9783828206182

Citation preview

Determinants and Economic Consequences of Youth Unemployment at the Beginning of the 21st Century

Edited by Bernd Fitzenberger, Nicole Giirtzgen and Friedhelm Pfeiffer

With Contributions by van den Berg, Gerard J., University of Mannheim Cabane, Charlotte, University of Sankt Gallen, Switzerland Fitzenberger, Bernd, University of Berlin Lechner, Michael, University of Sankt Gallen, Switzerland Licklederer, Stefanie, University of Freiburg Mäder, Miriam, University of Erlangen-Nuremberg Möller, Joachim, Institute for Employment Research (IAB) Nuremberg Mohrenweiser, Jens, Centre for European Economic Research (ZEW), Mannheim Müller, Steffen, Halle Institute for Economic Research and Magdeburg University

Lucius &c Lucius • Stuttgart 2015

Pfeiffer, Friedhelm, Centre for European Economic Research (ZEW), Mannheim Riphahn, Regina T., University of Erlangen-Nuremberg Sachs, Andreas, Centre for European Economic Research (ZEW), Mannheim Schwientek, Caroline, University of Erlangen-Nuremberg Smolny, Werner, Ulm University Tertilt, Michèle, University of Mannheim Umkehrer, Matthias, Institute for Employment Research (IAB) Nuremberg Zwick, Thomas, University of Würzburg, Centre for European Economic Research (ZEW) Mannheim, and ROA Mastricht

Guest Editors Prof. Bernd Fitzenberger, Ph.D. Humboldt-Universität zu Berlin School of Business and Economics Spandauer Strasse 1 1 0 0 9 9 Berlin [email protected] PD Dr. Nicole Gürtzgen Centre for European Economic Research (ZEW) L 7, 1 6 8 1 6 1 Mannheim [email protected] PD Dr. Friedhelm Pfeiffer Centre for European Economic Research (ZEW) L 7, 1 6 8 1 6 1 Mannheim [email protected]

Bibliografische Information der Deutschen Nationalbibliothek Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet über http://dnb.d-nb.de abrufbar ISBN 9 7 8 - 3 - 8 2 8 2 - 0 6 1 8 - 2

© Lucius Sc Lucius Verlagsgesellschaft mbH • Stuttgart - 2 0 1 5 Gerokstraße 51, D - 7 0 1 8 4 Stuttgart Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlags unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Ubersetzungen und Mikroverfilmungen sowie die Einspeicherung und Verarbeitung in elektronischen Systemen.

Satz: inMEDIAlo, Digital- und Printmedien UG, Plankstadt Druck und Bindung: Neumann Druck, Heidelberg Printed in Germany

Jahrbücher! Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2015) Bd. (Vol.) 235/4+5

Inhalt / Contents Guest Editorial

352-354

Abhandlungen / Original Papers Mäder, Miriam, Steffen Müller, Regina T. Riphahn, Caroline Schwientek, Intergenerational Transmission of Unemployment Evidence for German Sons Cabane, Charlotte, Michael Lechner, Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects Sachs, Andreas, Werner Smolny, Youth Unemployment in the OECD: The Role of Institutions Mohrenweiser, Jens, Thomas Zwick, Youth Unemployment After Apprenticeship Training and Individual, Occupation and Training Employer Characteristics Fitzenberger, Bernd, Stefanie Licklederer, Career Planning, School Grades, and Transitions: The Last Two Years in a German Lower Track Secondary School Mohrenweiser, Jens, Friedhelm Pfeiffer, Coaching Disadvantaged Young People: Evidence from Firm Level Data Möller, Joachim, Matthias Umkehrer, Are there Long-Term Earnings Scars from Youth Unemployment in Germany? Tertilt, Michèle, Gerard J. van den Berg, The Association Between Own Unemployment and Violence Victimization Among Female Youths

355-375 376—402 403-417

418-432

433—458 459-473 474—498

499-513

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2015) Bd. (Vol.) 235/4+5

Guest Editorial In 1982, Wolfgang Franz published his treatise on youth unemployment starting with the sentence: "One of the major labor market problems during the 1970's was constituted by the relatively high and increasing rate of unemployment of young persons relative to adult members of the labor force." Nowadays, in a number of European economies youth unemployment has increased again to unexpected and unwanted levels. It has become one of the pressing labor market problems that many countries are currently facing, not only Greece or Spain, where youth unemployment rates are higher than 50 percent, but also France, Portugal, Italy and other European countries where it exceeds 20 percent. Youth unemployment can result in a permanent reduction of individual human capital and earnings and in a rise of poverty, violence and social inequality to undesired levels. This special issue involves a collection of current research and new findings. The goal of the special issue is to improve our understanding of the determinants and economic consequences of youth unemployment and to discuss implications for policies to combat youth unemployment. Specifically, the contributions to this special issue deal with (i)

the determinants of youth unemployment such as the role of institutions and family background, (ii) issues related to early career labor market transitions and finally (iii) the consequences of youth unemployment. Investigating the (i) determinants of youth unemployment, Mader, Miiller, Riphahn and Schwientek (2015) examine the intergenerational association between the unemployment experiences of fathers and their sons. Their innovative study is based on German survey data that cover the last decades. In agreement with a still small international literature they do not find a positive causal effect for intergenerational unemployment transmission. Their results therefore suggest that improving the employment situation of adults will presumably not reduce youth unemployment. Cabane and Lechner (2015) survey the literature on the determinants of participation in sports or physical exercise and possible beneficial effects of sports on labor market and health related outcomes. The authors show that there appears to be a consensus in the literature that individual leisure sports participation and physical activity have mainly positive effects not only on employment, but also on health, life satisfaction and earnings. Sachs and Smolny (2015) investigate the role of labor market institutions for youth unemployment, as contrasted to total unemployment based on data for seventeen OECD countries from 1982 to 2005. According to their empirical findings, employment protection for regular jobs and the combined effects of powerful unions with coordinated wage bargaining tend to increase youth relative to adult unemployment rates. Regarding (ii) early career labor market transitions, Mohrenweiser and Zwick (2015) analyze the risk of unemployment, unemployment duration and the risk of long-term unemployment immediately after apprenticeship graduation. They show that individual productivity assessment of the training firm, initial selection into high reputation firms and occupations, and adverse selection of employer moving graduates are correlated with unemployment after apprenticeship graduation. Fitzenberger and Licklederer (2015) study the first transition after graduation for students from lower track secondary schools, based on data from repeated surveys conducted in the city of Freiburg. In their samples only 10 percent of students start an apprenticeship immediately after graduation. The majority of students with poor school grades continue with pre-vocational training, while students with good grades chose higher schooling. The authors conclude that career guidance programmes should not

Guest Editorial • 353

focus solely o n t h e i m m e d i a t e start of a n a p p r e n t i c e s h i p a f t e r g r a d u a t i o n . M o h r e n w e i s e r a n d Pfeiffer (2015) investigate w h e t h e r a p r o g r a m m e t h a t s u p p o r t s firms t o train d i s a d v a n taged y o u t h can r e d u c e recruiting difficulties in apprentice training firms. Based o n u n i q u e firm-level d a t a f r o m the metal a n d electronic i n d u s t r y in B a d e n - W ü r t t e m b e r g f r o m 2 0 1 0 t o 2 0 1 3 , they apply i n s t r u m e n t a l variable a n d difference-in-differences estimations a n d their findings d o n o t s h o w a significant s h o r t - t e r m causal i m p a c t of the p r o g r a m m e . A l t h o u g h t h e investigated p r o g r a m m e did n o t reduce r e c r u i t m e n t difficulties in a p p r e n t i c e training firms it helped t h e d i s a d v a n t a g e d y o u t h to find a job. E x p l o r i n g (iii) the consequences of y o u t h u n e m p l o y m e n t , M ö l l e r a n d U m k e h r e r (2015) analyze p o t e n t i a l scarring effects f r o m early career u n e m p l o y m e n t . Based o n large-scale G e r m a n administrative d a t a , the a u t h o r s f o l l o w y o u n g w o r k e r s ' l a b o r m a r k e t careers after t h e c o m p l e t i o n of a p p r e n t i c e s h i p training over a period of 2 4 years. Overall, t h e study establishes non-negligible long t e r m effects of y o u t h u n e m p l o y m e n t o n later earnings, w h i c h are particularly p r o n o u n c e d a m o n g those at the b o t t o m of t h e earnings d i s t r i b u t i o n . Finally, Tertilt a n d v a n d e n Berg (2015) a d d r e s s t h e association b e t w e e n individual u n e m p l o y m e n t a n d t h e p r o p e n s i t y of being subject to violence a m o n g y o u n g w o m e n . Based o n Swedish health care register d a t a , the a u t h o r s d e m o n s t r a t e t h a t female victimization is m o r e prevalent a m o n g u n e m p l o y e d w o m e n as c o m p a r e d t o their e m p l o y e d c o u n t e r p a r t s . T h e i r study suggests t h a t m u c h of t h e established difference m a y be a t t r i b u t e d t o a higher prevalence of n o n - d o m e s t i c violence a n d l o n g - r u n abuse a m o n g u n e m p l o y e d female y o u t h s . T h e a u t h o r s dedicate their peer-reviewed c o n t r i b u t i o n s t o their colleague, m e n t o r , a n d / o r friend W o l f g a n g F r a n z on the occasion of his 7 0 t h b i r t h d a y , o n J a n u a r y 7 t h 2 0 1 4 . While W o l f g a n g F r a n z h a s m a d e c o n t r i b u t i o n s to m a n y aspects of empirical labor m a r k e t research, m u c h of his early career w o r k - including his h a b i l i t a t i o n thesis, the 1 9 8 2 treatise on y o u t h u n e m p l o y m e n t - dealt w i t h y o u t h u n e m p l o y m e n t . W o l f g a n g F r a n z ' o u t s t a n d i n g interest in t h e subject a l o n g w i t h t h e high relevance m a k e s the overall t h e m e of this special issue a n excellent choice for a t r i b u t e t o his scientific achievements. W e t h a n k the C e n t r e for E u r o p e a n E c o n o m i c Research ( Z E W ) , M a n n h e i m - especially Clemens Fuest a n d T h o m a s K o h l - for their initiative a n d s u p p o r t in hosting the S y m p o s i u m o n " D e t e r m i n a n t s a n d E c o n o m i c C o n s e q u e n c e s of Y o u t h U n e m p l o y m e n t at the Beginning of t h e 2 1 s t C e n t u r y " o n J a n u a r y 17th 2 0 1 4 , w h e r e t h e articles in this issue w e r e presented a n d discussed. T h e seminal c o n t r i b u t i o n of W o l f g a n g F r a n z a n d its lasting relevance in o u r discipline s h o u l d e n c o u r a g e y o u n g colleagues t o never s t o p struggling for excellent research.

References M ä d e r , M . , S. Müller, R.T. Riphahn, C. Schwientek (2015), Intergenerational Transmission of Unemployment - Evidence for German Sons. Journal of Economics and Statistics 235(4+5): 355-375. Cabane, C., M . Lechner (2015), Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects. Journal of Economics and Statistics 235(4+5): 3 7 6 ^ - 0 2 . Sachs, A., W . Smolny (2015), Youth Unemployment in the O E C D : The Role of Institutions. Journal of Economics and Statistics 235(4+5): 4 0 3 ^ - 1 7 . Mohrenweiser J., T. Zwick (2015), Youth Unemployment after Apprenticeship Training and Individual, Occupation and Training Employer Characteristics. Journal of Economics and Statistics 235(4+5): 4 1 8 - 4 3 2 . Fitzenberger, B., S. Licklederer (2015), Career Planning, School Grades, and Transitions: The Last T w o Years in a German Lower Track Secondary School. Journal of Economics and Statistics 235(4+5): 4 3 3 - 4 5 8 .

354 • Guest Editorial

Mohrenweiser, J., F. Pfeiffer (2015), Coaching Disadvantaged Young People: Evidence from Firm Level Data. Journal of Economics and Statistics 235(4+5): 4 5 9 - 4 7 3 . Moller, J., M. Umkehrer (2015), Are There Long-term Earnings Scars from Youth Unemployment in Germany? Journal of Economics and Statistics 235(4+5): 4 7 4 - 4 9 8 . Tertilt, M., G. van den Berg (2015), The Association between Own Unemployment and Violence Victimization among Female Youths. Journal of Economics and Statistics 235(4+5): 4 9 9 - 5 1 3 .

Bernd

Fitzenberger

Nicole Friedhelm

Gürtzgen Pfeiffer

Jahrbücherf. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2015) Bd. (Vol.) 235/4+5

Intergenerational Transmission of Unemployment Evidence for German Sons Miriam Mäder, Regina T. Riphahn, and Caroline Schwientek Friedrich-Alexander-University Erlangen-Niirnberg (FAU) Steffen Müller Halle Institute for Economic Research and Magdeburg University JEL J62; C21; C26 Youth unemployment; non-employment; intergenerational mobility; causal effect; Gottschalk method.

Summary This paper studies the association between the unemployment experience of fathers and their sons. Based on German survey data that cover the last decades we find significant positive correlations. Using instrumental variables estimation and the Gottschalk (1996) method we investigate to what extent fathers' unemployment is causal for offsprings' employment outcomes. In agreement with most of the small international literature we do not find a positive causal effect for intergenerational unemployment transmission. This outcome is robust to alternative data structures and to tests at the intensive and extetisive margin of unemployment.

1

Introduction

Unemployment of young individuals is one of the most pressing labor market problems of our times. Recently, some of the crisis ridden European economies faced youth unemployment rates well beyond 20 percent which instigate not only poverty and a sense of desperation but also waves of emigration and delays in family formation. The literature shows that the early experience of unemployment can be influential for lifetime labor market opportunities (e.g., Gregg 2001; Schmillen/Umkehrer 2013). However, while most commentators agree on the significance of early unemployment there is surprisingly little discussion and evidence on some of its key determinants especially the family background. In this paper we study the intergenerational transmission of unemployment experience, describe its patterns, and investigate causal relationships. A number of mechanisms may relate parent and child unemployment. They comprise correlated observable characteristics of parent and child, correlated unobservable characteristics, and true causal effects of parent unemployment on child unemployment. Clearly, observable characteristics such as formal education, choice of industry, occupation, region of residence, or social networks are correlated across generations and may affect employment outcomes. Similarly, it is plausible that unobservables such as ability, motivation, attitudes, beliefs, or personality traits are shared between parents and their children and may affect the risk of experiencing an unemployment spell. However, causal connections between parent and child unemployment are of particular interest. Such causal mechanisms may generate both positive and negative effects: the

356 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

experience of parental unemployment may affect household and family tastes and attitudes and reduce the perceived stigma of unemployment. Also, it may reduce child h u m a n capital investments as a consequence of reduced household income or unemploymentrelated stress in the family. These mechanisms suggest a positive correlation between parent and child unemployment. O n the other hand it is possible that the additional leisure of an unemployed parent benefits the offspring and that the family values h u m a n capital more after experiencing a loss of employment. In that case one might as well expect a negative correlation between parent a n d child unemployment. The literature on the intergenerational transmission of unemployment has studied the situation for Canada (Corak et al. 2 0 0 4 ; Oreopoulos et al. 2008), the U.K. (Johnson/Reed 1996; O'Neill/Sweetman 1998; Macmillan 2 0 1 0 ; Gregg et al. 2012), N o r w a y (Bratberg et al. 2 0 0 8 ; Ekhaugen 2009), and Sweden (Corak et al. 2004). While almost all studies yield positive intergenerational correlations of unemployment, the evidence on true causal effects of parent on child unemployment is mixed. Only the two studies on C a n a d a appear to support a causal intergenerational effect while all others find insignificant effects. W e add to this inconclusive literature by offering evidence for Germany, a country for which intergenerational transmission of unemployment has not been studied before. Germany is a particularly interesting case because on the one hand it is well k n o w n for its low youth unemployment (Riphahn/Zibrowius 2014) and on the other h a n d it features low intergenerational mobility and high intergenerational correlation of economic outcomes e.g. compared to Scandinavian countries (see, e.g., C o u c h / D u n n 1997; or Schnitzlein 2 0 1 4 and studies cited there). This provides a unique setting that has not been studied before. We take advantage of long running panel data f r o m the G e r m a n Socio-Economic Panel (SOEP) to investigate the correlation and causation patterns between fathers' and sons' unemployment experience. We are interested in both intergenerational correlation patterns and causal parent-child effects which we identify based on an instrumental variables a p p r o a c h and the Gottschalk (1996) method. The evidence on correlation patterns yields the gross impact of family background and parental unemployment on child unemployment risks. This is of interest in itself and in its heterogeneity across population groups; certainly, the relevance of high intergenerational unemployment correlation (i.e., of low intergenerational mobility) differs if correlations are strong in families with low as opposed to high unemployment risk. Such patterns can be evaluated independent of causal analyses that separate family unobservables f r o m the true causal parental unemployment effect. As the t w o components of intergenerational correlation have different policy implications it is important to clarify their relative importance. O u r analyses yield three key results: first, the unemployment experience of fathers and sons is significantly positively correlated; second, there is n o evidence in favor of positive causal intergenerational effects; third, most of the intergenerational unemployment correlation is associated with paternal characteristics such as age and education. This paper is structured as follows. W e first summarize key findings and approaches of the literature on the intergenerational transmission of labor market outcomes and discuss our empirical methods. T h e n we describe our data. The results section presents findings of least squares regressions, instrumental variables analyses, an application of the Gottschalk (1996) method, and robustness tests. In section 5 we conclude with a summary of our findings.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 357

2

Literature and empirical approach

2.1

Existing evidence on intergenerational transmission of labor market outcomes

Several empirical studies investigate the relation between the outcomes of parents and their children with a focus on unemployment and welfare receipt. Studies on unemployment transmission look at the relation of father and son outcomes (e.g., O'Neill/Sweetman 1998; Ekhaugen 2009) while studies on welfare receipt in the U.S. typically analyze transmission from mother to daughter (e.g., Antel 1992; Gottschalk 1990, 1996). Gottschalk (1990) shows a strong positive intergenerational correlation in welfare receipt using U.S. data and speculates whether this correlation is a causal effect or explained by family background. Antel (1992) and Gottschalk (1996) report a causal effect of mothers' welfare receipt on daughters' welfare receipt. In a more recent study, Beaulieu et al. (2005) analyze the relation between parents' and children's receipt of social assistance in Canada and report similar results, i.e., a strong positive correlation that can be interpreted causally. These studies' results rely on untestable identifying assumptions, e.g., assumptions on the joint distribution of unobservables or the validity of exclusion restrictions. A new paper using Swedish data and comparing siblings (Edmark/Hanspers 2012) finds no causal relation between parental welfare use and welfare use of the next generation. The literature on the transmission of unemployment from father to son yields a more homogenous picture. Studies for the U.K. (Johnson/Reed 1996; O'Neill/Sweetman 1998; Macmillan 2010), Norway (Ekhaugen 2009), Canada, and Sweden (Corak et al. 2004) report a strong positive intergenerational correlation in the incidence of unemployment, but no study finds clear evidence for a causal mechanism. Studies exploiting father's displacement due to mass layoffs or plant closures yield mixed results. Oreopolous et al. (2008) find a higher unemployment risk for children of displaced fathers in Canada. Similarly, Gregg et al. (2012) report a 1.5 percent higher youth unemployment duration for children of fathers w h o worked in industries with adverse employment shocks during the 1980 recession in the U.K. If, however, father's job displacement is related to his unobserved characteristics, these estimates might mix the effect of family background with the causal effect of parental unemployment. Using Norwegian data, Bratberg et al. (2008) find no effect of father's displacement on child's later earnings. Taken together, international evidence points at observed and unobserved family background characteristics as predominant drivers of the intergenerational correlation of unemployment. 1 To the best of our knowledge there exists no single study for Germany systematically exploring the intergenerational transmission of unemployment. Franz et al. (2000) analyze the transition from vocational training to permanent jobs and find a prolonged unemployment duration for children from households where the head of the household is unemployed. Franz et al. (2000) do not distinguish empirically between causality and the influence of family background. A recent paper by Pinger (2012) reports a negative causal effect of paternal unemployment on the probability of upper secondary school choice. Pinger (2012) also finds negative effects on child self-confidence and mental health and a more external locus of control for affected children. All in all, there is some evidence on negative effects of parental unemployment for Germany but no systematic study on the intergenerational transmission of unemployment.

1

O n alternative intergenerational transmission mechanisms see De Paola (2013) or Blomeyer et al. (2013).

358 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

2.2

Model and relevant estimation methods

A regression of son's unemployment experience in the observation period (t\) on father's unemployment history in a previous period (fO) (and a vector of son and father characteristics) yields a measure of the correlation between father's and son's unemployment outcomes. 2 This is interesting as it shows whether sons of unemployed fathers are more or less likely to become unemployed themselves. The correlation can be interpreted as the causal effect of father's unemployment history if the latter is uncorrelated with the error term in the son's unemployment equation. This is unlikely because the reasons for father's and son's unemployment may have a common component shared by all family members. Family background may include biological factors, ability, or similar tastes and preferences concerning work. Consider the following model: unsia

= unfit oP + x'sit\Y

un

=

fM

x

'f,tOS

+

e

W

+

£

«il



where s denotes sons, f fathers, i families, fO and i l refer to the past and ongoing time periods, and P , y , and S are parameter vectors. Son's unemployment unsjt\ is affected by the father's unemployment history utifitQ and a vector of control variables x. The error terms are defined as fisitl = asi + hitl

(3)

and s

f i t O = a f i + rfitO
0. This correlation generally biases OLS estimates of equation (1) in the sense that /3 is not reflecting the causal effect of paternal unemployment history, only. The biased estimate, instead, mixes the effects of family background and paternal unemployment. The challenge is to determine which part is causal and which reflects the influence of family background. Both effects are interesting but have different policy implications. In previous studies three methods have been used to disentangle family background and true causal effects. Ekhaugen (2009) compares siblings w h o have been at different ages at the time of parental unemployment. On the basis of assumptions about the age after which parental unemployment does and does not affect a child's employment outcomes, sibling differences can net out the effect of family background. Other scholars estimate the system of equations (1) and (2) and either model SfitQ) within a bivariate probit framework (e.g., Antel 1992; O'Neill/Sweetman 1998) or apply a two-stage least squares approach (2SLS) (e.g., Macmillan 2010). The 2SLS approach requires that at least one instrumental variable which strongly affects father's unemployment risk is exogenous (conditional on covariates) in equation (1). cov(sslti;

2

In our empirical application we will consider the son's age 10-15 to represent period tO, and son's age 17-24 to represent period i l .

