Labour Market Institutions : Amplifiers Or Attenuators?

Labour Market Institutions – Ampliﬁers or Attenuators?

INAUGURALDISSERTATION zur Erlangung der Wurde¨ eines Doktors der Wirtschaftswissenschaft der Fakultat¨ fur¨ Wirtschaftswissenschaft der Ruhr-Universitat¨ Bochum

Kumulative Dissertation, bestehend aus 5 Beitragen¨

vorgelegt von Rahel Felder, M.Sc. aus Huttwil, Kanton Bern, Schweiz 2019 Dekan: Prof. Dr. Michael Roos Referent Prof. Dr. Thomas K. Bauer Korreferent: Prof. Dr. Ronald Bachmann Tag der mundlichen¨ Prufung:¨ 22. Oktober 2019 Contents

List of Tables iii

List of Figures v

1 Introduction 1

2 Labour Market Transitions, Shocks and Institutions 9 2.1 Introduction...... 10 2.2 Unemployment and Labour Market Flows Over the Cycle ...... 13 2.3 Labour Market Institutions and their Interaction with Shocks...... 19 2.4 Methodology ...... 24 2.5 Results...... 27 2.5.1 Main Analysis...... 27 2.5.2 Robustness ...... 35 2.6 Conclusion...... 37 2.A Figures and Tables ...... 40 2.B Data Description and Imputation Methods ...... 51

3 Job Stability in Europe Over the Cycle 54 3.1 Introduction...... 55 3.2 The European Union Labour Force Survey (EU-LFS) Data...... 58 3.3 The Aggregate Evidence...... 59 3.4 Worker and Job Heterogeneities...... 65 3.4.1 Worker Characteristics: Age...... 66 3.4.2 Job Characteristics: Contract Type...... 73 3.5 Econometric Findings: Regression Analysis...... 80 3.6 Conclusion...... 87 3.A Supplementary Figures and Tables...... 90

4 Potential Beneﬁt Duration Eﬀect on Reemployment Timing and Wages 93 4.1 Introduction...... 94 4.2 Empirical Strategy ...... 97 4.2.1 Identifying Local Average Treatment Effects ...... 97 4.2.2 Wage Decomposition...... 100 4.3 Heterogeneity and Its Consequences...... 102 4.3.1 Dynamic Differences and Dynamic Treatment Effects ...... 103 4.3.2 Heterogeneity in Duration Effects Changes Dynamic Selection . . . . 105 CONTENTS ii

4.3.3 Identifying Effects on Wages and Other Outcomes Affected by Dura- tion...... 107 4.4 Background, Data and Prior Results ...... 109 4.4.1 Institutional Framework and Data...... 109 4.4.2 RDD Validity and Prior Results...... 111 4.4.3 The Reemployment Wage Path, Its Importance and Problems . . . . 113 4.5 Dynamic Selection ...... 117 4.5.1 Heterogeneous Duration Effects and Reemployment Probability and Timing...... 118 4.5.2 Dynamic Selection on Observable and Time-Invariant Characteristics 121 4.6 Isolating Wage Effects ...... 126 4.6.1 Effects on Wage Components...... 127 4.6.2 Dynamics of Wage Components ...... 129 4.7 Conclusion...... 133 4.A Identiﬁcation of UI Eligibility in the Data ...... 136 4.B Wage Decomposition...... 137 4.C RDD Validity and Prior Results...... 140 4.D Additional Figures and Tables...... 143

5 Labour Market Participation and Atypical Employment 145 5.1 Atypical Employment in Germany: Extent and Importance ...... 146 5.2 Data and Deﬁnition of Atypical Employment...... 150 5.3 Life Cycle Employment Proﬁles by Birth Cohort and Gender ...... 152 5.4 Typical Employment Trajectories: Results of a Sequence Analysis ...... 162 5.4.1 Education and First Job Entry...... 164 5.4.2 Early Main Employment Period ...... 168 5.5 Summary and Conclusion...... 172 5.A Figures...... 175

6 Worker Motives for Multiple Job Holding in Germany 179 6.1 Introduction...... 180 6.2 Data...... 182 6.3 Aggregate Trends and Institutional Background ...... 184 6.4 Worker Motives...... 188 6.4.1 Determinants...... 189 6.4.2 Consequences...... 192 6.5 Conclusion...... 195 6.A Tables...... 197

Bibliography 198

Acknowledgements 210

Curriculum Vitae 211 List of Tables

2.1 Unobserved shocks model ...... 28 2.2 Observed shocks model...... 29 2.A.1 Descriptive statistics of labour market institutions ...... 42 2.A.2 Pairwise correlations across labour market institutions, total variation . . . . 42 2.A.3 Descriptive statistics of the regression sample...... 42 2.A.4 Heterogeneity of main economic recessions ...... 42 2.A.5 Observed shocks model, annual data...... 43 2.A.6 Observed shocks model, economic growth periods...... 44 2.A.7 Observed shocks model, economic recession periods...... 45 2.A.8 Observed shocks model, lagged institutions...... 46 2.A.9 Observed shocks model, ﬁxed institutions...... 47 2.A.10 Observed shocks model, without Portugal...... 48 2.A.11 Observed shocks model, without Germany...... 49 2.A.12 Observed shocks model, with temporary employment...... 50 2.A.13 Observed shocks model, total E outﬂows...... 51

3.1 Shift-share analysis of change in mean tenure, according to age ...... 71 3.2 Shift-share analysis of change in mean tenure, according to contract type . . 79 3.3 Results of regression analysis of individual tenure before and during the crisis 81 3.A.1 Summary of sample before and during the crisis, percentage ...... 90 3.A.2 Results of regression analysis of individual tenure before and during the crisis 92 3.A.3 Importance of estimated components for the explanatory power of the model, percentage of the predicted variance ...... 92

4.1 Effect of PBD on unemployment and nonemployment duration and reemployment wages...... 113 4.2 Effect of PBD on reemployment wage components and pre minus post changes in wage components...... 128 4.B.1 Wage decompostition...... 137

6.1 Worker and job characteristics of multiple and single job holders by sex, 2003- 2014 ...... 190 6.2 Determinants of multiple job holding by sex, 2003-2014 ...... 191 6.3 Differences in labour market mobility between multiple and single job holders by sex, 2003-2013...... 193 6.A.1 Differences in labour market mobility between multiple and single job holders by sex, 2003-2013, complete decomposition ...... 197 List of Figures

1.1 Unemployment rate, 2001-2013 ...... 2

2.1 Relationship between annual GDP growth and labour market reactions at the beginning of the Great Recession, 2007 and 2008...... 15 2.2 Unemployment rate by country, 1999-2013...... 17 2.3 Annual transition rates between employment and unemployment by country, 1999-2013...... 18 2.4 Relationship between union density and labour market reactions, 1999 and 2013 ...... 33 2.A.1 Annual GDP growth by country, 1999-2013...... 40 2.A.2 Employment protection legislation by country, 1999-2013 ...... 41 2.A.3 Union density rate by country, 1999-2013...... 41

3.1 Mean tenure and unemployment rate in the EU, 2002-2012 ...... 60 3.2 Mean tenure and unemployment rate by EU Member State, 2002-2012 . . . . 63 3.3 Mean tenure before and during the crisis by EU Member State ...... 64 3.4 Mean tenure by age group in the EU, 2002-2012...... 67 3.5 Mean tenure by age group for selected EU Member States, 2002-2012 . . . . 68 3.6 Mean tenure by age group before and during the crisis by EU Member State 69 3.7 Mean tenure by contract type in the EU, 2002-2012...... 74 3.8 Mean tenure of temporary workers in selected EU Member States, 2002-2012 75 3.9 Mean tenure of temporary workers before and during the crisis by EU Mem- ber State...... 77 3.10 Relationship between mean tenure and the EPL index for EU Member States, 2007 ...... 85 3.11 Relationship between the change in mean tenure during the recession and the EPL index for EU Member States ...... 86

4.1 Reemployment rate ...... 113 4.2 Reemployment wages...... 115 4.3 “Effect” of PBD on worker fixed effects...... 116 4.4 “Effect” of PBD on worker fixed effect by transition paths ...... 117 4.5 Quantile treatment effect ...... 119 4.6 Dynamic selection on worker fixed effects: Aggregated PBD groups . . . . . 122 4.7 Dynamic selection on time-invariant characteristics and time-varying observables...... 124 4.8 Reemployment paths: Wage components...... 130 4.C.1 RDD validity around the age threshold...... 140 LISTOF FIGURES v

4.C.2 Effect of PBD on unemployment and nonemployment duration and reemployment wages...... 141 4.D.1 Reemployment paths: Education level ...... 143 4.D.2 Wages in previous employment by PBD and transition paths ...... 143 4.D.3 Effect of PBD on wage components: Graphical RDD evidence ...... 144

5.1 Share of persons in education by age, year of birth and sex ...... 154 5.2 Employment rates by age, birth cohorts and sex ...... 155 5.3 Share of regular and atypical employees as a proportion of all employed persons by age, birth cohorts and sex...... 157 5.4 Type of atypical employment by age and cohort (men) ...... 159 5.5 Type of atypical employment by age and cohort (women) ...... 160 5.6 Average duration of employment state by employment type (age range 16-30) 165 5.7 Dominant employment state by age and type of employment (age range 16-30)166 5.8 Share of employment types by cohort and sex (age range 16-30) ...... 166 5.9 Average duration of employment state by type of employment (age range 16-40) ...... 170 5.10 Dominant employment state by age and employment type (age range 16-40) 171 5.11 Share of employment types by cohort and sex (age range 16-40) ...... 171 5.A.1 Share of regular and atypical employees as a proportion of the total population by age (women)...... 175 5.A.2 Employment rates by age, sex and region (age group 1974-86) ...... 176 5.A.3 Share of regular and atypical employees as a proportion of total employment by age, sex and region (born 1974-86)...... 177 5.A.4 Type of atypical employment by age, region and sex (age group 1974-86) . . 178

6.1 Number of employed and the share of multiple job holders , 1999-2014 . . . 185 6.2 Number of minijobbers, 1999-2014...... 187 6.3 Types of multiple job holding, 1999-2014...... 187 1. Introduction

The last Great Recession, which started in late 2007, provoked substantial labour market turbulence. The stark economic downturn preceded by a ﬁnancial crisis involved high numbers of job losses. The unemployment rate in the EU and the US rose strongly and persistently from 6.4 percent in 2008 to 10.7 percent in 2013 (see Figure 1.1). However, this aggregate increase masks large divergences in labour market reactions across countries.

Germany, for example, experienced in the same period a decrease in the unemployment rate from 7.6 to 5.5 percent.

Two important sources for this heterogeneity are country-speciﬁc exposure to the recession as well as national labour market institutions. The recession affected some countries relatively mildly, such as Norway and Germany, other countries got strongly hit, such as

Italy and Spain. Moreover, national labour market institutions differ signiﬁcantly across countries, e.g. with regard to employment protection legislation and generosity of unemployment beneﬁts.

Theoretical and empirical economists have been debating on the inﬂuence of labour market institutions on labour markets for quite some time. On the one hand, the seminal study of Nickell, Nunziata and Ochel[2005] suggests that institutional changes directly affect the labour market. On the other hand, the inﬂuential paper of Blanchard and Wolfers[2000] attributes labour market institutions a more indirect role. In particular, the prevailing institutional framework might amplify or attenuate the impact of economic turbulences on the labour market.

Moreover, Figure 1.1 reveals that the German labour market has performed exceptionally well during and after the last Great Recession. Generally, the German economy evolved 1. INTRODUCTION 2

“from the sick man of Europe” in the early 2000s to an “economic superstar”, as famously phrased by Dustmann et al.[2014]. This development is strongly linked to the Hartz reforms, which increased work incentives and trimmed labour administration [among others

Jacobi and Kluve 2007; Rinne and Zimmermann 2013], as well as to the ﬂexible management of working time during the last Great Recession [among others Balleer et al. 2016;

Burda and Hunt 2011].

Figure 1.1.: Unemployment rate, 2001-2013 12 10 8 Unemployment Rate 6 4 2001 2003 2005 2007 2009 2011 2013 Year

Globe Germany

Globe includes the following countries: Austria, Belgium, Czech Republic, Germany, Denmark, Estonia, Spain, Finland, France, Greece, Hungary, Italy, Luxembourg, Norway, Poland, Portugal, Sweden, Slovenia, Slovak Republic, United King- dom and United States of America. Vertical lines mark the introduction of the Hartz reforms in Germany (dotted) and the start of the Great recession (solid). Source: EU-LFS, CPS, own calculation.

The present dissertation contributes twofold to the debate on the role of labour market institutions in shaping labour markets: First, it takes a cross-country perspective on the last Great Recession and analyses how differences in labour market institutions relate to country-speciﬁc labour market reactions. Second, it takes a German perspective and investigates the association of labour market reforms with the development in the German labour market before the last Great Recession.

Labour market institutions are highly policy relevant because of their implications for workers, ﬁrms and ﬁscal accounts. For workers, the institutional framework determines, among others, job security, career prospects and unemployment assistance. As the last

Great Recession strongly indicates, governments seek to influence the labour market to prevent job losses and to foster reemployment. Prominent examples are the introduction 1. INTRODUCTION 3 of short-term working schemes in Germany and the expansion of unemployment benefit duration in the US. In their specific contexts, the dissertation chapters answer the question as to whether a reform in labour market institutions exerts a beneficial impact for workers and the state.

Part I examines cross-country heterogeneities in labour market institutions (Chapter 2 and

Chapter 3). Chapter 2 investigates the transmission of macro-economic shocks to national labour markets. In particular, it analyses the interaction between labour market institutions and shocks for EU Member States and the US in a period including the last Great

Recession. The measures for national labour market reactions contain, in addition to the unemployment rate, worker transitions between employment and unemployment. Chap- ter 3 focuses on job tenure and compares its level before and after the Great Recession across EU Member States. Part I suggests that trade unions attenuate while employment protection legislation ampliﬁes macro-economic shocks.

Part II examines how labour market institutions and reforms relate to the recent development of the German labour market (Chapter 4, Chapter 5 and Chapter 6). Chapter 4 considers an extension in unemployment beneﬁt duration in the 1990s, whereas Chapter

5 and Chapter 6 regard the increased ﬂexibility of the institutional framework, e.g. by the

Hartz reforms. The former evaluates reemployment, the latter analyses forms of atypical employment, respectively. Part II implies that a more generous unemployment insurance system exerts negative effects on reemployment timing and wages and that a flexible institutional framework amplifies the prevalence of atypical employment, in general, and of multiple job holding, in specific.

In the following, the ﬁve dissertation chapters are brieﬂy summarized. 1. INTRODUCTION 4

Chapter 2 (co-authored with Ronald Bachmann) investigates the reasons for cross-country differences in labour market dynamics before and during the Great Recession for a large number of EU Member States as well as the US.

We follow the methodology of Blanchard and Wolfers[2000] by investigating the interaction of macro-economic shocks and labour market institutions.

The European Labour Force Survey (EU-LFS) enables us to calculate transition rates and stocks by year and EU country. It consists of a large number of representative national household surveys, using harmonized concepts and deﬁnitions. US data comes from the Labor Force Statistics released by the Bureau of Labor Market Statistics (BLS), which are based on the Current Population Survey (CPS). Measures for institutions are taken from numerous OECD data sets and the Database on Institutional Characteristics of Trade

Unions, Wage Setting, State Intervention and Social Pacts (ICTWSS).

Within a search and matching model, labour market institutions may increase wage rigidities, thereby reducing the sensitivity of wages to economic condition [among others Blan- chard 1999]. Higher wage rigidities lead to higher dismissals in the presence of economic shocks. Also, labour market institutions such as employment protection legislation may reduce worker ﬂows. In recessions and in the presence of high employment protection legislation, ﬁrms might adjust the productivity level downwards without decreasing the employment level.

Our contribution is twofold. First, we apply the empirical framework of Blanchard and

Wolfers[2000], who examined the unemployment rate, to an analysis of worker ﬂows. Our second contribution is that we extend their framework to the period of the Great Recession.

This chapter suggests that institutions play an important channelling role of macro-economic shocks to national labour markets. It provides evidence for the signiﬁcance of trade unions.

Speciﬁcally, union density, i.e. the share of union members, relates to more moderate labour market reactions to shocks. Also, an attenuating link between the tax wedge and unemployment materializes in recessions. 1. INTRODUCTION 5

Chapter 3 (co-authored with Ronald Bachmann) examines changes in job stability measured by job tenure before and during the last Great Recession for 26 EU Member States.

Job stability is of paramount meaning for workers. Workers may fear that globalization and technological progress, such as advances in communication technologies, have induced changes in the labour market that require them to be more ﬂexible.

We compute descriptive statistics on job tenure, apply a shift-share analysis to focus on the importance of speciﬁc job or worker characteristics and perform multivariate regression analyses in order to control for cross-country differences in the composition of the workforce. Also, we correlate our results regarding job tenure with the country-speciﬁc indicator of employment protection legislation.

Job tenure is derived from the European Labour Force Survey (EU-LFS), whereas the measure for employment protection legislation is from the OECD. The EU-LFS constitute representative household surveys with harmonized concepts and deﬁnitions and is conducted in many EU Member States.

While there is detailed research on job tenure in the pre-crisis period across European countries [Auer and Cazes 2000; Cazes and Tonin 2010], the impact of the Great Recession on job tenure has not been examined for European countries yet.

We show that, at the EU level, average job tenure increased slightly from 2002 to 2012.

Abstracting from cyclical effects, no evidence suggests that job tenure decreased between

2002 and 2012 as suspected by the public. The shift-share analysis shows that job destruction due to the Great recession as well as an ageing workforce have increased job tenure.

However, there seems to be a long-term trend towards shorter job tenure for given age groups. Moreover, our analysis conﬁrms the results in Cazes and Tonin[2010] that shows a strong positive correlation between the degree of employment protection legislation and job tenure.

This chapter implies that job stability was relatively stable across EU Member States and that employment protection ampliﬁed changes in job tenure during the last Great Reces- sion. 1. INTRODUCTION 6

Chapter 4 (co-authored with Hanna Frings and Nikolas Mittag) evaluates the impact of unemployment benefit duration on reemployment timing and wages. We follow Schmieder, von Wachter and Bender[2016] and investigate an age discontinuity in potential benefit duration (PBD) from the German unemployment insurance system in the 1990s. Specifi- cally, PBD was extended from 12 to 18 months for unemployed older than 42 years.

We apply a regression discontinuity design to identify the effect of benefit duration on unemployment duration and reemployment wages. We decompose the wage effect into worker, firm and job specific wage components and analyse dynamic treatment effects.

Hence, we study which wage components cause PBD wage effects and at which points in the unemployment spell they accumulate.

The empirical analyses are based on the LIAB Mover Model, an administrative linked employer-employee dataset that is provided by the Research Data Center of the German

Federal Employment Agency. Besides its richness and high precision, the LIAB Mover

Model has the additional advantage of a high degree of realized worker mobility between

ﬁrms (i.e. connectedness). This is crucial for an unbiased wage decomposition in worker,

ﬁrm and job ﬁxed effects [Andrews et al. 2008].

PBD extensions are a popular policy tool. For example, the US has increased PBD during the last Great recession. To understand how PBD affects reemployment timing and wages is important to evaluate PBD extensions. Moreover, it is informative on the general causal relation between unemployment, job search and wages.

This chapter adds to evidence on the importance of the role of ﬁrms in wage setting [Card,

Heining and Kline 2013]. The adverse wage effect from PBD extensions operates through the firm fixed effect. Since changes in firm fixed effects are partly offset by changes in the dynamics of the other components, the chapter reconciles the diverging findings of

Schmieder, von Wachter and Bender[2016] and Nekoei and Weber[2017]. Moreover, it suggests that heterogeneity in sample selection in transitions to self-employment and moving abroad may drive the wage effects. The duration effect mainly stems from workers prolonging unemployment from the old to the new exhaustion point. Thus, a more generous unemployment insurance exerts negative effects on reemployment timing and wages. 1. INTRODUCTION 7

Chapter 5 (co-authored with Ronald Bachmann and Marcus Tamm) studies the rise of atypical employment in Germany. The share of atypical employed increased from 12.8 percent in 1991 to 20.8 percent in 2015 [Statistisches Bundesamt 2016a]. Atypical employment may, for example, facilitate access to the labour market at the ﬁrst entry into the labour market or after career breaks, be a stepping stone to regular employment or represent the beginning of a permanent period of such employment. The latter may have adverse consequences for an individual’s career. Moreover, atypical employment is related to a higher risk of unemployment and disadvantages in pay in Germany [among others Kvasnicka and Werwatz

2002; RWI 2016].

We provide descriptive evidence on cohort-speciﬁc labour market proﬁles of the birth cohorts 1944 to 1986. Thereby, we focus on individual employment trajectories and typical sequences of different employment forms over the life cycle.

We use a novel data set that combines the survey data of the National Education Panel

(NEPS) with administrative data, the SIAB, from the Research Data Center of the German

Federal Employment Agency. Atypical employment is usually deﬁned as employment in

ﬁxed-term, part-time, marginal employment or temporary employment. The linked data further allow us to include freelance work and self-employment.

The analyses provide important implications for economic policy. For example, if involvement in atypical types of employment at early career stages is accompanied by a significantly reduced probability of ever achieving a stable full-time employment relationship, regulating atypical employment more strictly will be justiﬁed.

This chapter shows that employment patterns strongly altered across birth cohorts. Younger cohorts are characterized by a longer period in education and a corresponding later entry into the labour market. There has also been a fundamental change in the employment behaviour of women. For women, the growing prevalence of atypical employment is closely linked to the development of labour market participation, as much of the increase in employment takes place in the form of atypical employment, mainly part-time and minijobs.

Moreover, entry into the labour market is increasingly more difﬁcult for men from younger than from older cohorts and is characterised by longer periods of atypical employment. 1. INTRODUCTION 8

Chapter 6 investigates worker motives behind the recent strong increase in multiple job holding in Germany. From 2003 to 2014 the number of registered workers with more than one job rose from 1.0 to 2.2 million [Klinger and Weber 2017] and thus has become an important labour market state. The economic literature distinguishes two categories of motives for multiple job holding. The ﬁrst is related to immediate monetary intentions and the second roots in expected pay-offs from job heterogeneity.

In line with previous studies, I descriptively compare the populations of multiple to single job holders and apply logit regressions to study the worker motives for multiple job holding. Moreover, I conduct Blinder-Oaxaca decompositions to quantify how raw differences between the two worker types are due to differences in the worker compositions.

I exploit the richness of administrative employment biography records provided by Re- search Data Center of the German Federal Employment Agency, namely the SIAB, to differentiate forms of multiple job holding, to investigate a wide range of socio-economic information of multiple job holders and to study their labour market transitions. Thus, I identify worker motives for multiple job holding by examining its determinants as well as its consequences for career progression.

The analyses provide answers for the policy relevant questions of who decides to hold more than one job, why and with which implications.

This chapter ﬁnds that multiple job holding in Germany actually corresponds to dual job holding. In particular, after the Hartz II reform in 2003 it is of the type primary job subject to social security contributions with a secondary subsidized minijob. Furthermore, the results suggest that individuals hold multiple jobs due to immediate monetary intentions.

Multiple job holding is more prevalent among primary part-time employed, females, low educated and routine workers, indicating the need to increase earnings with an additional subsidized job. At the same time, multiple job holding relates to career progression. Multi- ple job holders are more likely to switch jobs, sectors and tasks as well as to receive a wage increase relative to single job holders, which supports an investment motive. I ﬁnd strong unobserved heterogeneities between multiple and single job holders. All ﬁndings prevail when separating the analyses by sex. 2. Labour Market Transitions, Shocks and Institutions in Turbulent Times: A Cross-Country Analysis*

Abstract: This paper analyses the impact of the business cycle on labour market dynamics in EU Member States and the US during the ﬁrst decade of the 21st century. Using unique measures of labour market ﬂows constructed from worker-level micro data, we examine to what extent macro shocks were transmitted to national labour markets. We apply the approach by Blanchard and Wolfers[2000] to analyse the role of the interaction of macroeconomic shocks and labour market institutions for worker transitions in order to explain cross-country differences in labour market reactions in a period including the

Great Recession. Our results suggest a significant influence of trade unions in channelling macroeconomic shocks. Specifically, union density moderates these impacts over the business cycle, i.e. countries with stronger trade unions experience weaker reactions of the unemployment rate and of worker transitions.

*This chapter is co-authored with Ronald Bachmann. It contains a substantially revised version of: Bach- mann, Ronald and Rahel Felder. 2017. “Labour Market Transitions, Shocks and Institutions in Turbulent Times: A Cross-Country Analysis.” Ruhr Economic Paper No. 709. The chapter is under revision for Empir- ica. We thank Thomas K. Bauer, Daniel Baumgarten, Giuseppe Bertola, Romain Duval, Hanna Frings, Matthias Giesecke, Philipp Jager,¨ Lisa Leschnig, Pedro S. Martins, Christian Merkl, Battista Severgnini as well as participants of the 30th Annual Congress of the European Economic Association, the IZA/OECD Employment Seminar, the World Bank DIME seminar, the “Conference in honor of Christopher A. Pissarides” at SciencesPo, Paris, the FAU/IAB-Seminar Macroeconomics and Labor Markets in Nurnberg,¨ the 9th RGS Doctoral Confer- ence in Economics, the 21st SMYE, the Jahrestagung of the Verein fur¨ Socialpolitik 2016 and seminars at RWI for helpful comments and suggestions. We are grateful to Fernanda Martinez Flores for excellent research assistance. All remaining errors are our own. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 10

2.1. Introduction

The recent Great Recession was associated with job losses and displacements for a substantial number of persons, and a strong and persistent increase in unemployment in many

European countries. The average unemployment rate in the Eurozone rose from an average of 7.0% in 2008 to 10.9% in 2013 [Eurostat 2014b]. This ﬁgure, however, masks large divergences in labour market reactions across the EU and associated countries [European

Central Bank 2012; OECD 2013]. In Austria, Belgium, Germany, Luxembourg and Norway the unemployment rate hardly increased during the crisis, whereas in Estonia, Greece and

Spain, it rose strongly, reaching a level of 25%. Worker flows, which determine the level of unemployment, exhibit substantial heterogeneity as well [Bachmann et al. 2015]. For certain countries, an increase in job losses during the Great Recession led to large outflows from employment, while for others a decline in job creation led to small outflows from unemployment. These cross-country differences are not only likely to be strongly influenced by cross-country differences in the magnitude of economic shocks, but also by the institutional framework of national labour markets.

In this paper, we investigate the role of labour market institutions for the transmission of macroeconomic shocks to labour markets looking at both the unemployment rate and worker ﬂows. In particular, we apply Blanchard and Wolfers’[2000] empirical method, which was originally used to examine the causes of diverging development of US and

European unemployment from the 1960s until the mid-1990s.2 Our analysis focuses on cross-country differences in unemployment and labour market dynamics between 1999 and 2013, a period covering the Great Recession, for a large number of European countries as well as the US. We enhance their model by allowing for changes in institutional variables over time, which accounts for the variation of institutions within countries as motivated by Nickell[1997]. Speciﬁcally, we analyse the impact of shocks and the interaction of shocks and labour market institutions. We separately identify (i) the direct impact of macroeconomic shocks and (ii) how shocks of a given size were transmitted to the national

2Note that this methodology does not yield causal effects. The word “effect” should therefore be broadly interpreted in the following. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 11 labour markets through the prevailing institutional framework. The latter thus measures the indirect effect of institutions on labour market dynamics.

The main result of our analysis addresses the role of trade unions in shaping macroeconomic shocks for labour market dynamics. In particular, higher union density is associated with more moderate labour market reactions in recessions as well as in economic upturns. One explanation is the objective of trade unions to provide job security to their members, which leads to both lower employment growth in economic upturns and lower job destruction in recessions. As this result has not been found by the preceding literature, it seems to be a particular phenomenon of the time period leading up to and including the Great Recession. Furthermore, our results lend support to ﬁndings from the literature that employment protection legislation becomes more important for labour market ﬂows when the trend growth of the economy is low [e.g. Bentolila and Bertola 1990; Messina and

Vallanti 2007].

Our analyses are related to two strands of the economic literature. On the one hand, there is a considerable and rapidly growing literature on worker flows, focusing on the mechanisms underlying the cyclical behaviour of the unemployment rate. These studies investigate the relative importance of the inflows into and the outflows from unemployment, with the most recent articles establishing a relatively balanced role of inflows into and out-

ﬂows out of unemployment [e.g. Elsby, Michaels and Solon 2009; Yashiv 2008; Fujita and

Ramey 2009].

On the other hand, our paper is connected to a large body of theoretical and empirical literature examining heterogeneity in the unemployment rate caused by institutions across and within countries. An overview of these studies is provided by Boeri and Van Ours

[2013]. In theory, labour market institutions can have ambiguous effects on labour market performance as they play two contrasting roles. First, they may worsen labour market outcomes by forming rigidities which distort price- and wage-setting mechanisms [Layard,

Nickell and Jackman 1991, 2005; Blanchard 1999]; second, they may have positive effects by disseminating information and increasing coordination [Traxler and Kittel 2000] in the labour market. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 12

Bassanini and Duval[2006], Eichhorst, Feil and Marx[2010], Orlandi[2012], Flaig and

Rottmann[2013], de Serres and Murtin[2013], Gal and Theising[2015] and Bertola[2017] represent recent examples of empirical research applying cross-country comparisons. They provide evidence for an adverse effect of generous unemployment insurance systems and large tax wedges on unemployment. In contrast, high levels of wage bargaining coordination and active labour market programmes exert a favourable inﬂuence. Only Eichhorst,

Feil and Marx[2010] cannot ﬁnd support for the relevance of classical labour market institutions, but they attribute a key role to internal labour market ﬂexibility, in particular to working time adjustments.

Our contributions to the literature are as follows. First, we extend the framework of Blan- chard and Wolfers[2000] which has been extensively used to examine differences across countries in labour market stocks, especially unemployment, to an analysis of labour market transitions. Taking into account worker transitions allows us to investigate the behaviour of national labour markets over the business cycle more precisely, since ﬂows are generally more sensitive to macroeconomic shocks and respond more quickly than it is the case for stocks, which could especially be seen in many European countries during the

Great Recession [Bachmann et al. 2015]. The analysis of worker flows yields insights into the mechanisms underlying the dynamic components of employment and unemployment, which is at the core of the “ins vs. outs” debate. Our study presents indications regarding institutional reasons for cross-country differences in worker flows. Moreover, labour market transitions are measures of employment security (for worker flows from employment to unemployment) and of unemployment duration (for worker flows from unemployment to employment). Therefore, our study provides welfare implications of labour market institutions.

Second, we investigate labour market behaviour during the time period 1999 to 2013, i.e. the Great Recession and the preceding decade.3 This period is particularly interesting as it includes a long expansion with strong employment creation in many industrialised coun-

3The analysis of the evolution of unemployment by Bertola[2017] also includes the time period of the Great Recession. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 13 tries, as well as the Great Recession, which featured an economic shock much larger than what could be seen in previous recessions. Finally, in contrast to many existing studies, we analyse the entire ﬁrst decade of the 21st century. Therefore, we take into account the

(medium-run) effects of important changes in labour market institutions that had taken place in or just before this decade in many Southern European countries such as Italy and

Spain, but also Central European countries such as Germany.

The remainder of the paper is structured as follows. Section 2.2 describes the data used to construct the measures for labour market dynamics and for macroeconomic shocks and provides descriptive evidence for the 21 countries in our sample. Section 2.3 illustrates labour market institutions and reviews their potential impact on labour market dynamics from a theoretical point of view. Section 2.4 explains the empirical identiﬁcation strategy.