Intergenerational Transmission of Unemployment - Evidence for German Sons • 359

Although the bivariate probit can identify fi without exclusion restrictions, corresponding estimates are typically not robust to slight changes in specification. Hence, also for the bivariate probit at least one exclusion restriction is recommended. Finally, based on Gottschalk (1996) we add future parental unemployment to equation (1) yielding: unsin

= unfit0p

+

unfitle + x'sin

y + ssin.

(5)

The idea behind the inclusion of future paternal unemployment in period t2 (e.g. when the son is aged 2 5 - 3 0 ) is that it should have no causal impact on a son's unemployment if it occurs after the son is old enough to be unaffected by the father's labor market outcomes. If this is true, the parameter associated with future paternal unemployment (un^ t 2 ) captures family background only. Subtracting it from the coefficient on prior paternal unemployment {un^tQ) estimates the causal effect of interest if an effect of son's unemployment on father's unemployment is ruled out. The obvious advantage of Gottschalk's (1996) method is that there is no need to find exclusion restrictions. 3 2.3

Empirical approach

As the Gottschalk (1996) method and the methods relying on exclusion restrictions have different advantages and shortcomings, we will apply both types of models and compare the results. We start by estimating equation (1) via OLS. T o extract as much family background f r o m the error term as possible, we also add information on the father to equation (1). Based on empirical results for other countries, we expect a positive sign for p. A negative ¡3 is theoretically possible if, e.g., the experience of having an unemployed father motivates the son to avoid own future unemployment. A negative sign is, however, unlikely as the negative causal effect would have to overcompensate the expected positive effect of family background. Before turning to the causal methods, we point out that the vector of son's control variables ( x ' j t i ) does not contain information on son's education, or industry. These variables are themselves likely to be affected by father's unemployment. Including them constitutes a case of over-controlling, i.e., of extracting explanatory power originally belonging to father's unemployment iun^ t 0 ). Later, we will add son's characteristics in order to test whether our (positive) estimate of /S becomes smaller. If it does, education and occupational choice are transmission channels for the intergenerational correlation in unemployment. O u r 2SLS instrumental variables approach relies on the availability of an instrumental variable that is strongly correlated with parental unemployment (un^ t 0 ) but unrelated to £S(il- We opt for industry level labor market conditions in tO because these should be related to father's unemployment propensity. 4 In particular, we generate indicators of the annual industry-specific risk of a transition to unemployment and of the annual 3

4

Ekhaugen (2009:101) points out that it has additionally to be assumed that parents becoming unemployed after their offspring reaches the critical age are not systematically different from parents becoming unemployed before (identifying fi). The author discusses that the approach may underestimate the causal effect if parental unemployment in t2 is correlated with child outcomes for other than family background mechanisms, e.g., due to shared regional labor markets. We will be more explicit about the exact time structure in section 3.

3 6 0 • M i r i a m M ä d e r , Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

industry-specific stock of unemployment. As exogeneity of instruments cannot be tested, we must assume that unemployment in the father's industry is uncorrelated with unobserved determinants of son's unemployment years later. The exogeneity assumption is violated, e.g., if family background characteristics systematically cause fathers to be in certain industries. This might pose a problem in regions with only a h a n d f u l of employers but should be less of a challenge in metropolitan areas where the choice of an industry is less restricted. Also, the instruments are invalid if there is a direct partial effect of the paternal industry characteristics on youth employment outcomes which we observe 9 - 1 6 years later based on mechanisms other t h a n family unobservables. W e additionally implement Gottschalk's (1996) approach by adding father's unemployment experience in tl, i.e. after the son exceeds age 2 4 to the OLS regression. 2.4

Interpretation of overall, causal, and family background effects

The OLS estimate of ft, i.e., the overall effect, measures whether sons of unemployed fathers are more or less likely to become unemployed themselves. It therefore adds to our understanding of the sources of intergenerational (economic) mobility in Germany, which has typically been analyzed with respect to wage or education outcomes (e.g., Schnitzlein 2 0 1 4 ; Heineck/Riphahn 2009). A high positive value of fi indicates low overall mobility and vice versa. W e will study the heterogeneity of the overall effect, as the relevance of a high fi for sons of high-risk fathers differs f r o m that for sons in low-risk families. However, the overall effect does not tell us much about the sources of the intergenerational transmission of unemployment and appropriate policy interventions. The interpretation of fi depends on whether it reflects the effect of family background or the effect of paternal unemployment per se. If ft reflects the effect of family background, the sons' unemployment perspectives cannot be shaped by policy interventions that reduce paternal unemployment such as active labor market policies. Effective policies would then have to reduce the influence of family background, e.g., by offering special training or educational programs to children of unemployed parents. Contrarily, if /S reflects a positive causal effect, reducing paternal unemployment reduces unemployment of the future generation. Then, the costs associated with today's unemployment extend beyond the direct financial and indirect social costs of paternal unemployment.

3 3.1

Data Sample

O u r analysis exploits data f r o m the G e r m a n Socio-Economic panel (SOEP), a longitudinal survey conducted annually since 1984 (Wagner et al. 2007). We use all available annual waves ( 1 9 8 4 - 2 0 1 2 ) and all samples. The advantage of the SOEP is the long observation period and the availability of detailed information on family background and labor force status. W e use retrospective biographical as well as annually collected survey information. Compared to administrative data the SOEP comprises relatively small samples. At the same time the SOEP overcomes an important d r a w b a c k of administrative data: it covers all unemployed persons, independent of whether they are officially registered. This allows a more flexible definition of unemployment which is particularly appropriate for the analysis of youth unemployment. Since youths are typically not eligible for unemployment benefits they tend not to register with the unemployment insurance.

I n t e r g e n e r a t i o n a l Transmission of U n e m p l o y m e n t - Evidence f o r G e r m a n Sons • 3 6 1

We study youth unemployment a m o n g male respondents aged 17 to 24. 5 We d r o p observations with missing information on own labor force status (0.1% of the sample) and without information on fathers ( 2 8 % of the sample). T o evaluate the impact of past paternal unemployment we collect information on fathers' unemployment for their sons' age range 10 to 15 using the annual self-reported employment status at the time of the interview. W e d r o p observations of sons for w h o m we d o not observe the father at least once in this age range. For our IV strategy we need to observe the father when the son was 8 years old and we require information on the last industry of fathers' employment, at least once. 6 In the end, these sample selection criteria leave us with a sample of 2 , 1 7 5 sons. This is our primary sample for OLS and IV estimations. Table 1 shows our sample selection procedure in detail. For the application of the Gottschalk (1996) method we additionally need to observe fathers after their sons turn 25. For these analyses our sample size declines further to 1,266 observations. 7 T a b l e 1 S a m p l e selection

Persons

Person-Years

Male respondents aged 1 7 - 2 4 - Missing labor f o r c e s t a t u s - Father n o t o b s e r v e d or n o biographical q u e s t i o n n a i r e a n s w e r e d - Father n o t o b s e r v e d a t s o n ' s a g e 1 0 - 1 5 - Father n o t o b s e r v e d a t s o n ' s a g e 8 - Missing f a t h e r ' s industry

7,614 8 2,156 2,412 759 104

31,339 20 5,329 9,734 4,750 613

= OLS a n d IV S a m p l e

2,175

10,893

Note: "-" stands for minus; the number of cases in the first row describes the magnitude of the initial raw sample. Each row provides the number of observations lost for the row-specific selection criterion. The numbers depend on the order of applied criterions. The last row provides the sample sizes available for the analysis. Source: SOEP 1984-2012, own calculations.

Since the additional information that can be gained f r o m a panel structure is limited, we use only cross-sectional information; the key explanatory variable - father's years of unemployment at son's age 10 to 15 - does not vary over time. Consequently, considering panel data would shift weights in favor of individuals w h o are observed more often in the considered age range (17-24). As non-response and panel attrition at this age are potentially selective, we use each person only once in the estimation sample to limit the influence of confounding factors. W e will exploit the panel structure of the data as a robustness check. 3.2

Key variables

O u r dependent variable comprises the number of years during which the son has been registered unemployed or has been non-working between age 17 and 24, i.e. the classic age range considered in the definitions of youth unemployment. The main explanatory 5

6

7

Female respondents would also be of interest. The intergenerational transmission of unemployment may differ for males and females. W e leave the analysis of these differences for future work. Deleting persons from our basic sample w h o never reported an industry is potentially endogenous. However, given that we lose only 37 persons this has minor consequences for our estimates. For this subsample we omit the selection on observing the father when the son was 8 years old.

362 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

variable is father's registered unemployment in years at the son's age 10 to 15, i.e. in late childhood as collected from surveys of the fathers. We use sons' years of worklessness and fathers' years of registered unemployment, i.e., a broad definition of unemployment for the son and a more narrow definition for the father. 8 In both cases we do not regard individuals as workless or unemployed if they are in full- or part-time employment, vocational training, tertiary education, or military and substitute service. Due to missing information we do not observe all fathers and sons in all years. Therefore, we control in our model for the number of years without information on labor market participation, both for the son and the father to avoid confounding effects of selective panel attrition.' As discussed above our instrumental variable describes the industry-specific unemployment risk. This is based on the assumption that while paternal unemployment may be endogenous to sons' unemployment this endogeneity does not exist between the paternal choice of an industry when the son is a child and sons' unemployment outcomes as a young adult. To the extent that paternal choice of industry directly affects sons' youth unemployment our instrument is invalid. The measure is calculated on an annual basis and measures for each father the unemployment risk in his industry of employment when his son was 8 years old. 10 This is used to instrument the father's unemployment when his son is aged 10 to 15. More specifically, we code by industry the share of the number of employed workers in t — 1 w h o enter unemployment in t relative to the sum of those employed in the specific industry in t plus those who entered unemployment (one year unemployment risk). As second measure of industry-specific unemployment risk we consider not entry to unemployment, a flow measure, but an indicator of the stock of unemployment (five year unemployment risk). We consider the number of prior industry employees w h o have been unemployed for between one and five years relative to the sum of employed workers in that industry in year t plus those unemployed. Both measures are calculated based on a two-digit industry code. Table 2 describes the key variables by paternal unemployment status (one or more years unemployed when son was 10-15 years old vs. employed). 11 The first row shows that sons' unemployment exposure is substantially longer if the father was unemployed at least once: while in total sons are unemployed for about 0.32 years in the age range 17 to 24 this figure amounts to 0.29 years for sons of fathers without past unemployment and almost double that period, i.e., 0.54 years for sons of fathers with past unemployment. Also, sons with an unemployed father tend to have lower educational attainment, a higher number of older siblings, and more often a migration background (first or second generation). As expected, we observe higher education among fathers who did not experience unemployment. 12

8

9

10

11

12

About one third of the unemployed sons indicate worklessness whereas two thirds report to be registered unemployed. About 35% of the sons in our final sample are observed for 8 or 9 subsequent years, 65% are available 3 to 7 years in sequence, and 2 0 % are only observed once or twice. 5 5 % of fathers are observed for the full period, 2 6 % are observed 3 to 5 times, and 19% are only observed once or twice. In cases where fathers' industry was unobservable for this period we used information for earlier (or if those were not available either, for later) periods. If fathers' industry was never observed the observation was dropped. To avoid selective sample reductions due to item non response in control variables we consider missing value categories in the specification. We consider the highest educational attainment observed over the age years 17-24 for each youth.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 363

Table 2 Descriptive statistics

Number of years son workless Years father unemployed while son aged 10-15 Number of years son not observed age 17-24 Number of years father not observed (son aged 10-15) Sons' characteristics Sons'year of birth Lower secondary school degree (Hauptschulabschluss) Intermediate school degree (Mittlere Reife) Upper secondary school degree (Abitur)/Technical school dregree (Fachhochschulreife) Other degree/No school degree/Missing information Currently in school Sons' number of siblings Sons' birthorder Migration background No migration background Direct migration background Indirect migration background Fathers' Characteristics Fathers' year of birth Father lived in East Germany at sons' age 10 Secondary Schooling Lower secondary school degree (Hauptschulabschluss) Intermediate school degree (Mittlere Reife) Technical school degree (Fachhochschulreife) Upper secondary school degree (Abitur) Other degree

Father never unemployed

Years father unemployed > 0

Full Sample

0.286 (0.719) 0 (0.000) 2.996 (2.523) 0.951 (1.569)

0.538 (0.988) 2.047 (1.306) 2.964 (2.437) 0.588 (1.199)

0.318 (0.763) 0.263 (0.829) 2.992 (2.511) 0.904 (1.531)

1986.139 (5.047) 0.170 (0.376) 0.259 (0.438) 0.253

1985.695 (4.789) 0.240 (0.428) 0.280 (0.450) 0.115

1986.082 (5.016) 0.179 (0.383) 0.262 (0.440) 0.235

(0.435) 0.022 (0.147) 0.296 (0.457) 1.642 (1.335) 1.830 (0.949)

(0.319) 0.050 (0.219) 0.315 (0.466) 2.036 (1.385) 2.140 (1.184)

(0.424) 0.026 (0.158) 0.299 (0.458) 1.692 (1.348) 1.869 (0.988)

0.806 (0.395) 0.180 (0.384) 0.014 (0.116)

0.642 (0.480) 0.344 (0.476) 0.014 (0.119)

0.785 (0.411) 0.201 (0.401) 0.014 (0.117)

1956.109 (6.895) 0.267 (0.443)

1956.140 (8.274) 0.409 (0.492)

1956.113 (7.085) 0.286 (0.452)

0.313 (0.464) 0.322 (0.467) 0.050 (0.218) 0.205 (0.404) 0.080 (0.271)

0.330 (0.471) 0.308 (0.463) 0.004 (0.060) 0.068 (0.252) 0.176 (0.381)

0.315 (0.465) 0.320 (0.467) 0.044 (0.205) 0.187 (0.390) 0.092 (0.289)

364 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

Table 2 Continued.

No school degree Postsecondary education No postsecondary education Other vocational training Industrial/commercial/health care apprenticeship Technical college, civil servant training University degree Missing information Number of observations

Father never unemployed

Years father unemployed > 0

Full Sample

0.031 (0.172)

0.115 (0.319)

0.041 (0.199)

0.084 (0.278) 0.116 (0.320) 0.415 (0.493) 0.133 (0.340) 0.249 (0.433) 0.004 (0.061)

0.226 (0.419) 0.208 (0.407) 0.441 (0.497) 0.047 (0.211) 0.075 (0.264) 0.004 (0.060)

0.103 (0.303) 0.127 (0.333) 0.418 (0.493) 0.122 (0.327) 0.227 (0.419) 0.004 (0.061)

1,896

279

2,175

Note: Table shows means and standard deviations in parentheses of key variables. Source: SOEP 1984-2012, o w n calculations.

4

Results

We present our results in four steps: we start with the least squares perspective which combines any causal and family background effects in the coefficients of paternal unemployment background. In step two we apply estimators that intend to strip off any endogeneity from the paternal unemployment indicator either by means of instrumental variables estimation or by use of the Gottschalk (1996) approach. Once we understand the causal character of the observed correlation patterns, it is of interest to study heterogeneities and transmission channels in greater detail in step three and to undertake robustness tests as step four of our analysis. 4.1

Conditional correlation patterns

The first two columns of Table 3 present the coefficient estimate that results when we regress the number of years of sons' worklessness between ages 17 and 24 on the number of years fathers were unemployment when their sons were aged 10 to 15. Column 1 describes the raw correlation, column 2 accounts for a set of family characteristics (i.e., year of birth of father and son, paternal education and migration background, sons' state of residence at age 10, sons' birth order, number of siblings, and the number of years with missing information on son and father). 13 The unconditional correlation amounts to 0.103 and is highly statistical significant. Overall, the intergenerational unemployment correlation is thus positive and one additional year of paternal unemployment is associated with five additional weeks of sons' worklessness between ages 17 and 24. Given a mean duration of sons' worklessness of 16.5 weeks 13

For the full specification and results of the linear regressions please see the Appendix.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 365

Table 3 Estimation results using O L S and IV methods IV-Results

OLS

Without controls (1) Years father unemployed First stage results First stage F-statistic First stage coefficient Number of observations Number of controls

With controls (2)

1 year unemployment risk Without With controls controls (3) (4)

0.103**'» 0.053** - 0 . 0 2 0 (0.026) (0.285) (0.026) -

-

-

-

2,175 1

2,175 51

8.64 1.470*** (0.500) 2,175 1

-0.126 (0.400) 4.28 1.109** (0.536) 2,175 51

5 year unemployment risk Without With controls controls (5) (6) -0.106 (0.410)

-0.225 (0.595)

3.51 0.522* (0.278)

1.78 0.433 (0.325)

2,175 1

2,175 51

Note: Columns (3)-(6) show IV-results with two instruments of unemployment risk. Each coefficient represents a separate linear regression. Dependent variable is years a son experienced worklessness between ages 17 and 24. Standard errors in parentheses are clustered at fathers' person number, control variables are dummies for year of birth (son and father), fathers' education, fathers' migration background, sons' state of residence at age 10, sons' birth order, sons' number of siblings, number of years son is not observed (age 17-24), number of years father is not observed (sons age 10-15). Source: SOEP 1984-2012, own calculations.

the relevance of paternal unemployment is limited. Once additional controls are considered the correlation drops by about half. While the estimate is still significantly different from zero the magnitude of the conditional correlation is small also by international comparison; O'Neill and Sweetman (1998) find that sons' unemployment experience between ages 2 1 and 31 increases by about three months if their father experienced any unemployment when the son was aged 11 or 16. Ekhaugen (2009) shows that youths with at least one unemployed parent as a teenager had an unemployment propensity that was 5 7 to 95 percent higher than that of their peers without unemployed parents. Just as in our case, Ekhaugen (2009) finds that accounting for observed family heterogeneity reduces the gross intergenerational correlation in unemployment by half. Next, we investigate the evidence with respect to causal effects.

4.2

Causality of conditional correlation patterns

We apply two methods to inspect the evidence in favor of causal effects, instrumental variables and the Gottschalk (1996) method. Table 3 shows the estimation results for the IV approach (see columns 3 - 6 ) . The first stage results for the one-year unemployment risk (i.e., unemployment entry) in columns 3 and 4 yield a significant positive correlation of aggregate unemployment risks with paternal unemployment. The five year unemployment measures are also positively associated with fathers' unemployment experience but the coefficients are estimated much less precisely. The first stage F-statistic reaches a value above 5 only in column 3 when no control variables are considered. Overall, our instruments are rather weak and the evidence has to be interpreted with caution.

366 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

The IV estimate of the effect of fathers' on sons' unemployment is negative in all four columns. Therefore, it provides no evidence in favor of a positive causal intergenerational transmission of unemployment. This suggests that the positive OLS coefficients exclusively reflect the effects of family background and of correlated observable or unobservable characteristics between fathers and sons but no causal effects. These findings match prior findings in the international literature. As an example, Macmillan (2010) instrumented paternal unemployment with being associated with a hard hit industry; the author finds an insignificant estimate of the causal paternal unemployment effect. Our second strategy to separate the true causal paternal unemployment effect from general family background correlation patterns follows Gottschalk (1996). Table 4 shows our estimation results. Because the Gottschalk specifications are estimated only on the subsample of observations for which we have evidence on paternal unemployment after

Table 4 Estimation results using the Gottschalk method (1) OLS no controls

(2) OLS with controls

(3) Gottschalk no controls

(4) Gottschalk with controls

Panel A: Father observed at least once Years father unemployed while son aged 10-15 Years father unemployed while son aged 25-30

0.152** (0.060)

Difference Number of observations Number of controls

0.046 (0.065)

-

' 266 1

1,266 47

0.123** (0.061) 0.101*** (0.027)

0.046 (0.065) 0.052* (0.029)

0.022 (0.069)

-0.006 (0.074)

1,266 2

1,266 48

Panel B: Father observed at least 3 times Years father unemployed while son aged 10-15 Years father unemployed while son aged 25-30 Difference Number of observations Number of controls

0.091 (0.072)

-

719 1

-0.046 (0.080)

-

719 45

0.049 (0.073) 0.116*** (0.031)

-0.045 (0.080) 0.054* (0.032)

-0.067 (0.083)

-0.099 (0.092)

719 2

719 46

Note: Each column represents a separate linear regression. Dependent variable is years son experienced worklessness between ages 17 and 24. In the sample we use in Panel A fathers who are observed at least one year both in the before (son age 10-15) and the after period (son age 25-30), whereas in Panel B the fathers are observed at least three times respectively. Columns (1) and (2) show OLS-results for the respective samples, columns (3) and (4) present results by using the Gottschalk-method. Columns (1) and (3) exclude and columns (2) and (4) include control variables. Standard errors in parentheses are clustered at fathers' person number, control variables are indicators for year of birth (son and father), fathers' education, fathers' migration background, sons' state of residence at age 10, sons' birth order, sons' number of siblings, number of years son is not observed (age 17-24), number of years father is not observed (sons age 10-15 and age 25-30). Source: SOEP 1984-2012, own calculations.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 367

the son reaches age 25, we re-estimated the OLS models on this subsample. The results in columns 1 and 2 confirm prior findings in Panel A, which uses 1,266 observations of sons for which the father was observed at least once both in the period when the son was aged 10-15 and when the son was aged 2 5 - 3 0 . Once we require at least three observations on paternal employment outcomes during the sons' childhood and after age 25 the sample size drops to 719 (see Panel B). In Panel B we no longer obtain significant positive correlations between father and son unemployment in the least squares estimations. Thus, the results in Panel A may be more informative. We show the estimation results of the Gottschalk (1996) approach in columns 3 and 4 without and with control variables. The estimated coefficient differences are never significant and in three out of four cases they are negative. 14 Thus, after accounting for the family background effect no positive causal effect remains. This evidence confirms the IV results and suggests - in agreement with the international literature (e.g., Macmillan 2010; Ekhaugen 2009) - that there is no positive significant causal effect of father on son unemployment. 4.3

Heterogenities and transmission channels

Next we study potential heterogeneities in the observed correlation patterns. Table 5 presents the coefficients of least squares regressions that condition on similar sets of control variables as before. The results suggest that the intergenerational unemployment correlation is larger in West than in East Germany (see column 1). One possible explanation for this difference is the generally higher unemployment incidence in East Germany which renders unemployed families (and their unobserved characteristics) more similar to the average. In columns 2 and 3 we compare correlation patterns for natives and immigrants. Due to the small number of immigrants in East Germany the results cannot be presented for this subsample. Overall, the results suggest that intergenerational correlations are tighter in the native population. This confirms prior evidence on higher educational mobility among immigrants than natives (Bauer/Riphahn 2007). In our sample of fathers the gradient of unemployment by postsecondary training is steep, with an average of 0.73 years of measured unemployment among fathers with low, 0.29 among fathers with medium, and 0.07 among fathers with high levels of education for the period when their sons were aged 10-15 (figures not presented). So clearly, a high intergenerational unemployment correlation would generate the worst outcome for sons of low educated fathers and would be beneficial in the case of highly educated fathers. Interestingly, the estimated correlation patterns in columns 4 - 6 of Table 5 yield that the intergenerational correlation of unemployment is high and statistically significant only in the medium education category. Therefore, neither do the sons of low educated fathers suffer nor do the sons of highly educated fathers benefit in any particular way. Instead the overall intergenerational correlation of unemployment outcomes is borne by the largest population group of medium educated fathers, which in our sample account for 52 percent of all fathers. In Table 6 we present estimation results that describe the transmission channels between paternal and youth unemployment. In Panel A we commence by presenting the raw cor14

In a robustness test we redid all estimations based on the Gottschalk approach when using paternal unemployment as measured at sons' age 10-13. The results of no significant positive causal effect are confirmed and the estimates of the difference in parameters are rather similar to those in Table 4.