Section 2.5 presents the results, including a set of robustness tests. The last section summarizes the main ﬁndings and concludes the discussion.

2.2. Unemployment and Labour Market Flows Over the Cycle

Our sample of analysis consists of 20 European countries4 as well as the US. We obtain labour market dynamics of European countries from the European Labour Force Survey

(EU-LFS). It includes all EU Member States without Croatia (EU 27) as well as Norway,

Iceland and Switzerland. The Labour Force Surveys are conducted by the national statistical agencies applying harmonized concepts and deﬁnitions, which enables us to perform a cross-country comparison. From a person’s current and previous labour market status, we compute the stock of employed, unemployed and non-participating individuals, along with transition rates between every labour market state by year and country. In the data, an individual’s current labour market status is deﬁned according to the ILO standard.5 By

4The countries are Austria, Belgium, Czech Republic, Germany, Denmark, Estonia, Spain, Finland, France, Greece, Hungary, Italy, Luxembourg, Norway, Poland, Portugal, Sweden, Slovenia, Slovak Republic and the United Kingdom. 5This means that a person is defined as employed if he or she performed some work for wage/salary or for profit or family gain, or – if temporarily not at work – had a formal attachment to his or her job or was with an enterprise; and as unemployed if he or she was without work, currently available for work, and seeking work [ILO 1988]. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 14 contrast, the labour market status in the previous year is based on self-perception of the in- terviewed person. Although these two definitions might not overlap perfectly, using both to identify labour market flows from one year to the next is preferable to alternative approaches, which would not allow for a consistent approach across countries (see Appendix

B).

US data on labour market status and worker transitions are taken from the IPUMS-CPS database [Flood et al. 2015], which is derived using the Current Population Survey (CPS).

In order to make worker ﬂows comparable with the EU-LFS, we impose the time structure of the EU-LFS data set on the CPS data. In particular, for each month of the observation period, we construct stocks and ﬂows for individuals observed in the same month one year before. For each year, the monthly values are averaged yielding one measure for each labour market outcome per year.

In our analysis, we focus on the time period 1999 to 2013. This time period corresponds to the largest number of available country-year combinations for which information on labour market transitions is available in the EU-LFS. As explained in detail in Appendix

2.B, we exclude several European countries from the EU-LFS because of limited data availability; for the same reason, we need to impute some missing values. The ﬁnal data set includes the unemployment rate and transition rates between employment and unemployment at the country-year level. At the individual level, we restrict the sample to dependent- status employees, and omit individuals living in institutional households (e.g. retirement homes or military barracks), working for the military as well as children under the age of

15 and adults aged 65 and over.

The initial observation motivating our analysis is that changes in the unemployment rate and worker ﬂow rates between employment and unemployment show large cross-country variation over the business cycle, especially during the Great Recession, a period of strong labour market turbulence. Figure 2.1 illustrates these differences by relating the changes in annual GDP growth to changes in unemployment and the transitions between employment and unemployment from 2007 to 2008. This makes clear that countries with very 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 15 similar changes in GDP, such as Austria, Italy and Spain – with a reduction in GDP growth of about 3 percentage points – differ considerably in their labour market reactions.

Figure 2.1.: Relationship between annual GDP growth and labour market reactions at the beginning of the Great Recession, 2007 and 2008 (a) Unemployment rate 4

ES 2

LU EE IT UK HU US SE

0 NO DK FI SI FRPTBE GR AT Change in U CZ DE SK −2 PL −4 −15 −10 −5 0 Change in GDP Growth

(b) Transition rate from employment to un- (c) Transition rate from unemployment to employment employment

3 FI

5 SK PL HU CZDK ES IT GR EE BE 2 ES

0 DEFR AT US SI EE SE NO

1 PT −5 UK

UK Change in EtoU Change in UtoE SK IT HU LU SE US −10 0 GR NOFRDKBE SI CZPT PL ATDE FI −15 LU −1 −15 −10 −5 0 −15 −10 −5 0 Change in GDP Growth Change in GDP Growth Country codes: AT: Austria, BE: Belgium, CZ: Czech Republic, DE: Germany, DK: Denmark, EE: Estonia, ES: Spain, FI: Finland, FR: France, GR: Greece, HU: Hungary, IT: Italy, LU: Luxembourg, NO: Norway, PL: Poland, PT: Portugal, SE: Sweden, SI: Slovenia, SK: Slovak Republic, UK: United Kingdom, US: United States of America. Source: EU-LFS, CPS, ICTWSS, own calculation.

This suggests that similar macroeconomic shocks were transmitted heterogeneously to national labour markets. Furthermore, the response in the transition rate from unemployment to employment was of comparable size, but the reverse worker ﬂow and therefore the unemployment rate exhibited extremely contrasting trends. This is an indication that the answer to the “ins vs. outs” debate is likely to differ between countries [see e.g. Petron- golo and Pissarides 2008 and Elsby, Hobijn and Sahin 2013 for an explicit cross-country analysis of these issues]. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 16

Looking at our variables of interest in more detail, we ﬁrst focus on macroeconomic shocks.

In order to proceed as parsimoniously as possible, we use the most aggregate measure of the business cycle available, the annual growth rate of real GDP.6 The annual GDP time-series for countries provided by the OECD allows us to compute economic growth rates. Figure 2.A.1 shows that all countries experienced country-speciﬁc stable growth levels with moderate business cycle movements in the time period 1999 to 2007, except for a dip in the early 2000s driven by the new economy recession which followed the dot-com bubble. In 2007/2008, when the Great Recession hit economies, an extreme reduction in

GDP growth in most countries is visible. The extent of the fall was heterogeneous across countries. While the recession was relatively mild in countries such as Norway, the US,

France, Belgium and Germany, where the decrease in the growth rate is at most 3 percentage points, it was rather strong in Slovenia, Slovakia and Estonia with a decline of up to 11 percentage points. Furthermore, in the aftermath of the recession, some countries such as

Estonia and the US, recovered relatively quickly, whereas other countries such as Greece and Spain, faced a protracted recession.

Turning to the detailed descriptive evidence on the unemployment rate and labour market transitions, clear cyclical features become apparent for the time period 1999 to 2013

(Figures 2.2 and 2.3): The unemployment rate and the transition from employment to unemployment were countercyclical, whereas the transition from unemployment to employment was procyclical. This pattern was especially strong in 2007 and 2008, the beginning of the Great Recession. In addition, the ﬁgures indicate that the extent of labour market reactions at the start of the recession and their persistence in the following years varied remarkably across countries.

In particular, the adverse responses in the unemployment rate during the Great Recession range between one percentage point in the Czech Republic to 14 percentage points in Spain

(Figure 2.2). In countries where the unemployment rate rose considerably, it remained stubbornly high until 2013. An exceptional case is Estonia. After the rate peaked in 2010, it

6Alternative measures of the business cycle are for example the output gap, the real interest rate and total factor productivity growth, which we apply in robustness tests (see Section 2.5.2). 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 17

Figure 2.2.: Unemployment rate by country, 1999-2013

AT BE CZ DE DK .3 .2 .1 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

EE ES FI FR GR .3 .2 .1 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

HU IT LU NO PL .3 .2 .1 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

PT SE SI SK UK .3 .2 .1 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

US .3 .2 .1 0 1999 2002 2005 2008 2011

See Figure 2.1 for the country codes. Source: EU-LFS, CPS, own calculation. decreased quickly, almost going down to its initial level. Only a few countries experienced hardly any change or even a decrease in unemployment during the observation period.

For example, the unemployment rate was stable in Austria, Belgium, Finland, Norway and Poland; in Germany it fell during most of the recession years.7

A heterogeneity of similar extent emerges for the transition rates from employment to unemployment and unemployment to employment. Changes in worker ﬂows from employment to unemployment were especially big in 2008, which is in line with expectations

(Figure 2.3): At the start of the recession, the large adverse shock raised job destruction, increasing the transition rates from employment to unemployment. Countries which were strongly hit by the economic downturn, such as Spain, Greece, Portugal and Estonia, experienced a substantial rise in the corresponding rate of up to 7 percentage points. However,

7See e.g. Burda and Hunt[2011] and Burda and Weder[2016] for an analysis of the German experience during the Great Recession. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 18

Figure 2.3.: Annual transition rates between employment and unemployment by country, 1999-2013

AT BE CZ DE DK .1 .6 .4 .05 .2 0 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

EE ES FI FR GR .1 .6 .4 .05 .2 0 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

HU IT LU NO PL .1 .6 .4 .05 .2 EtoU UtoE 0 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

PT SE SI SK UK .1 .6 .4 .05 .2 0 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

US .1 .05 0 1999 2002 2005 2008 2011

EtoU UtoE

See Figure 2.1 for the country codes. Transitions from employment to unemployment are plotted against the primary axis. Transitions from unemployment to employment are plotted against the secondary axis. Source: EU-LFS, CPS, own calculation. they display very different speeds of recovery. The transition rate in Estonia dropped to almost its pre-recession level in 2011, whereas in Greece it was still above the respective value in 2013. Again, Germany was an exception as the transition rate from employment to unemployment decreases slightly during the Great Recession.

The evolution of worker ﬂows from unemployment to employment shows pronounced trends in a number of countries, together with some business cycle turbulence (Figure 2.3).

In economic downturns job creation and, hence, hirings are lower. Therefore, the transition rate from unemployment to employment decreases. Indeed, this is the case for most of the countries during the Great Recession. The initial drop was the highest in Spain, Italy and

Norway where it equalled roughly 20 percentage points. However, some countries, such 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 19 as Austria and Poland, experienced an increase in this transition rate. Since the variation of the rate within a country over time is high, it is not feasible to draw conclusions on the speed of recovery.

2.3. Labour Market Institutions and their Interaction with Shocks

The main question arising from the descriptive evidence above is whether and how national labour market institutions are responsible for differences in the labour market reactions to macroeconomic shocks. Looking at the indicators for employment protection legislation and union density supports the intuition that labour market institutions exert an indirect, rather than a direct, inﬂuence on national labour markets, since the measures vary little over the observation period within countries (Figures 2.A.2 and 2.A.3 respectively). Moreover, the level of the indicators differs substantially across countries. To go back to the previous example: While Italy and Spain were similarly hit by the Great Re- cession, they display very dissimilar labour market reactions, which could be explained by union density being substantially higher in Italy compared to Spain. Taken together this indicates that institutions amplify or diminish the impact of economic turbulence.

Speciﬁcally, labour market institutions inﬂuence the transmission of economic shocks to national labour markets in a twofold manner. Initially, they affect the intensity of a shock hitting the labour market and later the adjustment process back to the steady-state level.

Institutions describing the ﬂexibility of labour markets by creating wage or employment rigidities exhibit both attributes, whereas institutions inﬂuencing the reservation wage and job search intensity relate mainly to the adjustment process.

Following Blanchard and Wolfers[2000] we capture the institutional setting of national labour markets by using eight indicators.8 They cover the unemployment insurance system, employment protection legislation (EPL), the collective bargaining system, active labour market policies (ALMP), and the tax burden of employees for each country. Ta-

8The institutional measures are the replacement rate of unemployment beneﬁts and their length, employment protection legislation, union coverage, union density, the level of wage bargaining, active labour market policies and the tax wedge. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 20 ble 2.A.1 reports descriptive statistics of our variables and illustrates their size, variation and availability.9 The correlation between the various institutions is overall strong and positive (Table 2.A.2). For instance, strict EPL is signiﬁcantly positively correlated with all indicators except for union density. This can introduce multicollinearity in our empirical model, which we address using different strategies in our benchmark models as well as in robustness tests (see Section 2.4 for a discussion).

We next describe the economic rationale for each of these indicators. From a theoretical point of view, trade unions take centre stage in the determination of wages. In a right-to- manage model [Nickell and Andrews 1983], unions and employer organisations negotiate over wages, which are then taken as given by individual ﬁrms in their decision over employment. Thus, higher trade union power is associated with higher wages. By contrast, in an efﬁcient bargaining model [McDonald and Solow 1981], both wage and employment are bargained over, which implies that an increase in trade union power does not necessarily cause higher wages.

Nevertheless, trade unions can create wage rigidities being especially relevant for the re- sponsiveness of wages to a change in aggregate economic conditions. In particular, with higher downward wage rigidity initially more job matches are destroyed as a reaction to a large adverse shock, leading to higher worker transitions from employment to unemployment and a stronger increase in unemployment [Bertola and Rogerson 1997; Goette, Sunde and Bauer 2007]. Yet, the overall inﬂuence of trade unions in economic turbulence is not clear-cut, because trade union’s aim to protect the jobs of their members [Freeman 1978;

Medoff 1979], which can mitigate extreme reactions of labour markets. On the one hand, the job security motive generates a scope for modiﬁcation at the intensive margin, e.g. reducing working hours in order to prevent job losses. This goes along with lower adverse labour market reactions and a faster recovery of the economy. On the other hand, the same motive fosters the segregation of labour markets making it harder for outsiders, the unemployed, to enter employment. Thus, during an economic upturn strong trade unions might

9See the appendix 2.B for a detailed description of the shocks and institutions variables, as well as the respective data sources. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 21 affect the buildup of the employment stock, thereby hindering the reduction of unemployment. Taken together, these effects lead to low variation in the unemployment and employment stocks over the business cycle. This is in line with evidence presented by Pierse and

McHale[2015], Goerke and Pannenberg[2011] and Ivlevs and Veliziotis[2017] who find that union membership decreases the probability of dismissal for the UK, Germany and countries from Central and Eastern Europe, respectively. Also, Hijzen and Martins[2016] investigating government-issued extensions in collective bargaining coverage confirm this relationship. Overall however, not only the size of the trade unions is important for wage formation and employment, but also the structure of collective wage bargaining [Traxler and Kittel 2000] which influences the capability of national labour markets to internalize detrimental effects caused by asymmetric information. We therefore consider three measures of wage-setting institutions in our empirical analysis: Union coverage, union density and the coordination of the bargaining process.

Employment protection legislation (EPL) measures employment rigidity. EPL represents the costs that arise for firms in case of the dismissal of an employee and is an indicator for the flexibility of a labour market. In a simple steady-state search-and-matching model of the labour market, the following mechanism holds [Mortensen and Pissarides 1999a]: The stricter EPL, the more costly it is for employers to lay off workers, which reduces worker outflows from employment. Because employers are forward-looking, it also decreases vacancy creation and therefore inflows to employment. Therefore, EPL lowers labour turnover with ambiguous effects on unemployment. There exists empirical evidence in line with this theory: Higher EPL is associated with lower aggregate labour market flows, and there is no clear association between EPL and the unemployment rate [Scarpetta 1996;

Nunziata 2002]. However, this picture changes when considering varying business cycle conditions as the impact of EPL on labour market dynamics has been shown to be more important under lower trend growth than under higher trend growth, both theoretically

[Bentolila and Bertola 1990] and empirically [Messina and Vallanti 2007].

The relation of EPL, economic turbulence and the labour market is thus straightforward.

In economic upturns ﬁrms will hire less employees if EPL is high, which leads to lower 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 22 employment growth. Similarly, at the beginning of recessions, strict EPL is associated with lower unemployment since ﬁrms cannot adjust properly to the situation. This mechanism also hinders the recovery process. Thus, over the business cycle, EPL affects both the initial impact of an economic shock as well as its persistence on the labour market. However, this effect is likely to vary over the business cycle. In our empirical analysis, we therefore study the link between EPL and labour market dynamics for the whole time period investigated, as well as separately by boom and recession periods. Furthermore, we use the EPL measure for regular employment as this is generally the most prevalent employment type. It applies to workers with permanent contracts.

Next, we concentrate on labour market institutions that inﬂuence individuals’ job search intensity and the reservation wage of the unemployed. In theory, both affect only the adjustment process of the labour market back to the steady-state after economic turbulence.

One of the most important institutions in this context is the unemployment insurance system. The likelihood of taking up a job decreases when unemployment benefits are higher and when benefit entitlements are longer, since these factors lower the incentives to search for work. At the same time a generous unemployment insurance pushes up the reservation wage due to lower opportunity costs of unemployment. Indeed, empirical evidence suggests that unemployment benefits have a significant adverse effect on unemployment and on worker flows from unemployment to employment [among others Nunziata 2002;

Nickell, Nunziata and Ochel 2005; Schmieder, von Wachter and Bender 2016].

As for cyclical features in this context, by the same reasoning we expect that during economic upturns, a more generous unemployment insurance system goes together with a higher unemployment rate and lower unemployment to employment transitions. During recessions a similar mechanism is likely to apply. However, the job market perspectives of unemployed persons worsen irrespective of the generosity of the unemployment insurance system, which means that the moral hazard induced by a generous unemployment insurance system may be lower in recessions than in booms [Schmieder, von Wachter and

Bender 2012]. Among the variables characterizing the unemployment insurance system, we choose the benefit replacement rate and the duration of unemployment benefits. While 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 23 the replacement rate captures the level of unemployment benefits relative to previous earnings, benefit duration measures how long individuals are entitled to unemployment bene-

ﬁts.

Taxes relate to both job search intensity and the reservation wage, too. The labour tax wedge measuring the difference between the labour costs to the employer and net take- home pay of the employee increases the reservation wage and reduces the efforts of an unemployed individual to search for a job. Therefore, it is associated with lower transitions from unemployment to employment and higher unemployment. However, Blanchard and

Wolfers[2000] argue that this effect is small, because the labour tax wedge contains, among others, payments for health beneﬁts and retirement. Nevertheless, many empirical studies

ﬁnd a strong adverse relationship between the tax wedge and unemployment [Belot and

Van Ours 2004; Nickell 1997].

On the labour supply side, the tax wedge can be expected to have a similar inﬂuence as the unemployment insurance system since both increase reservation wages. That is, during economic booms the tax wedge is adversely related to outﬂows from unemployment and to the unemployment rate. During recessions the interaction is expected to be negligible.

In addition, however, one can expect effects on the labour demand side, with a high tax wedge implying labour costs and thus reducing labour demand. This could amplify the negative labour-market effects in a recession.

Active labour market policies (ALMP) inﬂuence the labour market mainly via changes in job search intensity. ALMP programmes aim at reducing unemployment by improving the job matching process and by enhancing opportunities for unemployed to accumulate skills and work experience affecting their job search behaviour. Thus, unemployed individuals become more employable. In theory, programmes of this type lower the unemployment rate as transitions from unemployment to employment increase. In practice, this effect has been shown to depend strongly on the speciﬁc programme design [Card, Kluve and

Weber 2010, 2018]. Concerning the channelling and persistency property with respect to adverse macroeconomic shocks, ALMP does not inﬂuence the initial depth of a downturn, but in contrast exerts a positive impact on the recovery of the labour market by supporting 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 24 recently unemployed to get back into work. In an economic upturn the impact is expected to be small since the programmes are not very effective for long-term unemployed.

2.4. Methodology

The aim of our empirical analysis is to examine the medium-term developments of Eu- ropean and US labour markets dynamics over the time period 1999 to 2013. We apply the empirical methodology of Blanchard and Wolfers[2000] to investigate the importance of the interaction of macroeconomic shocks and labour market institutions for the unemployment rate and worker flow rates. Therefore, we estimate two models: The first model assumes that shocks are unobservable but common across countries, whereas the second model allows for observable and country-specific shocks. While the former uses a very general shock measure which incorporates the correlation of prevailing economic conditions between countries, the latter takes into account differences in GDP growth between countries, which includes the depth of the Great Recession.

The unobserved shocks model reads as follows:

j Λit = ci + dt + ∑ bj(dt ∗ Xit) + eit (2.1) j

where Λit is the dependent variable which is either the unemployment rate or a worker

flow rate in country i at time t, ci are country dummies and dt represent time dummies. As explained in more detail below, t represents a time period of three years. The time dummies serve as common unobserved shocks across countries. Their coefficients measure the direct j effects of these common economic shocks on national labour markets. Furthermore, Xit is the value of institution j in country i at time t. The coefficient of interest, bj, quantifies the interaction between shocks and institutions. In particular, the estimate captures the transmission property of the corresponding institution and, thus, depicts the indirect effect of institutions via shocks on the outcome variable. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 25

In the second model, the common observed shocks are replaced by an explicit country- speciﬁc macroeconomic shock measure, the annual growth rate of GDP. It has the following form: j Λit = ci + θYit + ∑ bj ∗ (θYit ∗ Xit) + eit (2.2) j where the notation is the same as before, only that Yit denotes the shock in terms of GDP growth.

Note a distinct feature of the main coefficient. Institutions enter the model only in the interaction term representing their transmission property for a given shock. We perform non-linear least squares estimations because the shock coefficient is simultaneously estimated as coefficient for the macroeconomic shock alone and for the interaction with institutions. Both empirical models account for the theoretical mechanisms invoked in Section

2.3. Hence, the same macroeconomic shock may generate very heterogeneous labour market reactions in countries with different institutions.

The regression sample is deﬁned as follows. It covers 20 European countries and the US.

We split the observation period from 1999 to 2013 into ﬁve three-year sub-periods, i.e. 1999 to 2001, 2002 to 2004, 2005 to 2007, 2008 to 2010, and 2011 to 2013. For each sub-period, we compute averages of yearly unemployment and transition rates, as well as of real annual

GDP growth rates. This implies three advantages compared to the use of annual data. First, as argued in Blanchard and Wolfers[2000], the slow movement of institutions only justi-

ﬁes a model in which all variables are summarized over a longer period of time. Second, it diminishes autocorrelation, i.e. it reduces the degree of ﬁrst-order autocorrelation in the error term, which would lead to wrong standard errors and inference. Finally, business cycle effects are smoothed, allowing us to abstract from short-run labour market reactions. The eight labour market institution measures described in detail in Section 2.3 are constructed as deviations from the cross-country mean following Blanchard and Wolfers[2000]. Table

2.A.3 illustrates descriptive statistics of all variables employed in the benchmark regressions. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 26

The main concerns for identification are endogeneity of labour market institutions and shock measures, multicollinearity of the institutional set, autocorrelation and heteroscedasticity. First, endogeneity arises from reverse causality between the evolution of labour markets on the one hand and institutions and shocks on the other hand. Put differently, labour market reforms can take place as a reaction to adverse or advantageous labour market conditions. Second, changes in GDP may be driven by labour market reactions. Therefore, our benchmark specification estimating equations (2.1) and (2.2) using time-varying institutional measures, is potentially subject to an endogeneity problem. In order to deal with these issues, we follow four strategies. Foremost, we reduce the endogeneity of GDP by using 3-year averages. With respect to institutional endogeneity, we restrict the variation in the institutional variables by considering only their values in the first year for each time window. In addition, we perform a robustness test in which we fix institutions to their level in the first years of our observation period, 1999. Thus, changes in labour market institutions are eliminated after 1999. Finally, we check whether the estimates are sensitive to substituting the measures with their respective 3-year-lagged values.

Multicollinearity between institutional measures arises if the indicators are strongly correlated with each other. As Table 2.A.2 illustrates, this is clearly an issue here. Moreover, institutions change very slowly over time. Therefore, the value of one institution in period t in a country is correlated with the same institution in the adjacent periods. Typically, the consequences of multicollinearity are particularly sensitive estimates and inﬂated standard errors. Hence, we run both models on ﬁve subsets of institutions to check the stability of our estimates.

Finally, autocorrelation and heteroscedasticity are of concern in our regression speciﬁca- tion. The application of 3-year-intervals of all variables should reduce the severity of this identiﬁcation threat. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 27

2.5. Results

2.5.1. Main Analysis

We begin to examine the role of labour market institutions for the reaction patterns of national labour markets over the business cycle with the unobserved shocks model introduced in Section 2.4. Table 2.1 displays the regression results of our three labour market outcomes, the unemployment rate, the transition rate from employment to unemployment and the transition rate from unemployment to employment.

The analysis indicates that the level of wage bargaining and the degree of union coverage account for a notable part of the heterogeneity in the trajectory of the unemployment rate across countries. In particular, a coordinated wage bargaining process reduces the impact of shocks to national labour markets; furthermore, there is some (weak) evidence that a higher union coverage rate amplifies the magnitude of the reactions. Both relationships are in line with Blanchard and Wolfers’[2000] findings and with theoretical predictions and consistent with the view that trade unions influence labour markets ambiguously. On the one hand, strong trade unions establish wage rigidities reinforcing negative shocks. On the other hand, they relate to more coordination providing a basis for internalising detrimental effects, which turns out to be the dominant influence in the subsequent analysis.

Furthermore, the results suggest that the other labour market institutions considered in our analysis cannot explain the observed movements in the unemployment rate.

Turning to labour market transitions between employment and unemployment, we get the general picture that the interaction of shocks and institutions is a negligible factor for these

flows in the unobserved shocks model. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 28 ) ) ) ) 0.12 0.137 0.79 0.10 0.005 0.71 − − − − )( ) )( )( 0.17 0.16 1.123 0.02 0.09 0.065 0.011 − − − − )( )( )( 0.00 0.60 0.86 0.027 0.388 − − )( ) )() ) )( )()( 0.75 1.50 0.030 0.000 1.30 0.238 0.56 0.359 0.29 0.10 0.38 0.73 0.034 − − − − − − − − )( )( )()( 0.1/0.05/0.01. = 0.46 0.66 0.010 0.015 0.24 0.006 0.76 α − − − − )( )( )( )( )( refers to 0.59 0.28 0.440 1.05 0.94 1.13 0.025 ∗∗∗ − − − − / ∗∗ / )( )( )()( ∗ 0.32 0.30 0.25 1.63 0.034 − − )( )( )( 0.56 0.008 0.004 0.020 0.23 0.77 0.006 0.008 0.017 − − − − ∗∗ )( )()( )( )( )( )( )( 0.37 0.94 2.08 0.28 0.0840.14 0.318 0.1311.761 0.215 0.1260.20 0.315 0.169 1.241 0.157 1.147 0.022 1.66 0.41 ( − − − − − − − − ( )( )( )( )( 1.47 0.010 0.55 0.054 0.02 1.07 0.021 0.067 ( − − − − ∗∗ ∗ )( )( )( )( )( Table 2.1. : Unobserved shocks model 2.25 1.811 1.00 1.82 0.73 0.160 0.28 − − − − ( )( )( )( )( 1.26 0.290 0.54 0.048 0.026 0.0341.43 0.027 0.014 1.66 − − − − − − ( )( )( )( 1.49 0.017 0.055 0.04 1.31 ( − − − − ( ( Nonlinear least-squares estimation. Time and country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSpecifications contribution III of and institutions V and shocks includingin to the the the measure table shared for above. interaction coordination coefficient. Source: do EU-LFS, not OECD converge Economic for Outlook the 95, dependent ICTWSS, variable own transition calculations. from unemployment to employment and are therefore not displayed Employment Protection Benefit Length 0.063 0.070 Dependent VariableSpecificationReplacement Rate I IICoordination UALMP IIITax WedgeObservations IVR-squared V 105 0.708 I 105 0.718 105 II 0.736 E to U 0.709 105 III 0.024 0.742 105 IV 0.016 0.711 102 0.725 V 102 0.726 I 102 0.712 U to E 0.736 II 102 0.699 102 IV 0.017 0.726 104 0.004 0.700 104 104 Union Density Union Coverage 0.025 0.041 0.061 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 29 ∗ ∗ ) ) ) ) ) ) ) ) ) 0.74 0.16 0.013 0.16 0.552 1.71 1.96 0.012 1.04 1.18 0.98 0.104 0.91 0.082 − − − − − − − − − − )( )()( )( )( 0.497 0.592 0.01 1.63 0.46 0.027 0.75 0.65 0.040 − − − − ∗ )( )( )( )()()( 0.49 0.549 0.74 1.85 0.32 0.585 0.013 1.14 1.54 0.088 − − − − ∗ )( )( )( )( )( 1.85 0.66 0.36 1.32 0.512 1.53 0.084 − − − − )( )( )( )( 0.502 0.546 1.65 0.44 0.01 0.68 0.025 0.0160.008 0.014 − − − − ∗∗∗ ∗∗∗ )( )( )( )( )()( )()()( 0.93 1.65 1.13 0.020 0.242 6.32 0.32 0.54 0.51 1.02 0.012 0.115 3.71 0.047 − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.226 5.46 1.01 0.007 0.002 0.016 0.038 0.033 0.018 0.026 0.18 0.48 0.18 − − − − 0.1/0.05/0.01. ∗∗∗ ∗∗∗ )( )( )( )( )( )( = α 0.233 6.17 0.56 1.54 0.63 0.53 0.145 0.123 3.45 0.043 − − − − − − refers to ∗∗∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 0.233 6.19 0.87 1.59 0.54 0.121 3.47 0.041 / − − − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.226 5.49 1.00 0.006 0.005 0.004 0.49 0.17 − − − − ∗∗ ∗∗∗ )( )( )()( )( )( )( )()( 0.507 3.92 0.56 0.11 0.73 0.06 0.0040.68 0.027 0.266 0.125 0.002 0.003 0.020 0.020 0.003 1.3331.39 0.963 1.818 2.43 0.08 0.002 0.003 0.052 ( − − − − − − − − − − − − ( Table 2.2. : Observed shocks model ∗∗∗ )( )( )( )( )( 0.82 0.459 3.51 0.0010.07 0.006 0.13 0.01 0.050 0.000 ( − − − − − − − − ∗∗∗ ∗∗ )( )( )( )( )( )( 0.478 3.76 0.85 0.88 0.84 0.339 0.57 2.22 0.046 ( − − − − − − ∗∗ ∗∗∗ )( )( )( )( )( 0.477 3.77 1.18 0.79 0.298 0.90 2.19 0.043 − − − − − − ( ∗∗∗ )( )( )( )( 0.457 3.50 0.07 0.10 0.039 0.05 0.001 0.019 0.018 ( − − − − − − ( ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Union Coverage 0.001 0.013 0.010 Coordination 0.152 0.194 0.085 0.162 0.272 0.675 ALMP Union Density Tax Wedge 0.022 Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 105 I 0.947 105 0.951 II E 105 to U 0.951 III 0.947 105 IV 0.952 105 0.955 V 102 0.963 102 I 0.964 102 0.955 II 0.965 102 U to E III 0.975 102 0.977 IV 104 0.977 V 104 0.975 0.978 104 104 104 Benefit Length Employment Protection Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 30

As the unobserved shocks model displays a general lack of significant explanatory variables, the question arises whether this model is well-specified. In this context, it is worth noting that all executed regressions with the unobserved shocks model exhibit an R-squared value ranging from 0.70 to 0.74 (see Table 2.1). By contrast, the R-squared values of the observed shocks model lie in the interval from 0.95 to 0.98 (see Table 2.2). This implies that the assumption that the business cycle is the same across countries is not adequate for the period analysed, 1999 to 2013. Furthermore, with respect to the fit, the observed country- specific shocks model outperforms the unobserved shocks model.10

Contrary to this result, Blanchard and Wolfers[2000] conclude that the empirical model with unobserved shocks predicts the trajectory of unemployment patterns better. As they investigate the time period 1965 to 1994, one explanation for this discrepancy could be that the heterogeneity of recessions and economic upturns across countries has increased over time. Indeed, Table 2.A.4 shows that this holds true. The variation in annual GDP growth rates across countries in our sample during the latest recessions rose from one recession to the next. Therefore, we proceed with the observed shocks model as our benchmark model for the further investigation.