368 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

Table 5 Heterogeneities in linear regression results All

Panel A: West Germany Years father unemployed Number of observations Number of controls

0)

Fathers' migration background

Fathers' postsecondary education

no

yes

low

medium

high

(2)

(3)

(4)

(5)

(6)

0.025 (0.051) 444 33

-0.036 (0.058) 235 31

0.060* (0.037) 1,554 36

0.111** (0.051) 1,110 33

0.037 (0.037) 621 36

0.039 (0.038) 598 33

23

39

0.053** (0.026) 2,175 51

0.082*** 0.016 (0.030) (0.049) 1,708 467 34 33

0.001 (0.047) 274 32

0.148** (0.057) 775 32

0.007 (0.074) 542 32

0.045 (0.044) 360 29

0.005 (0.066) 216 30

Panel B: East Germany Years father unemployed Number of observations Number of controls Panel C: Full sample Years father unemployed Number of observations

0.114*** 0.017 (0.038) (0.047) 1,135 758 32 32

Note: Each coefficient represents a separate linear OLS regression. Dependent variable is years son experienced worklessness between ages 17 and 24. Standard errors in parentheses are clustered at fathers' person number, control variables are indicators for year of birth (son and father), fathers' education, fathers' migration background (not for subgroups by migration background), sons' birth order, sons' number of siblings, number of years son is not observed (age 17-24), number of years father is not observed (sons age 10-15). We only present results for subgroups with at least 100 observations. Fathers' postsecondary education is defined as follows, low: without tertiary degree or still in education; medium: apprenticeship training or in-firm training; high: master, technical college, university. Source: SOEP 1984-2012, own calculations.

relation between the two unemployment measures conditional on only a few covariates such as region, son's year of birth, and number of missing observations. We then add covariate groups based on their relevance for the considered correlation patterns: we start with paternal characteristics, then enter family characteristics and finally allow for youth characteristics that may be confounded by paternal unemployment. Already considering paternal year of birth and education (in column 2) reduces the correlation coefficient by one third. The family characteristics in column 3, i.e., migration background, birth order, and number of siblings does not add much to the explanation of the correlation patterns. In fact, Panel B shows that by themselves even the most basic paternal characteristics are more relevant to the unemployment correlation than the family indicators: compared to column 1 the coefficient declines more in column 2 than in column 3. When we consider additional characteristics of the son such as education and industry of employment the correlation coefficient declines further and loses statistical significance (see Panel A).15 Overall, all considered groups of covariates yield jointly statistically significant coefficient

15

This agrees well with the finding of Pinger (2012) w h o shows that parental u n e m p l o y m e n t significantly affects youth educational outcomes.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 369

Table 6 Linear regression results on transmission mechanisms Panel A: Inclusion jointly

(1)

Years father unemployed

0.105*** (0.025)

0.065** (0.026)

0.053** (0.026)

0.037 (0.026)

Basic controls (26) Father characteristics (14) Family characteristics (10) Son characteristics (46) Number of observations Number of controls

F=8.44

F=6.58 F=6.15

F=6.84 F=4.96 F=2.49

F=6.23 F=3.03 F=2.03 F=6.06 2,175 97

-

2,175 27

(2)

-

2,175 41

(3)

-

2,175 51

(4)

Panel B: Inclusion pairwise Years father unemployed

0.105*** (0.025)

0.065** (0.026)

0.073*** (0.026)

0.068*** (0.024)

Basic controls (26) Father characteristics (14) Family characteristics (10) Son characteristics (46) Number of observations Number of controls

F=8.44

F=6.58 F=6.15

F=7.92

F=6.31

-

-

-

2,175 27

-

F=3.45 -

2,175 41

2,175 37

-

F=12.04 2,175 73

Note: Each coefficient represents a separate linear OLS regression. Dependent variable is years son experienced worklessness between ages 17 and 24. Standard errors in parentheses are clustered at fathers' person number. The control variable groups are defined as follows: (a) basic controls: sons year of birth, number of years son (father) missing when son was 17-24 (10-15), state dummies; (b) fathers characteristics: year of birth dummies, education (secondary and tertiary); (c) family characteristics: migration background, number of siblings, birth order; (d) sons characteristics: industry (2 digit), education (highest completed). Source: SOEP 1984-2012, own calculations.

estimates and are correlated with sons' and fathers' unemployment outcomes; however, paternal characteristics are the most influential transmission mechanism.

4.4

Robustness checks

We submit our key results, i.e., a significant positive overall correlation between paternal and child unemployment but no positive causal effect to three robustness checks. First, we use the available panel data for sons' dichotomous annual worklessness outcomes between ages 17 and 24 instead of an aggregate count of the total number of years that we studied so far. Panel A of Table 7 shows the results of applying least squares and instrumental variables estimators to the now much larger sample of 10,893 observations. Here, we use 2,175 different observations on sons with about five annual observations on average and a mean annual unemployment probability of 0.06 percent. The estimation outcomes confirm prior results: the parent-child unemployment correlations in columns 1 and 2 are significantly positive, and the instrumental variables estimates yield insignificant negative coefficients, again based on potentially weak instruments. Next, we return to the cross-sectional sample but replace the continuous unemployment measures for sons and fathers by dichotomous measures that describe whether son or father ever experienced at least one spell of unemployment in the respective considered

370 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

Table 7 Robustness tests OLS

Without controls (1)

IV-Results

With controls (2)

1 year unemployment risk Without With controls controls (3) (4)

5 year unemployment risk Without With controls controls (5) (6)

Panel A: Panel structure - dependent var.: dichotomous indicator of annual son unemployment Years father unemployed First stage results First stage F-statistic First stage coefficient Number of observations Number of controls

0.022*** 0.012** (0.005) (0.005) -

10,893 1

10,893 51

-0.012 (0.063)

-0.022 (0.088)

-0.041 (0.117)

-0.067 (0.209)

7.57 3.30 1.310*** 1.003' (0.476) (0.552)

1.44 0.368 (0.307)

0.57 0.267 (0.354)

10,893 1

10,893 1

10,893 51

10,893 51

Panel B: Cross-section data - dependent var.: dichotomous indicator whether son ever unemployed Father ever unemployed (0/1) First stage results First stage F-statistic First stage coefficient Number of observations Number of controls

0.040*** 0.030* (0.015) (0.016) -

2,175 1

2,175 51

-0.147 (0.133)

-0.212 (0.205)

-0.093 (0.264)

-0.305 (0.518)

11.06 5.54 0.740*** 0.568' (0.223) (0.242)

2.69 0.168 (0.102)

1.01 0.121 (0.120)

2,175 1

2,175 51

2,175 1

2,175 51

Panel C: Cross-section data - measuring paternal unemployment at sons' age 10-13 only Years father unemployed First stage results First stage F-statistic First stage coefficient Number of observations Number of controls

0.118*** 0.050* (0.039) (0.038) -

1,960 1

1,960 51

-0.101 (0.382)

-0.159 (0.475)

8.87 5.71 1.146*** 0.961' (0.385) (0.402) 1,960 1

1,960 51

-0.146 (0.475)

-0.193 (0.660)

4.63 0.463** (0.215)

2.37 0.380 (0.246)

1,960 1

1,960 51

Note: Columns ( 3 H 6 ) show IV-results with two instruments of unemployment risk. Each coefficient represents a separate linear regression. Dependent variable is years son experienced worklessness between ages 17 and 24. Standard errors in parentheses clustered at fathers person number, control variables are dummies for year of birth (son and father), fathers' education, fathers' migration background, sons' state of residence at age 10, sons' birth order, sons' number of siblings, number of years son is not observed (age 17-24), number of years father is not observed while son aged 10-15 (and 10-13 in Panel C). Source: SOEP 1984-2012, own calculations.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 371

periods, i.e., for sons at age 1 7 - 2 4 and for fathers at the time when their sons were aged 1 0 - 1 5 . This shifts the focus to the extensive margin of the unemployment experience. Panel B of Table 7 shows the estimation results, which confirm prior findings: in columns 1 and 2 we obtain significant positive estimates, while the IV results yield negative insignificant unemployment coefficients throughout. T h e robustness test corroborates prior results. In our final robustness test we investigate the sensitivity of the results to the definition of the period in which paternal unemployment is measured. Instead of focusing o n fathers' unemployment when the son is aged 1 0 - 1 5 we n o w investigate the outcomes w h e n using paternal unemployment in the sons' age range 1 0 - 1 3 only. This causes a slight decline in the number of observations, however, the overall outcome as shown in Panel C of Table 7 is robust: again we find positive correlations in the least squares results, but n o evidence for a positive causal effect in the instrumental variables estimation (nor in the Gottschalk (1996) approach, which we do not present save space). Thus, our estimations are robust to three different robustness tests.

5

Conclusions

Even though youth unemployment is the most pressing problem in many European labor markets we k n o w very little a b o u t the mechanisms behind it. While some authors address the role of d e m a n d and supply for young workers and the patterns of incidence and duration of youth unemployment we address one factor that so far has been neglected in m a n y discussions and certainly in the literature on G e r m a n youth unemployment: the impact of family background. This paper studies the intergenerational correlation between the unemployment experiences of fathers and their sons. The international literature features only few contributions on the subject. Several mechanisms may generate a correlation between the employment outcomes of fathers and their sons; we can think of observable characteristics that r u n in the family and of unobservable traits and attitudes that m a y be transmitted f r o m parent to child. In addition, it is of particular interest to determine whether there is a causal effect that makes sons more (or less) likely to experience unemployment once they have seen their fathers unemployed. A variety of reasons may be behind such causal mechanisms and their relevance is obvious: if there is a causal intergenerational transmission of unemployment this provides an additional rationale for labor market policy supporting the employment opportunities of parents. If n o such causal connection can be established the fight against youth unemployment may be more successful if it focuses on youths themselves. This paper shows that the unemployment experience of G e r m a n fathers and sons is positively correlated. W e apply standard empirical approaches to test whether the character of these correlations is causal. O u r results are robust to the application of instrumental variables techniques, the application of Gottschalk's (1996) m e t h o d , to considering outcomes at the intensive and the extensive margin, and to applying data in a cross-sectional and panel data setting: while parent a n d child unemployment experiences are significantly positively correlated, this correlation does not go back to a causal effect. Instead, family background affects the unemployment risks of both fathers and their sons. O u r results agree with most of the literature on intergenerational unemployment transmission, which confirms positive correlations but rejects causal mechanisms.

372 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

Appendix Full regression results (1)

Number of years father unemployed Number of years son not observed age 17-24 Number of years father not observed (son aged 10-15) Sons' year of birth 1977-1978 1979-1980 1981-1982 1983-1984 1985-1986 1987-1988 1989-1990 1991-1992 1993-1994

(2)

0.103*** 0.053* (0.026) (0.026) -0.067*** (0.007) 0.006 (0.012) 0.001 (0.086) -0.002 (0.077) 0.140 (0.078) 0.011 (0.082) 0.093 (0.071) -0.032 (0.065) -0.043 (0.058) -0.059 (0.052) -0.052 (0.044) Reference

1995 Fathers' year of birth

Reference

1929-1939

-0.131 (0.199) -0.123 (0.192) -0.035 (0.199) -0.196 (0.206)

1940-1949 1950-1959 1960-1969 1970-1975 Fathers' secondary schooling Lower secondary school degree Intermediate school degree Technical school degree Upper secondary school degree

-0.106 (0.122) -0.170 (0.125) -0.162 (0.140) -0.192 (0.123)

Intergenerational Transmission of Unemployment - Evidence for German Sons • 373 (Continued.) (1)

Other degree

-0.332** (0.123) Reference

No School Degree Fathers' postsecondary education

Reference

No postsecondary education

0.006 (0.086) -0.143* (0.071) -0.263*** (0.072) -0.232** (0.072) -0.329* (0.158)

Vocational training Industrial/commercial/health care apprenticeship Technical college, Civil servant training University Missing Sons' number of siblings

-0.146** (0.148) -0.263** (0.090) -0.212* (0.091) -0.215* (0.102) Reference

No siblings 1 sibling 2 siblings 3 siblings > 4 siblings

-0.133 (0.134) -0.055 (0.138) -0.103 (0.141) Reference

Sons' birth order 1 st born 2nd born 3rd born > 4rd born

0.022 (0.176)

Missing Fathers' migration background

Reference

No migration background

0.145* (0.060) 0.260 (0.194) Yes (p=.097)

Direct migration background Indirect migration background Additional controls: state dummies (15) Number of observations

(2)

2,175

2,175

Note: Each column represents a separate linear regression. Dependent variable is years son experienced worklessness between ages 17 and 24. Standard errors in parentheses are clustered at fathers' person number. Source: SOEP 1984-2012, own calculations.

374 • Miriam Mäder, Steffen Müller, Regina T. Riphahn, and Caroline Schwientek

References Antel, J.J. (1992), The intergenerational transfer of welfare dependency: Some statistical evidence. The Review of Economics and Statistics, 74: 467^473. Bauer, P.C., R.T. Riphahn (2007), Heterogeneity in the Intergenerational Transmission of Educational Attainment: Evidence from Switzerland on Natives and Second Generation Immigrants. Journal of Population Economics 20: 121-148. Beaulieu, N., J.-Y. Duclos, B. Fortin, M. Rouleau (2005), Intergenerational reliance on social assistance: Evidence from Canada. Journal of Population Economics 18: 539-562. Blomeyer, D., M. Laucht, K. Coneus, F. Pfeiffer (2013), Early Life Adversity and Children's Competence Development: Evidence from the Mannheim Study of Children at Risk. Jahrbücher für Nationalökonomie und Statistik (Journal of Economics and Statistics) 233: 467-485. Bratberg, E., O. Anti Nilsen, K. Vaage (2008), Job losses and child outcomes. Labour Economics 15: 591-603. Corak, M., B. Gustafsson, T. Österberg (2004), Intergenerational influences on the receipt of unemployment insurance in Canada and Sweden. Pp. 245-288, in: M. Corak (ed.), Generational income mobility in North America and Europe, Cambridge University Press, Cambridge. Couch, K.A., Dunn T.A. (1997), Intergenerational Correlations in Labor Market States. A Comparison of the United States and Germany, Journal of H u m a n Resources 32: 210-232. De Paola, M . (2013), The determinants of Risk Aversion: The Role of Intergenerational Transmission. German Economic Review 14: 214-234. Edmark, K., K. Hanspers (2012), Is welfare dependency inherited? Estimating the causal welfare transmission effects using Swedish sibling data. IFN working paper N o . 894, Stockholm. Ekhaugen, T. (2009), Extracting the causal component from the intergenerational correlation in unemployment. Journal of Population Economics 22: 97-113. Franz, W., J. Inkmann, W. Pohlmeier, V. Zimmermann (2000), Young and Out in Germany. On Youths' Chances of Labor Market Entrance in Germany. Pp. 381—426 in: D. Blanchflower, R. Freeman (eds.), Youth Unemployment and Joblessness in Advanced Countries. NBER Cambridge, MA. Gottschalk, P. (1990), AFDC participation across generations. American Economic Review 80: 367-371. Gottschalk, P. (1996), Is the correlation in welfare participation across generations spurious? Journal of Public Economics 63: 1 - 2 5 . Gregg, P. (2001), The impact of youth unemployment on adult unemployment in the NCDS. Economic Journal 111: F626-F653. Gregg, P., L. Macmillan, B. Nasim (2012), The impact of fathers' job loss during the recession of the 1980s on their children's educational attainment and labour market outcomes. Fiscal Studies 33: 237-264. Heineck, G., R.T. Riphahn (2009), Intergenerational Transmission of Educational Attainment in Germany - The Last Five Decades. Journal of Economics and Statistics (Jahrbücher für Nationalökonomie und Statistik) 229: 36-60. Johnson, P., H. Reed (1996), Intergenerational mobility among the rich and poor: results from the National Child Development Survey. Oxford Review of Economic Policy 12: 127-42. Macmillan, L. (2010), The Intergenerational Transmission of Worklessness in the UK. C M P O Working Paper Series N o . 10/231. Bristol. O'Neill, D., O. Sweetman (1998), Intergenerational mobility in Britain: evidence from unemployment patterns. Oxford Bulletin of Economics and Statistics 60: 431—447. Oreopoulos, P., M. Page, A. Huff Stevens (2008), The intergenerational effect of worker displacement. Journal of Labor Economics 26: 455-483. Pinger, P.R. (2012), Intergenerational Effects of Economic Distress: Paternal Unemployment and Child Secondary Schooling Decisions, mimeo. University of Mannheim.

Intergenerational Transmission of Unemployment - Evidence for German Sons • 3 7 5

Riphahn, R.T., M. Zibrowius (2014), Apprenticeship, Vocational Training and Early Labor Market Outcomes - Evidence from East and West Germany, mimeo, University of ErlangenNuremberg. Schmillen, A., M. Umkehrer (2013), The scars of youth. Effects of early-career unemployment on future unemployment experience, IAB Discussion Paper No. 6/2013, Nürnberg. Schnitzlein, D. D. (2014), How important is the family? Evidence from sibling correlations in permanent earnings in the USA, Germany, and Denmark. Journal of Population Economics 27: 69-89. Wagner, G.G., J.R. Frick, J. Schupp (2007), The German Socio-Economic Panel Study (SOEP) Scope, Evolution and Enhancements. Schmollers Jahrbuch (Journal of Applied Social Science Studies) 127: 139-169. Corresponding author: Prof. Regina T. Riphahn, Ph.D., Friedrich-Alexander University Erlangen-Nürnberg (FAU), Lange Gasse 20, 90403 Nuremberg, Germany. [email protected] Dr. Miriam Mäder; [email protected] Caroline Schwientek; [email protected] Prof. Dr. Steffen Müller, Halle Institute for Economic Research and Magdeburg University; [email protected]

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2015) Bd. (Vol.) 235/4+5

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects Charlotte Cabane and Michael Lechner* Swiss Institute for Empirical Research (SEW), University of St. Gallen JEL 112; 118; J20; J30; J68; L83 Physical activity; leisure time physical activity; sports participation; labour market effects; unemployment; earnings.

Summary We survey the literature on the link of labour market related outcomes to individual physical activity and sports participation. The first part of the survey is devoted to the individual participation decision and is based on papers from various disciplines. The second part summarises parts of the epidemiological literature on health effects and the economic literature on the labour market effects as well as on the effects on well-being and social capital. Somewhat surprisingly, at least for studies in empirical economics, all the papers seem to agree that individual leisure sports participation and physical activity has positive effects for adults.

1

Introduction

Improving the productivity of the labour force is an important policy goal of m o d e r n states. It helps to improve the position of the economy in the international competition and increases welfare and usually reduces unemployment. M a n y of such policies, like schooling and vocational training for younger individuals, and active labour market policies for the unemployed have been more or less thoroughly investigated by empirical economists, sometimes with rather mixed results. An aspect that has been somewhat overlooked by economists are the productivity enhancing features that are attributed to individual sports activities, as well as to individual physical activities in general. Such activities are likely to foster non-cognitive skills (as well as some cognitive ones), like self-discipline, better coping with stress, the ability to w o r k in teams etc. Moreover, the individual's productivity also depends on his/her health status which is directly affected by the level of physical activity.

* Michael Lechner is also affiliated with CEPR and PSI, London, CESIfo, Munich, IAB, Nuremberg, and IZA, Bonn. A draft of a related paper was presented at the ZEW workshop in honour of Wolfgang Franzs 70th birthday in Mannheim. We thank participants, in particular Wolfgang Franz, as well as Nicole Guertzgen and two anonymous referees, for helpful comments and suggestions. The usual disclaimer applies.

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 377

Therefore, in this survey we try to take stock of the relevant literature in economics, but also in epidemiology, sports sciences, and other social sciences. 1 The aim is to understand which individuals participate in sports and physical activities for w h a t reasons, as well as to understand w h a t is k n o w n a b o u t the effects of such participation. Concerning the latter, we consider those effects that can be related to productivity, namely health and well-being, social capital, and in particular labour market outcomes. Since the literature on this topic is huge and increasing, in particular in epidemiology, we have to restrict ourselves to specific groups of individuals and specific types of individual activities. Concerning groups of individuals, we focus on adults in working age for w h o m labour market performance can be directly measured (of course, this is not all meant to imply that physical activities for children are less relevant for their later labour market productivity). 2 Concerning physical activity, we d o not restrict ourselves to individual sports participation, but include also papers that use broader concepts of physical activity. However, those papers usually consider such activities conducted during leisure time (leisure time physical activity, LTPA), because data on the level of physical activity on the job are very rare. Since physical activities are considered a key determinant of individual health, various public bodies issued recommendations for m i n i m u m levels of such activities. With some institution specific variation, they recommend that all adults aged 18 to 65 need "moderate-intensity aerobic physical activity for a minimum of 30 minutes on five days each w e e k " (Haskell et al. 2007: 1083). This is the current guideline adopted by the W o r l d Health Organization ( W H O 2010), the Canadian Society for Exercise Physiology (CSEP 2011), and the US Surgeon General (U.S. D e p a r t m e n t of Health and H u m a n Services 2010), a m o n g many other national health organizations. There is also a consensus that, f r o m the point of view of health effects, more physical activity is better. This recommended m i n i m u m level of physical activity corresponds to a daily energy expenditure of at least 1.5 kilo-calories per kilogram of body weight (unit: kcal/kg, so-called M E T ) f r o m all LTPAs. 3 For example, in Switzerland, which is a country with rather high activity levels, still more than a third of the adult population does not achieve this level (HEPA 2013), which is an indication of the considerable possibilities to further increase population activity levels. The paper proceeds with a section reviewing the factors that are associated with LTPA participation as well as considering theoretical explanations for participation. Then, the literature on the effects of LTPA on health, well-being, labour market outcomes, and social capital is discussed. Finally, we identify some avenues for future research.

1

1 3

Since the fields and journals where sports related papers appear are huge, we certainly have overlooked important contributions and may have misrepresenting other work (in particular, but not only, outside of economics). To all authors who rightly have reasons to complain we apologize in advance. The literature reviews for these groups will be referred to some other place. Individuals may meet this goal with various types and duration of sports and exercises. Examples are daily walking for 30 minutes with a speed of 2.5 miles per hour on a firm surface, or 3-times a week running for 25 minutes or longer with a speed of 5 miles per hour (for other examples, see Ainsworth et al. 2000).

378 • Charlotte Cabane and Michael Lechner

2

Determinants and correlates of participation in sports and physical activity

Baumann et al. (2002) underline the importance of distinguishing between determinants and correlates of physical activity ('PA' henceforth). Indeed, although most of the empirical papers on physical activity agree on the list of characteristics associated with physically active individuals or with sports-friendly environments, few of these papers are able to identify a causal relationship between these characteristics and physical activity. Clearly, correlates of physical activity are interesting and worth investigating. However, in order to get a better understanding of the underlying causal mechanisms, it is necessary to attempt to identify 'determinants' of physical activity. Such understanding, for example, helps to develop policy recommendations. Below, we strive to differentiate between papers that are successful in identifying factors that are likely to causally influence participation in physical activity and papers that merely highlight correlates (although this distinction is not possible without some subjective judgement on the side of the writers of this survey). In doing so, in the next subsection we first review simple stylized theoretical models that explain individual involvement in physical activity and empirical studies that test these models. In Sect. 2.2, we present the studies that mainly propose and analyse diverse potential correlates of physical activity without systematically relying on economic theory.

2.1

Why do individuals participate in physical activities?

Downward (2007) outlines two types of competing economic theories to explain sports participation: the 'neoclassical' theory and a set of alternative theories, which he labels as 'heterodox'. The neoclassical theory refers to the allocation of time framework (Becker 1965) where individuals choose leisure time and consumption in order to maximize their utility. In this type of model, the main drivers of physical activity are hours of work and income. The 'heterodox' or behavioural theories draw from other sciences such as sociology and psychology. They weaken the assumption of full rationality by assuming bounded rationality instead. Individuals are subject to many non-financial influences (e.g. social pressure, learning-by-doing or spill-over effects) which modify their preferences throughout life. Therefore, sports participation is primarily determined by social relations while financial constraints, which are emphasized by the neoclassical framework, are second order. However, the great majority of the sports economics literature nevertheless uses the neoclassical framework (e.g. Colman/Dave 2013a; Humphreys/Ruseski 2007, 2011). In neoclassical models there are two main reasons to engage in physical activities: taste for sports and desire to maintain health. In the first case physical activity is modelled as consumption good, e.g. physical activity directly increases the level of utility without any further lasting impact. In the second case, it is modelled as an intermediary (investment) good that is used to improve current and future health (Grossman 1972). In this case, the link between physical activity and utility is indirect because individuals derive utility from health and not from physical activity directly. Both characteristics of sports can be modelled simultaneously. Of course, treating all important aspects of physical activity as an investment good requires a dynamic model. However, the dynamic sports participation literature is not yet well developed. Therefore, we present (only) a static framework with physical activity as consumption good, and alternatively as intermedi-

Physical A c t i v i t y of A d u l t s : A Survey o f Correlates, D e t e r m i n a n t s , a n d Effects • 3 7 9

ary good.4 It will be seen that even such simplified models reveal the major trade-offs involved. In Cawley (2004) Sleep-Leisure-Occupation-Transport-Home-production model, which we use as our benchmark model, individuals' utility depends on time allocation (which gives the model its name), health, body weight, and consumption.5 Physical activity is defined as time dedicated to leisure (recreational sport) and time dedicated to commuting (active transportation such as biking and walking). The time spent on physical activities affects the level of utility directly and indirectly via its impact on health and body weight. In other words, physical activity is modelled as consumption as well as intermediary (investment) good. Five channels link physical activity to utility: the consumption of physical activity, the increase in health due to physical activity, the decrease in weight due to physical activity, the increase in health due to a decrease in weight, as well as an effect that comes by increased (or reduced) transportation time. The individual is constrained in terms of time, financial resources and faces fixed biological conditions. 2.2.1 Physical activity as a consumption good

One strand of the literature considers physical activity as consumption good only. Hence, individuals derive utility from physical activities directly, rather than indirectly through an intermediate factor. We use the SLOTH model as a point of departure, but ignore the role of physical activity as an intermediary good in health and weight production. Thus, weight and health are considered as exogenous.6 These simplifications lead to the following model: Max U = U{S, L, O, T, HP, H, W, Y) under the constraints: i.

Y = w* O

ii. S + L + O + T + HP = 24 The arguments of the utility function are sleep (S), leisure time (L), working hours (O), transport (T), home production (HP), health (H), body weight (W), and consumption (Y). This setting corresponds to a leisure consumption model. The demand of leisure time physical activity (LTPA) is included one-to-one in the demand for leisure time (L). The first order conditions (FOC) show that the demand for leisure time physical activities (LTPA) depends on the opportunity cost of time at work w, i.e. the earnings per unit of time. FOC: ULTPA/UY

=

W

with LTPA = L - time dedicated to other leisure activities. 4

5

6

A short section on dynamic models is included in the earlier version of the survey that is available in the discussion paper series of the Economics Department of the University of St. Gallen (Cabane/Lechner 2 0 1 4 ) . We choose the S L O T H model as a departure point because the great majority of the literature builds on this model. The discussion paper version of this survey includes a section that extensively reviews this model. I.e. health and weight production are assumed not to be influenced by physical activity.

3 8 0 • Charlotte Cabane and Michael Lechner

Garcia et al. (2011) use this type of model to explain participation in physical activity in Spain. In line with the above analysis, they predict that the opportunity cost of time determines the participation in physical activities. As expected, the demand for physical activity is negatively correlated to hourly earnings. Furthermore, the relation between age and physical activity forms a U-shaped curve with a minimum at an age of 33. According to the authors, the age effect can be attributed to the fact that the opportunity cost of time varies with age (this suggest that either older people have lower opportunity cost of time or that the effects of physical activity increase with age, for example). Their claim is however not easy to reconcile with the increasing wage-experience profiles which are frequently observed. Humphreys and Ruseski (2011) propose a model that combines elements of the SLOTH and the so-called recreation demand model (McConnell 1992). The main idea of the recreation demand model is to take into account the impact of travel costs on the demand for leisure. Humphreys and Ruseski (2011) key assumption is that engaging in physical activity has a fixed cost as well as a variable monetary cost. The fixed cost depends on the decision to engage in physical activity (extensive margin). The variable cost depends on the duration of the participation (intensive margin). The intensity of the physical activity refers to the duration of the physical activity. For example, the fixed cost comprises the yearly membership fee of a fitness club while the variable costs include expenditures for running shoes (the number of pairs needed depends on the frequency). This aspect of Humphreys and Ruseski (2011) model is very much in the spirit of many labour supply models (see Heckman 1993, for example) in which individuals are modelled first to decide on the entry into the labour force (extensive margin) and subsequently to choose their working hours (intensive margin). An interesting feature of their model is that they treat individuals' wages as endogenous and instrument incomes with the unemployment rate. The idea behind this strategy is that physical activity can influence incomes via health (and potentially other channels discussed in section 3). Although allowing for a more realistic setting, it is not clear that the instrument used is valid because the unemployment rate might have a direct effect on incomes as well. Humphreys and Ruseski (2011) use the Behavioural Risk Factor Surveillance System (BRFSS) data to estimate their model. They find results in line with their theoretical predictions: income has a positive and significant effect on the extensive margin but a negative and significant effect on the intensive margin. Downward et al. (2009) suggest the use of the New Household Economics model (Becker 1974). In this model, the focus is on the interaction between the household members and their joint consumption and production. Indeed, the individual maximizes her/his utility within the household by deciding on the amount of consumption and production of goods which maximize the household utility under the household constraints (in terms of time and money). Therefore, intra-household optimization influences the choice to engage in physical activity. Moreover, the past consumption of physical activity and the other household members' consumption of physical activity are relevant. If another member of the household used to go running, for example, then the costs to engage in running may be lower for the other members of the household than the costs to engage in a different physical activity. Relevant costs in this respect may be transportation costs, learning costs and informational costs, among others. Also, the utility of sharing the same activity for both members might be greater than the sum of the utility of each member doing a different activity (peer effect).

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 381

2.1.2 Physical activity as an intermediary good for health production

There is ample evidence that physical activities impact health (see Section 3.2). Hence, it appears natural to model the impact of physical activity on overall utility by letting it influence health (only). In line with the S L O T H model, Meltzer and Jena (2010) consider physical activity as an intermediary good in health production. However, they also take into account the different intensity levels in physical activity a n d its impact on health production. In particular, utility is assumed to depend on health (H), the consumption of a composite good (Y), and the intensity of the physical activity (I). The health production function depends on the intensity of the physical activity (I) and on the n u m b e r of hours spent in physical activity per day {PA). It is w o r t h noting that the intensity (I) relates to the a m o u n t of energy or effort required to engage in physical activity during a specific time unit. This differs f r o m H u m p h r e y s a n d Ruseski f r a m e w o r k . Using the above notation, the optimization p r o g r a m m e in the spirit of Meltzer and Jena (2010) can be written as follows: Max U = U(H, I, Y) with H = H(PA,

I)

subject to the following constraints: i.

Y =

w*0

ii. O + PA = 24. The key difference with respect to the previous section is that health depends n o w on the physical activities while the S L O T H activities themselves enter the utility function only indirectly via their effect on health. 7 Meltzer and Jena (2010) derive the following relationship (using the above notation): FOC: HPA/HI

=

w*UY/(-UI)

with HPA denoting the marginal health benefit of the time spent in physical activity a n d H I denoting the marginal health benefit of increasing exercise intensity. Hence, this framew o r k highlights the role of physical activities in increasing health, w represents the cost of increasing the time engaged in physical activity which is equivalent to the opportunity cost of time. Finally, the ratio Uy/(— Uj) represents the cost of the increasing intensity of physical activity. Meltzer and Jena (2010) emphasize the fact that a wage increase has opposing income a n d substitution effects. Assume that health and distaste for intensive exercise are normal 'goods'. Then, for a constant level of health, a higher wage increases the intensity of exercise. The authors use the National Health and Nutrition Survey (NHANES) to test their model predictions empirically. Their analysis suggests that higher opportunity costs of time are indeed correlated with higher intensity of exercise. 7

According to the SLOTH model individuals are physically active either during leisure time (L) or transport time (T). We use the simplification suggested by Meltzer and Jena (2010) which implies that health only depends on physical activities. Therefore, we consider the case in which time is allocated either to work or to physical activities.

382 • Charlotte Cabane and Michael Lechner

2.1.3 Heterodox theories As it is well known, the typical neoclassical models presented in the last sections have several caveats. The aim of neoclassical theories is to build models that predict the demand of physical activity. By contrast, heterodox frameworks are restricted to the description and explanation of the demand. The neoclassical predictions are made under a set of assumptions which imply, among others, stable preferences and rationality. Both assumptions are rather strong and might not realistically describe the individuals' decision making. The heterodox theories address these concerns by incorporating preferences that evolve over time according to consumption skills and social interactions, for example. Furthermore, a strand of research argues that individuals are not strict utility maximisers but also chose heuristically. Heterodox authors may distinguish between wants and needs and may assume that individuals are able to establish a hierarchy in their needs. In other words, bounded rationality implies that individuals choose within subsets that are ranked according to their needs (rather than choosing among all goods). Hence, the heterodox framework allows understanding facts and believes about physical activities that are not in line with the neoclassical model predictions (as underlined by Downward 2007). These frameworks help to shed light on various aspects previously neglected. For example, sport as a way to socialize and integrate or the effect of peers and the social environment on sports behaviours and habits. The following articles investigate similar questions with respect to the demand of physical activity using a heterodox framework. Downward (2007) tests versions of both the neoclassical theory and the heterodox theory empirically. Using English data (General Household Survey of UK households), he estimates the probability of being engaged in physical activities. His results are in favour of the heterodox economic theory. Indeed, he finds a very small impact of income while characteristics such as participation in other leisure activities and volunteering are highly and significantly correlated with participation in physical activities. 8 He concludes that, rather than facing an income-leisure trade-off, individuals rank their needs and participate in physical activities once higher needs are satisfied. In order to incorporate these heterodox findings, the author's subsequent articles include variables that reflect the social environment of the individual. According to Downward et al. (2011), individuals engage in sports or physical activity in order to socialize and to "live life to the full". In this case physical activity appears to be a consumption good which involves social interactions and impacts others' satisfaction/ utility. Downward and Riordan (2007) use cluster analysis on the General Household Survey (GHS) in order to estimate a model of social interaction. Cluster analysis is somewhat similar to matching: individuals are allocated to the cluster which fit them the best (in terms of individuals' characteristics). The number of clusters is identified from the data. It is however possible that some individuals are not allocated to any cluster (because of being too different). Downward and Riordan (2007) identify three clusters: sport, recreation and leisure. The clusters contain 2 . 9 % , 20.7% and 76.4% of the overall sample allocated to clusters. In the next step, the authors estimate the probability to be engaged in a specific sport given the specific cluster and given the individual's characteristics. They find similar results as Downward (2007): participating in a set of activities is positively correlated to being engaged in physical activities while income has a marginal effect. These results

8

Education is also positively and significantly correlated with participation in physical activity.

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 383

support the heterodox theories, although the identification strategies may be subject to scrutiny. Stempel (2005) uses U.S. data to test Bourdieu's theory applied to sports. The main idea is that members of the d o m i n a n t class (in terms of cultural and economic capital) use sports as a way to distinguish themselves f r o m the others classes. In other words, individuals choose their type of physical activity (sport) according to the class they belong to. The a u t h o r estimates the probability that individual i does a specific type of sport given that she/he belongs to a specific class. In a second step, he uses these estimations to define which type of sport is considered as increasing the cultural capital (from the d o m i n a n t class's point of view). The author argues that the d o m i n a n t class engages more in strenuous aerobic activities (fitness sport). Also, Stempel (2005) underlines the fact that the d o m i n a n t class is omnivore, i.e. individuals f r o m this class are more likely to participate in all the sports than individuals f r o m the middle or f r o m the lower class. 9 Last, fitness sports are relatively more favoured by the culturally d o m i n a n t class while the economically d o m i n a n t class invests relatively more in competitive sports. The decision to engage in physical activity is complex. The neoclassical models propose t w o motivations: the taste for physical activity and the will to maintain or increase health. 1 0 The main cost is the opportunity cost of time. The heterodox theories underline the importance of the social environment and non-rational decision making. It turns out that income has merely a threshold effect. T h e influence of education, health and demographic characteristics (e.g. age or gender) are acknowledged and studied by both theories. Both f r a m e w o r k s are backed by empirical support. Therefore, it seems reasonable to take into account motives for physical activity that come f r o m the neoclassical as well as heterodox reasoning. 2.2

Who participates?

A substantial n u m b e r of papers study the differences between physically active people and physically inactive ones. Several papers (e.g. D o w n w a r d et al. 2 0 1 1 ; Garcia et al. 2011) suggest disentangling the t w o decisions related to participation in physical activity: whether to engage in physical activity at all and for h o w long (and in which intensity). W e distinguish different types of correlates: individual characteristics, weather conditions, neighbourhood characteristics, as well as the influence of peers, habits and commitments a n d the role of incentives. 2.2.1 Individual characteristics The literature on the determinants of physical activity broadly agrees on the same set of relevant correlates with some rare variations concerning the sign of the correlations: W o m e n are less likely to engage in physical activity than men are, and the type of sports is gender specific (e.g. Breuer/Wicker 2 0 0 8 ; D o w n w a r d 2007; Lechner/Downward 2013). Studies tend to agree on the fact that age has a non-linear effect on physical activity but there is n o consensus on the exact f o r m of the relationship. Part of the literature considers participation to increase with lower age i.e. to decrease after youth (e.g. D o w n w a r d et al. 2 0 1 1 ; Eberth/Smith 2010; Garcia et al. 2 0 1 1 ; Humphreys/Ruseski 2010). The same 9 10

Stempel (2005) tests a total of 15 different sports. Given the empirical results in Sect. 3.3 and 3.4 below, it is somewhat surprising that these models do not address the feature that sports may also directly improve skills and thus earnings capacity.