As expected, the shock measure in the observed shocks model, the annual GDP growth rate, is highly signiﬁcant and negatively correlated with the unemployment rate and transitions out of employment (Table 2.2). The correlation between GDP growth and worker

flows from unemployment to employment is positive. The significance of this correlation is not as clear-cut as for the worker flows in the opposite direction, which is not surprising as the matching process on the labour market weakens the influence of the business cycle.

For countries with mean values for all institutions (X = 0) unemployment decreases by about 0.5 percentage points as GDP growth rises by one percentage point. This is very similar to reported coefﬁcients from Okun’s Law [Perman, Stephan and Tavera´ 2015].

10The root-mean-square error of the unemployment rate regression including the full set of institutions corresponds to 2.099 for the unobserved and to 2.088 for the observed model. The values of the employment to unemployment transition are 0.768 and 0.600, respectively. A comparison of the unemployment to employment transition is impossible since the unobserved model does not converge. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 31

With respect to the interaction of shocks and institutions, the unemployment beneﬁt system

(the replacement rate and the length of benefit entitlements) does not play a role neither for the unemployment rate, nor for worker transitions between employment and unemployment. This is probably due to the fact that our period of analysis is strongly influenced by the Great Recession, i.e. the results are mainly driven by the labour market reactions during an economic downturn. As spelt out in detail in Section 2.3, this means that the impact of the unemployment benefit system can be expected to be relatively weak. The same explanation is likely to apply to the non-significance of ALMP. It is much more important during an economic upswing than during a recession. Again, as our results are mainly driven by the Great Recession, this is a probable explanation for the lack of significance of the variable. The tax wedge is also hardly significant. As it turns out below, this picture changes when looking at different business cycle phases.

The interaction of GDP and employment protection also does not seem to play a role for the evolution of labour markets, both with respect to stocks and ﬂows. For the unemployment stock, this is in line with the literature which does not yield consistent results with respect to the effects of EPL on the unemployment stock [e.g. Bertola 1990; Boeri

1999; Nickell, Nunziata and Ochel 2005]. As for ﬂows, a standard steady-state search-and- matching model such as Mortensen and Pissarides[1994], predicts a clear negative correlation between worker ﬂows and EPL. However, recent cross-country evidence shows that employment protection does not necessarily reduce transitions from employment to non-employment [Bassanini and Garnero 2013].

Furthermore, the effect of EPL on labour market dynamics seems to be weaker when trend growth is higher, and stronger when trend growth is lower [Bentolila and Bertola 1990;

Messina and Vallanti 2007]. Our results lend support to this conclusion because the time period analysed includes a relatively long period of sustained growth (i.e. the time period before the Great Recession), which dampens the effect of EPL on worker ﬂows. This picture is corroborated by our separate analysis of upturns and downturns, which is described below. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 32

As for the importance of trade unions, the signiﬁcant results for coordination and union coverage in the unobserved shocks model do not materialize in the observed shocks model.

Nevertheless, the direction of the respective estimates are broadly the same. Furthermore, another variable capturing the role of trade unions, union density, is generally significant at conventional levels for all outcomes and specifications. In particular, union density is negatively correlated with the unemployment rate, worker flows from employment to unemployment, as well as worker flows from unemployment to employment (although only weakly so in the latter case). Evaluating the equation for unemployment at a GDP growth of -1 percent implies that an increase in union density of 1 percentage point is associated with a decrease in unemployment of about 0.02 percentage points. Given that union density varies between 6.9% and 81.6% in our country sample (see Table 2.A.1), this seems like a large impact.

This relationship can be illustrated by correlating union density with changes in unemployment (Figure 2.4, Panel a) and with changes in the transition rate from employment to unemployment (Figure 2.4, Panel b), where the change considered is between the years

1999 and 2013. From these correlations, it becomes apparent that countries with higher union density experienced a lower increase in unemployment, and lower transition rates from employment to unemployment were a major contributing factor to this. Our results, in turn, show that this conclusion remains intact when considering the size of the economic shock, particularly in interaction with labour market institutions.

Since it is very consistent across our model specifications, we examine the role of union density in more detail. The estimations suggest that higher union density in a country is associated with more moderate labour market reactions to shocks. The observed pattern may be driven by trade unions aiming to protect employed workers from unemployment, which leads to segregated labour markets. These are characterised by a situation where insiders, i.e. workers employed in stable jobs, gain and outsiders, i.e. persons who are not employed or very rarely so, face difficulties to enter the labour market at all. An alternative mechanism by which unions influence labour market dynamics is an adjustment at the intensive margin such as the reduction of working hours during the Great Recession, 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 33

Figure 2.4.: Relationship between union density and labour market reactions, 1999 and 2013 (b) Transition rate from employment to un- (a) Unemployment rate employment

GR 4 15 PT

PT SK SI GR ES ES AT 10 2 PL LU

IT 5 BE US LU HU HU US DK Change in U SI UK Change in EtoU

UK DK 0 FR SE IT AT NO SE NO 0 BE CZ CZ SK FR PL DE FI EE DE FI EE −5 −2 0 20 40 60 80 0 20 40 60 80 Union Density Union Density

See Figure 2.1 for the country codes. Source: EU-LFS, CPS, ICTWSS, own calculation. implemented for example in Germany by a short-time work scheme. However, for our main ﬁnding, the indirect inﬂuence of union density, using working hours provided by

Ohanian and Raffo[2012], we find in the Great Recession period neither a negative correlation between union density and working hours nor a statistically significant negative relationship using fixed effects OLS regression techniques. Therefore, the intensive margin does not play an important role in our context. Following either of these lines of reasoning – segregated labour markets or adjustment along the intensive margin –, we expect to

find that union density relates to lower unemployment outflows during economic growth periods and lower employment outflows in recessions.

To test this hypothesis, we conduct separate regressions for periods of economic boom and bust. Accordingly, we replace the interaction term with a growth or recession dummy, respectively.11 Hence, the model still comprises the intensity of the shock. In the preceding analysis we have used 3-year-intervals instead of yearly observations to account for potential endogeneity in the shock and institution measures between the years. For this

11 j j This means that in equation 2.2, we use Yit ∗ Xit ∗ Dit instead of Yit ∗ Xit, where Dit takes the value 1 in case of a recession, and 0 otherwise, or the reverse in case of an economic boom. A recession is defined as a negative or zero yearly GDP growth rate. Accordingly, an economic boom is defined by a positive GDP growth rate. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 34 investigation we now rely on annual data in order to have sufficient observations.12 In- deed, Tables 2.A.6 and 2.A.7 support the notion that union membership is related with more robust labour market reactions over the business cycle. The coefficients on union density imply thatcountries with a high union membership display a smaller reduction in unemployment during an economic upturn. The same result materialises in recessions, i.e. strong union density reduces the adverse impact on the unemployment rate. The results for the worker flows confirm these findings.

Furthermore, the separate analysis with respect to business cycle periods shows that in economic growth periods, the level of wage coordination ampliﬁes national labour market reactions. The opposite holds true for recession periods. Therefore, a high level of wage coordination is a favourable inﬂuence, which is in line with the literature, among others

Bassanini and Duval[2006], Gal and Theising[2015] and Bertola[2017]. Moreover, during recessions the benefit length and the tax wedge intensify, whereas EPL weakly cushions the adverse effects on the labour market. The first result can be explained by a longer benefit length slowing down the unemployment exit rate; however, this result seems particular to the Great Recession, as research for earlier time periods has shown that the unemployment benefit system plays a more important role in boom than in recession periods [Schmieder, von Wachter and Bender 2012]. The negative link with the tax wedge is in line with existing research [Belot and Van Ours 2004; Nickell 1997]. Interestingly, this result only materialises in the recession sample, which can be viewed as an indication that the negative labour- demand effects described in Section 2.3 play a dominant role in this context. The result on

EPL stands in contrast to the non-significant effect of EPL on labour market flows indicated above for the entire time period. However, our finding of a significantly negative effect of

EPL for recession periods is in line with and conﬁrms the result by Bentolila and Bertola

[1990] and Messina and Vallanti[2007] that the impact of EPL is stronger the weaker trend growth is, as well as evidence on the stabilising role of EPL on job tenure during the Great

Recession in Europe [Bachmann and Felder 2018].

12Table 2.A.5 displays the estimates of the benchmark model using annual data. The results are very similar compared to applying three-year windows. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 35

Overall, our investigation indicates that institution variables related to trade unions exhibit explanatory power for the prevailing patterns in labour market dynamics across countries.

Speciﬁcally, we ﬁnd that the interaction of shocks with union density weakens labour market reactions. Strong trade unions tend to lower employment growth in economic upturns and employment contractions in recessions, which overall results in more moderate reactions of the unemployment rate over the business cycle.

2.5.2. Robustness

In order to support our conclusions with respect to the determining role of trade unions in shaping unemployment and worker transitions between employment and unemployment, we run a battery of robustness tests.

First, we present estimates testing for endogeneity in our benchmark regression. Endo- geneity poses a threat for identification, because of the potential for reverse causality between the evolution of national labour markets on the one hand, and institutions and shocks on the other hand. Institutional reforms may be induced by unfavourable labour market conditions, and changes in labour market dynamics may influence the business cycle, respectively. We therefore run regressions with (i) institution measures lagged by one period and (ii) institution measures fixed at their values for the year 1999 instead of using contemporary values.

The results of these sensitivity tests show that generally, our benchmark conclusions are robust (Tables 2.A.8 and 2.A.9). In particular, the moderating role of trade unions for labour markets is a consistent result. However, in the specifications controlling for union density, the lagged model indicates that union coverage adversely influences worker flows from employment to unemployment. This points to the two opposing features of trade unions discussed in Section 2.3. On the one hand, trade unions establish wage rigidities which reinforce negative macroeconomic shocks, while on the other hand trade unions aim at making jobs more secure, which has effects in the opposite direction. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 36

To investigate the endogenous nature of the GDP growth rate with unemployment we run Granger causality tests. The results suggest that reverse causality is not an important concern for most countries since we can reject that unemployment Granger causes GDP growth for 15 countries. Furthermore, we replace the GDP growth rate with its one period lagged value as well as with alternative measures for macroeconomic conditions, i.e. the output gap, the real interest rate and total factor productivity growth. These robustness tests do not change our results and conclusions signiﬁcantly (results available from the authors upon request).

Second, we assess the sensitivity of our benchmark model to changes in the sample. We exclude countries that represent extreme cases with respect to the intensity of the Great

Recession. Portugal, which was hit strongly by the economic recession and which features very strong employment protection, is omitted in Table 2.A.10. The smoothing behaviour of trade unions on labour market dynamics over the business cycle is conﬁrmed by the regression. Additionally, the estimates without the observations of Portugal imply that the effects of a negative economic shock on national labour markets are less pronounced in countries with higher EPL, which is line with the investigation of economic upturns only. Excluding Germany (Table 2.A.11), a country where the labour market was hardly affected by the Great Recession, also leaves the conclusions of the main speciﬁcation intact. Additionally, we run separate regressions where we remove two country groups (the

Nordic and the East European Countries), and individual countries (Belgium, Denmark,

France, Spain and the US) from the sample. The ﬁndings are robust to these alterations

(results available from the authors upon request). Moreover, we estimate the inﬂuence of institutions on labour markets separately by gender and for young and old individuals.

Institutions have a higher relevance for men and young individuals (results available from the authors upon request).

Third, we check the robustness by accounting for temporary employment. Since temporary workers are less costly to lay off for firms than regular workers, the share of temporary employees is an indicator for the flexibility of labour markets. We expect that countries with a high rate of temporary workers experience higher adverse reactions in unemployment 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 37 and in worker outflows from employment. Indeed, Table 2.A.12 provides evidence for this perception. Controlling for temporary employment, the transition rate from employment to unemployment is not adversely influenced by union coverage as suggested by other specifications. This may be due to the additional collinearity imposed on the model.

Nevertheless, the importance of union density in inﬂuencing national labour markets is strongly reﬂected.

Fourth, we assess whether including movements out of the labour force change our main results. This is of interest as it has been argued that such worker ﬂows are an important aspect of the cyclical features of the labour market Ebell[2011]. We therefore extend our analysis to aggregate employment outﬂows to both unemployment and nonparticipation.

The corresponding results presented in Table 2.A.13 are similar to the estimation results for transitions from employment to unemployment. Therefore, taking into account the participation margin does not alter our main results.

2.6. Conclusion

In this paper, we examine the reasons for cross-country differences in labour market dynamics for a large number of European countries as well as the US for the time period

1999 to 2013. Thereby, we focus on both the unemployment rate and worker ﬂows between employment and unemployment, which we compute from micro data at the worker level. In our analysis, we employ the methodology of Blanchard and Wolfers[2000] to separately identify the impact of shocks on the one hand, and the transmission of shocks to national labour markets through the institutional framework on the other hand. We thus extend the existing literature by (i) explicitly analysing worker ﬂows, in addition to the unemployment rate, and (ii) analysing the time period of the Great Recession as well as the preceding decade.

Our results suggest that institutions play an important channeling role of macroeconomic shocks to national labour markets. While this is not the case for the unemployment benefit system, the results of our empirical analysis provide evidence for the importance of trade 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 38 unions in this context. Specifically, union density, i.e. the share of union members, relates to more moderate labour market reactions to shocks. This result is more pronounced for worker flows from employment to unemployment than for worker flows from unemployment to employment. In particular, the analysis shows that trade unions tend to reduce employment growth in economic upturns and employment contractions in recessions. As a result, the unemployment rate in countries with stronger trade unions features a lower variation over the business cycle. Our results furthermore confirm findings from the literature that employment protection legislation becomes more important for labour market flows when the trend growth of the economy is low [e.g. Bentolila and Bertola 1990;

Messina and Vallanti 2007]. Finally, we show the negative link between the tax wedge and unemployment only materializes in the recession and is therefore likely to be driven by labour-demand considerations.

In order to analyse the robustness of our results, we conduct a number of tests. It turns out that the potential endogeneity of both shocks and institutions, country outliers with respect to institutions and the depth of the recession, and the importance of temporary employment in national labour markets do not signiﬁcantly affect our analysis. Further- more, including the participation margin, i.e. worker transitions out of the labour market, also does not alter our main conclusions.

Our results with respect to trade unions, i.e. that union density is associated with lower unemployment and lower inﬂows into unemployment, stand in contrast to Blanchard and

Wolfers[2000] and Nickell[1997] who show that up to the mid-1990s strong unions were positively associated with unemployment. However, the results are partially in line with

Bassanini and Duval[2006] who study a more recent period from 1982 to 2003 and who also ﬁnd a dampening role of unions. Moreover, there is recent evidence from a number of European countries which shows that union membership reduces a person’s layoff probability [Goerke and Pannenberg 2011; Ivlevs and Veliziotis 2017; Pierse and McHale

2015].

The welfare implications of our results with respect to trade unions are not clear-cut. On the one hand, lower volatility of unemployment (i.e. lower unemployment inflows and 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 39 outflows) is associated with higher subjective well-being [Wolfers 2003], which means that unions would be welfare-enhancing. On the other hand, lower volatility of unemployment is likely to go together with a lower permeability of labour markets. This implies the existence of segregated labour markets where a part of the workforce benefits from stable employment relationships while another part of the workforce has great difficulties entering the labour market or only attains low-paid and/or unstable jobs. In this respect, unions would be welfare-decreasing. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 40

2.A. Figures and Tables

Figure 2.A.1.: Annual GDP growth by country, 1999-2013

AT BE CZ DE DK .15 0 −.15 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

EE ES FI FR GR .15 0 −.15 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

HU IT LU NO PL .15 0 −.15 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

PT SE SI SK UK .15 0 −.15 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

US .15 0 −.15 1999 2002 2005 2008 2011

Country codes: AT: Austria, BE: Belgium, CZ: Czech Republic, DE: Germany, DK: Denmark, EE: Estonia, ES: Spain, FI: Finland, FR: France, GR: Greece, HU: Hungary, IT: Italy, LU: Luxembourg, NO: Norway, PL: Poland, PT: Portugal, SE: Sweden, SI: Slovenia, SK: Slovak Republic, UK: United Kingdom, US: United States of America. Source: Economic Outlook No. 95, own calculation. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 41

Figure 2.A.2.: Employment protection legislation by country, 1999-2013

AT BE CZ DE DK 5 2.5 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

EE ES FI FR GR 5 2.5 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

HU IT LU NO PL 5 2.5 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

PT SE SI SK UK 5 2.5 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

US 5 2.5 0 1999 2002 2005 2008 2011

See Figure 2.A.1 for the country codes. Source: OECD Indicators of Employment Protection (2013).

Figure 2.A.3.: Union density rate by country, 1999-2013

AT BE CZ DE DK .8 .4 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

EE ES FI FR GR .8 .4 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

HU IT LU NO PL .8 .4 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

PT SE SI SK UK .8 .4 0 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011 1999 2002 2005 2008 2011

US .8 .4 0 1999 2002 2005 2008 2011

See Figure 2.A.1 for the country codes. Source: ICTWSS. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 42

Table 2.A.1.: Descriptive statistics of labour market institutions

Institution Measure Obs. Mean Std. Dev. Min Max

Replacement Rate 105 37.26 15.51 19.44 71.67 Beneﬁt Length 105 17.37 12.21 5.00 48.00 Employment Protection 105 2.36 0.77 0.26 4.58 Union Coverage 105 64.63 28.02 0 100.00 Union Density 105 32.76 20.50 6.89 81.58 Coordination 105 2.83 1.31 1.00 5.00 ALMP 105 0.62 0.45 0.07 2.21 Tax Wedge 105 31.94 8.24 9.88 44.32 Source: OECD Economic Outlook 95, ICTWSS, own calculations.

Table 2.A.2.: Pairwise correlations across labour market institutions, total variation Replacement Rate Beneﬁt Length Employment Protection Union Coverage Union Density Coordination ALMP

Beneﬁt Length 0.78∗∗∗ 1.00 Employment Protection 0.12 0.02 1.00 Union Coverage 0.63∗∗∗ 0.47∗∗∗ 0.37∗∗∗ 1.00 Union Density 0.48∗∗∗ 0.49∗∗∗ −0.012 0.52∗∗∗ 1.00 Coordination 0.58∗∗∗ 0.49∗∗∗ 0.20∗∗ 0.74∗∗∗ 0.67∗∗∗ 1.00 ALMP 0.67∗∗∗ 0.57∗∗∗ 0.10 0.56∗∗∗ 0.59∗∗∗ 0.51∗∗∗ 1.00 Tax Wedge 0.27∗∗∗ 0.27∗∗∗ 0.16 0.40∗∗∗ 0.17∗ 0.43∗∗∗ 0.36∗∗∗ ∗ / ∗∗ / ∗∗∗ refers to α = 0.1/0.05/0.01. Source: OECD Economic Outlook 95, ICTWSS, own calculations.

Table 2.A.3.: Descriptive statistics of the regression sample

Institution Measure Obs. Mean Std. Dev. Min Max

Unemployment rate 105 8.63 4.15 2.19 24.41 Employment to Unemployment Rate 102 2.85 1.50 0.30 7.97 Unemployment to Employment Rate 104 31.11 9.46 9.67 55.72 GDP Growth 105 1.991 2.49 −5.98 8.81 Replacement Rate 105 0.00 15.51 −17.82 34.42 Beneﬁt Length 105 0.00 12.21 −12.37 30.63 Employment Protection 105 0.00 0.77 −2.11 2.22 Union Coverage 105 0.00 28.02 −64.63 35.37 Union Density 105 0.00 20.50 −25.87 48.81 Coordination 105 0.00 1.31 −1.83 2.17 ALMP 105 0.00 0.44 −0.55 1.59 Tax Wedge 105 0.00 8.24 −22.06 12.38 Source: EU-LFS, OECD Economic Outlook 95, ICTWSS, own calculations.