384 • Charlotte Cabane and Michael Lechner

and other studies find an increase of physical activities with older ages (e.g. Garcia et al. 2011; Stamatakis/Chaudhury 2008)). Garcia et al. (2011) define the d e m a n d for physical activities as U-shaped with a minimum at 33 years of age. These results are in line with one of the predictions of the neoclassical models: d e m a n d for leisure time physical activities depends on the opportunity cost of time which varies with the employment status and thus with age. H o w e v e r , age is also closely related to health. Older individuals are likely to be less physically active due to relatively worse health conditions than younger individuals. Kokolakakis et al. (2011) find a negative relationship between age and physical activities. They find regional differences: the differences in the physical activity rates that are due to age are much larger a m o n g English people than a m o n g Spanish people. Wicker et al. (2009) highly recommend to study the physical activity by age band rather than globally in order to properly capture the age-specific physical activity behavioural patterns. For example, older people are more likely to walk while young adults prefer to go to fitness centres. Being married is negatively correlated to engage in physical activity (e.g. Eberth/Smith 2010; Garcia et al. 2 0 1 1 ; Rapp/Schneider 2013). And, w o m e n living in a household with young children are less likely to participate in physical activity (e.g. Eberth/Smith 2010; Garcia et al. 2011). D o w n w a r d et al. (2011) find that parents' participation to physical activity might affect the individual's own participation to physical activity. T h e socio-economic background also plays an important role: higher incomes or earnings and higher level of education increase the probability to engage in sport (e.g. Downward/Rasciute 2 0 1 0 ; Fridberg 2 0 1 0 ; Hovemann/Wicker 2 0 0 9 ; Humphreys/Roseski 2010; Lechner 2 0 0 9 ; Meltzer/Jena 2010). However, higher incomes are associated with a decrease in the a m o u n t of time spent doing sport and an increase in the intensity (or the frequency) of the sport (e.g. D o w n w a r d et al. 2011; Meltzer/Jena 2 0 1 0 ; Taks et al. 1994). This illustrates the concept of opportunity costs of time. In the same line of argument, C o l m a n and Dave (2013a) suggest that a decrease in employment leads to a decrease in total physical activity. Indeed, the decrease in physical exertion (PA done while working) for unemployed individuals is not compensated by an increase in their leisure time physical activity. In western countries, belonging to specific ethnic minorities m a y be associated with less sports participation (e.g. Lechner 2009) as well. An explanation for this finding can be f o u n d in the heterodox theories in which the social environment has an important influence on the physical activity. Finally, individuals' health is also related to their sports participation as p o o r health reduces participation to physical activities (e.g. Bauman et al. 2002). Participation in physical activity is negatively associated with smoking but positively associated with drinking ( D o w n w a r d 2007). 2.2.2 Weather conditions (and daylight)

Weather conditions can prevent individuals to engage in physical activity by decreasing (increasing) the utility derived of being physically active. W i t h a m et al. (2014) investigate the impact of day light and weather conditions on the physical activity of older people in Scotland (PA measured using accelerometers). They find a small positive relationship between day length a n d daily physical activity (1 hour more of day light leads to 1 . 5 % more physical activity) as well as between an increase in the m i n i m u m temperature and daily physical activity (1 additional degree Celsius translates into 0 . 9 % additional physical activity). However, the authors d o not have information on where the physical activity

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 385

takes place (indoor or outdoor) and the results are, of course, specific to the particular population. Eisenberg and Okeke (2009) analyse the impact of unexpected changes in weather conditions on LTPA in 48 American states between 1993 and 2 0 0 0 . They find that a decrease in low range temperatures ( < 1 5 . 5 ° C ) is related to a decrease in LTPA (a d r o p of 3 ° C is associated with a d r o p of the level of physical activity of 0 . 6 % ) . They emphasise the fact that individuals with different socio-economic status (SES) have different levels of elasticity t o w a r d s LTPA. Indeed, individuals with lower SES appear to be more sensitive to changes in temperatures, which can be explained by the fact that they are less able to substitute o u t d o o r with indoor LTPA. The authors venture some further explanations according to which transport costs, schedule constraints, and preferences are different a m o n g SES. This is very relevant in terms of policy implications because it emphasises the importance of population targeted interventions. 2.2.3 Neighbourhood characteristics The characteristics related to the individual's environment are often presented as potential determinants of physical activity. Indeed, a substantial n u m b e r of studies analyse the relationship between the presence of green spaces and sports infrastructures on individuals' physical activity. M a n y of them differentiate between the perceived environment (i.e. individual's perception of the environment) and more objective (external) measures of the environmental characteristics. Perceived or factual, the environment is likely to have an impact on either the monetary costs of being physically active or on the level of utility derived f r o m being active. D u n c a n et al. (2005) and Kaczynski and H e n d e r s o n (2008) survey papers using the perceived environment and conclude that it is related to physical activity but only to a small extend. 4 0 % of the papers surveyed in Kaczynski and H e n d e r s o n (2008) report a statistically significant association " t h a t is entirely or primarily positive" and according to D u n c a n et al. (2005) perceived environment explains 4 to 7 % of the variation in physical activity. Using the N e i g h b o u r h o o d Environment Walkability Survey (NEWS) measured near San Diego, Saelens et al. (2003) argue that high-walkability environments positively affect individuals' physical activity (measured using an accelerometer). H u s t o n et al. (2003) note that if some specific neighbourhood characteristics (e.g. presence of trails) are associated with higher levels of physical activity, the perceived environment and access to places for physical activity are highly correlated with race, education, and income (study based on N o r t h Carolina data). People reporting less favourable environments and less access are more likely to be Blacks or American Indians with lower income and education (Huston et al. 2003). Several studies use measures of environmental characteristics either collected by the research team itself Wicker et al. (2009, 2013); Kaczynski et al. (2008); Kumagai (2013) or f r o m administrative data (Humphreys/Ruseski, 2 0 0 7 ; Richardson et al., 2013). Kumagai (2013) and Richardson et al. (2013) associate the presence of green spaces or public sports infrastructures with a higher level of health (which is not entirely mediated by physical activity according to Richardson et al. 2013). They use the Japanese General Social Survey and the N e w Zealand Health Survey respectively. In their study about O n t a r i o (Canada), Kaczynski et al. (2008) argue that the distance to parks and the size of parks are not related to an increase in physical activity, contrary to the presence of a trail or facilities (rather than amenities). H u m p h r e y s and Ruseski (2007) highlight a positive relationship between governmental spending on parks and o u t d o o r physical activity

3 8 6 • Charlotte Cabane and Michael Lechner

(Behavioural Risk Factor Surveillance System, US data). Wicker et al. (2009) and Wicker et al. (2013) insist on the specificity of each age g r o u p concerning their use and d e m a n d for sports specific infrastructures in Germany (Stuttgart and M u n i c h respectively). Wicker et al. (2009) pinpoint the infrastructures that matter by age: swimming pools and playground areas for the 3 - 1 8 years old, diverse sports facilities for the 1 9 - 2 8 years old, 11 fitness centres for the 2 9 - 3 5 years old, none for the 36—44 and 4 5 - 6 4 years old, a n d forests for the 65+ years old. There is some literature on the impact of the vicinity of sports infrastructures a n d sports clubs on children sports participation (e.g. Reimers et al. 2014; Steinmayr et al. 2011). However, few studies focus on adult participation. O n the one h a n d , it is problematic to assume that adults choose their place of leaving independently of the infrastructures available in the area. O n the other h a n d , many physical activities such as walking, cycling, a n d hiking do not require formal facilities which may be available publicly or in sports clubs. Finally, the current insufficient information on sports clubs and sports infrastructures in general in m a n y countries prevents researchers to investigate these questions more extensively. T o sum up, there is a consensus on the fact that parks and sports infrastructures are positively related to physical activity, although this relation appears not to be t o o relevant. Furthermore, the literature sheds light on t w o important points: the perception of the environment and the requirements in terms of facilities depend on individuals' SES and on age. In particular, Wicker et al. (2009) underline the relevance of sports facilities for very young adults and the relevance of fitness centres for the 2 9 - 3 5 years old. Saelens and H a n d y (2008) provide an interesting survey a b o u t surveys a n d studies on that topic. They underline the fact that if the use of objective measures of the environment leads to an improvement, there is still w o r k to d o concerning the potential substitution effect between transportation walking (walking for commuting) and other forms or physical activity. They also argue that recent studies are able to associate pedestrian infrastructure with recreational walking but not with walking for commuting. Another aspect of the neighbourhood, which could also have an impact on the participation to physical activity, is safety. Janke et al. (2013) analyse the impact of violence on physical activity in England by using a difference-in-difference f r a m e w o r k (DiD) based on pooled cross-sections of 0.9 million individuals observed quarterly over 6 years (22 periods) in 3 2 3 local authorities. They argue that self-reported physical activity (over the last 4 weeks) is negatively correlated with the quarterly rate of violent crimes with injury (recorded by the police in the last 4 quarters before the interview). They also use the Riots in August 2011 1 2 as a natural experiment and still find a deterring effect for w o m e n . H o w ever, surprisingly, they find the opposite effects for men. According to them, the men's answer to local crime is to " m a n u p " by going out more and exercise more. 1 3 Caruso (2011) studies the interactions between sports a n d crime rates in Italy. H e finds negative correlations between sports participation and the crime rate concerning property and juvenile crime but a positive correlation between sports participation and violent crime. Finally, there are features of the individuals' surroundings determined at the aggregate level. H u m p h r e y s et al. (2012) suggest that government's sports policies and success of

11

12

13

The authors explain that this age group is more likely to participate in regular sport activity that requires a supply of gymnasia, sports fields, public playgrounds and fitness centres. Six days of riots happened after a man was shot dead by the police in London. The riots spread very quickly to other cities and towns in England and led to looting, arson, and a mass deployment of police. This explanation holds for the riots and not for the regular quarterly variation in local crime rate.

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 387

the national team could also influence individual's physical activity in a given country. In a study using a cross section of 34 countries, they analyse the impact of the success of the national team at the Olympics (in Athens) and of hosting sports mega-events on individuals' physical activity. According to their results, both are negatively correlated with the participation in physical activity. Their interpretation is that elite sports have ousted mass sports participation in terms of provision of resources and use of sports facilities (in countries hosting sports mega-events or being successful at the Olympics). The authors also investigate the role of economic freedom (using the related index) and the position of w o m e n in the society (using female labour force participation rates and w o m e n ' s suffrage year). They argue that institutional structures that favour gender equality and economic freedom are positively related to higher individuals' participation to physical activity. These findings are supported by Kokolakakis et al. (2011) w h o argue that although sports participation correlates appear to be the same for England a n d Spain, their influence is country specific and due to differences in institutions and culture. Indeed, education is a more important driver in Spain than in England and the impact of gender in Spain is twice its impact in England. 1 4 Conversely, age appears to matter more in England than in Spain meaning that sports participation in England decreases faster with age. This result is in contradiction with Garcia et al. (2011) w h o find an increase of physical activity with age (after the age of 33 years old) in Spain. Kokolakakis et al. (2011) also underline other differences that they consistently interpret as consequences of the culture a n d institutions. 1 5 2.2.4 Peers There is a growing recent literature on peer effects. The theoretical link between the individual d e m a n d of physical activity and peers can be established by extending the concept of a household in the context of the N e w Household Economics (NHE), or by relying on the 'heterodox theory'. According to the N H E , the consumption of physical activity is easier if another member of the individual's household already participated in such physical activity, for example. If the individual's peers are considered as part of an extended household a n d are physically active they directly impact the individual's level of physical activity. ' H e t e r o d o x theory' also takes into account social pressure, social interaction and desire of integration in one's consumption choices. Therefore, if the peers are physically active it is very likely that the individual will also be physically active, and vice-versa. In contrast to the previous studies, m a n y of those papers are based on an experimental design. In many cases, although not in all, this implies that the reported correlates are most likely to have a causal interpretation. Carrel et al. (2011) uses the randomised assignment of college students of the US Air Force Academy to squadrons t o analyse the impact of peers on individual fitness. They find that the effect of the initial fitness levels of the peer is a b o u t 4 0 to 7 0 % as large as the effect of the o w n initial individual fitness level. Furthermore, according to their results, individuals in such groups tend to converge towards their less fit peers.

14

15

Higher levels of education lead to higher sports participation rates and men are more likely to do sport than women are. For example, the fact that being a student (instead of a worker) has a positive impact on PA in Spain but not in England is justified by the fact that the great majority of the students in England work while attending college.

388 • Charlotte Cabane and Michael Lechner

Babcock and Hartman (2010) and Leslie and N o r t o n (2012) find similar results in their experiments: individuals tend to converge towards the lowest individual level of physical activity. In their pay-to-exercise experiment, Babcock and Hartman (2010) find that paidto-exercise college students (i.e. treated students) who have relatively more non-treated friends exercise less than treated college students w h o have relatively more treated friends (University of California). These results suggest an individual who is in a group of peers with a higher number of physically active people than inactive ones will exercise more. Leslie and N o r t o n (2012) randomly assigned people to solo, duo or quintet groups and gave them information about the other members of the group performance (except for solo). They observe that in duos and quintets the physical activity decreases (converges) towards the level of activity of the less active group member. However, Leslie and Norton (2012) highlight the fact that information on the peers' behaviour can be strategically used in order to reach the desirable results (increase physical activity participation for example). In their experiment, they choose to give information on all the group members' performance and they argue that the top-performers prevent the group to decline towards the lowest individual level. Therefore, they suggest that sharing information concerning only the top-performance might have a very different impact on the group members' performances. These two experiments lead to believe that selective information sharing in the peer group and peer group structure (share of physically active members) can be used to increase individuals' physical activity. However, it is important to recall that these are results from experiments conducted on specific samples (i.e. there is no external validity). Johannesson et al. (2010) set up an experiment in a Swedish hospital in order to observe the impact of contests and symbolic rewards on physical activity. The contest is a step contest: individuals have to wear a step counter and report their number of steps for a certain period of time. The best performers (teams or individuals) are entitled to receive a price or to be part of a lottery. Former studies underline the fact that step contests per se increase physical activity participation (e.g. Bravata et al. 2007). For their experiment, Johannesson et al. (2010) build different groups to which they gave different rewards and also different amounts of information with respect to the performance of the others. They find that step contests with symbolic rewards are an even larger incentive to exercise than regular step contests. They do not comment on the potential existence of a peer effect leading to downward convergence. Rapp and Schneider (2013) investigate the different types of partnership between couples on the physical activity and suggest three channels. On the one hand, being in a relationship should be associated with a decrease in the partners physical activity level because of "being released from the pressure of the marriage market" and because both partners experience a reduction in their discretionary time. On the other hand, they argue that the partner can exert social control and social support of healthy behaviours and thus increases her partner's level of physical activity. Using the German Socio-Economic Panel, (SOEP) they find that, for men, the negative correlation between physical activity and marriage becomes positive when they get older (for women the correlation remains negative but decreases with age). In Ruseski et al. (2014), the authors use the "sense of belonging to the community" as an instrument for physical activity arguing that it is positively correlated to individual's physical activity. Wilcox et al. (2000) underline the fact that rural older women report more barriers to leisure time physical activity than urban older women do. They associate it with the lack of physically active role model in rural areas. Finding important differences according to the ethnicity they also argue that cultural norms on physical activity might influence

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 389

individual physical activity. Mutter and Pawlowski (2014) investigate the impact of role models in Germany and find a positive correlation between the success of the national German soccer team and the motivation to do sports for young amateur soccer players (males and females). This result concerns only the motivation to do sports and also it refers only to people who are already engaged in sports participation. 2.2.5 Habits, commitment, and economic incentives Several papers observe the role of commitment and habits of participation in physical activity. The underlying idea is that the demand of physical activity as an investment good depends on time preferences. People with high preference for the present are less likely to engage in physical activity. Habits or commitment directly affects the time preference of the individual by creating external / artificial constraints (et vice versa) and thus affect physical activity. DeliaVigna and Malmendier (2006) use monthly panel data from health clubs over three years to test the predictions of the profit-maximization contract. According to these predictions, the consumer chooses the utility-maximising contract under rational expectations about her future consumption frequency. Those predictions do not correspond to the data: they observe that the individuals buy inadequate contracts in term of length and gym attendance frequency. In order to explain this phenomenon they suggest that people overestimate their future self-control or future efficiency when buying health clubs' contract. This explains why it seems that individuals do not act rationally when they face contracts with immediate costs and delayed benefits. 16 At least three experiments have been done subsequently in order to highlight the link of participation in physical activity and commitment and habits. First, Charness and Gneezy (2009) run two experiments with North-American university students in which they investigate the effects of two different forms of pay-to-exercise incentives. In one experiment the students are paid $100 if they go eight times to the gym within four weeks. In the other experiment the students w h o meet the requirements are paid $175. In this second experiment, for one group the requirement is to go to the gym at least once within a month, for the other group the requirement is to go at least eight times within a month. The results suggest that students w h o were financially encouraged to exercise at least eight times within four weeks - and w h o were not regular gym attendants before - increase their participation to physical activity during the experiment and also during some weeks after the experiment. This increase in physical activity is associated with an improvement in various biometric indicators such as body-fat and pulse rate. The variation in body-fat is more than 2 percentage points lower and the variation of the pulse rate is higher by 5.15 beats per minute for the control group (when comparing with the - intensively treated group). Given that the observed effects tend to be temporary only, Royer et al. (2012) and Acland and Levy (2010) started experiments that focus on the 'after-treatment period' of the previous studies. In other words, they try to understand why people do not continue

16

Goldhaber-Fiebert et al. (2010) study the impact of nudges and anchoring on individual's decision concerning the type of exercise commitment contract by making an experiment in the US. They find that the default values of duration influence the length of the contract especially for first time users (information on financial stake does not). Interestingly, this change in the contract duration does not seem to impact the frequency of the participation to physical activity. However, they do not evaluate compliance with the exercise commitments.

390 • Charlotte Cabane and Michael Lechner

exercising and how to encourage them to do so even when payments ended. Royer et al. (2012) set up a structure in which an own self-funded commitment contract has been coupled with an initial incentive programme (in a Fortune 500 company in the Midwest of the USA). The own self-funded commitment contract is similar to the pay-to-exercise incentive except that the participant use her own money. The participant chooses the amount of her own money that she will gain (loose) by complying (not complying) with the contract requirements. 17 They underline the fact that this type of joint intervention does improve the long-term effect of the pay to exercise programme, even if individuals might have issues with time consistency in a first place. Acland and Levy (2010) build an experiment similar to the one proposed by Charness and Gneezy (2009) but with a much longer follow-up period and focusing exclusively on people who do not regularly attend the gym before the experiment (university students from Berkeley). They also model habit formation including naivety arguing that the individuals are biased concerning their habit formation and also naive with respect to this bias. In other words, the students overestimate their change of behaviour towards physical activity and they are overoptimistic concerning the realisation of their predictions. This literature allows us to gain a better understanding of individuals' decision to engage in physical activity and the impact of commitment and structure of the physical activity offer. It is worth noting that naivety and over-confidence towards capacity or will to change physical activity behaviours seem to be well-spread among individuals who are not physically active. Therefore, counselling, encouragement and supervision appear to be appealing interventions. Active transportation is a way to engage in physical activity while commuting by walking or cycling, for example. Some studies focus on diverse ways to encourage active transportation (and thus physical activity) and physical activity via different types of incentives. Brockman and Fox (2011) analyses the evolution of active transportation among Bristol university staff between 1998 and 2007 while the university reduced the parking opportunities. They find that during this period - and contrary to national trends - the share of people who reported that walking and cycling to work increased (from 19 to 30% and from 7 to 12% respectively). Furthermore, those 70% of the people walking or cycling to the university achieved more than 80% of the recommended levels of physical activity according to the official guidelines for physical activity by doing so. Davis and Jones (2007) report on the success of incentives given by companies to increase physical activity (and productivity) among their employees. They find evidence in the literature that health promotion programmes and exercise programmes do increase physical activity. Also, physical activity counselling sessions are associated with higher selfreported levels of physical activity and higher observed fitness in the short term. They conclude that employer supported interventions - at the workplace or elsewhere - do increase physical activity level among their employees. Sockoll et al. (2009) review a large number of papers analysing the effectiveness of workplace health promotion and prevention. They highlight the fact that physical activity programmes do increase - to a limited extent - workers' physical activity. Aside to physical exercise courses and counselling, low-cost interventions such as motivating signs for encouraging stair use, initiation of jogging groups or substituting personal visits instead of using a telephone positively influence employees' physical activity.

17

The commitment is fixed to "not missing more than 1 4 days in a row at the gym".

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 391

Sen (2012) is interested in the effect of changes in the gasoline price (due to Hurricane Katrina) on physical activity in the US (using the American Time Use Survey which allows calculating the M E T ) . Indeed, t w o effects can be figured out: the substitution effect (people opt for active transportation rather than motorized one) and the income effect (people stop part of their leisure activities because of budget restrictions). Sen (2012) finds that an increase in gasoline price is associated with an increase in both: participation in physical activity and duration of physical activity. However, the concerned physical activity is moderately energy intense and corresponds to housework. Furthermore, physical activity of individuals with the lowest and highest SESs is not influenced. Therefore, the author concludes that the income effect is nonlinear. While middle SES groups stop hiring housekeepers (domestic help) and do their moderately intensive housework themselves, individuals with low SES did not have any housekeeper, and individuals with high SES can still afford to have some. Therefore, he suggests that taxes on gasoline price are not an efficient way to increase physical activity of all SES groups.

3

Returns of physical activity for working age adults

According to the theories on the individual d e m a n d for physical activity, involvement in physical activities m a y be motivated by a desire to increase utility directly, a desire to maintain and improve health, a desire to build and or at least to signal some specific skills, and/or a desire to be socially integrated. In this section we consider the corresponding outcome variables, beginning with the attempt to measure changes in utility directly by various indicators for individual happiness. Health can also be analysed by different sets of objective and subjective indicators. The presence of a skill related signal is usually tested by job application experiments sending out CV's with and w i t h o u t that information. Skill effects may be approximated by measuring long term labour market outcomes. Social inclusion, however, is a more complex concept that is mainly studied f r o m a qualitative perspective. 3.1

Happiness

Hoecke et al. (2014) implement an experiment in order to evaluate the impact of needsupportive physical activity counselling on physical well-being of Flemish sedentary adults w h o are to be motivated to engage in physical activity. They find a positive relationship, which increase with the level of physical activity. They also conclude that it is relevant for individuals to be advised and supported by experts with respect to their o w n decision concerning physical activity in order to increase their well-being. This explanation relates to the Self-Determination Theory (e.g. Deci/Ryan 2002), which takes into account individuals' self-motivation and self-determination in order to explain their behaviour. Using the Behavioural Risk Factor Surveillance System data (BRFSS), H u a n g and H u m p h r e y s (2012) estimate a positive relationship between participation in physical activity and self-reported life satisfaction. According to their results, this effect is partially mediated by an improvement in health and the overall impact is greater for men. They use an instrumental variable a p p r o a c h for the participation in physical activity in order to c o m p u t e this effect (based on 1.5 million individuals). The instrument is the number of sports facilities in the county where the individual lives. In Ruseski et al. (2014), the authors also have recourse to IV estimations in order to measure the impact of physical activity on well-being for a sample of individuals living in Rheinberg (Germany). They

392 • Charlotte Cabane and Michael Lechner

use two instruments: the distance to sports facilities (computed using geo-codes) and the answer to a question asking whether participating in physical activity is important. They include an indicator of disability in their estimations but no further information on the individual health status. They argue that physical activity has a positive impact on wellbeing. Rasciute and Downward (2010) analyse the link between physical activity and happiness using data from the English Taking Part Survey. They insist on the fact that happiness and health are interdependent and they differentiate the impact of physical activity by type and motivation. In particular, they distinguish cycling and walking from other physical activities. According to them, physical activity for recreation or health does have a positive impact on health and well-being while cycling for the same reasons (or used as a mean of transportation) is negatively correlated to well-being and positively correlated to health. They suggest that the disutility of cycling might be due to safety issues on the roads. 3.2

Health

In this section, we survey papers analysing the effect of physical activity by health outcomes: specific conditions (e.g. diabetes mellitus, cardiovascular diseases - CVD henceforth), health costs, and self-reported health status. Most of the studies look at short terms effect and are able to compute the metabolic equivalent task (MET). The M E T values are used to build categories such as inactive, moderately and active for example. 18 The positive impact of physical activity on health is well established in the medical literature (e.g. Warburton 2006), in the sports science literature (e.g. Reiner et al. 2013), as well as in economics (e.g. Humphreys et al. 2014; Sari 2009, 2014). The surveys on the impact of physical activity on health find evidence of the existence of a positive relationship (e.g. Hillman et al. 2008; Reiner et al. 2013; Shephard 1996; Sockoll et al. 2009; Warburton 2006). According to Warburton (2006), physical activity is effective in the primary and secondary prevention of a substantial number of physical conditions (e.g. CVD, diabetes mellitus, osteoporosis, breast and colon cancer)." They also underline the fact that if the relation between physical activity and health is linear, the highest improvement in health is observed when people w h o are least fit become physically active. Hillmann et al. (2006) review studies which analyse the relationship between physical activity (and more specifically aerobic fitness training) and brain function and cognition. They find positive results but underline the lack of precision with respect to the optimal design of the exercise intervention (e.g. type, duration, and schedule over the lifespan). Sari (2009, 2014) uses Canadian data (Canadian Community Health Survey -CCHS- and National Population Health Survey -NPHS- respectively) which allow him to calculate the M E T and analyses the corresponding healthcare use. He finds that being moderately active rather than inactive has a greater impact in reducing healthcare use than being active rather than moderately active. More precisely, the length of a hospital stay decreases by 35 to 4 1 % (Sari 2014). The impact is larger for women and also for people w h o have a chronic disease (i.e. diabetes, heart disease, cancer; stroke or high blood pressure). Results 18 19

The use of 3 categories is the most c o m m o n in the presented studies. Primary prevention is health promotion and specific protection while secondary prevention consists in dealing with latent disease and preventing progression of disease (or preventing asymptomatic disease to become symptomatic).