Table 2.A.4.: Heterogeneity of main economic recessions Recessions SD in GDP growth 1973-1975 0.962 1980-1982 1.115 2001-2003 1.593 2008-2009 2.242 SD (standard deviation) of the country average of yearly growth rates in the corresponding recession periods. Source: OECD Economic Outlook 95, own calculations. 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 43 ) ) ) ) ) ) ) ) ) 0.78 0.81 0.376 2.378 0.79 0.87 0.468 0.17 0.71 0.67 0.181 0.81 0.154 0.88 − − − − − − − − )( )( )( )( )( 0.61 0.65 1.967 0.186 0.36 0.28 0.75 0.041 − − − − − − )( )( )()()( )( 0.68 0.88 0.01 0.098 0.59 0.88 2.192 0.97 0.093 − − − − − − )( )( )( )( )( 0.59 0.88 2.192 0.76 0.98 0.88 0.093 0.097 − − − − − − )( )( )( )( 0.151 0.1950.62 0.1950.66 0.1490.34 0.176 0.76 1.973 0.186 − − − − ∗∗∗ ∗ ∗∗∗ )( )()( )()( )( )()()( 1.38 0.63 3.80 0.28 0.36 0.035 0.187 7.86 0.45 1.93 0.09 0.009 − − − − − − ∗∗∗ )( )( )( )( )( 0.14 1.27 0.001 0.009 0.020 0.048 0.047 0.021 0.058 7.41 0.15 0.177 0.98 0.002 − − − − 0.1/0.05/0.01. ∗∗∗ ∗∗∗ ∗ )( )( )( )( )( )( = α 0.007 1.13 8.06 0.22 0.191 1.66 0.80 3.72 0.018 0.031 − − − − refers to ∗ ∗∗∗ ∗∗∗ ∗ )( )( )( )( )( ∗∗∗ 3.63 0.030 8.03 1.76 0.189 0.31 1.67 / − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.0000.05 0.010 0.176 7.39 1.31 0.22 − − − − ∗∗∗ ∗∗∗ )( )()( )( )( )( )()()( 0.284 4.78 0.57 1.05 0.80 0.71 0.260 0.286 0.0620.90 0.044 0.277 0.019 2.99 0.8981.03 0.57 0.146 8.524 0.049 0.023 0.003 0.018 ( ( − − − − − − − − ∗∗∗ )( )( )( )( )( 1.18 0.253 4.29 0.09 0.33 0.91 0.020 ( − − − − ∗∗∗ ∗∗∗ Table 2.A.5. : Observed shocks model, annual data )()( )( )( )( )( 0.283 4.80 0.84 0.0111.06 0.003 0.011 0.24 0.273 0.032 1.08 2.90 0.044 ( − − − − − − ∗ ∗∗∗ ∗∗∗ )( )( )( )( )( 0.278 4.72 0.71 1.78 0.235 0.25 2.76 0.041 − − − − − − ( ∗∗∗ )( )( )( )( 0.251 4.26 0.14 0.43 0.84 0.019 0.005 0.005 ( ( − − − − ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Union Coverage 0.004 0.017 Coordination 0.233 0.197 0.102 0.083 0.007 0.197 ALMP Union Density Tax Wedge 0.029 0.013 0.014 0.005 Dependent VariableSpecificationGDP Growth Benefit Length I II U III IVObservationsR-squared V 315 I 0.927 315 0.019 0.930 II E 315 to U 0.930 III 0.928 315 IV 0.931 315 0.887 V 315 0.893 315 I 0.893 315 0.888 II 0.040 0.895 315 U to E III 0.925 315 0.926 IV 315 0.926 V 315 0.925 0.928 315 315 315 Employment Protection 0.050 Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 44 ) ) ) ) ) ) ) ) ) 0.86 1.24 1.10 1.714 0.62 1.02 1.06 0.50 0.207 1.21 0.214 0.380 1.31 − − − − − − − − )( )( )( )( )( 0.69 0.68 2.500 0.12 0.32 0.068 0.84 0.251 − − − − − − )( )()()( )( )( 0.55 0.95 0.6430.45 0.713 0.183 0.90 0.65 2.910 1.03 0.138 − − − − − − )( )( )( )( )( 0.89 0.67 0.95 0.64 2.741 1.04 0.174 0.129 − − − − − − )( )( )()( 0.09 0.71 0.161 0.196 0.195 0.159 0.248 0.85 0.69 2.556 0.257 − − − − ∗∗∗ ∗∗ ∗∗∗ )( )( )( )( )()()()( )( 1.02 0.402 2.40 3.89 0.1610.26 0.88 8.763 0.019 0.050 0.60 0.186 8.28 1.24 0.00 0.000 − − − − − − − − ∗∗∗ )( )( )( )( )( 0.42 1.40 7.84 0.01 0.179 0.000 0.19 − − 0.1/0.05/0.01. ∗∗ ∗∗∗ ∗∗∗ )( )( )( )( )( )( = α 0.96 0.008 0.003 0.008 0.006 0.050 0.039 0.008 0.027 0.52 0.381 2.33 8.42 0.187 3.85 1.16 0.047 − − − − refers to ∗ ∗∗∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 1.84 3.50 0.042 8.35 0.88 0.187 1.49 / − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.43 1.41 0.179 7.86 0.20 − − ∗∗ ∗ ∗ ∗∗∗ ∗∗∗ )( )( )( )( )( )()()( )( 0.70 0.779 0.281 5.07 0.35 2.57 0.0110.28 3.48 1.846 1.83 1.93 0.031 0.288 0.072 0.084 0.41 0.016 0.003 0.023 0.018 0.003 ( − − − − − − − − − − ( ∗∗∗ )( )( )( )( )( 0.87 0.033 0.63 0.248 4.41 0.98 0.76 0.023 − − − − − − ( ∗∗∗ ∗∗ ∗∗∗ )( )( )( )( )( )( 0.85 0.012 0.008 0.009 0.003 0.014 0.267 4.85 0.04 2.30 3.11 0.075 0.17 0.005 ( − − − − − − ∗ ∗∗∗ ∗∗∗ Table 2.A.6. : Observed shocks model, economic growth periods )( )( )( )( )( 1.74 0.268 4.81 0.43 2.83 0.064 0.17 − − − − ( ∗∗∗ )( )( )( )( 0.56 0.250 4.44 0.91 0.91 0.028 0.005 ( ( − − − − ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Union Coverage 0.007 0.023 Coordination 0.713 ALMP Employment Protection 0.472Union Density 0.203 0.019Tax Wedge 0.518 0.156 0.411 0.238 0.139 0.410 0.164 Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 315 I 0.928 315 0.930 II E 315 to U 0.932 III 0.928 315 IV 0.933 315 0.888 V 315 0.893 315 I 0.895 315 0.888 II 0.896 315 U to E III 0.926 315 0.927 IV 315 0.927 V 315 0.926 0.929 315 315 315 Benefit Length Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 45 ) ) ) ) ) ) ) ) ) 0.72 0.58 3.660 0.29 1.835 0.01 0.51 0.487 3.431 0.64 0.75 0.79 0.128 0.295 1.00 − − − − − − − − − − )( )( )( )( )( 0.42 0.15 0.56 1.216 0.46 1.24 0.105 − − − − )( )( )( )()( )( 0.57 0.66 1.735 0.35 1.824 0.82 0.48 1.20 0.056 0.065 − − − − − − − − )( )( )( )( )( 0.62 0.68 0.36 1.825 0.56 1.32 0.060 0.061 − − − − − − )( )( )()( 0.44 0.0330.48 0.053 0.120 0.032 0.176 0.240 0.258 0.237 0.240 0.204 1.25 0.57 1.271 0.106 − − − − ∗∗∗ ∗∗∗ ∗∗ ∗∗∗ ∗∗∗ )( )()()( )( )( )( )( )( 1.37 2.77 2.64 1.32 0.043 0.025 0.003 0.70 0.159 6.78 0.132 2.21 1.051 1.51 0.387 3.06 0.078 0.098 − − − − − − − − − − ∗ ∗∗∗ )( )( )( )( )( 0.050 1.75 0.0090.79 0.058 0.75 7.58 0.174 0.22 − − − − 0.1/0.05/0.01. ∗∗ ∗∗ ∗∗ ∗∗∗ )( )( )( )( )( )( = α 2.35 2.56 0.27 7.59 0.175 0.906 1.26 0.135 0.358 2.14 0.039 − − − − − − − − refers to ∗∗∗ ∗∗ )( )( )( )( )( ∗∗∗ 0.75 0.44 7.90 0.183 0.209 2.37 1.22 0.042 / − − − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.0070.63 0.009 0.042 0.52 0.175 7.57 0.244 0.14 − − − − ∗∗∗ ∗ ∗∗ ∗∗ ∗∗ ∗∗∗ ∗ )()( )( )( )( )()( )( )( 0.19 0.156 2.66 0.567 2.690 2.38 2.07 0.248 4.30 1.86 1.92 1.392 1.783 2.29 0.93 0.096 0.094 0.003 0.029 0.031 0.005 ( ( − − − − − − − − − − ∗∗∗ ∗∗∗ )( )( )( )( )( 2.79 0.005 0.087 0.29 0.269 4.81 0.83 0.607 0.35 0.013 ( − − − − − − − − ∗∗ ∗∗∗ ∗∗ ∗ ∗ )( )( )( )( )( )( 2.13 2.02 0.272 4.73 1.80 1.211 1.438 1.73 0.44 0.051 − − − − − − − − ( ∗ ∗∗∗ ∗ )( )( )( )( )( Table 2.A.7. : Observed shocks model, economic recession periods 0.288 1.07 5.02 1.92 1.460 1.93 0.42 0.054 − − − − − − ( ∗∗∗ )( )( )( )( 0.01 0.271 4.75 1.27 0.944 0.48 0.019 0.016 0.018 ( − − − − − − ( ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. ALMP Tax Wedge 0.152 Coordination Union Density Union Coverage 0.000 0.020 0.065 Dependent VariableSpecificationAnnual GDP Growth I II U III IVObservationsR-squared V 315 I 0.928 315 0.929 II E 315 to U 0.930 III 0.930 315 IV 0.937 315 0.887 V 315 0.889 315 I 0.892 315 0.888 II 0.897 315 U to E III 0.925 315 0.925 IV 315 0.925 V 315 0.925 0.925 315 315 315 Benefit Length 0.187 Employment Protection Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 46 ) ) ) ) ) ) ) ) ) 0.57 0.068 0.446 0.58 0.49 0.060 1.60 1.43 0.82 1.11 1.20 0.111 0.37 0.119 − − − − − − − − )( )( )( )( )( 0.453 0.486 0.62 0.34 1.48 0.043 0.91 0.85 0.070 − − − − ∗ )( )( )( )()()( 1.31 1.46 0.34 0.091 0.509 0.09 0.3660.43 0.352 1.70 0.004 − − − − − − ∗ )( )( )( )( )( 0.08 1.73 0.282 0.35 0.004 1.38 1.49 0.087 − − − − − − )()( )( )( 0.458 0.514 1.50 0.61 0.25 0.0260.89 0.048 0.046 0.027 0.044 0.042 − − ∗∗∗ ∗ ∗ ∗∗∗ )( )( )()( )( )( )()()( 1.94 0.34 3.99 1.19 0.88 0.018 0.053 0.227 6.26 1.24 1.90 0.11 0.025 0.254 0.030 − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.217 5.17 0.31 0.99 0.13 0.001 0.013 0.005 0.24 − − − − − − 0.1/0.05/0.01. ∗ ∗∗∗ ∗∗∗ )( )( )( )( )( )( = α 0.226 6.23 1.03 0.03 0.0080.013 0.288 1.89 0.20 4.00 0.049 − − − − − − refers to ∗∗∗ ∗∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 4.13 0.048 0.226 6.28 0.02 1.05 2.05 / − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.216 5.19 1.02 0.32 0.0010.12 0.013 0.005 0.013 0.013 − − − − − − ∗∗∗ ∗∗∗ )( )( )( )( )( )( )()()( 0.460 3.66 0.97 0.46 1.10 0.184 0.294 0.005 1.51 0.048 1.24 2.75 1.4971.59 0.68 0.627 0.766 0.024 0.005 0.070 ( ( − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.86 0.421 3.20 0.53 0.53 0.80 0.014 ( − − − − ∗∗∗ ∗∗ )( )( )( )( )( )( 0.440 3.54 0.41 0.24 0.173 0.249 0.0191.53 0.009 0.018 1.18 2.52 0.058 ( − − − − − − Table 2.A.8. : Observed shocks model, lagged institutions ∗ ∗∗ ∗∗∗ )( )( )( )( )( 0.448 3.59 0.08 0.032 0.30 1.89 2.46 0.051 − − − − − − ( ∗∗∗ )( )( )( )( 0.419 3.19 0.61 0.54 0.82 0.014 0.007 0.005 ( ( − − − − ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Union Coverage 0.009 0.023 Coordination 0.287 0.326 0.026 0.048 0.165 0.541 ALMP Union Density Tax Wedge 0.028 Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 105 I 0.947 105 0.953 II E 105 to U 0.954 III 0.948 105 IV 0.956 105 0.956 V 102 0.967 102 I 0.967 102 0.956 II 0.970 102 U to E III 0.976 102 0.977 IV 104 0.977 V 104 0.976 0.978 104 104 104 Benefit Length 0.047 0.056 Employment Protection 0.288 Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 47 ∗ ) ) ) ) ) ) ) ) ) 1.77 0.017 0.19 0.043 0.123 1.58 0.40 0.16 1.27 1.24 0.76 0.143 0.140 0.33 − − − − − − − − )( )( )( )( )( 0.477 0.557 1.50 0.22 0.53 0.82 0.005 0.007 0.074 0.21 − − − − ∗ )( )( )( )()( )( 0.13 1.54 0.95 0.97 0.102 0.573 1.85 0.44 0.098 0.550 − − − − ∗ )( )( )( )( )( 1.81 0.53 0.15 1.48 0.81 0.074 − − )( )( )( )( 0.498 0.559 1.58 0.11 0.006 0.024 0.020 0.012 0.47 0.23 − − ∗∗∗ ∗ ∗∗ )( )( )( )()( )()()()( 0.60 0.10 0.59 2.50 1.06 0.38 0.010 0.041 0.227 5.61 1.21 0.036 1.68 − − − − − − − − ∗∗∗ )( )( )( )( )( 0.218 5.08 0.18 1.10 0.60 0.13 0.004 0.004 0.005 0.016 0.020 − − − − 0.1/0.05/0.01. ∗∗ ∗∗∗ )( )( )( )( )( )( = α 0.227 5.61 1.08 0.64 0.10 0.25 2.39 0.035 − − − − refers to ∗∗∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 0.24 2.97 0.034 0.227 5.65 1.10 0.70 / − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.218 5.11 0.23 1.12 0.0050.72 0.001 0.002 − − − − ∗∗ ∗∗∗ )( )()( )( )( )()( )( )( 0.458 3.44 0.0230.49 0.004 0.016 0.016 0.003 0.70 0.24 1.42 1.9021.59 1.12 0.731 1.754 0.010 0.83 2.25 0.051 0.003 0.069 ( ( − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.014 0.24 1.16 0.418 3.12 0.89 0.77 ( − − − − ∗∗∗ ∗ )( )( )( )( )( )( 0.442 3.36 1.42 0.22 1.20 0.007 0.68 1.89 0.050 Table 2.A.9. : Observed shocks model, fixed institutions ( − − − − − − ∗∗∗ )( )( )( )( )( 0.435 3.30 1.52 0.54 0.010 0.93 1.60 0.031 − − − − − − ( ∗∗∗ )( )( )( )( 0.016 1.35 0.416 3.13 1.04 0.79 ( ( − − − − ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Benefit Length 0.038 0.059 Coordination 0.322 0.430 0.016 0.017 0.499 0.955 Employment ProtectionUnion Coverage 0.375 0.237ALMP 0.097 0.365 0.100 0.315 0.172 0.165 0.312 0.150 0.454 0.116 Tax Wedge 0.008 Union Density Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 105 I 0.949 105 0.950 II E 105 to U 0.951 III 0.949 105 IV 0.954 105 0.956 V 102 0.962 102 I 0.962 102 0.956 II 0.964 102 U to E III 0.975 102 0.977 IV 104 0.977 V 104 0.975 0.978 104 104 104 Replacement Rate 0.026 0.037 0.034 0.024 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 48 ∗ ∗ ) ) ) ) ) ) ) ) ) 1.65 1.70 1.05 0.001 0.02 2.976 0.123 1.56 0.71 1.80 0.125 3.010 1.24 0.119 1.35 − − − − − − − − − − )( )( )( )( )( 1.20 0.90 2.382 0.94 1.32 0.111 0.030 0.41 − − − − − − )( )( )( )( )()( 1.39 1.53 0.93 1.2241.24 1.732 1.60 0.86 3.214 0.102 0.075 − − − − − − )()( )( )( )( 1.23 1.25 1.48 1.37 0.66 2.568 0.089 0.054 − − − − − − )( )( )( )( 0.386 0.416 0.441 0.3820.93 0.501 1.35 1.22 0.91 2.401 0.110 − − − − ∗∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗ )( )( )()( )()( )( )( )( 0.09 0.226 6.18 0.001 0.0333.89 0.426 2.20 0.055 0.030 0.035 0.019 1.19 2.60 1.62 0.756 1.249 0.051 2.12 0.043 0.027 1.52 − − − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.65 0.208 5.13 0.004 1.41 0.90 0.451 0.016 0.49 − − − − − − − − 0.1/0.05/0.01. ∗ ∗∗∗ ∗∗∗ ∗∗ )( )( )( )( )( )( = α 0.25 0.214 5.86 0.312 1.71 3.44 2.51 0.779 0.12 0.045 0.002 − − − − − − − − refers to ∗∗∗ ∗∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 0.212 5.70 1.08 3.19 2.00 0.584 0.28 0.041 / − − − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.207 0.59 5.14 0.004 0.007 0.002 1.38 0.439 0.91 0.016 0.004 − − − − − − − − ∗ ∗ ∗∗ ∗∗∗ ∗∗ )( )( )()( )( )( )( )( )( 0.473 3.85 0.17 0.632 2.55 1.89 0.02 0.0012.57 1.787 0.031 1.343 1.16 0.056 1.81 0.049 0.0120.43 0.009 ( − − − − − − − − − − − − ( ∗∗∗ ∗ )( )( )( )( )( 0.415 3.29 0.37 1.95 1.067 1.04 0.032 1.16 ( − − − − − − ∗∗∗ ∗ ∗∗ ∗∗ )()( )( )( )( )( 0.435 3.58 0.55 2.19 1.71 1.479 2.53 0.73 0.048 0.021 ( Table 2.A.10. : Observed shocks model, without Portugal − − − − − − − − ∗ ∗∗∗ ∗∗ )( )( )( )( )( 0.423 3.43 1.38 1.91 1.151 2.14 0.36 0.040 0.010 − − − − − − − − ( ∗∗∗ ∗ )( )( )( )( 0.410 3.25 0.55 1.029 1.87 1.04 0.032 ( − − − − − − ( ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Coordination 0.558 ALMP Union Density Tax Wedge 0.034 Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 100 I 0.952 100 0.955 II E 100 to U 0.957 III 0.953 100 IV 0.959 100 0.958 V 97 0.965 I 97 0.967 0.959 97 II 0.969 U to E 97 III 0.978 0.979 97 IV 0.980 99 V 0.978 99 0.981 99 99 99 Union Coverage 0.006 0.017 0.007 0.004 0.002 Benefit Length Employment Protection Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 49 ∗ ∗ ) ) ) ) ) ) ) ) ) 0.591 0.16 0.045 0.49 0.539 0.72 1.94 1.68 1.08 1.15 0.88 0.113 2.187 1.02 0.082 − − − − − − − − ∗ )( )( )( )( )( 0.525 0.40 1.69 0.022 0.013 0.02 0.73 0.63 0.038 − − − − ∗ )()( )( )()()( 1.11 1.51 0.51 0.091 0.553 1.83 0.49 0.72 0.578 0.020 − − − − ∗ )( )( )( )( )( 0.556 1.85 0.53 0.65 0.496 1.30 1.51 0.085 − − − − ∗ )()( )( )( 0.528 1.70 0.37 0.020 0.0230.00 0.022 0.65 − − ∗∗∗ ∗ ∗∗∗ )( )( )()( )( )( )()()( 1.12 3.80 1.88 1.12 0.019 0.051 0.244 6.44 0.08 0.49 0.14 0.002 0.47 0.108 0.003 − − − − − − − − − − ∗∗∗ )( )( )( )( )( 0.231 5.52 0.34 1.05 0.50 0.007 0.001 0.014 0.037 0.032 0.017 0.023 0.005 0.22 − − − − 0.1/0.05/0.01. ∗∗∗ ∗∗∗ ∗ )( )( )( )( )( )( = α 0.233 6.20 1.85 0.51 0.62 0.024 0.143 0.128 0.61 3.49 0.044 − − − − − − refers to ∗∗∗ ∗ ∗∗∗ )( )( )( )( )( ∗∗∗ 3.52 0.042 0.234 6.24 1.90 0.52 0.85 0.116 / − − − − − − ∗∗ / ∗ ∗∗∗ )( )( )( )( 0.231 5.55 0.33 0.51 1.03 0.007 0.005 0.004 − − − − ∗∗ ∗∗∗ )( )( )()()( )( )( )()( 0.514 4.01 0.40 0.62 0.70 0.36 0.027 0.013 0.252 0.130 0.91 1.6160.09 1.113 2.56 1.63 0.003 0.004 0.058 ( − − − − − − − − − − ( ∗∗∗ )( )( )( )( )( 0.85 0.474 3.59 0.14 0.08 0.0010.12 0.004 0.033 ( − − − − − − ∗∗∗ ∗∗ )( )( )( )( )( )( 0.484 3.79 0.87 1.08 0.80 0.332 0.65 2.26 0.048 ( − − − − − − Table 2.A.11. : Observed shocks model, without Germany ∗∗ ∗∗∗ )( )( )( )( )( 0.485 3.82 0.76 1.13 0.283 1.17 2.23 0.045 − − − − − − ( ∗∗∗ )( )( )( )( 0.472 3.58 0.06 0.09 0.02 0.022 ( ( − − − − ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Employment Protection Union Coverage 0.000 0.013 0.009 Coordination 0.178 0.268 0.098 0.200 0.293 0.769 ALMP Union Density Tax Wedge 0.023 Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V 100 I 0.946 100 0.951 II E 100 to U 0.951 III 0.947 100 IV 0.953 100 0.956 V 97 0.964 I 97 0.964 0.956 97 II 0.966 U to E 97 III 0.975 0.977 97 IV 0.977 99 V 0.975 99 0.978 99 99 99 Benefit Length Replacement Rate 0.002 0.024 0.023 0.003 0.015 0.005 0.025 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 50 ∗ ) ) ) ) ) ) ) ) ) ) 0.513 1.14 0.75 0.089 0.075 1.36 0.088 0.04 0.612 0.02 0.64 0.22 0.020 0.22 1.48 1.79 0.019 − − − − − − − − − − − − ∗ )( )( )( )( )()( 0.471 1.08 0.15 0.017 0.001 0.081 0.67 0.122 1.52 1.67 0.91 0.070 − − − − − − − − ∗ )( )( )( )()( )( )( 0.516 0.65 1.32 0.14 0.065 0.05 1.54 1.83 0.520 0.45 0.025 − − − − − − ∗ )( )()( )( )( )( 0.516 0.66 0.17 1.35 1.56 1.84 0.064 0.512 0.45 0.025 − − − − − − ∗ )( )( )( )( )( 0.481 0.75 0.3271.52 0.269 0.268 0.359 0.288 1.70 0.18 0.019 0.004 0.004 0.150 0.87 0.063 − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ )()( )( )( )( )()( )( )( )( 0.227 1.07 0.098 3.08 6.40 0.99 2.92 0.018 0.52 0.007 0.29 0.38 0.036 0.93 0.115 0.58 0.013 − − − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ )( )( )( )( )( )( 0.219 0.124 3.65 0.65 2.68 0.010 0.019 0.36 0.86 0.013 5.99 − − − − − − − − 0.1/0.05/0.01. ∗∗∗ ∗∗∗ ∗∗∗ )( )( )( )()( )( )( = α 0.226 0.100 3.41 6.58 1.04 0.007 3.00 0.0090.06 0.050 0.030 0.423 0.51 0.033 0.111 0.084 0.37 − − − − − − − − − − ∗∗∗ ∗∗∗ ∗∗∗ refers to )( )( )( )( )( )( ∗∗∗ 0.226 0.099 3.46 1.12 6.62 0.008 3.15 0.53 0.033 0.113 0.37 / − − − − − − − − ∗∗ / ∗∗∗ ∗∗∗ ∗∗∗ ∗ )( )( )( )( )( 0.220 0.121 3.65 2.72 0.020 0.35 0.0120.82 0.005 0.005 6.03 − − − − − − ∗∗∗ ∗∗ ∗ )( )( )()( )( )()( )( )( )( 0.477 0.127 3.84 0.48 0.01 2.21 0.006 0.17 1.71 0.44 0.442 0.235 0.73 0.035 0.0080.19 0.024 0.271 0.080 0.003 0.08 ( − − − − − − − − − − − − ( ∗∗∗ ∗∗ )()( )( )( )( )( 0.451 0.158 0.23 2.60 3.71 0.017 1.46 0.28 0.106 0.018 0.73 ( − − − − − − − − ∗∗ ∗∗∗ ∗ )( )( )( )( )( )( )( 0.466 0.137 0.006 2.47 0.45 0.09 1.72 0.79 0.032 0.292 0.001 0.06 3.89 ( − − − − − − − − − − ∗∗∗ ∗∗ ∗ )( )( )( )( )( )( 0.465 0.138 3.91 0.005 0.45 2.52 1.78 0.79 0.032 0.285 0.001 0.06 Table 2.A.12. : Observed shocks model, with temporary employment − − − − − − − − − − ( ∗∗∗ ∗∗∗ )( )( )( )( )( 0.450 3.73 0.017 1.46 2.65 0.28 0.103 0.019 0.76 ( − − − − − − − − ( ( ( ( Nonlinear least-squares estimation. Country fixed effects are included. T-statistics in parentheses. The estimates quantify the respectiveSource: contribution EU-LFS, of OECD institutions Economic and Outlook shocks 95, to ICTWSS, the own shared calculations. interaction coefficient. Dependent VariableSpecificationGDP Growth I II U III IVObservationsR-squared V I 105 0.955 II 105 0.957 E to U III 105 0.957 0.955 105 IV 0.957 105 V 0.966 102 0.970 I 102 0.970 II 102 0.966 U to E 0.971 III 102 0.979 102 IV 0.980 104 V 0.980 104 0.979 0.980 104 Temporary employment 0.160 Union Coverage CoordinationTax Wedge 0.024 0.050 0.006 0.000 Union Density ALMP Benefit Length Employment Protection Replacement Rate 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 51

Table 2.A.13.: Observed shocks model, total E outﬂows

Dependent Variable E to U and N

Speciﬁcation I II III IV V

GDP Growth −0.246∗∗∗ −0.253∗∗∗ −0.253∗∗∗ −0.246∗∗∗ −0.255∗∗∗ (−4.64)(−5.03)(−5.02)(−4.61)(−4.97) Replacement Rate −0.001 0.017 0.016 −0.001 −0.017 (−0.03)(1.05)(1.00)(−0.05)(−0.58) Beneﬁt Length 0.041 (1.22) Employment Protection 0.062 −0.173 −0.200 0.066 −0.202 (0.21)(−0.62)(−0.71)(0.22)(−0.71) Union Coverage −0.004 0.007 0.006 −0.003 0.006 (−0.52)(0.95)(0.66)(−0.44)(0.63) Union Density −0.039∗∗∗ −0.041∗∗∗ −0.045∗∗∗ (−2.74)(−2.74)(−2.81) Coordination 0.097 0.192 (0.49)(0.86) ALMP 0.475 (0.64) Tax Wedge −0.009 −0.031 (−0.46)(−1.35)

Observations 102 102 102 102 102 R-squared 0.980 0.982 0.982 0.980 0.983 Nonlinear least-squares estimation. Country ﬁxed effects are included. T-statistics in parentheses. ∗ / ∗∗ / ∗∗∗ refers to α = 0.1/0.05/0.01. The estimates quantify the respective contribution of institutions and shocks to the shared interaction coefﬁcient. Source: EU-LFS, OECD Economic Outlook 95, ICTWSS, own calculations.

2.B. Data Description and Imputation Methods

In order to construct labour market transitions, we use the information available in the variable ILOSTAT (labour market status in the current year according to the ILO deﬁnition), as well as the variable WSTAT1Y (self-declared labour market status in the year prior to the survey). An alternative would be to use the variable MAINSTAT (self-declared labour market status in the current year) instead of ILOSTAT. However, as this variable it not available at all for Germany and the UK. Therefore, we opt for ILOSTAT, which is consistent across our country sample.

There is also an issue of missing information for several countries. We therefore delete Bul- garia, Ireland, Iceland, and Switzerland from the sample due to low response rates in the data up to the year 2007. We also exclude the Netherlands from the analysis because information on the previous year’s employment status is predominantly missing until 2008.

For the same reason, a few yearly data points are not available for several countries. If this is the case, we impute the values by averaging transition rates close to the respective years.

Speciﬁcally, since we average the observations within 3 year windows in the analysis, we 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 52 use the available years in these windows for imputation. In three cases, extrapolation is based on only one year. This concerns France (2002 to 2004), Sweden (2005 to 2007) and Slo- vakia (1998 to 2000). The transition rates from employment to unemployment are missing for Austria in the period 1999 to 2001 and for Denmark from 1999 to 2001 and from 2002 to

2004. Furthermore, no information on worker ﬂows from unemployment to employment is available for Austria between 1999 and 2001. Therefore, these data points do not enter the regression sample, and the number of observations differs between the regressions for transitions from employment to unemployment and the regressions from unemployment to employment.

The GDP growth rate is calculated by annual GDP values in volume measured in market prices extracted from the OECD Economic Outlook No. 25. The OECD applies reference years deﬁned in national ofﬁcial publications which is 2005 for most countries.

The benefit replacement rate and the benefit duration are obtained from the OECD Benefit and Wages Statistics (2014). The former is a summary measure which is defined as the average of the net unemployment benefit replacement rate for two earnings levels, three family situations and 60 months of unemployment. It is available for between 2001 and

2013 for all countries of the sample. We impute values for 1999 from the measure in 2001.

Our variable for beneﬁt length captures for how many months 40-year old unemployed individuals are entitled to receive unemployment beneﬁts. The source provides measures only for the years 2002, 2005, 2007 and 2010. Since we are interested in the years 1999, 2002,

2005, 2008 and 2010, we use the values most closely to the corresponding year as a proxy for the measure in the three-year window.

Employment protection legislation (EPL) in a country is measured by an index and stems from OECD Indicators of Employment Protection (2013). Several indicators are available.

We select version 1, which measures the strictness of regulation of individual dismissal of employees on regular or indeﬁnite contracts. The index ranges from one to ﬁfe. For most countries it is available at a yearly frequency in the time period between 1999 and 2011.

However, it includes missing data for Estonia, Luxembourg and Slovenia before 2008. We 2. LABOUR MARKET TRANSITIONS,SHOCKSAND INSTITUTIONS 53 conduct a complete imputation for the previous years using the value of 2008 for these countries.

The dispersion of wage bargaining is characterized by three measures. We take all from the

ICTWSS 5.1 (Database on Institutional Characteristics of Trade Unions, Wage Setting, State

Intervention and Social Pacts). The first is the coordination of wage setting and represents an indicator ranging from one to five where five is the highest form of collective bargaining that is equal to centralized wage bargaining. It is complete for our country-year combinations. The second measure is union coverage calculated as the ratio of employees covered by collective bargaining agreements to the proportion of all wage and salary earners in employment with the right to bargaining, adjusted for the possibility that some sectors or occupations are excluded from the right to bargain. Since it is missing within countries at most for one year at once, we impute the previous year’s value. The last institutional variable to capture the dispersion of wage bargaining is union density, defined as the proportion of trade union members as a percentage of all employees. Here, mainly the same systematic missing structure exists as for the union coverage variable. Thus, we apply the same imputation method. Nevertheless, for Estonia, France, Poland and Greece up to three years are lacking in the data set. Thus, we use for 1999 the information of 2002.

The measure for active labour market policies (ALMP) comes from OECD Employment and Labour Market Statistics (2013). It refers to expenditures on programs that are aimed at helping unemployed individuals to get back into work. The time series is complete for all countries between 1999 and 2011.

The OECD reports a tax wedge indicator. We use the information given in 2015. The measure displays the difference between real labour costs and take home pay for a single- earner couple at 100 per cent of average earnings with two children. Since its values are missing for all countries in 1999, we use the data of 2000 instead. 3. Job Stability in Europe Over the Cycle*

Abstract: This article investigates the evolution of job tenure for the time period 2002-2012 using microdata from the European Union Labour Force Survey (EU-LFS). Overall, the data show a slight increase in average job tenure at the EU level, which can be explained by disproportional lay-offs of short-tenured workers during the crisis. When controlling for changes in the demographic composition of the workforce, an underlying negative trend in mean tenure becomes visible. Job tenure evolved very differently across the EU before and during the crisis, highlighting the importance of institutional frameworks, especially employment protection legislation (EPL).

*This chapter is co-authored with Ronald Bachmann. For this chapter, I made minor revisions on a previous version which is published as: Bachmann, Ronald and Rahel Felder. 2018. “Job stability in Europe over the Cycle.” International Labour Review 157(3):481-516. We thank the editors of that journal, two anonymous referees, Hanna Frings, Matthias Giesecke and Sylvi Rzepka for their helpful comments and suggestions. This article is based on the project “Labour market transitions in turbulent times”, which was carried out by RWI for Eurofound. We thank Donald Storrie and Carlos Vacas-Soriano from Eurofound for their support at various stages of the project. The views and opinions expressed in this article are those of the authors and do not necessarily reﬂect the ofﬁcial policy or position of Eurofound. Responsibility for opinions expressed in signed articles rests solely with their authors, and publication does not constitute an endorsement by the ILO. All remaining errors are our own. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 55

3.1. Introduction

Changes in the economic environment over recent decades have led to growing concerns about decreasing job stability. In particular, the potential decline in the prevalence of jobs that last for a long period of time (a ”job for life”) has been intensively discussed in both academic research and the media [Hall 1982 is a seminal article in this regard]. The fear is that globalization and technological progress, such as advances in communication technologies, have induced changes in the labour market that require employees to be more

ﬂexible. Workers have to adapt to more frequent transitions between jobs and intermittent spells of unemployment. These changes in the labour market are likely to adversely affect job satisfaction and worker well-being [European Commission 2001]. In this context, job tenure, or the length of time a worker has been continuously employed by the same employer, is of paramount interest to workers, as it can be interpreted as a measure of job stability [Neumark 2000].

Aside from long-term trends, evidence suggests that the labour turnover rate was strongly affected by the recent ﬁnancial and economic crisis, with potentially detrimental consequences for job tenure [Bachmann et al. 2015]. This crisis has led not only to a large and persistent increase in unemployment in many European countries, but also to a divergent development in labour markets across the European Union (EU). Since turnover is closely related to the length of time workers spend in a job, it is expected that the Great Recession also had an effect on job tenure.

Against this background, our article analyses the evolution of job tenure as measured by the length of uncompleted employment spells with the same employer2 for the time period 2002-2012 for a large number of European countries using worker-level data from the

European Union Labour Force Survey (EU-LFS). In so doing, we provide evidence both for longer-term trends and for recent developments which took place during the Great Re- cession. In particular, we provide aggregate evidence on the evolution of job tenure at the

European level, but also for speciﬁc countries. Furthermore, the richness of the EU-LFS

2This deﬁnition comprises accumulated durations of subsequent jobs at the current ﬁrm and is sometimes referred to as ”employment tenure” in the literature (for instance, by Auer and Cazes[2000]). 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 56 data allows us to analyse heterogeneities with respect to both worker and job characteristics. Finally, we analyse cross-country differences, pointing out the important role played by institutional frameworks, in particular when in the form of employment protection legislation (EPL). This study thus fully exploits the richness of the microdata from the EU-LFS from 2002 to 2012 in order to examine how job tenure evolved across EU countries during the pre-crisis and crisis periods.

Our analysis is related to several strands of the literature. Job tenure was analysed in an international context by Auer and Cazes[2000] and Cazes and Tonin[2010]. Both studies, the ﬁrst one for the 1990s and the second for the period from 1996 to 2006, ﬁnd that mean tenure remained relatively stable in most European countries and increased only slightly in a few countries during the observation period. However, the authors report pronounced differences in levels between countries, which they attribute to heterogeneous labour market institutions and workers’ labour market behaviours. Examining data for the mid-1990s for eight EU countries, Japan, the Russian Federation and the United States of America,

Burgess[1999] found that the United Kingdom and the United States had relatively low levels of tenure in the working relationship. These results suggest that tenure is generally low in countries that are characterized by ﬂexible labour markets. Additionally, results from Burgess[1999] suggest that EPL has a positive effect on mean tenure. Furthermore,

Boockmann and Steffes[2010] ﬁnd that labour market institutions play an important role in reducing mobility and thus prolonging tenure.

The relationship between tenure and job characteristics such as the prevalence of temporary contracts has also been examined. This is of paramount interest in the context of job stability because temporary contracts are encountered with increasing frequency in the

EU. However, this does not necessarily imply any immediate effects on mean tenure, since temporary contracts are designed differently with respect to termination time and contract renewals from one country to another. Auer and Cazes[2000] detect no clear pattern of association between temporary work and job tenure.

The literature identiﬁes different patterns and trends before the crisis in the relationship between socio-demographic characteristics and mean tenure, and these hold true for the 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 57 majority of EU countries. Cazes and Tonin[2010] show that young workers do not experience a systematic decline in tenure over time, except for those in Central and Eastern

European (CEE) countries. After controlling for age, however, they report some reduction in average tenure in a majority of EU countries. Against this background, our study explicitly analyses the role played by demographic change in relation to tenure.

As mentioned above, there is detailed research on job tenure in the pre-crisis period across

European countries. However, the impact of the Great Recession on job tenure has not been examined for European countries. Studies for the period before the Great Recession show that job tenure behaves countercyclically and moves with the unemployment rate

[Auer and Cazes 2000]. It decreases during an economic boom, when unemployment falls and job creation increases, leading to new hires and voluntary job-to-job transition [for the United States, see Shimer 2005; for Germany, see Bachmann 2005]. The opposite is found during a recession, when exit ﬂows from employment increase as ﬁrms dismiss workers. As a result, unemployment rises, and job tenure tends to increase [Eurofound

2014]. Importantly, workers are affected differently along the tenure distribution, with those who have little seniority being more likely to lose their job during a recession than high-tenured workers [Abraham and Medoff 1984; Jovanovic 1979].

Long-run trends and more recent developments such as the Great Recession may have adversely affected the job tenure of different subgroups, meaning that analysing tenure at the aggregate level may mask changes for subgroups of the population. In order to reveal variation in job stability within each group, we additionally investigate the evolution of job tenure for different worker groups and job types. We analyse changes both at the EU aggregate level and for speciﬁc countries, initially focusing on a comparison of mean job tenure across countries and subpopulations. Our study therefore contributes to the literature by creating a complete picture of the changes in job stability across countries and subgroups, taking both a longer-term perspective and looking at its evolution during the recent ﬁnancial and economic crisis.

The remainder of this article is structured as follows. The next section provides a brief description of the EU-LFS data set. The third section displays an overview of the aggregate 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 58 evolution of job tenure with respect to both trends and cyclical features for the EU at the aggregate level and for speciﬁc countries for the period from 2002 to 2012. Section four will then focus on heterogeneity in job tenure between socio-demographic groups and job types by drawing on the shift-share method to control for compositional changes over time. In order to explicitly analyse these heterogeneities, we conduct a regression analysis to provide econometric ﬁndings. Finally, the last section summarizes and concludes the discussion.

3.2. The European Union Labour Force Survey (EU-LFS) Data

In order to compute job tenure from European countries, we use the EU-LFS. The microdata set comprises a large number of representative national household surveys that provide quarterly and annual information on the labour force participation of persons. The

EU-LFS covers all EU Member States, with the exception of Croatia (EU-27), as well as Ice- land, Norway and Switzerland. The national labour force surveys are conducted by the national statistical agencies applying harmonized concepts and deﬁnitions, which enables us to perform cross-country comparisons both at the aggregate level and for subgroups.

The EU-LFS data set is provided by Eurostat, in the form of repeated cross-sections. The data include a variable for the time at which individuals began working for their current employers. We utilize this information to compute person speciﬁc job tenures. Due to the survey design, the data do not show a smooth distribution of tenure, and in particular contain implausible values of zero for speciﬁc tenure classes, for example for tenures between

37 and 43 months. We therefore recalculate job tenure in accordance with the EU-LFS user guide [Eurostat 2014a, p. 54].

At the individual level, we focus on dependent-status employees and omit individuals living in institutional households (such as those residing in retirement homes or military barracks), children under the age of 15 and adults aged 65 years and over. The study covers the period from 2002 to 2012. The starting point of this period was chosen because data availability is severely limited before 2002, with data on several countries (including Ger- 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 59 many) missing. Thus, we analyse an entire decade, which allows us to make a distinction between a ”pre-crisis period” (2002-2007) and a ”crisis period” (2008-2012) 3. The country sample includes 26 EU Member States, namely Austria, Belgium, Bulgaria, Cyprus, the

Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland,

Italy, Latvia, Lithuania, Luxembourg, the Netherlands, Poland, Portugal, Romania, Slo- vakia, Slovenia, Spain, Sweden and the United Kingdom (as mentioned above, Croatia is excluded from the EU-LFS data). In our study, Malta is not included because no data are available for that country before 2008.

The ﬁnal data set contains information on tenure for each country, as well as for subpopulations and job characteristics by country, which are largely available in the data set. It allows us to distinguish gender, age group, and type of employment contract (i.e. temporary or permanent), and to investigate how worker composition in terms of these characteristics changes over time. Given the lack of a panel dimension, life-cycle issues can not be taken into account, unfortunately. Table 3.A.1 in the appendix summarizes the sample separately for the pre-crisis and crisis periods.

3.3. The Aggregate Evidence

The analysis begins by looking at the EU aggregate tenure levels and then moves on to examine those of individual countries. Figure 3.1 illustrates the evolution of mean tenure and the unemployment rate in the EU between 2002 and 2012, including all EU-27 countries except Croatia and Malta. In 2002, mean tenure was 116.5 months (almost ten years).

From 2003 to 2005, it was somewhat higher at 118 months, only then to fall back to its previous level by 2008. Between 2008 and 2012, which covers the period of the Great Re- cession, mean tenure increased continuously, reaching its highest level of 123 months in

2012. These results are thus in line with the previous ﬁndings that tenure behaves counter cyclically.

3Since the crisis started in most countries in the third quarter of 2008, this leads to a slight underestimation of the effects of the crisis. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 60

Figure 3.1.: Mean tenure and unemployment rate in the EU, 2002-2012

124 12

122 10 Unemployment rate (percent)

120 8

118 6

116 4 Mean tenure (months) tenure Mean

114 2

112 0

mean tenure unemployment rate

Mean tenure is plotted against the primary axis; the axis starts at 112 months to make the variation visible. The unemployment rate is plotted against the secondary axis. Source: EU-LFS, authors’ calculations.

The strong correlation between mean tenure and the unemployment rate is positive and pronounced. These correlation coefficient corresponds to 0.78, which is statistically significant at the 1 per cent significance level. This strong relationship becomes especially visible in a recession, since during an economic crisis short-tenured jobs are more likely to be destroyed than long-tenured ones and fewer new (and thus short-tenured) jobs are created.

However, as the economy recovers, workers are rehired, thereby reducing mean tenure as by deﬁnition such workers have zero tenure.

Although Figure 3.1 suggests a common trend across countries at the EU level before and during the crisis, one can expect there to be variation between countries because of differences in population composition or the institutional framework of the labour market.

However, closer inspection reveals deep cross-country differences which call into question the general countercyclical nature of job tenure. This is illustrated in Figure 3.2, which 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 61 presents the mean tenure and the unemployment rate for each country, by year. This allows us to separate longer-term trends in mean tenure from the cyclical component. To the extent that the development of mean tenure was purely cyclical, both lines should move relatively closely together for each country. This is observable for Belgium, Estonia, Greece,

Ireland, Italy, the Netherlands, Poland, Portugal, Slovakia, Spain and the United Kingdom, where mean tenure and the unemployment rate are positively and signiﬁcantly correlated, at least at the 10 per cent signiﬁcance level. A few countries, such as Hungary and Latvia, experienced the reverse; we can see that job tenure and unemployment display different trends. The unemployment rate in Germany, in contrast to the rate in all other countries, even decreases slightly, yet tenure increases. Latvia and Lithuania, however, display a massive increase in unemployment combined with little change in tenure. The reason for these divergent behaviours may be the institutional frameworks in place during the observation period, an aspect analysed below in the section on worker and job heterogeneities.

In the following, we continue with a detailed discussion of pre-crisis trends and changes in mean tenure across countries during the Great Recession. However, it must be borne in mind that the period preceding the Great Recession was a time of strong growth, at least in some EU countries, which might in itself have been unusual. This should be taken into account when comparing the evolution of job tenure in the pre-crisis and crisis periods. During the pre-crisis period, trends in mean tenure appear to be relatively stable at the EU level. At a more disaggregated level, it becomes apparent that the 26 countries are split about equally with regard to increases and decreases in mean tenure: 12 had a higher level of mean tenure in 2007 compared with 2002, ten a lower level and four display virtually no change. Cyprus, France, Germany, Greece and Portugal show this most notable increases. For Greece and Portugal, higher unemployment is likely to explain this increase in mean tenure to an important extent4. This was probably due to the laying-off of many short-tenured workers and a reduction in hiring, leading to fewer new jobs and even fewer short-tenured workers. The other countries, by contrast, had stable or slightly de-

4However, other factors such as the ageing of the population or labour market institutions play a role as well, see the sections entitled ”Worker characteristics: Age” and ”Econometric ﬁndings: Regression analysis”. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 62 creasing unemployment rates, respectively. Since labour market reforms leading to lower turnover and hence to increased tenure levels are an unlikely explanation for cross-country differences (Germany, for example, passed large-scale labourmarket liberalization reforms during the observation period, which, if anything, led to a decrease in mean tenure), an ageing work force may be an alternative explanation. On the other hand, the massive fall in mean job tenure in Denmark and some CEE countries (Poland, Romania and Slovakia) can be explained by the decrease in the unemployment rate before the crisis. The marginal positive trend in mean job tenure in the Baltic States, accompanied by a strong decrease in the unemployment rate, may potentially be related to national labour market institutions.

This is because labour markets in these countries are characterized in the literature as being very ﬂexible and fostering voluntary worker turnover [Eamets et al. 2003], which implies that as the unemployment level falls mean tenure may rise slightly.

So far, the discussion in this article has focused on the evolution of mean tenure. However, countries differ not only in terms of the development of mean tenure but also in terms of their initial levels of tenure. Figure 3.3 illustrates mean tenure in 2002, six years before the

Great Recession. EU Member States display considerable variation: mean tenure is lowest in Latvia, at 86 months (i.e. just over seven years) and highest in Slovenia, at 137 months

(about 11.5 years). The Benelux countries, plus Austria, Germany and the Mediterranean countries exhibit a comparably high average tenure. By contrast, the CEE countries, together with Ireland and the United Kingdom are characterized by a low average tenure.

Spain and Slovenia are important exceptions to this pattern. Spain has a lower average tenure than the other Mediterranean countries, and Slovenia exhibits an extremely high average tenure compared with the other countries of Central and Eastern Europe. The

Scandinavian countries do not constitute a uniform group, but are instead scattered across the distribution of EU countries. On average, workers in Sweden are rather high-tenured, while the opposite is the case in Denmark. Cazes and Tonin[2010] draw a similar picture for mean tenure across the EU-24 countries prior to the crisis. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 63 Figure 3.2. : Mean tenure and unemployment rate by EU Member State, 2002-2012 Country codes: AT: Austria, BE:Estonia, Belgium, ES: BG: Spain, Bulgaria, FI: CY: Cyprus, Finland,LV: CZ: FR: Latvia, Czech France, NL: Republic, GR: The DE: Greece, Netherlands,UK: Germany, HU: DK: United Hungary, PL: Denmark, IE: Kingdom. Poland, EE: Ireland, PT: IT: Italy,Mean Portugal, LT: Lithuania, tenure RO: LU: is Romania, Luxembourg, plotted SE:unemployment against Sweden, rate the SI: is primary plotted Slovenia, axis. against SK:Source: the Note Slovak EU-LFS, secondary authors’ axis. that Republic, calculations. the axis starts at 80 months to make the variation visible. The 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 64

Figure 3.3.: Mean tenure before and during the crisis by EU Member State

140

120

100

60 Mean tenure (months) tenure Mean 40

0 SI BE FR IT LU DE SE PT GR AT NL EU SK FI PL CZ HU RO CY ES DK LT IE BG UK EE LV

pre-crisis crisis

Countries are sorted in order of their pre-crisis mean tenure level. See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations.

We now turn to the Great Recession, which is of predominant interest in this study. The reaction to this major event in terms of job tenure is likely to have been inﬂuenced by three components: the depth of the recession itself, the effect of national labour market institutions and the inﬂuence of institutional reforms undertaken in response to the crisis. Figure

3.3 clearly shows that job tenure increased in most European countries during the recession period. In total, 15 countries had a higher mean tenure in 2012 than in 2007. As one can see from Figure 3.2, those countries with the strongest reaction in the unemployment rate experienced the most sizeable increases in tenure (Bulgaria, Estonia, Greece, Ireland, Italy,

Portugal and Spain). The Netherlands exhibited no change in mean tenure levels, which can be explained by the fact that this country experienced no (or only slight) increases in the unemployment rate (and thus little change in hiring and lay-offs) during the crisis.

As for the remaining countries, which display a lower mean tenure in 2012 than in 2007, their increases in unemployment and decreases in mean tenure were both relatively small. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 65

Signiﬁcant decreases in mean tenure can be observed in Lithuania and Luxembourg. While

Luxembourg suffered little impact from the crisis, Lithuania was severely affected (Figure

3.2). Surprisingly in terms of timing, the reduction in mean tenure in Lithuania ended at exactly the same time as unemployment reached its peak. Put differently, in contrast to all other EU Member States, lay-offs in Lithuania appear to have affected longtenured workers proportionally more than short-tenured workers. This could be explained by a Lithuanian labour market reform undertaken in reaction to the crisis. This reform was targeted at fostering early retirement and enabling employers to more easily dismiss older workers.

In 2009, a law came into force that reduced workers’ protection against dismissal in the period before entitlement to an old-age pension from ﬁve to three years [Masso and Krillo

2011]. This is in line with Lithuania displaying extremely high ﬂows from employment to non-employment due to retirement, including early retirement, during the crisis period

[RWI 2014b].

3.4. Worker and Job Heterogeneities

Socio-demographic and job characteristics can potentially help to explain the evolution of mean tenure within individual countries, and hence the differences in mean tenure between the European countries. Therefore our goal in this section is to compare job tenure for subgroups of the working population before the Great Recession and the tenure changes occurring during the crisis. In particular, we aim to investigate whether subpopulations were affected by the Great Recession to varying extents. To do so, we employ two approaches. First, we analyse mean tenure by subgroup before and during the crisis, which makes it possible to identify trends and extreme cases. Second, we apply a shift-share analysis to examine to what extent evolutions of tenure within countries are caused by compositional effects. This method enables us to decompose the total difference observed over time into two components. The ﬁrst component is attributable to changes in the distribution of subgroups, holding tenure within groups constant, and the second 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 66 is caused by differences within subgroups, holding the distribution of groups constant.

These two components are obtained by

4Tenure = ∑ 4Sharet ∗ Tenuret + ∑ Sharet4Tenuret (3.1) t t where 4Tenure represents the difference in mean tenure between two time periods, i denotes the group and Sharet is the share of this group in the total workforce. The bars denote the mean over both time periods. The ﬁrst term on the right-hand side of the equation equals the difference in mean tenure attributable to shifts in employment shares between groups with different tenure, and the second term reports the difference in mean tenure within each group for ﬁxed employment shares. We focus on the results for age groups and contract types.

3.4.1. Worker Characteristics: Age

By its nature, the relationship between tenure and age is strong and positive; one further year of tenure is, by deﬁnition, also a further year of age. This becomes apparent in Figure

3.4, where we distinguish between individuals aged 15-24, 25-34, 35-54 and 55 years and older. Mean tenure is systematically higher for the older age groups in comparison with the younger age groups. The change over time, however, is not very pronounced when aggregating for EU countries; the oldest age group (55 and over) exhibits a slight increase in mean tenure during the crisis (2008-2012), while the 35-54 age group experiences a slight decrease in mean tenure. Mean tenure for the two youngest age groups (15-24 and 25-34 years, respectively) remains constant over the observation period.

A closer assessment of the data reveals various cross-country heterogeneities among age groups. Figure 3.5 provides discrete information on the situation in countries where the dispersion in mean tenure between age groups is large and in others where it is small.

There is a large dispersion for example in Germany, France and Sweden. For these countries, the oldest age group (55 years and over) exhibits mean tenures of roughly 250 months, while means in the second oldest age group (35-54 years) are approximately 100 months 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 67

Figure 3.4.: Mean tenure by age group in the EU, 2002-2012

300 70 Mean Mean tenure (months) of age15

60 250

54 and age > 54 > age and 54 200 -

40 150

- 24 24 andage 25 100 20

10 -

34 Mean tenure (months) of age 35 age of (months) tenure Mean

0 0

Age 35-54 Age >54 Age 15-24 Age 25-34

Mean tenure for age groups 35-54 and 55+ years is plotted against the primary axis. Mean tenure for the age groups 15-24 and 25-34 years is plotted against the secondary axis. Note that we distinguish between the primary and secondary axis to make the variation visible in all age groups. Source: EU-LFS, authors’ calculations. less. Bulgaria, Lithuania and the United Kingdom have a particularly low dispersion in mean tenure across age groups, with a maximum difference of around 40 months. One possible difference between low- and high-dispersion countries is the degree of job mobility and job security. In high-dispersion countries, job changes are apparently less frequent and tenure is thus more strongly related to age.

To concentrate on the importance of the crisis for the evolution of mean tenure, Figure 3.6 provides country-speciﬁc levels of mean tenure before and during the crisis, by age group.

Countries with a high mean tenure among older workers (55 years and over) before the crisis are Belgium, France, Italy, Luxembourg and Slovenia; in these countries, mean tenure is at least 250 months. Countries with a rather low mean tenure among older workers are

Estonia, Latvia, Lithuania and the United Kingdom (with roughly 150 months). 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 68 (b) Low wage dispersion (a) High wage dispersion Figure 3.5. : Mean tenure by age group for selected EU Member States, 2002-2012 Mean tenure for ageagainst groups the 35-54 secondary and axis; 55+ wegroups. years distinguish is between plotted the against primarySource: and the secondary EU-LFS, primary authors’ axes axis; calculations. to for make age the groups variation visible 15-24 in and all 25-34 age years 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 69 Figure 3.6. : Mean tenure by age group before and during the crisis by EU Member State For speciﬁc age groups,months the (for vertical the 25-34-year axis age startsSource: group) at EU-LFS, and authors’ 100 values calculations. months above (for zero the 35-54 to and make 55+ the -year age variation groups). visible. Values thus start at 40 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 70

For the youngest age group (15-24 years), the highest levels of tenure are apparent in Aus- tria, Greece, Italy, Portugal and Slovakia (all above 25 months before and during the crisis), while mean tenure is very low in Sweden and Finland (roughly 15 months). These ﬁnd- ings are in line with Cazes and Tonin’s(2010) results for age-speciﬁc mean tenure across

EU countries before the Great Recession.

A comparison of the pre-crisis period with the crisis period suggests that changes in mean tenure are small in all age groups when aggregating data across EU countries (Figure 3.6).

However, there are speciﬁc countries that deviate from this aggregated perspective. In Bul- garia and Romania, very young workers (15-24 years) exhibit a large relative increase in mean tenure (greater than 20 per cent). A large number of job losses among young workers during the crisis might account for this increase. In addition, both countries altered their regulations for atypical working contracts, in particular for young people, during the crisis [Clauwaert and Schomann¨ 2012], which might have caused the observed increases in mean tenure for young workers. In several countries, however, changes have resulted in a decrease in mean tenure (for example, in Austria, Hungary and the Netherlands). Among older workers (55 years and over), increases in mean tenure during the crisis were relatively large (about 5 per cent) in Cyprus, Spain and the United Kingdom; but in several

(mostly small) countries such as Lithuania, Luxembourg and the Netherlands, mean tenure decreased for this age group. In Lithuania, the decline in mean tenure for older workers is very high and might account for the exceptional decrease in mean tenure for the entire workforce during the crisis (discussed above in the section on the aggregate evidence). As mentioned previously, this phenomenon might be explained by either a reform of early retirement rules or legislation simplifying the dismissal of older workers

The results of a shift-share analysis are displayed in Table 3.1 and provide a more precise description of the underlying mechanisms. For the entire observation period (2002-2012), the increase in mean tenure at the EU level due to a compositional change in the age structure of the population is large (+9.3 months), which is in line with the values in Table 3.A.1 of the appendix. Before the Great Recession, the corresponding ﬁgure is 4 months, whereas for the crisis period it is 5 months. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 71 group group group 2002-2007 2007-2012 2002-2012 Table 3.1. : Shift-share analysis of change in mean tenure, according to age Due to Due to Total Due to Due to Total Due to Due to Total changing changes change changing changes change changing changes change composition within the composition within the composition within the The crisis period takes intomajority of account countries. changes See which Figure occurredSource:3.2 between EU-LFS,for 2007 authors’ country calculations. and codes. 2008 because the crisis started in 2008 in the EUATBEBG 4.19CY 4.32CZ 7.45DE 3.55 -3.62DK 3.15 -11.98EE 2.36ES 4.26 -5.70 -1.54 0.57 -7.13FI -7.66FR -1.50 2.18GR 1.75 0.40 4.29 -3.59HU 0.23 5.21 -12.91 4.88 2.32IE 5.18 5.33 -3.12 8.03IT 2.76 4.49 -14.45 5.26LT 2.32 -4.01 4.49LU -5.12 1.80 0.93 -3.37 1.80 -4.63LV 1.91 7.39 5.15 2.72 -1.30NL 0.28 3.41 -6.57 3.92 -4.73PL 7.96 5.70 -0.24 -1.05 6.14 3.05PT 1.67 -7.71 7.09 -4.14 6.73 5.88RO -2.08 -7.39 14.19 -1.15 0.53 5.09SE -2.33 -1.06 8.02 2.77 -0.74 6.81 9.29 -6.84SI -5.91 6.55 -2.34 5.89 9.31 8.98SK 1.09 2.33 8.47 11.87 4.00 -4.20UK 7.82 2.35 4.94 8.89 3.45 3.18 6.16 -8.31 1.12 10.80 -15.52 2.93 2.68 4.87 -2.58 12.03 2.90 2.76 0.53 10.13 -19.35 -3.05 -12.20 1.68 3.81 7.53 23.08 0.28 -10.64 1.45 4.99 8.16 -8.71 -6.70 -5.72 3.74 -1.73 5.46 6.71 6.04 -10.04 1.60 7.08 -6.38 -0.33 -1.87 2.84 3.72 -15.61 18.55 -1.01 -2.10 8.75 7.27 -2.16 -0.77 -1.76 -2.27 -19.18 -8.08 4.44 -0.49 16.84 -11.78 7.70 4.79 -3.45 -12.71 3.00 8.78 10.67 5.80 8.03 -14.19 16.78 0.51 1.65 6.75 4.81 10.36 1.14 -6.63 1.13 -1.37 -8.04 7.66 13.29 8.84 9.46 -2.48 -4.32 7.40 17.71 -0.74 8.55 -1.28 4.23 23.36 0.13 0.69 5.90 -1.30 -11.59 11.51 -4.92 -2.36 5.22 0.47 -1.94 -6.29 9.93 7.50 -1.23 -9.67 15.48 5.35 -20.40 4.87 4.52 -3.79 -18.43 6.59 10.93 6.90 14.33 3.16 -11.84 8.034 15.30 -6.92 4.54 1.63 7.28 -3.06 -9.26 13.12 12.05 -23.90 -0.72 6.98 1.80 -4.74 4.58 -10.60 -9.57 -18.92 -12.34 14.58 -6.06 -5.81 -0.29 4.35 8.93 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 72

This suggests that during both periods, the increase in mean tenure seems to be related to a change in the share of age groups. In this speciﬁc case - and taking into account demographic changes in terms of higher life expectancy and lower birth rates - the share of older workers has increased, and therefore seems to be a driving force for higher mean tenure.

Interestingly, the contribution made to the overall changes in mean tenure in the pre-crisis period (+0.57 months) by factors other than an ageing population was strongly negative

(-3.62 months). Cazes and Tonin[2010] report similar results. During the crisis, however, mean tenure isolated from the compositional aspects due to ageing increased by 0.93 months. This rise was driven in particular by increases in countries such as Bulgaria (approximately 6 months), Estonia (9 months), Ireland (6 months), Spain (9 months) and the

United Kingdom (7 months). Furthermore, the shift-share analysis indicates that the upturn in mean tenure in Germany during the Great Recession was driven by the ageing of the labour force. This explains why mean tenure rose even though the unemployment rate was unaffected by the crisis.

The shift-share analysis with respect to age thus yields two important ﬁndings. First, older age groups have much longer tenures, which to a large extent is due to the strong positive correlation between tenure and age. Second, mean tenure increased strongly across all age groups during the crisis. This might explain why overall mean tenure increased remarkably during the crisis. Yet the shift-share analysis provides evidence that a large part of this increase in tenure (aggregated across EU countries), especially in the pre-crisis period, was due to changes in the age composition of the labour force. In other words, aside from the crisis, the compositional aspect attributable to a growing share of older workers is relevant when explaining why there is a rise in overall tenure. In conclusion, our results suggest that exogenous impacts from the crisis may be responsible for a considerable upward shift in tenure, while it seems that there exists a long-run trend towards shorter job tenure. This

ﬁnding is remarkable because it indicates that there is a declining trend in tenure once age effects are taken into consideration. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 73

In the course of further analysis, we examined changes in job tenure over time and differences across countries by gender and skill group (the results are available from the authors, upon request). The results imply that mean tenure is somewhat greater for working men at the EU level. This is similar to the ﬁndings of Farber[2010] and Auer and Cazes[2000], which show that men enjoy longer tenures than women. However, for some countries, such as the Baltic States, where womens’ labour participation is high, the opposite relationship is found. Overall, an increase in job tenure can be observed during the crisis for both men and women.

Concerning skill levels, for the EU aggregate we find no strong differences in mean tenure across skill groups, and no trend is discernible over the total observation period, which is in line with the findings of Burgess[1999]. At the country level, we find examples where mean tenure for high-skilled workers is higher than for low-skilled workers, but we also find examples of the opposite. Nevertheless, with respect to the evolution of tenure over time, analysis indicates that those countries hit hardest by the recession experienced a larger increase in mean tenure among medium-skilled workers. This is, for example, the case for Portugal and Spain.

3.4.2. Job Characteristics: Contract Type

This section assesses mean tenure for permanent and temporary employment at the EU aggregate level and for speciﬁc countries, since cross-country differences in the prevalence of temporary contracts may explain cross-country differences in the development of mean tenure. This is particularly important because the increasing spread of temporary employment in recent decades goes hand in hand with a tendency towards dual or segmented labour markets [Boeri 2011; Cahuc, Charlot and Malherbet 2016].

Since permanent employees are by deﬁnition more likely to remain in their current employment relationship than temporary workers, the mean tenure of permanent workers can be expected to be higher. Additionally, in many EU Member States, temporary contracts are used as probation periods for newly hired workers. Figure 3.7 shows that the 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 74 average tenure is more than four times higher for workers with permanent contracts than it is those with temporary contracts.

Concerning time trends, average tenure remained rather stable during the pre-crisis period for temporary and permanent workers alike (Figure 3.7). However, for permanent workers, mean tenure increased with the beginning of the crisis in 2008/2009, growing from 131 months in 2008 to 138 months by 2012. At the same time, the mean tenure for temporary employees is seen to behave countercyclically, that is, it increases during a recession and decreases during a boom. That said, the overall changes are small in absolute terms; the mean tenure of temporary workers increases by only two months between 2007 and 2012.

Figure 3.7.: Mean tenure by contract type in the EU, 2002-2012

200 30

180 27 Mean tenure (months) Mean tenure (months) of temporary jobs 160 24

140 21

120 18

100 15

80 12

60 9

40 6 Mean tenure (months) of permanent jobspermanentof (months)tenure Mean 20 3

0 0

permanent job temporary job

Permanent jobs are plotted against the primary axis and temporary jobs are plotted against the secondary axis Source: EU-LFS, authors’ calculations.

As for temporary employment, Figure 3.8 depicts year-on-year changes during the observation period for selected countries. It displays a mixed picture of trends in the mean tenure of temporary workers across countries. Temporary workers in Denmark experienced a strong decline in mean tenure until 2005, followed by a rather stable level until 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 75

2012. By contrast, the mean tenure of temporary employment increased in Spain over the entire observation period. In Greece, mean job tenure levels for temporary employees behave counter cyclically, as they decrease during a boom and increase during a recession.

The opposite holds true for Lithuania, where it behaves pro-cyclically.

Figure 3.8.: Mean tenure of temporary workers in selected EU Member States, 2002-2012

months) 25

15 Mean tenure ( tenure Mean

DK ES GR LT EU

See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations.

In order to shed light on these differences, Figure 3.9 plots the mean tenure of temporary workers as well as the share of temporary workers among the working population by country during the pre-crisis period and the Great Recession. For both measures, considerable variation across countries and time becomes apparent. In Austria, Greece, the Czech

Republic, Italy and the United Kingdom, the mean tenure levels of temporary workers is comparably high, at around three years. By contrast, temporary workers have an especially low mean tenure in Estonia, Lithuania and Slovakia. One reason for this could be that these countries have temporary contracts with exceptionally low durations compared to other countries. Estonia, Lithuania and Slovakia also have very low shares of temporary employment among total employment (illustrated in the lower panel of Figure 3.9), 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 76 implying that in these countries temporary contracts are rare and that when workers are temporarily employed their status changes quickly. At the same time, those countries with the highest shares of temporary employment (Poland, Portugal and Spain) are characterized by higher average levels of mean tenure among temporary workers.

In general, temporary employment appears to be less common in the CEE countries (except for Poland). Auer and Cazes[2000] found a comparable distribution of temporary employment across the EU. This points to a positive relationship between EPL and the prevalence of temporary employment. In other words, in many countries with strict EPL (for example, Portugal), companies tend to use temporary employment to ensure ﬂexibility over the business cycle. This phenomenon is investigated in more detail in the next section of this article, on regression analysis.

However, the results do not directly conﬁrm the hypothesis that temporary workers are necessarily the ﬁrst to lose their jobs when a crisis hits. While the share of temporary workers was indeed strongly reduced during the Great Recession in Spain, for example, we actually see an increase in Portugal. In general, no clear relationship can be established between the share of temporary employment and the pre-crisis mean tenure level, the share of temporary employment prior to the crisis or its change during the crisis. In total, 15 countries experience a decrease and 11 countries show an increase in mean the tenure level

(upper panel of Figure 3.9). However, this result is not too surprising, because there is no clear hypothesis as to which temporary workers are likely to be the first to lose their jobs. Firm-specific human capital accumulation ought not to differ significantly, given the limited variation in tenure within temporary employment. Additionally, country-specific labour legislation governing temporary employment ought to also play an important role.

If, for example, legislation establishes that temporary workers must receive permanent contracts after a certain period, ﬁrms might actually be inclined to lay them off close to this time limit in order to maintain a certain degree of ﬂexibility during a recession.

The shift-share analysis makes it possible to study to what extent observed changes in aggregate mean tenure are caused by the changing mean tenure levels of permanent and temporary workers, or by a changing composition of the workforce with respect to these 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 77

Figure 3.9.: Mean tenure of temporary workers before and during the crisis by EU Member State (a) Mean tenure of temporary workers

months) 25

Mean tenure ( tenure Mean 15

0 SI BE FR IT LU DE SE PT GR AT NL EU SK FI PL CZ HU RO CY ES DK LT IE BG UK EE LV

pre-crisis crisis

(b) Share of temporary employment among total employment

10 Share of employment (percent) employmentof Share

0 SI BE FR IT LU DE SE PT GR AT NL EU SK FI PL CZ HU RO CY ES DK LT IE BG UK EE LV

pre-crisis crisis

Countries are sorted in order of their pre-crisis mean tenure level. See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 78 two contract types. The results for the EU aggregate level show that average tenure for given shares increased slightly, by 2.6 months, in the period from 2002 to 2007, but that the changing composition of temporary and permanent workers counteracted this increase.

This negative share component shows that temporary contracts became more common, reducing overall tenure. During the crisis, the share component was negligible, and tenure for a given share composition increased quite strongly (Table 3.2).

This ﬁnding that a change in the importance of temporary employment played hardly any role in the evolution of tenure during the Great Recession is surprising, but is con-

ﬁrmed when we analyse developments for individual countries. Regardless of whether the sign is positive or negative, a change in the composition of contract types (with either an increase or a decrease in temporary employment) is relatively unimportant for the majority of countries in this context. The only exceptions are Spain, where falling shares of temporary employment are correlated with increasing mean tenure levels, and the Nether- lands, where an increasing share of temporary employment is associated with a reduction in mean tenure. Instead in composition effects, it is actual changes in mean tenure for the two contract types that dominate the development. In this context, it is worth mentioning that the increase in the share of temporary workers in the Netherlands during the crisis was extraordinary (Figure 3.9, lower panel) and probably related to an accelerated move by the nation’s economy from an industry- to a more service-orientated economy [Gielen and Schils 2014]. In total, 20 of the 26 countries analysed experienced an increase in mean tenure during the crisis, holding the respective shares of contract types constant (Table 3.2).

Similar analyses for economic sectors and occupations indicate extremely large differences in the mean job tenure level across economic sectors (the results are available from the authors, upon request), thus conﬁrming the ﬁndings of Auer and Cazes[2000]. During the Great Recession, most economic sectors experienced a modest increase in mean tenure, mainly due to a change in tenure within sectors rather than a changing sectoral composition. Moreover, for occupations, a similar picture can be drawn. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 79 group group group 2002-2007 2007-2012 2002-2012 Due to Due to Total Due to Due to Total Due to Due to Total changing changes change changing changes change changing changes change composition within the composition within the composition within the Table 3.2. : Shift-share analysis of change in mean tenure, according to contract type The crisis period takes intomajority of account countries. changes See which Figure occurredSource:3.2 between EU-LFS,for 2007 authors’ country calculations. and codes. 2008 because the crisis started in 2008 in the EUATBE -2.06BG -1.31CYCZ -0.89DE 1.07 2.64 -3.48DK -6.35 -0.26EE -2.45 2.63ES -2.51 0.57 -4.66FI -7.66 8.81FR 0.17 3.02GR 1.75 2.35 -3.59 6.94HU 0.45 -11.93 1.01 -0.43 5.33IE -1.09 2.76 0.94IT -4.80 0.49 -14.45 0.16 4.49LT 1.19 -2.07LU -1.83 -3.65 -1.25 5.69 -0.62 -4.63 8.18LV -0.28 -3.38 0.31NL 5.79 0.28 3.61 0.22 -2.57 -2.68 0.37PL 6.83 -0.24 -1.05 6.14PT -0.51 -2.26 7.09 -1.15 4.15 -5.09RO 4.28 6.73 3.39 -2.08 6.98 -14.62 7.50SE 0.53 -1.26 8.02 -1.73 2.96 1.95 -1.62 3.80 -2.34 -5.91SI -0.49 -0.24 13.18 -0.66SK -1.39 4.00 0.00 -0.41 0.90 6.77UK 7.82 2.35 -0.70 -1.87 16.10 3.18 3.98 2.29 1.12 -5.16 -0.89 -6.98 -4.93 12.03 3.51 8.33 2.76 7.57 -6.32 3.08 -0.57 -0.70 -9.39 1.68 0.30 23.08 -10.64 -2.19 0.08 0.79 7.85 -2.25 -1.57 -8.71 -1.21 0.11 -1.14 5.46 17.73 2.14 6.71 8.15 7.08 -10.04 -0.35 1.48 2.84 -7.18 10.13 -6.95 -0.33 1.44 7.32 8.74 -2.27 8.75 1.36 -4.44 -1.76 -14.98 16.84 9.92 4.44 -6.83 3.00 3.03 -0.07 -3.45 -12.71 8.54 1.89 -1.32 4.58 6.75 -14.19 8.03 13.23 -6.63 1.41 7.07 1.65 1.75 -1.63 4.46 -4.58 -8.04 7.66 -1.77 1.70 7.40 0.54 2.19 4.23 11.25 23.36 5.61 4.43 -4.04 0.13 -0.08 5.90 -5.19 -3.80 13.72 15.51 0.40 0.47 5.22 9.93 8.67 1.46 7.50 3.88 -11.57 -16.28 -13.12 12.08 15.48 -3.79 10.93 -1.23 7.36 -3.12 6.90 -0.73 3.16 -11.84 1.30 8.04 13.38 3.10 0.56 8.38 7.28 -6.92 -8.50 -3.31 -8.84 1.80 6.98 13.29 -4.74 0.22 -6.61 -9.57 1.29 14.58 3.03 -6.06 8.71 -5.81 -0.29 8.93 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 80

3.5. Econometric Findings: Regression Analysis

In the final step of the analysis we perform a multivariate regression analysis of the relationship between tenure before and during the Great Recession with a wide range of socio- demographic and job characteristics, taking into account country-specific variation. This allows us to draw conclusions as to the influence on tenure of each observed and time- constant unobserved country specific factor. The regression includes socio-demographic variables such as age, gender and skill-level. Job characteristics are described by the economic sector and occupation. In order to focus on a country-specific variation of tenure, we exclude from the regressions the contract type, that is, information on whether a worker has a permanent or a temporary contract. Contract types are therefore subsumed under country-specific institutions, and country dummies must thus be interpreted in this regard.

Moreover, we focus on interpreting the characteristics discussed in the previous sections, using the additional variables as controls. The controls include place of birth, type of employment (full-time, part-time), work pattern (shift work) and company size.