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 393

concerning outpatient services are similar: inactive individuals use 12% more nurse services, 5 % more family physicians, and 13% more physician services than active people (Sari 2009). Furthermore, moderately active people compared to active people use 2.4 to 9.6% more physician (family and other respectively) services. Dunnagan et al. (1999) study the impact of work site based fitness programming on health-related costs. Using the University of Kentucky Wellness Program (UKWP) they build an experiment in order to compare individuals w h o are part of the program with individuals w h o are not. They argue that regular exercise reduces health care costs. Unfortunately, their sample contains less than hundred individuals and suffers from selection (selection of the participants and attrition). Humphreys et al. (2014) look at the impact of physical activity on the selfreported health status and several chronic conditions using the CCHS. They find the same pattern as Sari (2009) and Sari (2014): being moderately active rather than inactive is more rewarding than being active rather than moderately active. This results hold for the level of self-reported health as well as for diabetes, arthritis and high blood pressure. Colman and Dave (2013b) analyse the same outcomes using the National Health and Nutrition Examination Survey (NHANES1). They are not able to compute the M E T but they take into account the total physical activity as well as recreational physical activity instead of focusing on the latter like many other papers. According to their results, both types of physical activity have a protective effect on health via a reduction of the risk factors (BMI, high blood pressure and resting heart rate). A decrease in high recreational exercise and other physical activities explains 10 to 3 3 % of an increase in BMI and hypertension, 2 to 8% increase in diabetes and heart disease and 10 to 2 0 % in the increase of noted risk factors and illness conditions. In addition, Schulkind (2013) uses the so-called Title IX to evaluate the impact of physical activity on the intergenerational transmission of health. Title IX is a law amendment introduced in 1972 in the US. Its main purpose was to prohibit gender discrimination in federally funded activities such as education. Its compliance required a substantial increase in the supply of female high school and collegiate athletics. Schulkind (2013) uses the increase in female physical activity required to comply with Title IX regulations as an instrument and performs IV in order to estimate the impact of an increase in women's athletic participation (instrumented) on her child's birth weight. According to her results, an increase by 2 0 % in girls' athletic participation in 1970's translates into 6 (12) %-points reduction in low (very low) birth weight births. She attributes this effect to a change in the mothers' physical activity behaviours. The literature provides evidence supporting the hypothesis according to which physical activity improve health, self-reported health, and well-being. It is worth noting that most studies suggest that greater short term effects on health come from an increase from inactive to moderately active. By contrast, Lechner and Sari (2014) - who focus on long term effects - find that only an increase from moderately active to active has a positive and significant impact on health. The next relevant question concerns the transferability of the positive health effects of physical activity to the individual's human capital and labour market productivity. There is a consensus on the fact that more physical activity is always better for health, therefore individuals should spend more time engaging in physical activity. 20 Since the time is a limited resource dedicating more time to physical activity requires an adjustment (reduction) 20

This is a simplification of reality. Indeed, as mentioned before physical activity is also defined in terms of intensity suggesting that duration and intensity are substitutes to some extend (WHO

394 • Charlotte Cabane and Michael Lechner

of either leisure or work time. This change in the allocation of time has consequences on earnings. However, spending more time investing and maintaining health also leads, for example, to fewer sick days, which in turn increases the amount of time available for work and leisure. Moreover, physical activity may also build other skills (cognitive and non-cognitive skills) and facilitate social inclusion (e.g. Seippel 2 0 0 6 ; Frost et al. 2 0 1 3 ) . Non-cognitive skills such as team skills, self-discipline, and tenacity are associated with sports participation and higher level of productivity and thus, higher earnings. Thus, we continue this review by presenting papers on labour-market outcomes and social capital. 3.3

Labour market outcomes

In this section, we present empirical economic papers analysing the relationship between LTPA and labour market outcomes. The literature on this topic is limited, mainly because the data needed to analyse this topic is relatively poor. Indeed, while positive health effects are very well established by the literature, the evidence on labour market effects is more limited. We start with presenting the results concerning employability followed by the effects on the earnings and wages. 21 Rooth (2011) performs an experiment in Sweden and finds that the applicants who include a statement about being active in sports in their job application increase their call-back rate by 2%-points. The design of the experiment is such that more than 8 0 0 0 applications were sent to 3 8 2 1 employers in different sectors for 13 different occupations with different skills requirement and degree of costumer contact. In a non-experimental setting, Cabane (2014) uses the German Socio-Economic Panel (SOEP) to investigate the impact of physical activity on unemployment duration. She finds a positive correlation between physical activity (at least once a week) and exit from unemployment to employment for women who have at least 3 years of working experience. However, she argues that it might reflect lower psychological barriers to job search (such as bounded self-control) rather than an actual effect of being physically active. Another part of the literature on physical activity and earnings focuses on the short and long run effects of adults' participation to physical activity on their current earnings. Cornelissen and Pfeifer (2008) perform an analysis using the SOEP and a random effect regression strategy. They find that men who practise sport at least weekly earn 5 % more than men who do not (around 3 % more than men who do participate but less often). They test the impact of youth sport on earnings and find significant results. Women who declare having been involved in sports competition when they were 15 years old earn about 6 % more. Lechner (2009) has similar results using the same database (the SOEP) but adopting a different strategy (semi-parametric matching estimation using informative panel data in a specific way). He finds an increase by 1 2 0 0 euros p.a. over a 16-year period for adults who practise sports at least monthly (with respect to physically inactive or less active people). Lechner and Downward (2013) estimate the gain of different sports on annual household income between 4 3 0 0 and 6 5 0 0 GBP p.a. for 2 6 to 4 5 years old men and between 3 4 0 0 to 5 3 0 0 GBP p.a. for women of the same age (it varies according to type of sport). They find that for men outdoor sports and then fitness sports appear to have the highest association to earnings while it is racquet sports and then team sport for

21

guidelines). However, the material costs are different and the risk of injury higher, not everyone can use / benefit from such substitution. The literature on the characteristics of the jobs performed by physically active people focuses on adolescent physical activity and is therefore excluded from this survey.

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 395

young w o m e n (26 to 4 5 years old), and o u t d o o r sports for older w o m e n (46 to 64 years old) showing lower positive associations. Lechner and Sari (2014) analyse the impact on the earnings of Canadian adults changing f r o m being inactive to being moderately active and f r o m being moderately active to active based on an informative and long Canadian health panel. The method used is the same as in Lechner (2009). However, the data is m u c h more informative and therefore the identification strategy which relies on the conditional independence assumption together with a type of semi-parametric fixed effects approach is more credible. In Lechner and Sari (2014) it appears that the change f r o m inactivity to a moderate level does not lead to a significant increase in the earnings over time. However, the increase f r o m a moderate to more active level positively affects earnings by 10 to 2 0 % in the longer run ( 8 - 1 2 years). Kosteas (2012) analyses the same question and runs matching estimations on the NLSY79. According to his results, practising sport at least weekly increases the wage of 33 to 4 1 years old men and w o m e n by 6 to 1 1 % , respectively. In their report on "PA, absenteeism and productivity", Davis and Jones (2007) analyse the success of the firms as providers of incentives for physical activity. They find evidence that absenteeism can be substantially decreased (and physical activity increased) by introducing workplace health p r o m o t i o n programmes in which individuals commit for at least 12 m o n t h s and also by offering workplace exercise intervention programmes (interventions above 1 hour per week can decrease absenteeism by one third to one half). Effects are larger for people w h o are inactive. There is another channel (in addition to health and h u m a n capital) which could explain the link between success on the labour m a r k e t and physical activity: physical attractiveness. Indeed, several papers show that physical attractiveness (or beauty) is positively correlated to earnings (e.g. French 2 0 0 2 ; Hamermes/Biddle 1994; Mobius/Rosenblat 2006) a n d also to employment (Pfeifer 2011). Sports participation is related to physical attractiveness in many ways. First, it is generally admitted that healthier people 'look better' in terms of skin and body shape (fitness). Second, according to the Eurobarometer 4 1 2 (Eurobarometer 2014) 2 4 % of the people are engaged in physical activity in order " t o improve [their] physical appearance". This is the fifth most cited reason joint with " t o control your weight" which is also about physical appearance. This channel is rarely put f o r w a r d and to our knowledge the relationship between physical activity and physical attractiveness has not yet been studied in economics. As discussed in Sect. 2, the reasons to engage in physical activity might vary with a large set of individual characteristics. It would therefore be interesting to investigate the heterogeneity of the effect of being physically active on the labour market according to educational background, gender or age for example. Gender heterogeneity is the most commonly studied. 3.4

Social capital

As outlined before, social environment and peers influence individuals' participation in physical activity. Indeed, a substantial share of the physical activity involves social interactions, either directly like in team sports, or indirectly w h e n individual sports are done in sports centres for example. Therefore, it can be a way to socialize. Also, the sport world is supposed to be discrimination free and its rules are usually in line with accepted rule of social behaviour. In other words, sports clubs and centres are open to everybody and sports rules p r o m o t e collaboration, solidarity, fair play etc. This suggests that

396 • Charlotte Cabane and Michael Lechner

sports could be used to integrate people and improve social inclusion. However, according to Stempel (2005), the d o m i n a n t class w o r k s on differentiating itself f r o m the other and uses sport for that purpose. If this is the case, physical activity cannot be used to improve social inclusion a m o n g different SES groups. Nevertheless, rightly or wrongly sports is celebrated for its inclusive values, but, to the best of our knowledge, most of the literature on this topic is qualitative in nature. 2 2 In Krouwel et al. (2006) and in Frost et al. (2013) the authors use the opinions of survey respondents a b o u t sports and social inclusion in the Netherlands and in Australia, respectively. Krouwel et al. (2006) focus on the question of minorities and different ethnic groups. They ask four different social groups in R o t t e r d a m a b o u t their preferences and motivation for physical activity (mainly w h y they d o it and with w h o m they w a n t to do it). They conclude that, at least for adults' leisure time, physical activity is not necessarily the best method to increase integration of the minorities. Indeed, according to the survey, while young individuals enjoy being in mixed ethnicity sport clubs, adults prefer not to be mixed. Frost et al. (2013) analyse the role of rural football clubs in social inclusion of population w h o are living in remote Australian rural areas (in Victoria). Their data come f r o m reports presented to the Parliament of Victoria Rural and Regional Services and Development Committee in 2 0 0 4 at the occasion of the Inquiry into C o u n t r y Football. The reports claim that rural football clubs are beneficial to social inclusion in these areas. Delaney and Keaney (2005) and Seippel (2006) analyse the link between involvements in sports and trust in society. The underlying question is whether the values associated with sports participation (such as solidarity, responsibility, and trust) translate into political or societal involvement. Seippel (2006) uses data f r o m the N o r w e g i a n part of the J o h n s H o p k i n s Comparative N o n - p r o f i t Sector Project (a survey) to study the impact of participating in voluntary sport organizations - a m o n g other organizations - on trust and political commitment. H e argues that being a member of a voluntary sport organization is positively related to generalized trust, political interest and voting, but not to politicians' trustworthiness. Delaney and Keaney (2005) c o m p u t e descriptive statistics using several data sets (European Social Survey 2 0 0 2 , 2 0 0 0 UK H o m e Office Survey, 2 0 0 0 Time Usage Survey and polls f r o m the website M O R I ) in order to analyse the situation in the UK and also to characterise the position of the UK within the European countries. They emphasise the fact that after controlling for individuals' socio-demographic characteristics, sports participation is closely related to political trust, well-being and the frequency of socializing and meeting with friends. They suggest using sports in order to build up community networks and relationships. Both papers find a positive relationship between being engaged in sport (or sports related activities) and political trust but none is able to empirically identify a causal relationship. In summary, it is not clear yet which kind of effects physical activity has on social capital and trust a m o n g adults. Indeed, the studies presented here d o not plausibly identify causal relationships and differ in their conclusions.

22

An entire segment of the literature is dedicated to sports and violence but also sport and social inclusion among youth but since the focus of this paper is on adults we do not present this literature here (e.g. Hartmann/Depro 2 0 0 6 ; Parker et al. 2013).

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 397

4

Conclusion

This paper reviewed parts of the vast literature on the effects, determinants, and correlates of physical activity. Although, there seems to be a strong consensus in the literature that, with very few exceptions, being physically active improves almost all aspects of life, still many interesting questions remain unanswered. Let us take the example of labour market outcomes to consider a couple of open issues. The first issue is h o w exactly the different types of physical activities affect the different labour market outcomes (e.g. 1 h of cycling v.s. 1 h of football). This issue is also related to the question on h o w the effects come about. Which part is due to a health effect, which is due to increased physical attractiveness, additional social capital and so on? The current studies are often d o o m e d to treat these issues mainly as a black box to obtain an estimate of the overall effect. Indeed, it is very difficult to find existing data large and informative enough to be able to answer such important questions. In a similar vein, while the epidemiological literature seems to indicate that more physical activity is always better, this is not an appealing point of view for an economist. Since time is a limited resource, time devoted to physical activity cannot be devoted to other activities that also increase current or future utility (and consumption possibilities). Thus, since at some level the returns to physical activity should become smaller (jogging 24h a day cannot be healthy), it will usually not be optimal to live a life that consists only of sleeping, eating, and daily 16h w o r k o u t s . Unfortunately, close to nothing is k n o w n so far where this o p t i m u m is and h o w it varies with individual characteristics, like education and time preference rates, for example. T o understand such issues, we expect that developing theoretical models explicitly taking into account the dynamic nature of consuming and investing in physical activity will shed more light on mechanisms, expected heterogeneities, and (heterogeneous) optimal behaviour. Finally, developments in the dimensions outlined above should f o r m the basis for any reliable and robust investigation of the large a m o u n t of public subsidies spent and their returns to society (or the taxpayer) in general. 23 Overall, this indicates that the link between sports and labour economics is an under researched field which would benefit f r o m additional high quality data and the (hopefully) resulting reliable empirical studies as well as f r o m a p r o f o u n d theoretical analysis, and of course, the metamorphosis of the two. W e like to conclude that the large positive effects of physical activity with respect to almost all dimensions analysed (social integration may be the exception), in particular labour market performance, appears to indicate that increasing physical activity a m o n g workers and unemployed m a y offer a yet not fully explored avenue to raise their productivity, and thus make individual unemployment, for young as well as for older individuals, less likely.

23

For example, Pawlowski and Breuer (2013) estimate the annual level of (net) public subsidies directed to the sports sector in Germany to be around 8 bn EUR.

398 • Charlotte Cabane and Michael Lechner

References Acland, D., M . Levy (2010), Habit formation and naivete in gym attendance: evidence f r o m a field experiment. WP, mimeo. Ainsworth, B.E., W . Haskell, M . L . Whitt, M.L. Irwin, A.M. Swartz, S.J. Strath, W.L. O'Brien, D.R. Bassett, K.H. Schmitz, P.O. Emplaincourt, D.R. Jacobs, A.S. Leon (2000), C o m p e n d i u m of physical activities: an update of activity codes and M E T intensities. Medicien & Science in Sports & Exercise 32(9): 4 9 8 - 5 0 4 . Babcock, P.S. J.L. H a r t m a n (2010), N e t w o r k s and workouts: Treatment size and status specific peer effects in a randomized field experiment. W o r k i n g Paper 16581, National Bureau of Economic Research. Bauman, A.E., J.F. Sallis, D.A. Dzewaltowski, N . O w e n (2002), T o w a r d a better understanding of the inuences on physical activity: The role of determinants, correlates, causal variables, mediators, moderators, and confounders. American Journal of Preventive Medicine 23(2, Supplement 1): 5 - 1 4 . Becker, G. (1965), A theory of the allocation of time. The Economic Journal: 4 9 3 - 5 1 7 . Becker, G. (1974), Crime and punishment: An economic approach. Pp. 1 - 5 4 in: Essays in the Economics of Crime and Punishment, UMI. Bravata, D . M . , C. Smith-Spangler, V. Sundaram, A.L. Gienger, N . Lin, R. Lewis, C.D. Stave, I. Olkin, J.R. Sirard (2007), Using pedometers to increase physical activity and improve health. Journal of the American Medical Association 298: 2 2 9 6 - 2 3 0 4 . Breuer, C., P. Wicker (2008), Demographic and economic factors inuencing inclusion in the german sport system, a microanalysis of the years 1985 to 2005. European Journal for Sport and Society 5: 33—42. Brockman, R., K.R. Fox (2011), Physical activity by stealth? the potential health benefits of a workplace transport plan. Public Health 125(4): 2 1 0 - 2 1 6 . Cabane, C. (2014), Unemployment duration and sport participation. International Journal of Sport Finance 9(3): 2 6 1 - 2 8 0 . Cabane, C., M . Lechner (2014), Physical activity of adults: A survey of correlates, determinants, and effects. SSRN Scholarly Paper ID 2 5 2 3 3 7 6 , Social Science Research N e t w o r k , Rochester, NY. Caruso, R. (2011), Crime and sport participation: Evidence f r o m italian regions over the period 1 9 9 7 - 2 0 0 3 . The Journal of Socio-Economics 40(5): 4 5 5 - 4 6 3 . Cawley, J. (2004), An economic f r a m e w o r k for understanding physical activity and eating behaviors. American Journal of Preventive Medicine 27(3, Supplement): 1 1 7 - 1 2 5 . Charness, G., U. Gneezy (2009), Incentives to exercise. Econometrica 77(3): 9 0 9 - 9 3 1 . Colman, G., D. Dave (2013a), Exercise, physical activity, and exertion over the business cycle. Social Science 8c Medicine 93: 1 1 - 2 0 . Colman, G.J., D. M . Dave, (2013b), Physical activity and health. W o r k i n g Paper 18858, National Bureau of Economic Research. Cornelissen, T., C. Pfeifer, (2008), Sport u n d arbeitseinkommen - individuelle ertragsraten von sportaktivitaten in deutschland. J a h r b u c h fur Wirtschaftswissenschaften 59: 2 4 4 - 2 5 5 . CSEP (2011), CSEP guidelines' communique. Report, Canadian Society for Exercise Physiology, Ottawa. Davis, A., M . Jones (2007), Physical activity, absenteeism and productivity: an evidence review. Project Report T/102/07. Deci, E., R. Ryan (2002), H a n d b o o k of self-determination research. NY: University of Rochester Press, Rochester. Delaney, L., E. Keaney (2005), Sport and social capital in the united kingdom: statistical evidence f r o m national and international survey data. Dublin: Economic and Social Research Institute and Institute for Public Policy Research, 32. DellaVigna, S., U. Malmendier (2006), Paying not to go to the gym. The American Economic Review 96(3): 6 9 4 - 7 1 9 . D o w n w a r d , P. (2007), Exploring the economic choice to participate in sport: Results f r o m the 2 0 0 2 general household survey. International Review of Applied Economics 21(5): 6 3 3 - 6 5 3 .

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 399

D o w n w a r d , P., A. D a w s o n , T. Dejonghe (2009), Sports Economics: Theory, Evidence and Policy. Routledge. D o w n w a r d , P., F. Lera-Lôpez, S. Rasciute (2011 ), The economic analysis of sports participation. Pp. 3 3 1 - 3 5 3 in: International H a n d b o o k of Sports M a n a g e m e n t , London: Routledge. D o w n w a r d , P., S. Rasciute (2010), The relative demands for sports and leisure in england. European Sport M a n a g e m e n t Quarterly 10(2): 1 8 9 - 2 1 4 . D o w n w a r d , P., J. Riordan (2007), Social interactions and the demand for sport: An economic analysis. C o n t e m p o r a r y Economic Policy 25(4): 5 1 8 - 5 3 7 . D u n c a n , M.J., J.C. Spence, W.K. M u m m e r y (2005), Perceived environment and physical activity: a meta-analysis of selected environmental characteristics. International Journal of Behavioral Nutrition and Physical Activity 2(1): 11. D u n n a g a n , T., G. Haynes, M . N o l a n d (1999), Health care costs and participation in fitness programming. American Journal of Health Behavior 23(1): 4 3 - 5 1 . Eberth, B., M . D . Smith (2010), Modelling the participation decision and duration of sporting activity in Scotland. Economic Modelling 27(4): 8 2 2 - 8 3 4 . Eisenberg, D., E. Okeke (2009), T o o cold for a jog? weather, exercise, and socioeconomic status. The B.E. Journal of Economic Analysis & Policy 9(1). Eurobarometer (2014), Sport and physical activity. Special Survey EB-412, European Commission. French, M . T . (2002), Physical appearance and earnings: further evidence. Applied Economics 34(5): 5 6 9 - 5 7 2 . Fridberg, T. (2010), Sport and exercise in denmark, Scandinavia and europe. Sport in Society 13(4): 5 8 3 - 5 9 2 . Frost, L., M . Lightbody, A. Halabi (2013), Expanding social inclusion in community sports organizations: evidence from rural australian football clubs. W P 31/13, M o n a s h University. Garcia, J., F. Lera-Löpez, M.J. Surez (2011), Estimation of a structural model of the determinants of the time spent on physical activity and sport evidence for spain. Journal of Sports Economics 12(5): 5 1 5 - 5 3 7 . Goldhaber-Fiebert, J.D., E. Blumenkranz, A.M. Garber (2010), Committing to exercise: Contract design for virtuous habit formation. Working Paper 16624, National Bureau of Economic Research. Grossman, M . (1972), The demand for health: A theoretical and empirical investigation. NBER. H a m e r m e s h , D.S., J.E. Biddle (1994), Beauty and the labor market. The American Economic Review 84(5): 1 1 7 4 - 1 1 9 4 . H a r t m a n n , D., B. Depro (2006), Rethinking sports-based community crime prevention a preliminary analysis of the relationship between midnight basketball and u r b a n crime rates. Journal of Sport & Social Issues 30(2): 1 8 0 - 1 9 6 . Haskell, W., I.-M. Lee, R. Pate, K. Powell, S. Blair, B. Franklin, C. Macéra, G., Heath, P. T h o m p s o n , A. Bauman (2007), Physical activity and public health: Updated recommendation for adults f r o m the american college of sports medicine and the american heart association. Circulation 116: 1 0 8 1 - 1 0 9 3 . H e c k m a n , J.J. (1993), W h a t has been learned a b o u t labor supply in the past twenty years? The American Economic Review 83(2): 1 1 6 - 1 2 1 . HEPA (2013), Health-enhancing physical activity. Grundlagendokument, Bundesamt für Sport und Bundesamt für Statistik. Hillman, C.H., K.I. Erickson, A.F. Kramer (2008), Be smart, exercise your heart: exercise effects on brain and cognition. N a t u r e Reviews Neuroscience 9(1): 5 8 - 6 5 . Hoecke, A.-S.V., C. Delecluse, J. Opdenacker, F. Boen (2014), Year-round effectiveness of physical activity counseling on subjective well-being: A self-determination approach a m o n g emish sedentary adults. Applied Research in Quality of Life 9(3): 5 3 7 - 5 5 8 . H o v e m a n n , G., P. Wicker (2009), Determinants of sport participation in the european union. European Journal for Sport and Society 6: 5 1 - 5 9 . H u a n g , H., B.R. H u m p h r e y s (2012), Sports participation and happiness: Evidence f r o m US microdata. Journal of Economic Psychology 33(4): 7 7 6 - 7 9 3 .

400 • Charlotte Cabane and Michael Lechner

Humphreys, B.R., K. Maresova, J.E. Ruseski (2012), Institutional factors, sport policy and individual sport participation: an international comparison. WP 2012-01. Humphreys, B.R., L. McLeod, J.E. Ruseski (2014), Physical activity and health outcomes: Evidence from canada. Health Economics 23(1): 33-54. Humphreys, B.R., J.E. Ruseski, (2007), Participation in physical activity and government spending on parks and recreation. Contemporary Economic Policy 25(4): 538-552. Humphreys, B.R., J.E. Ruseski (2010), The economic choice of participation and time spent in physical activity and sport in canada. WP 2010-14. Humphreys, B.R., J.E. Ruseski (2011), An economic analysis of participation and time spent in physical activity. The B.E. Journal of Economic Analysis & Policy 11(1) Huston, S.L., K.R. Evenson, P. Bors, Z. Gizlice (2003), Neighborhood environment, access to places for activity, and leisure-time physical activity in a diverse north Carolina population. American Journal of Health Promotion 18(1): 58-69. Janke, K., C. Propper, M.A. Shields (2013), Does violent crime deter physical activity? WP 7545, IZA Discussion Paper. Johannesson, M., R. Östling, E. Ranehill (2010), The effect of competition on physical activity: A randomized trial. The B.E. Journal of Economic Analysis ÔC Policy 10(1). Kaczynski, A.T., K.A. Henderson (2008), Parks and recreation settings and active living: a review of associations with physical activity function and intensity. Journal of Physical Activity and Health 5: 619-632. Kaczynski, A.T., L.R. Potwarka, B.E. Saelens (2008), Association of park size, distance, and features with physical activity in neighborhood parks. American Journal of Public Health 98(8): 1451-1456. Kokolakakis, T., F. Lera-Lôpez, T. Panagouleas (2011), Analysis of the determinants of sports participation in spain and england. Applied Economics 44(21): 2785-2798. Kosteas, V.D. (2012), The effect of exercise on earnings: Evidence from the NLSY. Journal of Labor Research 33(2):225-250. Krouwel, A., N. Boonstra, J.W. Duyvendak, L. Veldboer (2006), A good sport? research into the capacity of recreational sport to integrate dutch minorities. International Review for the Sociology of Sport 41(2): 165-180. Kumagai, N. (2013), Physical inactivity of workers and its relation to uneven allocation of public sports facilities. DP 598. Lechner, M. (2009), Long-run labour market and health effects of individual sports activities. Journal of Health Economics 28(4): 839-854. Lechner, M., P. Downward (2013), Heterogeneous sports participation and labour market outcomes in england. WP 7690, IZA Discussion Paper. Lechner, M., N. Sari (2014), Labor market effects of sports and exercise: Evidence from Canadian panel data. SSRN Scholarly Paper ID 2444872, Social Science Research Network, Rochester, NY. Leslie, K.J., M.I. Norton (2012), Exercising to the lowest common denominator. WP, mimeo. McConnell, K.E. (1992), On-site time in the demand for recreation. American Journal of Agricultural Economics 74(4): 918-925. Meitzer, D.O., A.B. Jena (2010), The economics of intense exercise. Journal of Health Economics 29(3): 347-352. Möbius, M.M., T.S. Rosenblat (2006), Why beauty matters. The American Economic Review 96(1): 222-235. Mutter, F., T. Pawlowski (2014), Role models in sports can success in professional sports increase the demand for amateur sport participation? Sport Management Review 17(3): 324-336. Parker, A., R. Meek, G. Lewis (2013), Sport in a youth prison: male young o_enders' experiences of a sporting intervention. Journal of Youth Studies 17(3): 381-396. Pawlowski, T., C. Breuer (2013), Sport und öffentliche Finanzen: Die sportbezogenen Einnahmen und Ausgaben öffentlicher Haushalte in Deutschland. Wiesbaden: Springer Gabler. Pfeifer, C. (2011), Physical attractiveness, employment and earnings. Applied Economics Letters 19(6): 505-510.

Physical Activity of Adults: A Survey of Correlates, Determinants, and Effects • 401

Rapp, I., B. Schneider (2013), The impacts of marriage, cohabitation and dating relationships on weekly self-reported physical activity in germany: A 19-year longitudinal study. Social Science & Medicine 98: 1 9 7 - 2 0 3 . Rasciute, S., P. D o w n w a r d (2010), Health or happiness? w h a t is the impact of physical activity on the individual? Kyklos 63(2): 2 5 6 - 2 7 0 . Reimers, A.K., M . Wagner, S. Alvanides, A. Steinmayr, M . Reiner, S. Schmidt, A. Woll (2014), Proximity to sports facilities and sports participation for adolescents in germany. PLoS O N E 9(3): e93059. Reiner, M . , C. N i e r m a n n , D. Jekauc, A. Woll (2013), Long-term health benefits of physical activity a systematic review of longitudinal studies. B M C Public Health 13(1): 813. Richardson, E.A., J. Pearce, R. Mitchell, S. Kingham (2013), Role of physical activity in the relationship between urban green space and health. Public Health 127(4): 3 1 8 - 3 2 4 . Rooth, D.-O. (2011), W o r k out or out of w o r k the labor market return to physical fitness and leisure sports activities. Labour Economics 18(3): 3 9 9 - 4 0 9 . Royer, H., M.F. Stehr, J.R. Sydnor (2012), Incentives, commitments and habit formation in exercise: Evidence from a field experiment with workers at a fortune-500 company. Working Paper 18580, National Bureau of Economic Research. Ruseski, J.E., B.R. Humphreys, K. Hallman, P. Wicker, C. Breuer (2014), Sport participation and subjective well-being: Instrumental variable results from german survey data. Journal of Physical Activity and Health 11(2): 3 9 6 ^ 0 3 . Saelens, B.E., S.L. H a n d y (2008), Built environment correlates of walking: A review. Medicine and science in sports and exercise 40(7 Suppl): S550-S566. Saelens, B.E., S.L. Sallis, J.B. Blach, D. Chen (2003), Neighborhood-based differences in PA: an environment scale evaluation. American Journal of Public Health 93: 1 5 5 2 - 1 5 5 8 . Sari, N . (2009), Physical inactivity and its impact on healthcare utilization. Health Economics 18(8): 8 8 5 - 9 0 1 . Sari, N . (2014), Sports, exercise, and length of stay in hospitals: Is there a differential effect for the chronically ill people? C o n t e m p o r a r y Economic Policy 32(2): 2 4 7 - 2 6 0 . Schulkind, L. (2013), Getting a sporting chance: Title IX and the intergenerational transmission of health. W P 13-05, Trinity College Department of Economics. Seippel, R. (2006), Sport and social capital. Acta Sociologica 49(2): 1 6 9 - 1 8 3 . Sen, B. (2012), Is there an association between gasoline prices and physical activity? Evidence f r o m american time use data. Journal of Policy Analysis and M a n a g e m e n t 31(2): 3 3 8 - 3 6 6 . Shephard, R.J. (1996), Worksite fitness and exercise programs: A review of methodology and health impact. American Journal of Health Promotion 10(6): 4 3 6 ^ - 5 2 . Sockoll, I., I. Kramer, W. Boedeker (2009), Effectiveness and economic benefits of workplace health p r o m o t i o n and prevention: Summary of the scientific evidence 2 0 0 0 to 2006. IGAreport 13e. Stamatakis, E., M . C h a u d h u r y (2008), Temporal trends in adults' sports participation patterns in england between 1 9 9 7 and 2006: The health survey for england. British Journal of Sports Medicine. Steinmayr, A., C. Felfe, M . Lechner (2011), The closer the sportier? childrens sports activity and their distance to sports facilities. European Review of Aging and Physical Activity 8(2): 67-82. Stempel, C. (2005), Adult participation sports as cultural capital a test of bourdieus theory of the field of sports. International Review for the Sociology of Sport 40(4): 4 1 1 ^ 1 3 2 . Taks, M . , R. Renson, B. Vanreusel (1994), Of sport, time and money: An economic approach to sport participation. International Review for the Sociology of Sport 29(4): 3 8 1 - 3 9 5 . W a r b u r t o n , D.E. (2006), Health benefits of physical activity: the evidence. Canadian Medical Association Journal 174(6): 8 0 1 - 8 0 9 . W H O (2010), Global recommendations on physical activity for health. Report, W H O Press: Geneva, Switzerland. Wicker, P., C. Breuer, T. Pawlowski (2009), Promoting sport for all to age-specific target groups: the impact of sport infrastructure. European Sport M a n a g e m e n t Quarterly 9(2): 1 0 3 - 1 1 8 .

402 • Charlotte Cabane and Michael Lechner

Wicker, P., K. Hallmann, C. Breuer (2013), Analyzing the impact of sport infrastructure on sport participation using geo-coded data: Evidence from multi-level models. Sport Management Review 16(1): 5 4 - 6 7 . Wilcox, S., C. Castro, A.C. King, R. Housemann, R.C. Brownson (2000), Determinants of leisure time physical activity in rural compared with urban older and ethnically diverse women in the united states. Journal of Epidemiology and Community Health 54(9): 6 6 7 - 6 7 2 . Witham, M.D., P.T. Donnan, T. Vadiveloo, F.F. Sniehotta, I . K . Crombie, Z. Feng, M.E.T. McMurdo (2014), Association of day length and weather conditions with physical activity levels in older community dwelling people. PLoS O N E 9(1): e 8 5 3 3 1 . Charlotte Cabane, Ph.D., Assistant Professor, Universität Sankt Gallen, Swiss Institute for Empirical Economic Research (SEW), Varnbiielstrasse 14, 9 0 0 0 Sankt Gallen, Switzerland. [email protected] www.sew.unisg.ch Prof. Dr. Michael Lechner, Professor of Econometrics and Director of SEW, Universität Sankt Gallen, Swiss Institute for Empirical Economic Research (SEW), Varnbiielstrasse 14, 9 0 0 0 Sankt Gallen, Switzerland. [email protected] www.michael-lechner.eu

Jahrbücherf. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart2015) Bd. (Vol.) 235/4+5

Youth Unemployment in the OECD: The Role of Institutions Andreas Sachs* Centre for European Economic Research (ZEW), Mannheim Werner Smolny University of Ulm JEL E02; E24; J21; J68 Youth unemployment; labor market institutions; age-specific unemployment.

Summary This paper analyzes the role of labor market institutions for youth unemployment, as contrasted to total unemployment. The empirical results are basically consistent with an insider view of labor market institutions. Labor market institutions tend to protect (older) employees but might harm (young) entrants. Remarkable is especially the significant and very high effect of employment protection for regular jobs on youth unemployment. In addition, the combined effects of powerful unions and a coordinated wage bargaining system are beneficial for older people and detrimental to youth. Finally, the paper identifies a significant link between a demographic as well as an educational factor and both youth and total unemployment.

1

Introduction

In the aftermath o f the recent financial crisis, unemployment increased to rates o f more than 2 0 percent in some European countries. Y o u t h unemployment soared to rates of more than 4 0 or 5 0 percent. O n the one hand, this poor labor market outcome is driven by cyclical factors and the general economic conditions. O n the other hand, there is a large literature which highlights the role o f labor market institutions for unemployment. F o r instance, the generosity o f the benefit system or the strictness o f employment protection regulations are blamed for the magnitude and persistence of unemployment. In addition, some countries perform much better than others, both in terms o f total unemployment as well as in terms o f youth unemployment. T h e main contribution of our paper is the macroeconomic analysis of the effects o f labor market institutions on youth unemployment, i.e. people aged 15 to 2 4 , as contrasted to total unemployment or unemployment of people above 2 5 . Firstly, unemployment of youth is an especially important economic policy problem, since phases o f unemployment in the early labor market career are supposed to leave persistent scars and reduce earnings and employability over the entire life cycle (see, for example, Gregg 2 0 0 8 ) . Secondly, the

* W e thank two anonymous referees and the editor of this journal for very helpful comments and suggestions.

404 • Andreas Sachs and Werner Smolny

analysis of youth unemployment, as contrasted to unemployment of older people can yield more information on the impact of institutions on unemployment in general. One basic aspect of the literature on the effect of institutions on unemployment is the inconclusiveness of especially the empirical results (see, for instance, Howell et al. 2007). The analysis of age-specific unemployment provides several advantages. Most important are structural differences between youth and older people in the labor market. Youth can be interpreted to some extent as entrants, who often have to find their first job and have to gain job experience and job-specific human capital. Older people, in contrast, consist mainly of employees with job experience and tenure. It should therefore be expected that some labor market institutions affect youth and older people differently. One important example are employment protection regulations. On the one hand, those regulations protect (older) employees, typically according to tenure. On the other hand, Breen (2005) points out that it hinders the labor market entry of (young) entrants. The overall effect of employment protection on unemployment might be small or difficult to estimate, but the differentiating effect on youth and older people can give more information on the impact of those regulations in general. The advantage of the macroeconometric approach is that it allows cross-country and timeseries analyses of general aspects of the institutional framework. The empirical analysis is based on a panel of 17 OECD countries with annual data from 1982 to 2005. The data on institutions include indicators from five areas, i.e. tax system, bargaining system, employment protection, unemployment benefits, and product market regulation. We test, in addition, for demographic effects (the share of youth in the total population) and include an indicator of the education system (the share of the youth labor force in the youth population). The first indicator relates to the relative size of the entrants cohort and should capture labor supply effects. The second variable is an indicator for the labor market association of the schooling system. The role of the education system for youth unemployment is intensely studied in microeconometric analyses. The contribution of our paper is the macroeconomic cross-country and time-series analyses of the education and demographics, in combination with other labor market institutions. From the methodological side, a Bayesian estimator is employed which takes model uncertainty into account. The empirical analysis of the impact of institutions on unemployment is difficult, since a large number of indicators is available, and theoretical arguments provide only limited support regarding the specification of the model. The Bayesian model averaging approach (Sala-i-Martin et al. 2004) permits to test systematically for the relevance of a large number of explanatory variables within small data sets. The estimates yield inclusion probabilities for each of the variables. Section 2 gives a short review of the literature on the effects of institutions on unemployment, with a focus on youth unemployment. Section 3 discusses the data and the methodological approach. Section 4 presents the estimation results, and section 5 discusses the robustness of the findings. The final section concludes.

2 2.1

Literature Institutions and unemployment

Over the last two or three decades, a plethora of theoretical and empirical contributions focused on various labor market institutions to explain why labor market performance differs both over time as well as across countries. From a theoretical point of view, labor

Youth Unemployment in the OECD: The Role of Institutions • 405

market institutions influence the behaviour of either the labor demand and/or the labor supply side, thus affecting hiring and wage-setting decisions (see Nickel/Layard 1999). With the development of internationally comparable data, factors like the unemployment benefit system, employment protection or the labor tax system have moved at centre stage of empirical macroeconomic studies searching for sources of cross-country differences in labor market performance. Most of these studies focus on unemployment as the target variable since it is well-suited to reflect an economy's ability to avoid involuntary joblessness. Theoretical predictions on the link between a labor market institution and unemployment are partially ambiguous. For example, a generous unemployment benefit system lifts the unemployed's reservation wage, but can improve the job match quality (Holmlund 1998; Acemoglu/Shimer 2000). Further, the coverage and the duration of unemployment benefits can also substantially affect the macroeconomic impact on unemployment. A high level of employment protection is expected to lower both hirings and firings, and it is unclear which effect prevails (Ljungqvist 2002). More recently, the distinction between employment protection for permanent and for temporary contracts as well as the interplay between both aspects has gained attention (see, for instance, Boeri/Garibaldi 2 0 0 7 or Bentolila et al. 2012). Similarly, powerful unions can either negotiate such that the income of insiders is maximized, or that aggregate unemployment is minimized. This also partially depends on the level of bargaining coordination, that is the informal or formal coordination of the bargaining process. Empirical studies try to sort out which theories are most appropriate by estimating panel data models which explain unemployment by various labor market institutions and a set of control factors. Earlier studies tend to find a detrimental impact of labor market regulation (for instance Scarpetta 1996; Nickell 1997 or Blanchard/Wolfers 2000). More recent contributions build upon advances in both data quality and availability as well as methodology, and draw a less clear-cut picture on the influence of labor market regulation on unemployment. By using different specifications and estimators to check the robustness of this link, Baccaro and Rei (2007) state that only union density and unemployment exhibit a robust and significant positive correlation. Similarly, the significance of the findings of Bassanini and Duval (2006) depend to some extend on the chosen specification. In order to ensure robustness, Sachs (2012) applies a model averaging approach and pins down six institutional factors as significantly linked to the evolution of unemployment. A summary of the relevant literature is given in Arpaia and Mourre (2012).

2.2

Institutions and youth unemployment

Much less macroeconometric contributions have dealt with the relation between labor market institutions and different groups of unemployed. From a theoretical point of view, groups divided by age, sex, qualification or migration status exhibit differences in their labor market status or their labor supply decisions (Bertola et al. 2007). Due to these differences, a change in labor market regulation can affect groups differently. For instance, strengthening employment protection or establishing a minimum wage will probably favour older workers with permanent contracts and job experience over younger ones who will be blocked out of the labor market. This depressing effect of minimum wages on the opportunity to gain job experience is highlighted by Gorry (2013). Accordingly, Neumark and Wascher (2004) find that minimum wages reduce the employment

406 • Andreas Sachs and Werner Smolny

rate of the youth population more than of the prime-age population. Regarding job protection, Jimeno and Rodriguez-Palenzuela (2002) argue that young workers have, on average, a lower productivity than older workers. If strict employment protection leads to high firing costs (through severance payments, for instance), less productive young workers become unattractive for employers. However, job protection for temporary contracts can have a different effect since flexible temporary job contracts could lead to high transition rates between employment and unemployment (Blanchard/Landier 2002), and to a crowding out of regular jobs for temporary ones (Kahn 2010). This interplay between employment protection for fixed-term and open-ended contracts is also highlighted by Centeno and Novo (2012) who emphasize the substitutability between both types of contracts. More concretely, increasing protection for permanent contracts raises the relevance of fixed-term contracts for employment adjustments. Given that youth workers more often have fixed-term contracts they are affected more by changes in the level of employment protection. Finally, Bertola et al. (2007) report significant differences in the reaction of youth and prime-aged population rates to changes in the union bargaining power. More specifically, increasing unionization reduces employment rates of the youth population more strongly than employment rates of the prime-aged population. Breen (2005) focuses on the quality of the educational system to explain cross-country variation in youth unemployment. He concludes that a strong educational system which supports the integration of the youth population in the labor market can generally help to reduce unemployment by avoiding periods of youth joblessness. A comprehensive analysis of the advantages of vocational training systems, or in general combined schoolworkplace education is given by Biavaschi et al. (2012). Lopez-Mayan and Nicodemo (2012) find that apprenticeship training reduces the time of finding the first job in Spain. Mohrenweiser and Zwick (2009) discuss the costs and benefits of apprenticeship training for firms in Germany. Finally, Korenman and Neumark (2000) emphasize the role of demographic trends for youth unemployment. More concretely, a larger share of youth population to prime-age population pushes youth labor supply, leading to higher youth unemployment rates given that youth and prime-aged persons are not perfect substitutes in production.

3

Empirical specification

We make use of a comprehensive data set on different aspects of labor market regulation, demographic developments, and on the quality of the educational system. We further apply a model averaging approach in order to take model uncertainty into account and to ensure robustness of the findings.