Estimated coefficients from linear regressions that are fully interacted with the crisis dummy are reported in Table 3.3. The second column reports pre-crisis values, while the third column displays changes during the crisis. For their interpretation, both depend on the specific reference categories. The basic reference group consists of 35-54 year-old, medium- skilled, non-migrant men working in Austria,5 in manufacturing, as service and sales employees, with non-shift and full-time contracts, in companies with between 20 and 49 employees. This reference group, after controlling for all socio-demographic and job characteristics, has on average a mean tenure level of about 160 months (159.82 months) during the pre-crisis period. For the same group, job tenure decreased by 5.65 months during the crisis. This is a striking result, since it demonstrates that, ceteris paribus, there was a sizeable and significant negative impact made by the crisis on mean tenure in several European countries. In particular, the corresponding coefficient suggests that mean tenure decreased by 5.65 months on average during the crisis in the reference country, Austria. When de-

5A robustness test using Belgium instead of Austria as the reference country in the regression yields very similar results. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 81

Table 3.3.: Results of regression analysis of individual tenure before and during the crisis

Base Change coefﬁcient during crisis

Reference category (intercept and crisis dummy) 159.82 *** -5.65 ***

Age group

15-24 years -111.03 *** 1.57 *** 25-34 years -82.37 *** 1.16 *** 35-54 years ref ref 55+ years 78.83 *** 2.91 ***

Gender

Male ref ref Female -3.11 *** 1.79 ***

Skill level

Low-skilled: ISCED 0-2 1.62 *** -2.46 *** Medium-skilled: ISCED 3-4 ref ref High-skilled: ISCED 5-6 -20.47 *** 0.19 ***

Economic sector

A - Agriculture, forestry and ﬁshing 1.73 *** 0.1 ** B - Mining and quarrying 34.41 *** -0.73 *** C - Manufacturing ref ref D - Electricity, gas and water supply 31.8 *** -18.44 *** E - Construction -25.47 *** 1.79 *** F - Wholesale and retail trade; vehicle repair -13.63 *** 2.18 *** G - Hotels and restaurants -19.75 *** 0.88 *** H - Transport, storage and communications 5.06 *** -9.88 *** I - Financial intermediation 17.2 *** -2.81 *** J - Real estate, renting and business activities -28.06 *** 1.34 *** K - Public administration and defence 22.37 *** 4.07 *** L - Education 16.75 *** -4.90 *** M - Health and social work -3.59 *** -1.14 *** N - Other community, social and personal service activities -10.41 *** 0.14 *** O - Activities of households as employers -14.77 *** -5.68 *** P - Activities of extraterritorial organisations -3.42 *** 4.55 ***

Occupation

Armed forces occupations 33.94 *** 3.64 *** Managers, senior ofﬁcials and legislators 19.24 *** 1.86 *** Professionals 21.84 *** 0.75 *** Technicians and associate professionals 19.02 *** 3.13 *** Clerks 13.74 *** 3.42 *** Service and sales workers ref ref Skilled agricultural, ﬁshery, and forestry workers -9.58 *** 9.09 *** Craft and related trades workers 9.99 *** 1.76 *** Plant and machine operators, and assemblers -6.51 *** 1.88 *** Elementary occupations -21.93 *** 3.41 ***

Country

AT ref ref BE 5.66 *** 1.10 *** BG -49.2 *** 11.23 *** CY -13.6 *** 5.89 *** CZ -28.88 *** 2.63 *** DE -13.21 *** 3.69 *** DK -36.55 *** -0.73 *** EE -45.27 *** 5.23 *** ES -9.66 *** 9.39 *** FI -17.53 *** 5.99 *** FR 6.34 *** 1.27 *** GR -0.07 ** 4.40 *** HU -25.48 *** 0.99 *** IE -16.49 *** 8.51 *** IT -0.97 *** 6.87 *** LT -39.55 *** -10.76 *** LU 4.15 *** -3.47 *** LV -42.98 *** 7.89 *** NL -13.95 *** 5.36 *** PL -21.82 *** 0.4 *** PT 9.20 *** 0.33 *** RO -29.3 *** -0.36 *** SE -11.84 *** -0.89 *** SI 11.79 *** 8.65 *** SK -16.24 *** -1.2 *** UK -37.01 *** 10.01 *** R2 0.33 Number of observations 1945604584 Estimated coefficients are reported. The reference group (ref) has the following combination of characteristics: age 35-54 years, male, medium-skilled, full-time employed, no shift work, medium firm size (20-49 workers), occupation: service and sales workers, economic sector: manufacturing, country: Austria. The reference individual has a mean tenure of 159.82 months in the pre-crisis period and a corresponding mean tenure of 154.16 months during the crisis. Other controls: Place of birth (ref: national; others: born in another EU country, born outside the EU). Type of employment (ref: full-time employment; others: part-time employment). Company size (ref: 20-49 workers; others: 1-10, 11-19, 20-49, 50+, more than 10 but not sure). Work pattern (ref: no shift work, others: shift work). Significance levels are: *** p < 0.01, ** p < 0.05, * p < 0.10. See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 82 scriptive techniques were used, a decline in average tenure during the crisis period was reported in Austria, although of a much lower magnitude, of around 1 month (Figure 3.9);

16 countries depict an interaction term of less than 5.65 months, thus indicating a negative trend in job tenure too. Only nine countries (Bulgaria, Cyprus, Finland, Ireland, Italy,

Latvia, Slovenia, Spain and the United Kingdom) experienced a rise in mean tenure.

These fundamental differences suggest that the development of tenure in the period of the

Great Recession was largely determined at the country level. In addition, most of the countries with a rise in mean tenure have in common the fact that their overall economy suffered disproportionately strongly from the ﬁnancial crisis (see Figure 3.9 for comparison). This emphasizes that country-speciﬁc labour market performance, structure and institutions, social security systems and corresponding reforms are relevant in explaining country differences in tenure, as mentioned above in the section on worker and job heterogeneities.

The results for age groups conﬁrm the descriptive pattern according to which job tenure is substantially longer for older workers than for younger ones. The estimated coefﬁcients show that mean tenure for younger age groups (15-24 and 25-34 years, respectively) is much lower than for the reference group of prime-age workers (35-54 years) in the pre- crisis period, and that this difference narrows slightly during the crisis. For the oldest age group (55 years and over), mean tenure is considerably greater than for the reference group.

Therefore, the regression ﬁndings are in line with the implications of the shift-share analysis previously presented; they indicate that age effects are part of the explanation for the overall increase in mean tenure. Put differently, the underlying negative trend in mean tenure only becomes visible when controlling for the composition of the workforce in terms of age. A potential reason is demographic change, in the form of higher life expectancy and lower birth rate. The resultant higher share of older workers drives mean tenure higher.

Isolating the crisis-tenure relationship from other factors is thus central to revealing a potential structural trend towards shorter tenures.

Furthermore, the estimates conﬁrm the descriptive evidence that mean tenure is higher for men than for women. However, when controlling for other factors, the gender gap 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 83 in tenure is relatively small. It corresponds to three months before the crisis and to 1.3 months during the Great Recession. Interestingly, the pre-crisis level of mean tenure is more than 20 months higher for medium-skilled workers than for high-skilled ones. One likely explanation is that high-skilled workers more often look for a new job while still in employment and are thus more likely than medium- and low-skilled workers to change jobs voluntarily.

Turning to job characteristics, the relationship between tenure and economic sector highlights important differences between sectors. Again, the regression results in Table 3.3 are in line with the descriptive evidence, as a strong variation in mean tenure across economic sectors becomes visible. The estimates imply that tenure is very high for the energy and water supply sectors and very low for the hotel and restaurant sector. In general, sector specific changes in mean tenure during the crisis are considerably smaller once the terms interacting the crisis dummy with economic sectors are taken into account. This is also true for mean tenure with respect to occupations. This result is noteworthy because it implies that changes in tenure during the crisis seem not to have been strongly related to specific types of occupations, but rather to specific types of workers.

Table 3.A.2 in the appendix displays estimated coefﬁcients for the additional control variables. Natives experience substantially higher job tenures than migrants. Tenure is positively related with ﬁrm size and negatively associated with working hours. Lastly, shift work goes together with longer job duration. For all control variables, changes during the crisis are not very pronounced.

In order to get an impression as to the importance of the different groups of explanatory variables for job tenure, we conducted a variance decomposition. As one can see, the relative importance of worker characteristics in explaining job tenure is very high (Appendix,

Table 3.A.3). Two-thirds of the model’s explanatory power can be attributed to age, which partly reﬂects the fact that age naturally rises with tenure. Job characteristics and country speciﬁc factors account for 14 per cent and 8 per cent of the total variation, respectively.

Within the group of job characteristics, industry and occupation are the most important.

Lastly, the explanatory power of the crisis indicator is rather small. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 84

As stated previously, the country dummies represent country-speciﬁc labour market performance, structure and institutions - including the prevailing EPL in each country - social security systems and corresponding reforms during the observation time. In the literature [Auer and Cazes 2000], labour market institutions, in particular EPL, are identiﬁed as a driver of cross-country differences in job tenure. EPL measures the costs that arise for

ﬁrms in the case of a dismissal of an employee. The more stringent the EPL, the more costly it is for employers to lay off workers. Therefore, EPL reduces labour turnover [Blanchard and Portugal 2001], which leads to higher job tenures.

We will ﬁrst descriptively investigate the relationship prior to the Great Recession, and second utilize our regression results to shed some light on changes during the crisis. In the pre-crisis period, mean tenure in the Benelux countries, Austria, Germany and the

Mediterranean countries is comparably high, while the CEE countries plus Ireland and the

United Kingdom are characterized by low average tenure. Slovenia and Spain are notable exceptions to this pattern: Spain has a lower average tenure than the other Mediterranean countries, and Slovenia has an extremely high average tenure compared with the other

CEE countries. This cross-country pattern of mean tenure in the pre-crisis period broadly

ﬁts the diversity of the EPL index constructed by the OECD (Organisation for Economic

Co-operation and Development)[2004] governing regular employment across the EU.

As can be seen in Figure 3.10, countries with a low degree of employment protection, such as Ireland and the United Kingdom, display relatively low mean tenure in the pre-crisis period, while the opposite is true for Italy and Portugal. This is in line with the theoretical expectation that employment protection reduces worker turnover [Mortensen and

Pissarides 1999b] and conﬁrms the evidence provided in the ﬁndings of Cazes and Tonin

[2010] for the pre-crisis period. However, labour market institutions such as EPL are not the only determinants of mean tenure, as is clear from a comparison of Estonia with Poland, which share similar EPL indices but have very different mean tenure levels. In some countries, other labour market institutions are thus likely to have played an important role as well. This is true, for example, for short-time working schemes and working-time accounts, which helped to avoid many lay-offs, especially in Germany [Burda and Hunt 2011]. Still, 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 85 the correlation coefficient of mean tenure and the EPL index is 0.62, and it is statistically significant at the 1 per cent significance level.

Figure 3.10.: Relationship between mean tenure and the EPL index for EU Member States, 2007

160

140 BE FR SI IT LU DE GR PT 120 SE AT NL FI HU CZ PL SK 100 ES UK IE DK EE 80

60 Mean tenure (months)tenureMean

0 0 1 2 3 4 5 EPL Index

For the missing countries, the EPL index is unavailable. See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations.

For the period of the Great Recession, one could expect that the degree of EPL is also correlated with the change in tenure across countries. In a ﬂexible labour market with a low degree of EPL, companies could be expected to react quickly to the economic situation, and mean tenure closely follow the unemployment rate. By contrast, in labour markets with a high degree of EPL, companies could be expected to smooth their hiring behaviour over the business cycle and mean tenure would react, if at all, after a certain time lag

[Mortensen and Pissarides 1999b]. At the same time, in countries with high employment protection, workers would have greater difficulty finding a new job once they were no longer employed [Martin and Scarpetta 2012]. Finally, job-to-job transitions leading to new worker-firm matches could be expected to decrease [Boeri 1999]. All these factors 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 86

Figure 3.11.: Relationship between the change in mean tenure during the recession and the EPL index for EU Member States

10 UK ES IE SI 8 IT 6 FI EE NL GR 4 DE CZ 2 HU FRBE PL 0 PT 0 1 2 DK SE 3 4 5

Change in mean tenure (months) tenure mean in Change SK -2

LU -4

-6 EPL Index

For the missing countries, the EPL index is unavailable. The change in mean tenure is derived from regression results (see text). See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations. contribute to the expectation that increases in job tenure will be strongest in those countries where the EPL index is high.

These expectations are indeed borne out by our empirical evidence. As we are dealing with changes during the Great Recession, it is particularly important to control for GDP growth.

We therefore used the coefﬁcients of the country dummies, the country ﬁxed effects, for the period from 2008 to 2012 from the regression presented in Table 3.3, which additionally control for worker and job composition. Correlating the change in tenure measured in this way with the degree of employment protection yields a clear negative correlation

(Figure 3.11). This means that countries with a low degree of employment protection, such as Ireland and the United Kingdom, display a much stronger increase in tenure during the recession than countries with more stringent EPL, such as the Czech Republic and Portu- gal. This implies that in countries with lower employment protection, more short-tenured 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 87 jobs are destroyed, which leads to an increase in mean tenure. Furthermore, temporary employment seems to play some role as well, as the large increase in mean tenure seen in

Spain indicates. These results are thus in line with the evidence presented in the ﬁndings of Gnocchi, Lagerborg and Pappa[2015], who, using a panel of OECD countries, show that lowering employment protection increases the volatility of employment.

3.6. Conclusion

In summary, at the EU level, average job tenure increased slightly, from 116.5 months in

2002 to 123 months in 2012. As this observation period includes the Great Recession, cyclical factors are likely to provide an important explanation. First, short-tenured jobs were disproportionately destroyed during the crisis; second, there was less job creation during the crisis, leading to a lower number of newly created jobs [Bachmann et al. 2015]. This is in line with evidence from the United States Bureau of Labor Statistics (BLS), which reported that median tenure increased in that country between 2006 and 2014 [BLS (Bureau of Labor Statistics) 2014]. We may thus conclude that mean tenure is characterized by a strong cyclical component that has to be separated from long-term trends.

However, at the individual country level, strong heterogeneities prevail both before the

Great Recession and in the reaction to it. Possible reasons for diverging pre-crisis levels of mean tenure include composition effects in terms of the workforce or industry structure; different labour market institutions; and the country-speciﬁc preferences of workers in respect to the ”lifetime employment relationship”. In terms of labour market institutions, a particularly important role is apparently played by EPL, the prevalence of temporary contracts and, more generally, labour market ﬂexibility.

Even when abstracting from cyclical effects, there appears to be no evidence that mean tenure decreased in Europe between 2002 and 2012. However, it is of paramount importance to control for socio-demographic developments when comparing the evolution of mean tenure across countries. Once a control is introduced for an ageing workforce, an underlying negative trend towards shorter job tenure becomes apparent for many countries. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 88

The shift-share analysis suggests that lay-offs due to the crisis, together with the ageing of the workforce, may be responsible for a considerable upward shift in tenure. At the same time, there seems to be a long-term trend towards shorter job tenure for certain age groups.

The regression results conﬁrm that the age effect is apparently a force driving an increase in mean tenure, even during the crisis. The ageing of the workforce, combined with an underlying negative trend in mean tenure for speciﬁc age groups, result in the average tenure remaining stable once cyclical effects are abstracted.

Tenure also varied strongly according to job characteristics during the observation period.

Cross-country differences in the impact of the crisis on temporary employment suggest that the role of such employment in national labour markets varies considerably between

EU countries. However, large job losses among temporary workers and a corresponding increase in average tenure can be observed in only a very few countries, and most no- tably in Spain. Therefore, in our context, the conclusion of Bentolila et al.[2012] that the spread of temporary employment is an important predictive factor for labour market developments during the Great Recession cannot be generalized beyond a small number of countries. Overall, the results from the regression analysis suggest that the prime explanatory factor for tenure is age, followed by job characteristics, especially economic sector and occupation.

Moreover, our analysis conﬁrms the results in Cazes and Tonin[2010] showing a strong positive correlation between the degree of employment protection and mean tenure. In addition, our study shows that employment protection also played an important role during the Great Recession: countries with a low degree of employment protection experienced stronger increases in tenure during the Great Recession than did those with more stringent employment protection. There were more lay-offs in countries with little employment protection.

Our results can inform several policy debates. First, the underlying trend towards declining job tenure in Europe uncovered by our analysis is relatively sizeable. From the shift-share analysis displayed in table 5.1, we see that this effect amounts to -2.58 months over the period from 2002 to 2012, when we observe an overall increase in job tenure of 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 89

6.71 months. This can be seen as problematic because a decline in job stability is likely to adversely affect individual worker welfare. An analysis of potential factors explaining this underlying trend, such as increased voluntary job-to-job transitions (which would point to supply-side factors in the labour market) or more frequent dismissals (which would point to demand-side factors) constitutes an important ﬁeld of research. Second, while the widespread use of temporary contracts in a number of European countries continues to be a cause for concern because of the emergence of dual labour markets, the Great Re- cession does not seem to have increased this structural problem. Third, EPL apparently has had some stabilizing effect on European labour markets during the Great Recession.

The welfare effect of this stabilization is not clear-cut, however. Finally, as our analysis uses repeated cross-sections, it must remain silent about life-cycle issues. In particular, the question of whether the higher prevalence of temporary contracts among younger workers means that labour-market careers can be expected to be more unstable in the future is of great importance. In order to answer this question, one would have to conduct a cohort- based analysis, in the spirit of Ehrlinghagen and Knuth[2002] or Hanushek et al.[2017].

These issues are, however, beyond the scope of this study, and are therefore left for future research. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 90

3.A. Supplementary Figures and Tables

Table 3.A.1.: Summary of sample before and during the crisis, percentage Pre-Crisis Crisis Age group 15-24 years 11.4 10.3 25-34 years 26.1 24.7 35-54 years 52.2 52.1 55+ years 10.4 12.9 Gender Male 53.2 52.0 Female 46.8 48.0 Skill level Low-skilled: ISCED 0-2 21.0 18.1 Medium-skilled: ISCED 3-4 52.8 51.9 High-skilled: ISCED 5-6 26.2 29.9 Economic sector A - Agriculture, forestry and fishing 1.9 1.5 B - Mining and quarrying 0.5 0.5 C - Manufacturing 20.7 18.0 D - Electricity, gas and water supply 1.1 1.7 E - Construction -7.1 6.8 F - Wholesale and retail trade; 13.7 13.8 vehicle repair G - Hotels and restaurants 3.7 4.1 H - Transport, storage and 6.6 -8.0 communications I - Financial intermediation 3.4 3.4 J - Real estate, renting and 8.5 8.7 business activities K - Public administration and defence 8.6 8.6 L - Education 8.3 8.6 M - Health and social work 10.5 11.5 N - Other community, social and 4.2 3.5 personal service activities O - Activities of households as employers 1.2 1.4 P - Activities of extraterritorial 0.1 0.1 organizations Occupation Armed forces occupations 0.6 0.7 Managers, senior officials and legislators 5.7 5.5 Professionals 13.9 16.3 Technicians and associate professionals 17.4 17.4 Clerks 13.2 12.2 Service and sales workers 14.3 16.0 Skilled agricultural, fishery, 1.0 0.9 and forestry workers Craft and related trades workers 14.1 12.4 Plant and machine operators 9.7 8.6 and assemblers Elementary occupations 10.1 10.0 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 91

Pre-Crisis Crisis Country AT 1.9 2.0 BE 2.1 2.1 BG 1.5 1.5 CY 0.2 0.2 CZ 2.3 2.3 DE 18.6 19.2 DK 1.5 1.4 EE 0.3 0.3 ES 8.7 8.5 FI 1.2 1.2 FR 12.7 12.7 GR 1.6 1.5 HU 2.0 1.9 IE 0.9 0.9 IT 9.5 9.5 LT 0.7 0.7 LU 0.1 0.1 LV 0.5 0.5 NL 4.1 4.0 PL 6.1 6.7 PT 2.2 2.1 RO 3.4 3.4 SE 2.3 2.3 SI 0.5 0.5 SK 1.1 1.1 UK 14.3 13.7 Place of birth National 93.4 91.5 Born in other EU-28 countries 2.1 2.8 Born outside of EU-28 4.6 5.7 Work part-time No 82.7 80.9 Yes 17.3 19.1 Firm size 1-10 persons 22.8 22.9 11-19 persons 10.8 11.4 20-49 persons 16.2 16.2 50 and more 44.4 44.3 More than 10, but does not know exactly 5.9 5.3 Work in shiftwork No 82.6 82.8 Yes 17.4 17.2 See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations. 3. JOB STABILITY IN EUROPE OVERTHE CYCLE 92

Table 3.A.2.: Results of regression analysis of individual tenure before and during the crisis Base coefficient Change during crisis Place of birth National ref ref Born in other EU-28 countries -24.16*** -3.38*** Born outside of EU-28 -30.74*** -0.31*** Work part-time No ref ref Yes -23.67*** 0.55*** Firm size 1-10 persons -10.97*** -0.55*** 11-19 persons -4.44*** -0.22*** 20-49 persons ref ref 50 and more 17.26*** -1.58*** More than 10, but does not know exactly -8.20*** -1.45*** Work in shiftwork No ref ref Yes 5.42 *** 0.01 R2 0.33 Number of observations 1945604584 Estimated coefficients are reported. The reference group (ref) has the following combination of characteristics: age 35-54 years, male, medium-skilled, full-time employed, no shift work, medium firm size (20-49 workers), occupation: service and sales workers, economic sector: manufacturing, country: Austria. Note that the reference individual has a mean tenure of 159.82 months in the pre-crisis period and a corresponding mean tenure of 154.16 months during the crisis. Significance levels are as follows: *** p < 0.01, ** p < 0.05, * p < 0.10. See Figure 3.2 for country codes. Source: EU-LFS, authors’ calculations.

Table 3.A.3.: Importance of estimated components for the explanatory power of the model, percentage of the predicted variance Base coefficient Change during crisis Age group 66.78 0.02 Gender 0.05 0.01 Skill group 1.78 0.01 Nationality 1.29 0.00 Worker 69.90 0.05 Industry 5.23 0.21 Occupation 4.79 0.06 Part-time 1.65 0.00 Firm size 2.22 0.01 Shift work 0.10 0.00 Job 13.99 0.29 Country 8.29 0.26 Crisis 0.18 The importance of each estimated component corresponds to the share of the model’s predicted variance explained by the component. The shares are a function of sample variances and estimated effects. They are ad- ditive. The computation uses the results of the benchmark regression (table 3). Specifically, the predictive im- ˆ2 ˆ portance of a variable is equal to βi Var(Xi)/Var(Y). Source: EU-LFS, authors’ calculations. 4. How Does Potential Unemployment Insurance Benefit Duration Affect Reemployment Timing and Wages?*

Abstract: We examine dynamic treatment effects of a potential beneﬁt duration (PBD) extension from 12 to 18 months in the German unemployment insurance (UI) system. Under- standing dynamic effects is important to assess the value of UI reforms as a policy tool. We revisit the regression discontinuity design (RDD) of Schmieder, von Wachter and Bender

[2016], but use more detailed data and a wage decomposition to analyze the effects of PBD in a framework that allows for unrestricted heterogeneity of both duration and wage effects. Our results indicate treatment effect heterogeneity on durations, leading to a sample selection bias that complicates learning from LATEs as identified by the RDD. PBD almost exclusively prolongs unemployment spells ending close to exhaustion points and affects wages through the firm fixed effect, which adds to evidence on the importance of firms in wage setting. Moreover, dynamic selection on worker characteristics is non-monotonic and may create spurious treatment effects. Dynamic treatment effect heterogeneity can reconcile diverging findings on the effects of PBD.

*This chapter is co-authored with Hanna Frings and Nikolas Mittag. It is currently a mimeo. We thank Ronald Bachmann, Thomas K. Bauer, Stˇ epˇ an´ Jurajda, Jakub Kastl, Andrea Weber and participants at presenta- tions at CERGE-EI, RWI, UNCE and the Universities of Bath and Bochum, the 11th RGS Doctoral Conference in Economics, the 21st IZA Summer School in Labour Economics, the COMPIE Conference 2018 and the Jahresta- gung of the Verein fur¨ Socialpolitik 2018 for comments. All remaining errors are our own. Financial support by the Leibniz Gemeinschaft within the research grant “Worker Flows, Match Quality and Productivity - Evidence from European Micro Data” is gratefully acknowledged. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 94

4.1. Introduction

Unemployment insurance (UI) benefit extensions have been revived as a policy tool in the last Great Recession. Understanding the effects of UI is crucial to assess its value as a policy tool and can shed light on important labor market questions, such as the role of job search, the causes and consequences of unemployment duration and the implications of higher turnover. Recent studies follow Angrist, Imbens and Rubin[1996] in addressing the pervasive selection problems of the early literature by using quasi-randomized variation in PBD, which often varies discontinuously at age cutoffs, to identify a local average treatment effect (LATE) at the discontinuity. While these studies improve identification, they still disagree on key findings, including the direction of the effect of PBD on wages.2

A potential cause for the wide range of opposing estimated effects is that while LATEs are identiﬁed by quasi-randomization, they can be difﬁcult to interpret if treatment effects are heterogeneous. It is well known that in the presence of heterogeneity, LATEs by themselves may contain little information about the underlying causes and have limited external validity [Heckman, Urzua and Vytlacil 2006; Deaton 2009; Imbens 2010]. The problem of heterogeneity is particularly severe for dynamic analyses, such as responses to PBD, because there may be heterogeneity in the duration response, in the response of the outcome

(e.g. wages) to duration and in the response of the outcome to treatment.

This paper provides insights on how heterogeneity in the treatment effects on unemployment durations and wages affects the interpretation of LATEs of PBD on reemployment wages. We combine tools from the literature on treatment effect heterogeneity with a wage decomposition into worker, ﬁrm and job ﬁxed effects to develop a framework of counterfactual outcomes that is not restrictive on heterogeneity. This framework improves our understanding of how allowing for heterogeneous duration and wage responses can alter the interpretation of LATEs. It also provides novel empirical tools for estimating and (partly)

2These studies agree that an increase in PBD leads to longer unemployment durations, but the magnitudes vary substantially [Card, Chetty and Weber 2007; Van Ours and Vodopivec 2008; Tatsiramos 2009; Caliendo, Tatsiramos and Uhlendorff 2013]. For reemployment wages, several papers document a negative effect [e.g. Lalive 2007; Schmieder, von Wachter and Bender 2016], while Nekoei and Weber[2017] ﬁnd a positive effect in a very similar research design. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 95 purging dynamic selection effects3 from dynamic treatment effects. Applying these tools further enables us to study the patterns of dynamic selection and duration dependence, which are of interest independent of the effect of PBD extensions. The tools we develop are not only relevant for analyzing PBD extensions, but for any evaluation in which dynamic treatment effects are of interest in the presence of effect heterogeneity at the individual level.

We then apply our methods to a PBD extension from 12 to 18 months in Germany that has been studied by Schmieder, von Wachter and Bender[2016] with a regression discontinuity approach. Thereby, we estimate PBD extension effects on reemployment outcomes for middle-aged workers with a high labour market attachment. We use linked-employer- employee data (LEED), which allow us to decompose wages into time-varying observables and unobservables (i.e. the residuals), as well as an individual, firm and job component via a multi-way fixed effects model. In the following analysis we use these wage components, rather than wages, as the outcomes of interest. The worker fixed effect and the pre-determined observables isolate dynamic selection on observable and time-invariant characteristics. Thereby, our decomposition allows us to examine the nature of (a substantial part of) dynamic selection and how it changes with treatment. The remaining wage components better isolate the treatment effect, which is especially important in analyzing dynamics. Studying how treatment effects vary between wage components and over time can shed light on potential mechanisms through which PBD affect reemployment outcomes.

Our results reveal unexpected and interesting patterns in dynamic selection and duration dependence, which are two important explanations for the steep decline of reemployment wages with increasing unemployment durations independent of treatment. We show that dynamic selection indeed generates meaningful variation in wages over the unemploy-

3Dynamic selection refers to changes in average reemployment wages with unemployment duration, because workers with different characteristics leave unemployment at different durations. A common assumption is that high-wage workers exit unemployment early, while low-wage workers experience longer unemployment spells. Duration dependence refers to changes in the reemployment wage if the same worker exits at different durations, i.e. variation in reemployment wages caused by differences in unemployment duration. For example, wages are often thought to decline because human capital depreciates with unemployment duration or because stigma increases with unemployment duration. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 96 ment spell. Yet we do not confirm the common assumption of monotonically negative dynamic selection. Instead up until the respective exhaustion points, worker fixed effects and previous work experience do not seem to exhibit any trend. Thus, dynamic selection may cause wages to vary non-monotonically with unemployment duration, i.e. positive and negative dynamic selection is likely present over the unemployment spell. The pattern is more clear-cut at the respective exhaustion points, when low-wage workers with long experience tend to exit unemployment. Isolating the (dynamic) effects on the remaining wage components from dynamic selection shows that the steep wage drop over the unemployment spell is driven by the firm fixed effects and time-varying unobservables instead.

This result adds to the importance of ﬁrms in wage setting [Card, Heining and Kline 2013], pointing toward ﬁrm-level factors such as bargaining power and monopsony rather than worker-level causes, such as skill depreciation, as potential explanations for the adverse effects of unemployment.

After documenting the general pattern of dynamic selection, we estimate static and dynamic treatment effects of the PBD extension on unemployment duration (“duration effect”), dynamic selection (“dynamic selection effect”), and wages (“wage effect”). For the duration effect, we ﬁnd strong evidence of heterogeneity, which mainly stems from more frequent long unemployment spells. The results point toward longer PBD affecting unemployment duration mainly because workers close to the exhaustion point prolong unemployment to the new exhaustion point. The frequency of relatively short unemployment durations (less than 10 months) does not react to the PBD extension from 12 to 18 months.

The unchanged frequency of relatively short unemployment durations under treatment does not automatically imply that the PBD extension has no dynamic selection effect. First, even in the absence of systematic dynamic selection effects, slight differences in the non- monotonic pattern of dynamic selection between treatment and control group may generate spurious dynamic selection effects that would still bias any dynamic treatment effect.

Substantive differences, including differences in the estimated sign of the effect, may thus arise from simple choices of the analyst such as the length of the analysis window within studies of the same population. Second, our results suggest that the PBD extension induced 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 97 some high-wage workers with short work experience, who would have exited unemployment early under short PBD, to leave our sample which only covers employment subject to social security contributions. They may enter self-employment or migrate, among others.

Such systematic sample selection is difﬁcult to detect. It may bias not only the dynamic treatment effects, but also the sign of the LATE.

The results for the dynamic treatment effects on the firm fixed effects indeed suggest that they are confounded by sample selection bias. Six months of additional PBD lead to a decrease of 0.02 log points in the reemployment firm fixed effect. The magnitude is sizable; it implies that workers with longer PBD accept jobs at firms that pay 2 percent less to all their workers compared to the group with shorter PBD. Importantly, the effect appears to accumulate at short unemployment durations, raising doubts whether duration effects, that can only be observed at the exhaustion points, and wage effects are related. Instead the effect on the firm fixed effect appears to be caused by the workers in the treatment group with high worker fixed effects that would have moved to high-paying firms under short PBD.

The next section provides an overview of the literature, the RDD framework and our wage decomposition. Section 4.3 introduces a simple, but nonrestrictive, framework of counterfactual outcomes to illustrate the problems of heterogeneity. Section 4.4 reviews the institutional background, introduces our data, and shows that our data reproduce prior results. Section 4.5 establishes heterogeneity in duration effects and examines dynamic selection. Section 4.6 analyzes the remaining wage components after (partly) purging dynamic selection to learn more about the mechanisms though which PBD extensions affect reemployment wages. The ﬁnal section concludes.

4.2. Empirical Strategy

4.2.1. Identifying Local Average Treatment Eﬀects

The early literature established that unemployment insurance affects behavior by documenting strong exhaustion effects [see e.g. Hunt 1995; Meyer 1990]. Maximum UI beneﬁt 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 98 duration depends on the length of prior labor force participation, which makes the labor market history related to later labor market outcomes. To solve this problem, the more recent literature uses policy induced discontinuities of PBD at age cutoffs to identify LATEs of PBD on unemployment duration [e.g. Caliendo, Tatsiramos and Uhlendorff 2013; Card,

Chetty and Weber 2007; Tatsiramos 2009; Van Ours and Vodopivec 2008] and wages [Lalive

2007; Nekoei and Weber 2017; Schmieder, von Wachter and Bender 2016]. The discontinuities at age cutoffs locally identify the effects of potential beneﬁt duration, γ, the key parameter of interest, in the following standard regression discontinuity model:

∗ ∗ ∗ yi = α + γDai≥a + β0(1 − Dai≥a )ai + β1Dai≥a ai + εi (4.1)

where yi is an outcome, such as unemployment duration or reemployment wages. ai is

∗ the age of individual i and Dai≥a is an indicator for the individual being older than the relevant age threshold a∗ at the time of separation.

In addition to using the same identification strategy, these studies are also similar in that they are all based on social security records from a few European countries. Nonetheless, they vary substantially in their findings and conclusions. While they all find longer PBD to increase average unemployment duration, the size of the effect varies substantially across studies. For the effect of PBD on reemployment wages, even elementary results such as the direction of the effect deviate across studies. For example, Schmieder, von Wachter and

Bender[2016] ﬁnd a negative effect of PBD on reemployment wages in Germany. Despite many similarities in the institutional set-up and the identiﬁcation strategy, Nekoei and

Weber[2017] ﬁnd a positive effect of PBD on reemployment wages in Austria.

To reconcile these contradictory results it is important to note that PBD theoretically affects reemployment wages through two opposing mechanisms. First, PBD may help workers to overcome search frictions, thereby directly increasing match quality, wages and job stability. This function of PBD as a subsidy to search can also be understood in a broader sense as any factor increasing reemployment wages, for example by strengthening workers bargaining position, solving liquidity constraints or allowing workers to climb the job ladder 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES 99 by moving to high-wage ﬁrms. Second, PBD extensions may constitute a pure moral hazard problem. In this case PBD extensions act as a subsidy to leisure only inducing workers to exhibit less search effort, leading to longer unemployment durations without any positive effects on search behavior.

The empirical distinction between these two mechanisms is complicated by the established, but poorly-understood observation that reemployment wages decline steeply with longer unemployment durations [e.g. Blanchard and Diamond 1994; Schmieder, von Wachter and

Bender 2016]. A sizable share of this decline can be explained by dynamic selection, i.e. self- selection of low ability workers into longer unemployment spells. The existing evidence further suggests that human capital depreciation or stigma also plays an important role.

In what follows we will refer to this second part net of selection effects as duration dependence. While the wage decline over the unemployment spell is a robust empirical ﬁnding, differentiating between dynamic selection and duration dependence remains challenging.

The robust ﬁnding concerning the duration effect of PBD extensions can explain decreasing reemployment wages, all else equal. Thus, even in the absence of any moral hazard issues, the effect of PBD extensions on reemployment wages (wage effect) can take any sign, depending on its duration effect and the relative strength of positive search effects and negative duration dependence [Nekoei and Weber 2017].

The more recent literature on PBD extensions using RDDs acknowledges this complication and therefore analyzes reemployment wages conditional on completed unemployment duration in order to provide evidence against or in favor of search effects. The underlying idea is that reemployment wages at each unemployment duration can be compared between two groups exposed to different PBD independently of duration dependence. A positive gap is interpreted as an increase in reservation wages, which implies that PBD effectively subsidizes search. Schmieder, von Wachter and Bender[2016] ﬁnd no shift in reemployment wages conditional on unemployment durations, and conclude that PBD extensions prolong unemployment only by decreasing workers’ search effort. Thus, they suggests to use PBD extensions as instrument for unemployment duration. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES100

Despite the focus on the question if and to which extent reservation wages at each unemployment duration shift in response to PBD extensions, dynamic selection effects are generally ignored in the existing literature. While all studies conﬁrm that PBD extensions increase unemployment duration, they reject or are silent on the possibility that PBD extensions reshufﬂe workers along the reemployment path in a selective manner. Dynamic selection effects of this type still do not render the aggregate LATE invalid, but make it impossible to use shifts in reservation wages at each duration to make statements about the existence of search effects. We argue that decomposing wages into time-constant and time-varying observables and unobservables, and using these wage components as outcomes in the RDD, allows us to isolate dynamic selection effects to a large degree. Further, analyzing the wage components is informative about which channels cause changes in reemployment wages.

4.2.2. Wage Decomposition

We ﬁrst decompose the overall effect of PBD on reemployment wages into effects on wage components. We then analyze how these wage components evolve with unemployment duration and how these dynamic processes are affected by changes in PBD. For the wage decomposition we use the following match effect model:

wijt = xijtβ + θi + ψj + λij + τt + ηijt (4.2)

where wijt is the log daily wage of individual i working at firm j in year t. xijt is a vec- tor of time-varying control variables that includes work experience in days, following a basic Mincer[1974] approach, linearly as well as quadratic. Furthermore, the model comprises year fixed effect denoted as τt to account for the business-cycle (refer to Table 4.B.1 for an overview of the estimated coefficients). θi and ψj are worker and firm fixed effects. They are commonly used in two-way fixed effect models to decompose wages from linked employer-employee data [see Abowd, Kramarz and Woodcock 2008, for an overview]. The match effect model generalizes this two-way fixed effect model by additionally including 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES101

the interaction between the firm and the worker fixed effect, λij, which can be interpreted as a job or match fixed effect [see Jackson 2013; Mittag 2016; Woodcock 2015, for further discussion of the match effects model]. Identification of worker and firm fixed effects requires all firms in the sample to be connected by realized mobility. Even within such connected groups, a normalization is necessary. Hence, we normalize the average firm fixed effect to be zero. Job fixed effects nest in both firm and worker fixed effects. Therefore, the mean of the job fixed effects within each individual and firm is not identified. We normalize both to zero, so that estimated job fixed effects sum to zero for each individual and firm. For further discussion of identification, see Abowd et al.[2002] for the two-way fixed effects model and Woodcock[2015] for the match effects model.

The estimated worker fixed effect of individual i, θî, captures the wage component due to time invariant observed and unobserved worker characteristics. It is therefore informative about changes in dynamic selection, i.e. at which duration what type of individuals exit to employment. Including a job fixed effect in Equation (4.2) allows wages to differ systematically between jobs and thus allows pre- and post-unemployment wages to differ systematically. Thereby, its inclusion makes it more credible that the estimated worker

ﬁxed effect indeed isolates a permanent worker-speciﬁc wage component and is not biased by any treatment effects.4

The estimated firm fixed effect of firm j, ψˆj, indicates how much the daily wage of a worker at at firm j differs from what the average firm pays a worker with the same permanent wage component θ and time-varying observables x. The firm fixed effects – and the change in the firm fixed effect between pre- and post-unemployment jobs – is informative on how workers sort into firms and climb the job ladder.

The estimated job fixed effect, λˆ ij, captures the job-specific wage component. Such job- specific wage components likely arise if the productivity of workers differ across firms, for example due to specialization or firm-specific human capital. They may also arise from differences in bargaining power, differences in compensation over the life-cycle or any other

4Please refer to Appendix 4.B for a complete discussion on the possibility that treatment affects the estimated worker fixed effects and why this is not a serious issue. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES102 wage determinants that are constant within job, but not within firm or individual. The normalization of job fixed effects attributes average match quality to the respective individual or firm fixed effect, so that job fixed effects should be interpreted as deviations from the respective individual (or firm) average. This implies that a comparison of reemployment job fixed effects between workers with different PBD shows to which extent better matches are formed, potentially due to synergies between workers’ skills and firms’ production technologies.

4.3. Heterogeneity and Its Consequences

Recent empirical studies use shifts of the reemployment wage path to analyze if and to which extent PBD affects reemployment wages through subsidizing search. Such shifts in reemployment wages conditional on unemployment duration are interpreted as shifts in reservation wages, with a positive shift points towards PBD extensions increasing reservation wages. This strategy implicitly assumes that the effect of PBD on unemployment duration is homogeneous. Thus, while workers may select into shorter or longer unemployment duration in general, PBD extensions are assumed not to change the pattern of dynamic selection. The literature on heterogeneous treatment effects shows that this assumption is strong and non-trivial.

Homogeneous duration responses to PBD extensions are not only important for the validity of using shifts in the reemployment wage path to make statements about how UI affects the behavior of workers, but also for comparing LATEs across studies as well as for their connection to the average treatment effect [Deaton 2009; Heckman and Urzua 2010;

Imbens 2010]. Allowing for heterogeneous duration responses implies changes in dynamic selection by PBD. The LATE of PBD on wages can be positive or negative in this situation, so not even the sign of the average effect is identiﬁed.

In this section, we formalize these problems in a framework of heterogeneous treatment effects, focusing on the case of PBD, unemployment duration and wages. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES103

4.3.1. Dynamic Diﬀerences and Dynamic Treatment Eﬀects

Consider a simple framework of counterfactual outcomes to describe the relationship between PBD, unemployment duration d and an outcome Y, such as reemployment wages.5

Let YPBDi(d, X, ε) be individual i’s log reemployment wage in treatment state PBD as a function of unemployment duration d,6 other observable characteristics X and unobservables ε. This deﬁnes a function of (stochastic) outcomes for every individual and every possible UI regime (as deﬁned by its PBD).

We are interested in the relation between PBD, unemployment duration and reemployment wages, so we frequently use YPBDi(d) ≡ YPBDi(d, Xi(d), εi(d)) as a convenient shorthand. We compare unemployment durations and reemployment wages of individuals with 12 and 18 months of PBD. This allows us to further simplify notation, because we only compare four potential outcomes. Let dPBDi denote individual i’s realized unemployment duration when eligible for PBD months of UI beneﬁts. Then, Y12i(d12i) and Y18i(d18i) are the outcomes of individual i in the two possible PBD regimes, one of which is observed.

Y18i(d12i) and Y12i(d18i) are the outcomes if individual i had chosen the unemployment duration of one regime, but received the wage response of the other PBD regime. Unless there is no duration response or the wage response does not depend on PBD, these two outcomes are unobservable.

Evaluating the function corresponding to individual i’s actual potential beneﬁt duration

PBDi at the realized values (di, Xi, εi) yields the actual value of the outcome for individual i. This simple notation also allows us to ﬂexibly deﬁne counterfactual outcomes.

Evaluating functions that correspond to counterfactual PBD or evaluating them at counterfactual unemployment durations or (un)observable characteristics yields counterfactual outcomes. The difference between counterfactual outcomes deﬁnes treatment effects

[Heckman and Vytlacil 2000, 2001; Heckman 2008], such as the average treatment effect,

Y E[∆i ] = E[Y18i(d18i)] − E[Y12i(d12i)]. In order to make statements about the mechanisms

5In what follows, we will focus on PBD extensions as the treatment and reemployment wages as the outcome. Both, the econometric framework and the proposed solutions, apply to many other treatments and outcomes where dynamic treatment effects are of interest. 6Of course, unemployment duration d can be a function of other variables or PBD. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES104 leading to positive or negative average treatment effects, the literature relies on identifying dynamic treatment effects, i.e. the effect of PBD extensions on reemployment wages conditional on unemployment duration:

Y Y ∆ (d) = E[∆i |d12i = d] = E[Y18i(d18i)|d12i = d] − E[Y12i(d12i)|d12i = d] (4.3)

Note that the second term of Equation (4.3) is observed in the data, but the ﬁrst term conditions on unemployment duration in the untreated state, which is never observed along with unemployment duration in the treated state. One solution to estimating the dynamic treatment effect is to assume that it is equal to the dynamic difference, i.e. the difference between average treatment and control group outcomes at any value of unemployment duration.

DD(d) = E[Y18i(d18i)|d18i = d] − E[Y12i(d12i)|d12i = d] (4.4)

As can be easily seen from Equations(4.3) and (4.4), dynamic differences and dynamic treatment effects are not the necessarily identical. Speciﬁcally, the difference is the treatment effect ∆Y(d) plus a selection effect DSE(d):

DD(d) =E[Y18i(d18i)|d18i = d] − E[Y18i(d18i)|d12i = d]+

E[Y18i(d18i)|d12i = d] − E[Y12i(d12i)|d12i = d]

Y =∆ (d) + E[Y18i(d18i)|d18i = d] − E[Y18i(d18i)|d12i = d] (4.5) | {z } Dynamic Selection Effect DSE(d)

Equation (4.5) shows that the dynamic difference confounds the dynamic treatment effect with changes in the outcome (conditional on d) arising from individuals changing the timing of reemployment.7 Thus, if and to which extent dynamic differences are informative

7Note that our choice of 18 months of PBD as the treated state is arbitrary, but not irrelevant. Calling 12 month of PBD (or any other duration if there are more than two) the treated state leads to similar definitions. Changing this choice of label only flips the sign of treatment effects for scalar outcomes, but functions such as Y ∆ (d) can differ in more complex ways. For example, DSE(d) is the change in dynamic selection on Y18 when moving from 12 to 18 months of PBD. One can define a similar term for the change in dynamic selection for other outcomes, such as Y12. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES105 on dynamic treatment effects depends on the existence and the strength of an effect of PBD extensions on dynamic selection.

4.3.2. Heterogeneity in Duration Eﬀects Changes Dynamic Selection

Before turning to the dynamic selection effect of PBD, we ﬁrst formalize the phenomenon of dynamic selection independent of any treatment. Dynamic selection arises from individual differences in the timing of reemployment, i.e. from individuals sorting along the time dimension into different unemployment durations. Individuals exiting at a speciﬁc unemployment duration d may differ from the overall population:

DS(d) = E[Y18i(d18i)|d18i = d] − E[Y18i(d)] (4.6)

If the individuals who actually exit at some unemployment duration d˜ differ from the population of interest, DS(d˜) differs from zero even if we were able to assign the same unemployment duration d˜ to the entire population. This dynamic selection on Y18 makes outcomes to vary with duration for two reasons: Selection over time induced by the choice of reemployment timing, DS(d), and the effect of the chosen unemployment duration (i.e. duration dependence). This makes it difﬁcult to understand how unemployment duration affects wages and to understand the reasons why wages drop steeply with unemployment duration.

The literature on wage effects of PBD extensions mainly analyzes dynamic selection effects by examining if PBD extensions change the composition of workers exiting unemployment at each duration. The focus is on observable characteristics, such as age or education.

However, as Equation (4.6) shows, the key question is not whether workers differ in their characteristics, but whether they differ in their (counterfactual) wages or other outcomes of interest. Therefore, we take a slightly different approach and study dynamic selection 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES106 effects by examining the pre-determined wage components of the wage decomposition in

Equation (4.2). Conditional on duration, expected wages with 18 months of PBD become:

E[Y18i(d18i)|d] = E[xijtβ + θi + ψ18,j + λ18ij + τt + η18ijt|d]

= E[xijtβ + τt + θi|d] + E[ψ18,j + λ18ij + η18ijt|d] (4.7) | {z } | {z } Pre-determined Potentially affected by PBD

The wage components in the ﬁrst term in the second line of Equation (4.7) are clearly predetermined, i.e. they cannot be affected by the search behavior of workers that potentially changes after a PBD extension. Any change in the time-varying covariates xβˆ

(work experience as well as their non-linear terms) as well as the worker ﬁxed effects θˆi caused by a PBD extension therefore can only be explained by the dynamic selection effect.

E[xijtβ + τt + θi|d] can be estimated from the data, which allows us to study how PBD extensions affect dynamic selection. The wage components in the second term of Equation

(4.7) are potentially affected by workers’ search behavior, so this term contains the dynamic treatment effect of PBD extensions.

The wage decomposition does not only allow us to isolate the dynamic selection effect, it also enables us to describe its pattern in terms of heterogeneity. If there is no heterogeneity, we could simply subtract the (constant) duration effect from the durations of the treatment group to restore randomization conditional on d and thereby identify the dynamic treatment effect. However, if duration effects are heterogeneous, this strategy is not feasible, as it would require subtracting the (unobservable) individual duration effect. In the presence of heterogeneous duration effects, individuals can reshufﬂe the order in which they leave unemployment with and without treatment. Consequently, comparing individuals who are reemployed after unemployment duration d across treatment states amounts to comparing different, potentially uncomparable individuals.

Seeing the dynamic selection effect as arising from heterogeneous duration effects of PBD extensions allows us to test for its presence and analyze it using tools from the literature on heterogeneous treatment effects. As we argue above, the wage components in the ﬁrst 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES107 expectation of Equation (4.7) are predetermined and hence measure only dynamic selection. Therefore, we can estimate E[xijtβ + τt + θi|d = d12i] for any unemployment duration from the control group. The treatment group trivially identiﬁes E[xijtβ + τt + θi|d = d18i]. Thereby, we can estimate the part of the dynamic selection term that is driven by observable and time-invariant characteristics:

DSE(d) = E[xijtβ + τt + θi|d = d18i] − E[xijtβ + τt + θi|d = d12i] (4.8) | {z } Identiﬁed: dynamic selection effect from observable and time-invariant characteristics

+ E[ψ18,j + λ18ij + η18ijt|d = d18i] − E[ψ18,j + λ18ij + η18ijt|d = d12i] | {z } Unidentiﬁed: dynamic selection effect from time-varying unobservables

4.3.3. Identifying Eﬀects on Wages and Other Outcomes Aﬀected by Duration

Even if dynamic selection affects dynamic analyses, it does not necessarily affect static analyses, i.e. analyses of the overall treatment effect that do not condition on d [Ham and

LaLonde 1996]. However, some individuals may respond to longer PBD by dropping out of the labor market. This is a form of heterogeneous duration response (d18i − d12i = ∞) that leads to sample selection that affects the static treatment effect. Equation (4.5) clariﬁes the conditions under which this is problematic: If the treatment and control group are samples from the same distribution of the treated outcome Y18, then the expectation of DSE(d) is zero.8 However, if treatment makes the population in the treatment and control group to differ, possibly because it affects labor market participation, the two expectations may differ and even unconditional analyses will be biased. This clearly shows that while treatment may affect duration, consistency of unconditional estimates still requires treatment not to affect exit from the labor force, i.e. the composition of the treatment and control group. We return to this issue in our empirical analysis.

Further, dynamic selection is a problem for the analysis of dynamic treatment effects if treatment status is related to both duration d and the outcome of interest Y. If treatment

8 Taking the expectation of the ﬁrst term over the distribution of d18 (i.e. the treatment group) and the expectation of the second term over the distribution of d12 (i.e. the control group) both yields E[Y18i(d18i)], so that the dynamic selection terms cancel in the overall population. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES108

status does not affect duration, then dPBDi = di. Thus, DSE(d) = E[Y18i(d18i)|d = d18i] −

9 E[Y18i(d18i)|d = d12i] = E[Y18i(d18i)|d = d18i] − E[Y18i(d18i)|d = d18i] = 0. Therefore, if treatment does not affect duration, the dynamic difference between initially randomized treatment groups estimates the dynamic treatment effect.10 A similar argument clariﬁes why dynamic selection only becomes a problem if duration is also correlated (spuriously or causally) with the outcome of interest: If Cov(Y18i(d18i), d) = 0, then E[Y18i(d18i)|d] =

11 E[Y18i(d18i)], so that DSE(d) = 0. If these conditions do not hold, additional assumptions are required to identify dynamic treatment effects, because the dynamic difference DD(d) at any given duration d may re-

ﬂect a treatment effect (∆Y(d)) or a dynamic selection effect (DSE(d)), i.e. differences in the average outcomes of individuals that leave at time d in the treated and untreated state.

Several approaches from the program evaluation literature may be useful, such as controlling for differences in the counterfactual outcomes, matching on duration in the untreated state or instrumental variables. Our empirical approach below attempts to make individuals who exit at a given unemployment duration comparable across treatment states by conditioning on covariates and ﬁxed effects, which makes it a control function approach.12

The ideal control function would estimate E[Y18i(d18i)|d = d18i] − E[Y18i(d18i)|d = d12i]. As pointed out above, the second term is unobservable, so we cannot estimate these two

9The arguments presented here are easy to extend to the general case of multiple treatment states or a continuous treatment such as being assigned a PBD from a continuous distribution. If treatment does not affect duration, then E[Yai(dai)|d = dai] = E[Yai(dai)|d = dbi] for any pair of treatment states a and b and all durations d, which implies DSE(d) = 0. 10Thereby, dynamic differences and dynamic treatment effects are related in a similar way as quantile treatment effects and the distribution of individual treatment effects [Bedoya et al. 2017]: As quantile treatment effects combine unobserved individual treatment effects with the effect of individuals changing ranks in the distribution of outcomes, dynamic differences combine unobserved duration speciﬁc treatment effects with individuals changing ranks in the order of leaving unemployment. 11Studying conditional distributions or higher order moments requires the stronger assumption that Y ⊥⊥ d. 12The main alternative is matching. Rather than keeping unemployment durations as given and making the wages of individuals comparable, matching would keep wages as given and attempt to match treated individuals to individuals with the same (unobserved) untreated unemployment duration. This approach seems useful if a good model of unemployment durations and covariates that predict them is available. We do not have many covariates that predict unemployment durations, so we do not attempt to match individuals, but rather rely on the power of the wage decomposition to make wages comparable in a control function approach instead. See Heckman and Vytlacil[2007] for further discussion. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES109 terms. Combining Equations (4.4) and (4.7) shows how our wage decomposition allows us to make progress:

DD(d) = E[ψ18,j + λ18ij + η18ijt|d = d12i] − E[ψ12,j + λ12ij + η12ijt|d = d12i] | {z } Unidentiﬁed:∆Y (d)

+ E[xijtβ + τt + θi|d = d18i] − E[xijtβ + τt + θi|d = d12i] (4.9) | {z } Identiﬁed: dynamic selection effect from observable and time-invariant characteristics

+ E[ψ18,j + λ18ij + η18ijt|d = d18i] − E[ψ18,j + λ18ij + η18ijt|d = d12i] | {z } Unidentiﬁed: dynamic selection effect from time-varying unobservables

The first term of Equation (4.9) is the dynamic treatment effect, but it is confounded by two dynamic selection effect terms. We can estimate the dynamic selection effect in the second line (see Equation (4.8)), which enables us to study reemployment wages net of dynamic selection effect from observable and time-invariant characteristics. However, we cannot separate the dynamic treatment effect from the third term that captures dynamic selection on the firm and job fixed effect and the residual, because E[ψ18,j + λ18ij + η18ijt|d = d12i] is unobservable.

Thus, our approach isolates the dynamic treatment effect if the covariates and the individual ﬁxed effect completely control for changes in dynamic selection, i.e. if E[ψ18,j + λ18ij +

η18ijt|d = d18i] − E[ψ18,j + λ18ij + η18ijt|d = d12i] = 0. Under this assumption, the dynamic difference between the remaining three wage components (i.e. the ﬁrm ﬁxed effects, the job

ﬁxed effects and the residuals) is the dynamic treatment effect on the respective component and their sum is the dynamic treatment effect on wages.

4.4. Background, Data and Prior Results

4.4.1. Institutional Framework and Data

This study is based on the unemployment insurance regime that was in place in Germany between July 1987 and March 1997 with an additional two year phasing-out period. There- 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES110 fore, it corresponds the longest period without substantial changes to UI in Germany.13

PBD ranges from six to 32 months and is determined by a claimant’s age and labor market experience within the last seven years before unemployment.14 The monthly disburse- ment for claimants with children is 67 percent of the previous net earnings and 60 percent for claimants without children, respectively.15

Studying unemployment duration and reemployment wages requires the worker to have an employment spell after unemployment. Consequently, our estimates are the effect of

PBD conditional on ﬁnding a job again and do not capture effects of PBD on a permanent exit from employment. Further, the analysis focuses on middle-aged workers, since young and old employees are more idiosyncratic concerning job-to-job transitions and reemployment chances. For older individuals, reemployment is entangled with retirement decisions and those workers ﬁnding a new job after unemployment at that age are positively selected. In contrast, for younger individuals moving from a job to another job via unemployment is prevalent and periods spent in unemployment can be an intentional choice.

Therefore, we analyze claimants at the threshold age of 42 years. We employ a minimum of three years of work experience within the last seven years before unemployment, which results in a sample of labor market attached claimants. The minimum work experience of three years further ensures that PBD only varies with age for all workers. Finally, we restrict the sample to males working in West Germany to ensure homogeneity in the effects of PBD extensions and comparability over time. Thus, our main identiﬁcation strategy and sample restriction closely follows Schmieder, von Wachter and Bender[2016] and is based on an extension in PBD from 12 to 18 months at the age threshold of 42 years.

All empirical analyses are based on the LIAB Mover Model, an administrative linked- employer-employee dataset that is provided by the Research Data Center of the German

Federal Employment Agency. It covers more than 4.5 million individuals (with about 700

000 movers) employed at around 2 million ﬁrms. The LIAB Mover Model spans from 1993 to 2008 allowing us to investigate the UI system for individuals which were unemployed

13Bundesgesetzblatt, Volume 1987, Part I, p. 1542. 14For a detailed overview of the regime see Schmieder, von Wachter and Bender[2012]. 15Bundesgesetzblatt, Volume 1993, No. 72, p. 2357. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES111 between 1993 and 1999. Moreover, we can track these individuals at least nine years after the beginning of their unemployment spell. On the worker side, the data contain day- to-day information on employees covered by social security as well as on unemployment benefit recipients. Demographic characteristics, such as age and educational background, as well as gross daily wages are comprised. On the employer side the data include total employment, industry affiliation and salaries. For about 25 000 firms participating in a survey additional characteristics are available, such as detailed annual reports about the employee structure, workers hired, vacancies and revenues.

Next to its richness and high precision, the LIAB Mover Model has the additional advantage of a high degree of realized worker mobility between firms (i.e. connectedness). This is important for an unbiased wage decomposition in worker and firm fixed effects [An- drews et al. 2008]. Further, the estimation of a constant component between employer and employee, the job fixed effect, demands that at minimum two employees work at the same

ﬁrm and that those individuals are employed at least at one other ﬁrm in the sample.

Our empirical focus is on individuals who experience a period of unemployment between two full-time employment spells. Conceptually, we are interested in all types of employment before and after unemployment. However, the data does not include hourly wages, instead it provides daily wages that additionally differ due to working hours. We therefore focus on workers coming from and moving to full-time employment only, as differences in working hours are negligible for this group. Unemployment is observed in the data as receiving unemployment insurance beneﬁts. Therefore, unemployment duration is measured as the period an individual claims beneﬁts. On the other side, in the data, nonemployment duration is the length between two employment spells.

4.4.2. RDD Validity and Prior Results

As discussed in Section 4.3 the aim of this study is to gain insights in how treatment effect heterogeneity on unemployment duration and wages affects the interpretation of LATEs of PBD on reemployment outcomes. The answer to this question will further improve our 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES112 understanding through which channels PBD inﬂuences labour market outcomes. Given that we use the same institutional setting and the same PBD extension as Schmieder, von

Wachter and Bender[2016], we ﬁrst replicate their analyses of the validity of the regression discontinuity design and their basic results using our data. To ensure comparability, we use a bandwidth of two-years on each side of the threshold. Optimal bandwidth calculations

[Calonico, Cattaneo and Titiunik 2014; Imbens and Kalyanaraman 2012] suggest a smaller bandwidth, so we apply the bias correction of Calonico, Cattaneo and Titiunik[2014]. All standard errors are regression discontinuity robust [Calonico, Cattaneo and Titiunik 2014].

Despite the differences in sample composition and size, the results are remarkably similar.

Appendix 4.C shows all results and offers a detailed discussion. In summary, we confirm the validity of the RDD by ruling out that workers behave strategically around the age threshold (i.e. no bunching) and by showing that no discontinuity in pre-unemployment wages is observable at the age threshold. Thus, close to age cutoff of 42 years, those eligible for longer UI benefits only differ from those with shorter UI benefits in their age. Longer potential benefit duration is indeed quasi-randomly assigned close to the age cutoffs. Fur- ther, longer PBD increases time spent in unemployment and nonemployment as one would expect: The six months PBD extension leads to an increase of 1.51 months of unemployment duration or to an additional 2.17 months of non-employment. We find that longer

PBD decreases reemployment wages by 0.010 log points, but the effect is not signiﬁcantly different from zero (Table 4.1).

Schmieder, von Wachter and Bender[2016] report similar effect sizes on unemployment duration (1.77 months) and on wages (0.008 log points). In contrast to our results, their estimate on the wage effect is signiﬁcant at the 5 percent level using the much larger entire population of unemployed workers. Thus, the true effect is likely negative in both samples and our estimate only suffers from a lack of precision. The only substantial difference exists for the effect of PBD on nonemployment duration, where Schmieder, von Wachter and Bender[2016] report a much smaller estimate of 0.95 months. However, before the bias correction suggested by Calonico, Cattaneo and Titiunik[2014] that they do not apply, our estimate is close to theirs at slightly less than one additional month of nonemployment. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES113

Table 4.1.: Effect of PBD on unemployment and nonemployment duration and reemployment wages Unemployment Nonemployment Log Wage Duration Duration (1) (2) (3) RD estimate 1.51*** 2.17** -0.0102 [0.42837] [0.98779] [0.02087] Mean 6.87 12.75 4.25 Effect relative to mean 0.24 0.17 Observations 13392 13392 13392

Unemployment duration corresponds to beneﬁt duration, whereas nonemploymet duration measures the time between two jobs. Bias-corrected estimates with robust standard errors in brackets following Calonico, Cattaneo and Titiunik[2014]. */ **/ *** refers to α = 0.1/0.05/0.01. Source: LIAB Mover Model, own calculations.

4.4.3. The Reemployment Wage Path, Its Importance and Problems

A robust finding of the UI literature is that PBD extensions lead to longer unemployment duration [Caliendo, Tatsiramos and Uhlendorff 2013; Card, Chetty and Weber 2007; Tat- siramos 2009; Van Ours and Vodopivec 2008]. We confirm this result for our setting: The reemployment rate is lower for the group with longer PBD throughout the first 12 months and there are sizeable, but symmetric, peaks in the rates at the respective exhaustion points

(Figure 4.1). In addition to these well-established patterns, it is noteworthy that the difference between the reemployment rates is very small during the ﬁrst 7 months of unemployment.

Figure 4.1.: Reemployment rate 1 .5 Re−employment Rate (in %) 0 0 6 12 18 24 Duration in Months

Age below 42 Age above 42

Non-parametric local polynomial regression of reemployment on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Source: LIAB Mover Model, own calculations. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES114

While the effect of PBD extensions on unemployment duration is a settled issue, the question if UI is a subsidy to search or rather a subsidy to leisure is still unresolved (compare with Section 4.2.1). One key aspect in providing an answer is the shift in reservation wages at each unemployment duration in reaction to a PBD extension. Such a potential shift is an- alyzed by looking at average realized wages for each completed unemployment duration by PBD. To the extent that workers with longer PBD realize higher wages conditional on unemployment duration, the conclusion is that UI at least partly acts as a subsidy to search.