3.1

Data

We use an unbalanced panel data set for 17 OECD countries from 1982 to 2005. 1 The dependent variables are the total unemployment rate, the youth unemployment rate (15-24) measured as the share of youth unemployed to the youth labor force, and the

1

We cannot use more recent data due to to a structural break in the replacement rate data. More specifically, the O E C D provides new measures for the replacement rate starting in 2 0 0 1 which are not comparable to the historical time-series we use in this paper.

Youth Unemployment in the OECD: The Role of Institutions • 407

unemployment rate of the older population (25+) measured as the share of unemployed (25+) to the labor force (25+). Data on age-specific unemployment are taken from the ILO. For the selection of labor market regulation variables we follow Sachs (2012). More specifically, we use indicators of five categories of regulation: the labor tax system, employment protection, the unemployment benefit system, and the wage bargaining system. We consider product market regulation as an additional relevant category as indicators of this category proved to be relevant for unemployment (Feldmann 2008). As described in section 2, demographic developments as well as the quality of the educational system matter for unemployment. In total, we include 12 regulatory indicators, a demographic factor and an educational variable. We assign these 14 indicators to the group called Institutions. The specific indicators are briefly described in the following and more extensively in the Appendix. The labor tax system is represented by the average values of the payroll, the income and the consumption tax. Wage bargaining is captured by three indicators, the union coverage (the share of workers affected by union wage agreements to all workers), bargaining coordination (the level at which bargaining formally or informally takes place), and union density (the share of workers organized in unions to all workers). The unemployment benefit system is represented by two indicators. More specifically, the replacement rate during the first year of unemployment captures the generosity of the unemployment benefit system. We also use the share of unemployed entitled to unemployment benefits to control for the coverage of the unemployment benefit system. Two additional institutional indicators are considered: First, a minimum wage indicator is taken into account which describes the level at which a minimum wage is set (the indicator from 0 indicating no minimum wage to 8 indicating that the minimum wage is set by the government). Second, an indicator for the level of product market regulation is included. The degree of employment protection for regular and temporary employment represent the job protection system. More concretely, the former comprises information on severance payments, notice periods, notification procedures, the length of the trial period or the compensation following an unfair dismissal. The latter indicator covers information like the maximum number of fixed-term contracts that can be concluded successively, or whether employees from temporary work agencies and regularly employed workers are treated equally by regulation. These job protection variables are expected to capture the degree of labor market flexibility and the labor market access for entrants which probably affect youth and older workers differently according to their respective job experience. Finally, a demographic factor is created by calculating the share of the youth population in the total population. This indicator relates to the relative size of the entrants cohort and should capture labor supply effects. The characteristics of the education system are more difficult to display as they are not directly observable. We use the share of the youth labor force in the youth population as one easily measurable comprehensively available indicator. It largely correlates to the relevance of combined school-workplace education which is expected to facilitate the transition into work. Nevertheless, there could be other sources affecting this share not related to the quality of the education system. Hence, we are aware that this is a crude approximation, and we remain cautious when interpreting the findings of the education variable.

408 • Andreas Sachs and Werner Smolny

3.2

Econometric model

Panel data allows to exploit both variation across countries as well as over time. The basic model reads Uc,t,z = P Institutions

Cit

+ SControlscj

+ £c,t,z-

(1)

Uc,t,z's either total unemployment, youth unemployment (15-24) or unemployment (25+) in country c for time t, where z refers to the specific group of unemployed. We explain the dependent variable by the 14 indicators belonging to the group Institutions as explained in the data section, and some control factors. More concretely, we follow Nickell et al. (2005) and include the real interest rate, a productivity shock, a labor demand shock, and an import price shock in order to control for cyclical fluctuations. The construction of these shocks is described extensively in the Appendix. We also include an indicator capturing credit constraints since this appears to be a highly relevant factor for unemployment (Dromel et al. 2010). We further make use of the panel structure of our data and control for unobserved heterogeneity by including time- and country-specific fixed effects. The model described in equation (1) is generally well-suited to reveal significant correlations between institutions and unemployment. However, it might be prone to endogeneity resulting from omitted variables or reverse causality. It is impossible to consider all aspects which can influence unemployment in our model. However, the inclusion of fixed effects at least controls for any time-invariant idiosyncratic factor as well as for common shocks which affect all countries in the sample. Reverse causality could emerge when changes in unemployment cause changes in institutions. A government could be forced to implement labor market reforms when unemployment is particularly high. This issue is, to some extent, covered by the inclusion of fixed effects, as well. Nevertheless, a causal impact of a change in unemployment on an institutional change is less likely but still possible in our setup. Instrumental variable estimation would help but it is hard to find suitable instruments for institutions. Lagged values are probably a function of expected unemployment and are therefore inadequate instruments. GMM-type estimation based on lagged differences additionally suffers from rather persistent regulatory indicators. Similarly, political, demographic or macroeconomic factors which have been used as instruments (Amable et al. 2011) are likely correlated with both institutions and unemployment. Hence, we abstain to apply an instrumental variable estimator and interpret our findings as correlations, not as causal effects. Theory provides only limited support regarding the specification of the model. More specifically, we do not know a priori which institutional indicators should be included in the model. It appears straightforward to estimate (1) with all institutional indicators. If the number of observations is sufficient, such an approach is appropriate to determine the impact of institutional factors on unemployment. However, if degrees of freedom are restricted caused by a limited number of observations, or if explanatory factors are highly correlated, as in our case, a misspecified model could result in misleading inference. In order to explicitly control for model uncertainty, we apply a Bayesian model averaging approach following Sala-i-Martin et al. (2004). The basic idea of this method is to estimate and systematically evaluate variants of the basic model described in equation (1). More specifically, we estimate and evaluate all models which only differ in the combination of keK indicators (where K = 14) included in Institutionsc,t. In contrast, the set of control variables as well as the country- and time-specific fixed effects are considered in all models. The consideration of models consisting of all possible combinations of the 14 institutional

Youth Unemployment in the OECD: The Role of Institutions • 4 0 9

variables sums up to 2k = 16384 models estimated. The evaluation of the models is carried out on the basis of model weights, whereby models with a better fit receive a higher weight. The weight of a specific model M, is calculated as TJ-WSSE-^ , P(M,|A) = — ^P(M.0(E^-i »^c-l E / = i P(M,)(E?=,

Tc)/2

(2)

Tc)- i' SSE~ k

2

(EW

N is the number of cross-sections and Tc is the number of time-series observations for country c (this extension of the method provided by Sala-i-Martin et al. 2004 to unbalanced panels is provided by Moral-Benito 2012). Furthermore, kj is equal to the number of indicators contained in Institutionscj in model i. p(Mj) is the prior model probability, and SSEj is the sum of squared residuals of model i. The prior model probability is given by

**>- (If H f '

131

where k is the a priori specified model size, k, the number of indicators contained in Institutionsc t for a specific model, and K the total number of indicators which can be included in Institutions ct, i.e. K = 14. In the empirical application, we set the prior model size k to 6, but run the methods with prior model size values of 2 and 10 as a robustness check. P(Mj\ A) gives a measure for the quality of model M ( in comparison to all 2K considered models. A statement about the significance of the K explanatory variables can now be calculated. Suppose that one is interested in a variable x. Summing up the model weights (following equation (2)) of all models containing x gives the posterior inclusion probability. This measures the importance of the variable x for explaining the dependent variable. If the corresponding variable is often included in models with higher quality, the posterior inclusion probability of A: is comparably large. In order to derive statements about the significance of a variable, the posterior inclusion probability needs to be set in relation to the prior inclusion probability which serves as a threshold dividing significant from insignificant variables. The prior inclusion probability is just and thus depends on the prior model size. In the empirical application, a posterior inclusion probability value above the respective prior inclusion probability indicates significance of the corresponding variable. Furthermore, the model weights can be used to derive a posterior mean and a posterior variance for each of the K indicators. 2k

E(P\y) = J2P(MiWPi /= i 2k

>

a\"

« °

C

w ii

n 2

« ii

S

l/l l/l II U >> ^

^ C\

VI l/l 1

i/i dl

VI / II

'

i/i ai

o c

00 oC O in

* ,

*s>

°

£

o

o

o

c

o c

o c

o c

o c

-D «

n

•-r =

Ä °

c

! £ I VI

i-

CM

ti

X « .2 -2 Q. p i 5 ) a . Vt

n 2

o^ ~

0) — >, p O .¡3

O

,

Ol US Ol

0

5. « * " t •"fr

E

m rv

£ >,

00

o C

o c

o c

o c

o c

o c

o c

o c

S - 2 «

S

^

0)

rö

£ «

D .2

S

8

c o

r" >-

" KC äb-Si C XI td « - C "C u nj

Q- S ? 2 » £ 8 o n — 1 2 ^ re "§ ti s 5 ¡J O

!

r< in

—, m iv rv in

qj •O c

vi V

s

's.

s O

t/i V

c C

o

l

1

u Ol 11 1 - , tKW rçj O

5 .

o

T•. m Ii 1 =

vi II

o>

O

7

25

vi U

£

S

^ ro" IV CM

O m

ffl

o

vi D

0) Î5

o

O

, !

U m k e h r e r

41

.b

2 n ^ bß-2 -o Q. 3 o

4) i-n .. 0)

Î I in

fc.

c

£

Î5

15

5

m

!5

"2 - c fO ± i c 5 f0 " O

(0 o

K

o 2

E

^

E « Q. «

—

S •= •!= - 5

Are there Long-Term Earnings Scars from Youth Unemployment in Germany? • 487

There might be systematic sorting of individuals between training firms, based on the probability that a firm closes d o w n a few years later. W h e n we include our control variables and fixed effects for the training firm's location to control for initial sorting in column (2), the reduced-form effect decreases to statistically still significant 9 percent, and the first-stage effect drops to a b o u t t w o extra months of unemployment. This pattern suggests that individuals with a below-average earning potential are trained at firms with an above-average probability to close d o w n a few years later. Having controlled for this sorting effect, IV estimates a second-stage effect of early-career unemployment on prime-age earnings that is considerably smaller than the one presented in column (1). However, it is still statistically and economically significant. If it was possible to correctly predict the likelihood of a plant closure, there might be further r o o m for some strategic behavior in order to avoid starting an apprenticeship at a dying firm. T o rule out this possibility, we exclude apprentices with less t h a n t w o years of training in the specification displayed in column (3) of Table 4 . " While, relative to model (2) of the same table, the reduced-form effect hardly changes, the first-stage effect decreases to one and a half months, but remains statistically significant. As a result, the second-stage effect somewhat increases, and anticipation effects d o not seem to induce an u p w a r d bias. A further check of whether the instrument is as good as randomly assigned is to estimate specification (3) of Table 4 by firm size. As is shown by von Wachter and Bender (2006), wage losses caused by an early displacement tend to reflect losses of firm-size wage premia. Consequently, we should find strong reduced-form effects for apprentices w h o were displaced f r o m large training firms. Apprentices w h o had to leave small training firms, in contrast, should s h o w persistent reduced-form effects only if displacements were endogenous. Results f r o m an IV model, restricted to individuals w h o were trained at firms with at least 50 employees and five apprentices in the year of graduation, are shown in column (4) of Table 4. C o l u m n (5) contains the same results for those apprentices w h o graduated f r o m firms with less than 50 employees a n d less than five apprentices. For workers w h o were displaced f r o m large training firms, prime-age earnings losses are on average 14 percent. This effect is significant at the 5 percent level, although only 185 individuals are treated in this group. In the latter group, 2 , 1 2 4 individuals are treated. These workers suffer only 3.8 percent earnings losses in prime age. The effect is n o longer statistically different f r o m zero at the 5 percent level. This result not only supports exogeneity of the instrument but also suggests that losses of firm size wage premia are an important factor behind the local scarring effect of early-career unemployment identified by IV. Of course, identification is only achieved if the exclusion restriction holds. 5.2.2 Assessing excludability of the instrument The exclusion restriction might be problematic for at least three reasons: First, if exogenously displaced apprentices loose significant a m o u n t s of h u m a n capital specific to their training firm, they will suffer permanent earnings losses simply because they are n o longer as productive at other firms compared to similar apprentices w h o had the possibility to

13

Despite the absence of wage reductions, Fackler et al. (2013) report for Germany that exiting establishments somewhat shrink already more than t w o years before closure, relative to a matched sample of surviving establishments. Focusing on establishments that are still willing to train new apprentices two years before they close down should also mitigate this potential source of endogeneity.

488 • Joachim Möller and Matthias Umkehrer

stay with their training firm. However, G a t h m a n n and Schonberg (2010 demonstrate that h u m a n capital is highly transferable between occupations, and is rather b o u n d to carrying out a specific task. This is particularly the case for apprenticeship training, which develops, contrary to c o m m o n perceptions, general rather than firm-specific skills, cf. W i n k e l m a n n (1996); H a r h o f f and Kane (1997). Correlation of plant closure at the time of graduation with persistent initial match quality, stemming f r o m other sources t h a n early-career unemployment, constitutes a second potential violation of the exclusion restriction. As is pointed out by N e u m a r k (2002), the direction of this bias in the IV estimates depends on the sign of this correlation. If, for instance, displaced workers permanently accept lower-paid jobs as a direct consequence of an early layoff, IV estimates would absorb this additional effect, a n d would therefore be biased upwards. As an indirect test for whether directly induced differences in initial match quality persist over more than eight years of potential experience, N e u m a r k (2002) suggests to restrict the analysis to workers w h o have changed their employer during the course of the early career. The reasoning is that, if such persistent differences in initial match quality cause an u p w a r d bias in the IV estimates, this u p w a r d bias should be comparatively smaller in the sample of mobile workers. T o see whether this is the case, we re-estimate specification (3) of Table 4 for the sub-sample of workers w h o have experienced at least one employer change with an intervening period of joblessness n o longer than three weeks during the early career. Recalls are also excluded f r o m this definition of direct employer changes. Reflecting the early career's p r o n o u n c e d job dynamics, almost 54 percent of workers remain in the sample. T h e share of mobile workers a m o n g the initially displaced is, with almost 64 percent, remarkably higher. The OLS estimate of the scarring effect in the full sample is — 8.2 x 1 0 - 4 . T h e corresponding IV second-stage estimate is — 18.5 x 1 0 - 4 . While a one standard deviation increase of early unemployment would reduce average prime-age earnings in the former case by roughly 23 percent, this effect almost doubles to 4 5 percent in the latter case. In the sample of mobile workers, the estimated coefficient of earlycareer unemployment is — 8.4 x 1 0 - 4 for OLS, and - 2 0 . 7 x 1 0 " 4 for IV [column (6) of Table 4]. The implied effects of one standard deviation of early unemployment on primeage earnings are about minus 2 4 percent and minus 49 percent respectively. Consequently, the bias implied by the difference between the IV and the OLS estimates is not smaller in the sample of mobile workers than in the full sample. Third, almost 4 0 percent of apprentices w h o are involved in plant closures at the time of graduation are immediately reemployed, and t w o thirds are in covered employment within one m o n t h or less. 14 If displaced but immediately reemployed workers p e r f o r m similar as w o u l d have been the case in the absence of an initial plant closure, these never takers w o u l d not contribute to the IV estimate, and the exclusion restriction would not be violated. However, even w i t h o u t an intermediate spell of unemployment, some displaced workers might still suffer worse labor m a r k e t conditions later in life if they permanently lower their reservation wage as a consequence of the take-over experience itself. M o r e over, their displacement probabilities might n o w be higher because they do not have insider status and could not be screened during training.

14

For those 9 percent of apprentices who stay the longest time without employment after displacement, in contrast, it takes at least one year to find new employment.

Are there Long-Term Earnings Scars from Youth Unemployment in Germany? • 489

T o explore potential bias arising f r o m immediate take-overs, we include a d u m m y variable indicating the absence of an initial period of unemployment if the apprentice was displaced f r o m his training firm into specification (3) of Table 4. The results are displayed in column (7) of Table 4. The difference in average prime-age earnings between the displaced and the non-displaced is n o w a r o u n d 11 percent. Displaced workers are on average more than five m o n t h s longer unemployed during the early career. The IV second-stage estimate of early-career unemployment drops to —7.1 x 1 0 - 4 , which is even smaller in magnitude than the corresponding OLS estimate of —8.2 x 10~ 4 (not reported in the table). The implied effect of a one standard deviation increase of early-career unemployment on the mean of prime-age earnings is minus 2 0 . 7 percent, which is still substantial. N o t e , however, that unemployment at a later point in the early career could also be a delayed reaction to an initial plant closure. If this delayed unemployment experience in turn results in future earnings loss, it is part of the scarring effect, and the exclusion restriction would not be violated. Explicitly controlling for direct take-over therefore could also switch off some of the channels through which scarring actually operates. In any case, even if the exclusion restriction was violated, the associated bias does not appear to be large enough to explain the finding of economically significant scarring effects of early-career unemployment on prime-age earnings. 5.2.3 Interpreting the instrumental variable regression results IV produces point estimates of the scarring effect of early-career unemployment that are usually larger t h a n the corresponding OLS estimates. Three mechanisms could explain this finding: First, if plant closure of the training firm is a valid instrument for early-career unemployment, larger IV estimates imply d o w n w a r d biased OLS estimates. As outlined above, such a d o w n w a r d bias arises if, for instance, early job shopping activities involve temporary periods of unemployment but improve labor market outcomes in the future, cf. N e u m a r k (2002). An alternative explanation could be the presence of attenuation bias induced by classical measurement error in early-career unemployment. Second, in the case of heterogenous scarring effects, IV estimates a treatment effect for the subpopulation of compliers, i.e. an effect for those workers w h o experienced an elevated a m o u n t of early-career unemployment only because their training firm had to close down. In our application, it appears reasonable to assume that less skilled workers with lower reemployment probabilities are overrepresented a m o n g the g r o u p of compliers. As we will s h o w in the next section, scarring effects substantially decrease across the distribution of prime-age earnings. It is therefore not surprising that a local average scarring effect estimated by IV exceeds the corresponding OLS estimate. 15 Third, given the results presented above, we are quite confident that the instrument is as good as randomly assigned once sorting between training firms is taken into account. Although we find n o evidence that violations of the exclusion restriction invalidate the finding of significant and long-lasting scarring effects, we can never be completely sure that IV estimates are not biased upwards. N o t e , however, that the IV estimates are quite imprecise and that the null of exogeneity of early-career unemployment cannot be rejected

15

Appendix 8.4 of Schmillen and Umkehrer (2013) provides further evidence that scarring effects are not the same for everybody. In the presence of heterogenous scarring effects, however, the IV estimates presented in this paper may be interpreted as upper bounds for the average treatment effect on the treated.

490 • Joachim Möller and Matthias Umkehrer

at the 1 percent significance level in the majority of cases. 16 If early-career unemployment is in fact exogenous in a regression of prime-age earnings, however, O L S estimates are not only consistent but also efficient. We therefore treat early-career unemployment as exogenous throughout the remaining part of this paper. 5.3

Scarring effect heterogeneity across the earnings distribution

Workers who do generally well in the labor market might react very differently to early unemployment experiences compared to workers who perform rather poorly. T o explore this possibility, we now investigate whether the scarring effect of early-career unemployment varies across the unconditional distribution of prime-age earnings. For this purpose, we run regressions of the recentered influence function (RIF) of each unconditional decile of cumulative prime-age earnings, see Firpo et al. (2009). The models' specification is identical to the one in column (2) of Table 3. However, a coefficient estimated using RIF regression can be interpreted as the partial effect that a regressor has on the specified unconditional quantile of the prime-age earnings distribution. The estimated effects of early-career unemployment on each decile of prime-age earnings are presented in the upper panel of Table 5. Early-career unemployment has a significant negative effect on each decile. These effects are substantially larger at the lower than at the upper tail: At the first decile, an increase of early-career unemployment by one standard deviation (326 days) reduces prime-age earnings by 56 percent. The corresponding scarring effect on the median is 11 percent, and at the ninth decile it is roughly 7 percent. 17 We further restrict the quantile regression analysis to the 60 percent of workers who managed to completely avoid unemployment during their prime age period. This positive selection probably induces a downward bias. The true scarring effects of early-career unemployment on earnings, net of its effects on unemployment, should therefore be bounded by the effects for the full sample (first panel of Table 5), and for the selected sample (second panel of Table 5). At the bottom of the prime-age earnings distribution, earnings losses from unemployment experienced early in professional life are still substantial, even if exposure to unemployment during prime age is avoided. But the effects decrease rapidly as we move up the earnings distribution: An increase of early-career unemployment by one standard deviation (in the full sample) reduces the first decile of prime-age earnings for this selected group of workers by 45 percent, the median by 4.5 percent, and the ninth decile by 3.1 percent. Consequently, workers with a low earning potential, but without any prime-age unemployment, still suffer substantial earnings losses from early-career unemployment. On the contrary, workers with a high earning potential, and who avoid repeated exposure to unemployment, manage to almost completely offset these negative scarring effects.

16

17

Neither a potential d o w n w a r d bias in O L S estimates, nor the s u m of a complier effect and a potential u p w a r d bias in IV estimates a p p e a r to be - in absolute terms - large enough to lead to a rejection of the null in these cases. T h e corresponding effects on the conditional distribution of prime-age earnings are similar in m a g nitude, and also decrease a c r o s s this distribution. They are presented in T a b l e A 3 in the A p p e n d i x . Decreasing scarring effects within training firms, training o c c u p a t i o n s , a n d industry sectors provide s o m e evidence that the revealed pattern is the result of a f u n d a m e n t a l segmentation of the labor m a r k e t rather than of a persistent d e m a n d shock hitting firms, occupations, and industries in different ways.

Are there Long-Term Earnings Scars from Youth Unemployment in Germany? • 491

* * *

# * *

— ,^

* #

00 a\ o I

oo vo m ff\ m ^ P I o o

* iv v- m ro fN P ^ I o o

* in id m (N v- L2 N

> ,

>N

^

1

>N >N

C

l/l l/l l/l l/l « > L) 0) >N 1, °C

o o

c O i/i i/i i/i i/i . 'R