Schmieder, von Wachter and Bender[2016] show that if the distribution of reemployment wages does not change, the effect of PBD will only operate through unemployment duration. They indeed ﬁnd that no signiﬁcant differences exist in reemployment wages at each unemployment duration by PBD.

Figure 4.2 Panel A shows the reemployment path of log wages by treatment status, i.e. the realized reemployment wage conditional on observed exit time. Figures 4.2 Panel B and C give the momentary and cumulative differences at each duration. We define momentary differences as the contribution of those exiting at time di = x to the LATE, i.e. ( ) = −1 − −1 TE d N1 ∑is.t.d18i=d Y18i N0 ∑is.t.d12i=d Y12i. This graph shows which time periods are important to “generate” the LATE. The formula clarifies that a specific duration d may be important for the LATE for two reasons: Either because there is a large difference in the number of exits or because the average outcomes of those who exit at d differ substantially between treatment and control group. Cumulative differences examine whether any systematic differences accumulate over unemployment duration. They contain treatment effects for those who exit unemployment before a given duration. Cumulative differences therefore amount to the RDD estimate of the LATE for the subsample with di < x. Notice that we do not apply the bias correction of Calonico, Cattaneo and Titiunik[2014] for these differences, because it would impose additional heterogeneity across d. Hence, the cumulative differences do not converge to the RDD point estimates. However, the following interpretations are not affected, because the slopes are the same for both approaches.

We first confirm in Figure 4.2 Panel A that the reemployment wage path does not change significantly in our data either. Workers with longer PBD achieve lower reemployment 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES115 wages when finding a new job within the first 10 months of unemployment. This is contrary to theoretical expectations, which imply higher reemployment wages for the group with the lower reemployment rate. While this counter-intuitive result could be due to chance in our small sample, a downward shift in the reemployment wage path still provides evidence against reservation wage effects in the framework of Schmieder, von Wachter and Bender[2016].

Figure 4.2.: Reemployment wages (a) Reemployment Paths (b) Momentary Difference (c) Cumulative Difference .05 .005 .0002 0 0 .0001 −.005 −.05 0 −.01 Log Wage Log Wage Log Wage −.1 −.015 −.0001 −.02 −.15 0 6 12 18 24 Duration in Months −.025 −.0002 0 6 12 18 24 0 6 12 18 24 Age below 42 Age above 42 Duration in Months Duration in Months Non-parametric local polynomial regression of reemployment wages on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Panel A: Conditional on age. Panel B: Momentary treatment effects conditional on age and on the number of workers exiting in each group at unemployment duration d. Panel C: Cumulative treatment effects conditional on age and on the number of workers exiting in each group with unemployment duration smaller than d. Source: LIAB Mover Model, own calculations.

Schmieder, von Wachter and Bender[2016] argue that as long as dynamic differences are not positive, reservation wage effects are ruled out and there is no evidence in favor of an effect of job search. This test, however, assumes that PBD has no negative dynamic selection effect, which could offset positive reservation wage effects. As argued in Section 4.3, there is good reason to expect that PBD actually reshufﬂes workers in a selective manner along the reemployment path. In this case, the dynamic selection term will be positive and non-positive dynamic differences cannot be interpreted as evidence against reservation wage effects.

We provide two pieces of evidence that the dynamic selection term is indeed non-zero.

First, Figures 4.2 Panel B and 4.2 Panel C show in more detail how the negative shift of the reemployment wage path translates to the negative LATE of PBD on wages reported in Table 4.1. Workers with longer PBD who exit early in the unemployment spell realize considerably lower wages than workers with shorter PBD exiting at the same duration. If 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES116

Figure 4.3.: “Effect” of PBD on worker ﬁxed effects 2.85 2.8 2.75 Worker Fixed Effect 2.7 2.65 40 41 42 43 44 Age at Start of Unemployment

Estimated using the wage decomposition (Equation 4.2). Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Source: LIAB Mover Model, own calculations. this gap is interpreted as the dynamic treatment effect, it implies that longer PBD induces workers to take worse jobs early on. As such a causal effect appears highly unlikely, it is much more plausible that the dynamic selection effect is indeed non-zero and confounds the dynamic treatment effect at early durations. Workers leaving unemployment at the respective exhaustion points achieve, as expected, lower reemployment wages. Yet, as

Figure 4.2 Panel C shows, the LATE entirely accumulates during the ﬁrst six months of unemployment as the exhaustion points are highly symmetric and cancel each other out.

Second, Figure 4.3 that shows that the average worker ﬁxed effects drop at the age discontinuity, though the effect is not statistically signiﬁcant. This implies that workers with longer

PBD are negatively selected, which violates one of the core identification assumptions of any RDD. One possible explanation is that our estimated worker fixed effects are biased downwards for workers experiencing longer unemployment durations, which is systematically true for the group with longer PBD. As discussed in Appendix 4.B, we do not believe such a bias to be a serious issue due to the inclusion of job fixed effects. Figure 4.4 Panel

A provides additional evidence that the gap in worker ﬁxed effects in Figure 4.3 is not caused by an estimation bias by showing that there is no systematic difference between the treatment and control group when comparing the worker ﬁxed effects of all workers, i.e. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES117 all workers experiencing an unemployment spell independently of the labor market status following unemployment.

Figure 4.4.: “Effect” of PBD on worker ﬁxed effect by transition paths (a) “All” Transitions from Employment (b) Closing the Gap in Our Sample 3 3.2 2.9 3.1 2.8 3 2.7 2.9 Worker Fixed Effect Worker Fixed Effect 2.8 2.6 2.7 2.5 40 41 42 43 44 40 41 42 43 44 Age at Start of Unemployment Age at Start of Unemployment

All All minus EUE minus EN EN EUE plus EN EUE

Estimated using the wage decomposition (Equation 4.2). Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Observations meet the UI eligibility criteria. Panel A: Sample All includes all transitions from employment from individuals regardless of reemployment. Panel B: Subsample EUE corresponds to the RDD-baseline sample. Subsample EN includes transitions out of the data (e.g. self- employment or going abroad). Source: LIAB Mover Model, own calculations.

Rather, as Figure 4.4 Panel B shows, the difference in our sample arises from changes in the composition of our sample and is offset by the average worker ﬁxed effect of workers without any spells after unemployment. That is, those who become reemployed with long

PBD are slightly worse in their time-invariant wage component, while those who do not return to employment subject to social security are slightly better in their time-invariant wage components. This raises the question whether longer PBD affects the probability of reemployment or the probability of leaving our sample due to self-employment, informal employment or exits from the labor force. However, we cannot reject that the differences are due to chance. In addition, we do not ﬁnd an effect on the fraction of workers transi- tioning to other states (Figure 4.4 Panel A).

4.5. Dynamic Selection

In this section, we first test for the presence of a dynamic selection effect by showing that workers react heterogeneously to a PBD extension in their timing of reemployment. We 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES118 then use our wage decomposition (Section 4.2.2) to describe the different ways in which specific type of workers reshuffle their timing of reemployment. In doing so, we focus on pre-determined worker characteristics that cannot be affected by treatment or by unemployment, i.e. the worker fixed effect and previous labor market experience.

4.5.1. Heterogeneous Duration Eﬀects and Reemployment Probability and Timing

Considering the problem of a dynamic selection effect as arising from heterogeneous duration effects allows us to investigate whether it potentially biases dynamic treatment effects using tests for treatment effect heterogeneity [Bedoya et al. 2017]. The basic idea is that without any dynamic selection effect all workers under treatment should prolong unemployment by the same constant. As soon as some workers prolong unemployment more than others, this mechanically implies that the composition of workers leaving unemployment at each duration is not the same in the treatment and control group. This is the basic definition of the dynamic selection effect (Equation (4.8)). The test of Mittag[2019] for- malizes this idea by examining the null hypothesis that the two conditional distributions of unemployment durations do not differ. We can clearly reject the hypothesis that treatment does not affect dynamic selection. Thus, individuals reshuffle the ranks in which they leave unemployment between treatment and control group. However, the test is neither informative about the way in which they do so, nor whether the reshuffling matters for the outcome of interest.

To gain insights into the nature of the dynamic selection effect, we ﬁrst examine the distributions of unemployment durations and the timing of exit from unemployment in the two

PBD groups. Figure 4.5 plots the quantile treatment effect on unemployment duration, i.e. the horizontal distance between the cumulative distributions of unemployment duration in the treatment and control group. The quantile treatment effect is ﬂat and not signiﬁ- cantly different from zero until it slowly starts to rise between the 60th and 80th percentile.

Around the 80th percentile it becomes signiﬁcantly different from zero and quickly rises to 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES119 a difference of 5-6 months. The 80th percentile corresponds to an unemployment duration of slightly less than 12 months, so this result indicates that treatment does not change the frequency of unemployment spells that last less than 12 months. In the top 20 percentiles, the distribution of unemployment durations differs substantively by PBD, showing that long unemployment spells became more frequent. The quantile treatment effect is roughly constant at around 6 month for the top 20 percentiles. This means that for each spell lasting

12+x months in the control group, we observe a spell in the treatment group that lasts 18+x months, i.e. a large mass of workers (roughly 20 percent of our sample) exits close to the exhaustion point regardless of PBD duration.

Figure 4.5.: Quantile treatment effect

Months: 12 18 5 1 ) s h t n 0 o 1 M (

n o i t a r u 5 D

t n e m y o l p 0 m e n U 5 − 0 20 40 60 80 100 Percentile

QTE Uniform CI Bands

The quantile treatment effect on unemployment duration is the horizontal distance between the cumulative distributions of unemployment duration in the treatment and control group. Vertical lines mark beneﬁt exhaustion at 12 and 18 months for the control and treatment group, respectively. The conﬁdence interval level is 95 percent. Source: LIAB Mover Model, own calculations.

Our results imply that unemployment spells of 12 or more months became more frequent.

Concluding that long spells became longer, i.e. that the effect is due to some individuals (almost) exhausting beneﬁts of either potential duration, requires further assumptions on individual treatment effects. For the case of PBD and unemployment duration, the assumption that there are no negative individual treatment effects seems almost trivial.

Still, this is a powerful assumption, as it severely restricts rank changes and their implications. Without negative treatment effects, individuals can only change ranks by having treatment effects sufﬁciently larger than those ranked above them in the untreated state to “pass” them. This implies that the quantile treatment effect is strictly sloping upward, 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES120 which is consistent with Figure 4.5. The quantile treatment effect is ﬂat for the lower 80 percentiles, so unless there are negative duration effects, no one moved from the lower 80 percent of durations to the top 20 percent of durations.

Consequently, under the assumption of no negative duration effects, our results imply that the duration effect of PBD is entirely or almost entirely due to individuals with unemployment duration close to their PBD staying in unemployment about as much longer as the

PBD extension they get. One would expect such a pattern based on economic models and prior results [e.g. Caliendo, Tatsiramos and Uhlendorff 2013] indicating that PBD is primarily relevant for those workers for whom it presents a binding constraint, i.e. for those likely to exhaust their beneﬁts.

Empirical evidence supporting the conclusion that PBD extensions primarily affect duration via those who exhaust beneﬁts comes from the well-established fact that the haz- ards of leaving unemployment spike around the exhaustion points rather than changing more smoothly with treatment. Figure 4.1 reveals this pattern, too. Further suggestive evidence for this hypothesis can be obtained by examining average worker characteristics conditional on realized unemployment duration. Figure 4.D.1 shows average education by duration and treatment status, respectively, as an example. All education levels exhibit spikes around the relevant exhaustion point showing that those exiting at the ﬁrst exhaustion point with short PBD are similar to those exiting at the second exhaustion point with long PBD, rather than being similar to those exiting at the same unemployment duration.

Education is a pre-determined worker characteristic, so this change reflects a change in dynamic selection rather than a treatment effect. Thereby, Figure 4.D.1 suggests that duration effects indeed arise from individuals who become reemployed around the time their benefits exhaust regardless of whether they are eligible for 12 or 18 months of benefits.

Duration effects consisting almost exclusively of individuals moving from one exhaustion point to another suggests that the increase in PBD did not change binding constraints on search or UI duration for many individuals. Those who exit before exhaustion under short

PBD do not seem to change their behavior much, suggesting that PBD duration was not a relevant constraint for them. Most of those who would leave unemployment after 12 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES121 month with short PBD duration seem to exit at the new exhaustion point after 18 month. If this is related to job search, the fact that they still draw full beneﬁts suggests that PBD still constrains their job search. If the PBD extension did not change constraints on job search much, one would not expect it to yield large beneﬁts from job search. This may reconcile the negative effect of this PBD extension that Schmieder, von Wachter and Bender[2016]

ﬁnd with the positive effect of a PBD extension from a shorter base that Nekoei and Weber

[2017] document in Austria. Extensions at shorter UI durations may simply relax more binding search constraints and thereby amplify the positive effects of longer UI receipt.

4.5.2. Dynamic Selection on Observable and Time-Invariant Characteristics

Our analyses above clearly document heterogeneity in duration effects, which implies that dynamic selection changes with PBD. We examine next whether this affects what we learn from empirical analyses of reemployment wages, and how we can account for the presence of a dynamic selection effect confounding the treatment effect of interest. Specifi- cally, to study dynamic selection on time-varying observables, we examine how xijtβˆ from Equation (4.2) varies with unemployment duration.16 To study dynamic selection in time- invariant characteristics, we examine how the estimated worker fixed effect, θî, from Equa- tion (4.2) varies with unemployment duration. We are interested in two questions: First, is dynamic selection related to worker quality, i.e. does the average worker fixed effect vary with realized unemployment duration and how? In terms of the framework introduced in Section 4.3 this amounts to asking whether DS(d) = 0. Second, does dynamic selection change with PBD? If dynamic selection would not change with PBD this would imply

DSE(d) = 0∀d.

The ﬁrst question concerns variation in worker ﬁxed effects by realized unemployment duration, for which the size of the sample and its informativeness about the overall population appear more important than quasi-random PBD. Therefore, we examine four large groups of workers who encompass all PBD categories of the UI system instead of our RDD

16Since we do not allow β to vary with d or PBD, this is similar to plotting the characteristics themselves. Our approach converts them into a measure of wages by aggregating them with the returns to these characteristics as weights. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES122

Figure 4.6.: Dynamic selection on worker ﬁxed effects: Aggregated PBD groups 3.75 3.25 Worker Fixed Effect 2.75

0 6 12 18 24 Duration in Months

Short PBD 12 Age below 42 18 Age above 42 Long PBD

Non-parametric local polynomial regression of estimated worker ﬁxed wage components (Equation 4.2) on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Short PBD: PBD ranges between the overall minimum of 6 to 10 months. 12: PBD corresponds to 12 months. The control group is a subsample of this population since the maximal PBD for individuals aged below 42 is 12 months independent of work experience. Age below 42: Control Group (PBD is 12). 18: PBD corresponds to 18 months. The treatment group is a subsample of this population since the maximal PBD for individuals with work experience between 36 and 39 months is 18 months independent of age. Age above 42: Treatment Group (PBD is 18).Long PBD: PBD ranges between 20 to the overall maximum of 32 months. Source: LIAB Mover Model, own calculations.

sample. Speciﬁcally, Figure 4.6 plots the dynamics of time-invariant characteristics (θˆi) for everyone in our data for whom we can cleanly determine to be eligible for 12 or 18 months of PBD. Our treatment and control group are subsamples of these populations and included in the graph for comparison. In addition, we plot the pre-determined wage components for workers with shorter (6 to 10 months) and longer (20 to 32 months) PBD.

For the groups displayed in Figure 4.6, PBD is determined by age and work experience, which simultaneously determine wages. Therefore, differences in the level of the worker

fixed effects are endogenous and not informative. The large scale of the y-axis thereby masks the fact that average reemployment wages differ substantially across realized unemployment durations. These differences are higher than the treatment effects we find from the PBD extension from 12 to 18 months. Thus, dynamic selection generates meaningful variation in wages over the unemployment spell. Yet, up until the respective exhaustion points,17 the conditional expectations do not seem to exhibit a trend. Only the oldest age group with long benefit duration shows a clear pattern of negative dynamic

17Due to the small sample, the estimates become noisy after the exhaustion points. Our preliminary interpretation here rests on the less noisy part before exhaustion, we will add uniform conﬁdence bands to these graphs to clarify this point. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES123 selection, which may be driven by the known pattern of using UI to transit to retirement

[Grogger and Wunsch 2012].

Rather, the graphs suggest that dynamic selection induces a wavy pattern for the worker

fixed effects. The fact that the magnitude of the variation in the worker fixed effects is similar in our RDD samples and the larger PBD groups suggests that the changes over time in our RDD sample are not just noise. This gives rise to the conjecture that dynamic selection may cause wages to vary non-monotonically with unemployment duration, i.e. that both positive and negative dynamic selection are present over the unemployment spell. If so, the bias in most economic models can go in either direction, which would make results that do not address or purge dynamic selection difficult to interpret, simply because effects in either direction could also be spurious changes purely due to dynamic selection.

For analyses of dynamic treatment effects, we are less concerned with the general pattern of dynamic selection, but more interested in whether dynamic selection changes with PBD.

Figure 4.7 therefore plots the same conditional expectations of the worker ﬁxed effects as

Figure 4.6, but only for our treatment and control group. In addition, the lower panel of Figure 4.7 shows dynamic selection on time-varying observables, i.e. work experience and its quadratic term. Both time-invariant characteristics and time-varying observables exhibit systematic and large spikes at the respective exhaustion points. Those who exit around beneﬁt exhaustion are positively selected in terms of time-varying observables and negatively selected in terms of time-invariant characteristics. That is, low-wage workers with long work experience exit at the the respective exhaustion points. This further un- derlines the well-known fact that PBD affects exit from unemployment at the exhaustion points [Meyer 1990] and thereby changes dynamic selection at the longest durations.

Before the exhaustion points, the differences in the conditional expectations are likely to be small relative to sampling variation. Taking into account that negative duration effects are unlikely and that our analyses of unemployment durations rules out substantial changes in the number of individuals exiting at any given duration suggests that the differences prior to the exhaustion points are noise rather than a dynamic selection effect, i.e. that

DSE(d) = 0. However, there are two important caveats to this conclusion. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES124

Figure 4.7.: Dynamic selection on time-invariant characteristics and time-varying observables

Time-Invariant Characteristics: Worker Fixed Effect (a) Reeemployment Paths (b) Cumulative Difference 0 .1 −.005 .05 −.01 −.015 0 −.02 Worker Fixed Effect Worker Fixed Effect −.025 −.05 −.03 −.1

0 6 12 18 24 −.035 Duration in Months 0 6 12 18 24 Age below 42 Age above 42 Duration in Months

Time-Varying Observables: Xb (c) Reeemployment Paths (d) Cumulative Difference .1 .03 .05 .025 .02 0 Xb Xb .015 .01 −.05 .005 0 −.1 0 6 12 18 24 Duration in Months −.005 0 6 12 18 24 Age below 42 Age above 42 Duration in Months Non-parametric local polynomial regression of estimated time-varying observables and worker ﬁxed wage components (Equation 4.2) on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Panel A and D: Conditional on age. Panel B and E: Momentary treatment effects conditional on age and on the number of workers exiting in each group at unemployment duration d. Panel C and F: Cumulative treatment effects conditional on age and the number of workers exiting in each group with unemployment duration smaller than d. Source: LIAB Mover Model, own calculations.

First, even in the absence of systematic dynamic selection effects, small differences in the non-monotonic pattern of dynamic selection between treatment and control group may generate spurious dynamic selection effects that would still bias the estimation of dynamic treatment effects. Recall that the cumulative differences given by Figure 4.7 Panel A to

D amount to the RDD estimate of the LATE for the subsample with di < x. Clearly the contribution of the worker fixed effects and prior work experience to the LATE of an PBD extension on wages are economically significant and have a non-linear pattern. There are large differences at low realized unemployment durations. For both wage components, 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES125 the spurious effect quickly flips sign when including realized durations longer than three months. The sign changes again around the exhaustion points. Consequently, estimated effect sizes may not only reflect differences in the importance of selectivity and search in the population under study [Nekoei and Weber 2017]. Substantive differences, including differences in the estimated sign of the effect, may also arise from simple choices of the analyst such as the length of the analysis window within studies of the same population.18

Second, the absence on any systematic dynamic selection effect is difficult to reconcile with the observation that the average worker fixed effect drops at the age discontinuity (Fig- ure 4.3). Indeed, Figure 4.7 Panel B shows that this drop in the worker fixed effects is not related to the exhaustion points, but exists already at short unemployment durations and persists across all durations. One explanation is that the PBD extension induced some high-wage workers with short work experience, who would have exited unemployment early under short PBD, to leave our sample. This amounts to sample selection with the dynamic selection effect DSE(d) = ∞. The fact that we do not find any duration effects at short unemployment durations (Figure 4.5) implies that only few workers show such a reaction to the PBD extension. However, as long as these few workers have unusually high worker fixed effects, this type of sample selection could explain the drop in worker fixed effects with longer PBD very well.

Our data do not allow us to gain insights in the exact decision these workers took, but likely explanations include self-employment or going abroad. Aggregate statistics on self- employed individuals from the German Federal Ofﬁce of Statistics show that workers who become self-employed earn relatively high wages before doing so. We can replicate this result with our data: Workers moving from employment to non-employment have considerably higher wages in their previous job than workers moving from one job to another via unemployment or than workers making any transition from employment (Figure 4.D.2).

In addition, workers moving to non-employment are much more likely to have a foreign

18It should be noted that these spurious effects are unlikely to be signiﬁcant at any point. Yet, they systematically shift the average difference between treatment and control by a magnitude similar to or larger than the estimated treatment effects. This magnitude makes them relevant regardless of whether they are signiﬁcant themselves. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES126 nationality (13 percent) than workers in our RDD sample (8 percent). This supports the idea that some of these workers moved abroad.

In summary, PBD affects unemployment durations mainly through workers postponing unemployment from one duration point to the other. Some workers may also react by leaving employment subject to social security contributions altogether, but their number is too low and our sample is too small for the quantile treatment effect to pick up such effects on short unemployment durations. In contrast to the common assumption of negative dynamic selection across the unemployment spell, we find a non-monotonic pattern of dynamic selection at least up to the first exhaustion point of 12 months. Two potential problems for the identification of dynamic treatment effects are spurious dynamic selection effects created by the non-monotonicity of dynamic selection and a sample selection effect.

Our sampling variation is too big to draw any deﬁnite conclusions, but we provide the tools to revisit the literature and verify to which extent these problems exists and bias dynamic treatment effects as well as LATEs.

4.6. Isolating Wage Eﬀects

Our results from the previous section show that in order to analyze the effects of PBD on reemployment wages, we need to isolate wage effects from dynamic selection effects.

Thus, we study the reemployment wage net of dynamic selection on observables and time- invariant characteristics by purging the second term of Equation (4.9). We first investigate aggregate effects of PBD on the wage components in Section 4.6.1, where the main benefit is to gain insights into mechanisms by further decomposing the wage effect into a firm, job and residual component. Section 4.6.2 discusses dynamic effects on all three wage components. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES127

4.6.1. Eﬀects on Wage Components

Table 4.2 presents bias-corrected RDD estimates [Calonico, Cattaneo and Titiunik 2014] of the effect of PBD on the (estimated) firm fixed effect, job fixed effect and the residual.19

Even though the overall effect on wages is imprecisely estimated (Table 4.1), the effect on the firm fixed effect is significantly different from zero.20 Six months of additional PBD lead to a decrease of 0.02 log points in the reemployment firm fixed effect. This magnitude is sizable. It implies that workers with longer PBD accept jobs at firms that pay 2 percent less to all their workers compared to the group with shorter PBD. The effect on the firm fixed effect is of the same size as the effect of wages overall and potentially even larger. There is no effect on the job fixed effect (λij), or on time-varying unobservables (ηijt), as captured by the residuals. Those with longer PBD thus sort into firms that pay low wages to all employees rather than accepting low wage jobs (i.e. a lower job fixed effect) or accepting a temporary wage loss (i.e. a lower residual) to reenter employment.

As discussed in Section 4.3.3, the LATEs on firm fixed effects, job fixed effects and the residual are net of dynamic selection of workers. However, to the extent that dynamic selection also takes place on firm fixed effects, job fixed effects or the residuals themselves, the corresponding dynamic treatment effects are still biased. In addition, under the assumption of strong sorting of workers into firms or jobs, any dynamic selection on workers “spills over” to the worker and job fixed effects. Since this may also be true for the sample selection we document in Section 4.5, not only the dynamic treatment effects, but also the LATE of the PBD extension on firm fixed effects, job fixed effects and the residuals may be biased.

Table 4.2 provides tentative evidence on whether dynamic selection is still present. We show the RDD estimate of using the difference from the respective wage component of the previous job as the outcome variable. Analyzing changes in wage components improves our control function and thereby helps to detect or address problems of dynamic selection. A strong decrease in economic and statistical signiﬁcance in the coefﬁcient can be

19Figure 4.D.3 provides additional graphical evidence (without the bias correction). 20This is partly driven by a larger effect, but also by smaller standard errors, because there is less unexplained variation in wage components than in wages. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES128 interpreted as a sign that dynamic selection, including the special case of sample selection, still exists. Indeed, the point estimates for all three wage components decrease, but the decline is clearly strongest for the firm fixed effects with the coefficient more than halv- ing from −0.022 to −0.009. That the point estimate is much smaller when subtracting the previous firm fixed effect raises concerns about sample selection due to an effect on the reemployment probability, as it indicates a spurious effect on the previous firm fixed effect that cannot be affected by treatment.

Table 4.2.: Effect of PBD on reemployment wage components and pre minus post changes in wage components Firm Fixed Effect Job Fixed Effect Residuals Outcome: reemp. change reemp. change reemp. change (1) (2) (3) (4) (5) (6) RDD estimate -0.02171* -0.00878 0.00951 0.00508 -0.01267 -0.00615 [0.01237] [0.01252] [0.00806] [0.01338] [0.02455] [0.01163] ...conditional -0.01733 -0.00437 0.00981 0.00249 -0.00524 0.00206 on duration [0.01233] [0.01337] [0.00805] [0.01250] [0.00877] [0.01143] Estimated using the wage decomposition (Equation 4.2). Bias-corrected estimates with robust standard errors in brackets following Calonico, Cattaneo and Titiunik[2014]. */ **/ *** refers to α = 0.1/0.05/0.01. N: 13,392 Source: LIAB Mover Model, own calculations.

If we were to interpret the negative point estimate of the LATE of the PBD extension on the

firm fixed effects as the true treatment effect, the questions whether the treatment effect is driven by dynamic duration effects or dynamic wage effects still remains. However, negative dynamic wage effects of the PBD extensions on the firm fixed effects are implausible.

This would imply negative search effects of UI, i.e. workers accepting jobs at low-wage

ﬁrms because they have longer PBD. A more likely explanation are negative dynamic duration effects: Workers under treatment stay unemployed longer, and longer unemployment leads to reemployment in low-wage ﬁrms.

Yet, Table 4.2 provides first evidence that the decrease in average firm fixed effects under treatment is not due longer unemployment durations. In this case, the point estimate of the LATE should drop to zero in economic and statistical significance when unemployment duration is added as a control variable, because no difference exists between treatment and control group in the reemployment firm fixed effect conditional on duration. This is not 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES129 the case: The coefficient drops slightly from −0.022 to −0.017 and the standard error stays unchanged. Thus, the larger share of the negative difference in reemployment firm fixed effects between treatment and control group occurs at the same durations. The following section will analyse this result in more detail.

4.6.2. Dynamics of Wage Components

To further scrutinize the relation between unemployment duration and reemployment wages, in particular, the question whether the wage loss from longer PBD operates through duration or not, we study the reemployment wage net of dynamic selection bias from observable and time-invariant worker characteristics. As Equation (4.9) shows, this yields

E[ψ18,j + λ18ij + η18ijt|d = d12i] − E[ψ12,j + λ12ij + η12ijt|d = d12i] + E[ψ18,j + λ18ij + η18ijt|d = d18i] − E[ψ18,j + λ18ij + η18ijt|d = d12i], which better isolates the dynamic treatment effect than previous studies. Figure 4.8 presents three types of analyses for each of the remaining wage components. The graphs in the ﬁrst column plot the reemployment path of the respective wage component, the second column shows the momentary and cumulative differences. Recall from Section 4.4.3 that the momentary difference are informative about which durations are important to “generate” the LATE, while the cumulative differences

21 provide the LATE for the subsample with di < x. Similarly to the analysis of dynamic selection and dynamic selection effects in Section 4.5, it is helpful to differentiate between the general pattern of how wage components develop over the unemployment spell and dynamic treatment effects22 of the PBD extension. The

first column of Figure 4.8 shows that the firm fixed effects contribute to the steep decline of wages across unemployment durations up to 12 months. In contrast, there is no contribution of the job fixed effect, e.g. workers do not seem to wait for better matches or get worse matches by extending unemployment duration. This result extends recent evidence on

21Note that the momentary difference is not the derivative of the cumulative difference, but the momentary contribution to the estimate using the entire sample. Therefore, points may be far apart from each other, especially at short durations, e.g. Figure 4.8. 22As discussed in Section 4.6.1, we cannot rule out that dynamic selection effects still confound the analysis of dynamic treatment effects. We address this issue in the following when discussing potentially affected results. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES130

Figure 4.8.: Reemployment paths: Wage components

Firm Fixed Effect (a) Reemployment Paths (b) Differences .05 .015 .00015 .01 .0001 0 .005 .00005 0 0 −.05 Firm Fixed Effect Firm Fixed Effect Firm Fixed Effect −.005 −.00005 −.1 −.01 −.0001 −.15 −.015

0 6 12 18 24 0 6 12 18 24 −.00015 Duration in Months Duration in Months

Age below 42 Age above 42 Cumulative (Left Axis) Momentary (Right Axis)

Job Fixed Effect (c) Reemployment Paths (d) Differences .05 .015 .00015 .01 .0001 0 .005 .00005 0 0 −.05 Match Fixed Effect Match Fixed Effect Match Fixed Effect −.005 −.00005 −.1 −.01 −.0001 −.15 −.015

0 6 12 18 24 0 6 12 18 24 −.00015 Duration in Months Duration in Months

Age below 42 Age above 42 Cumulative (Left Axis) Momentary (Right Axis)

Residuals (e) Reemployment Paths (f) Differences .05 .015 .00015 .01 .0001 0 .005 .00005 0 0 −.05 Residual Wage Residual Wage Residual Wage −.005 −.00005 −.1 −.01 −.0001 −.15 −.015

0 6 12 18 24 0 6 12 18 24 −.00015 Duration in Months Duration in Months

Age below 42 Age above 42 Cumulative (Left Axis) Momentary (Right Axis)

Non-parametric local polynomial regression of estimated firm fixed, job fixed and residual wage components (Equation 4.2) on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Panels A, C and E: Conditional on age. Panels B, D and F: Both treatment effects are conditional on age. The momentary (cumulative) effect is conditional the number of workers exiting in each group at unemployment duration d (smaller than d). Source: LIAB Mover Model, own calculations. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES131 the importance of firms in wage dynamics [Haltiwanger et al. 2017] and dispersion [Card,

Heining and Kline 2013]. Substantively, an effect that operates only through the wage component common to all workers at the firm is hard to justify with a model in which the worker gets worse by unemployment, i.e. human capital depreciation, since that does not affect other workers at the firm. Similarly, our results on dynamic selection (Section 4.5.2) show that average worker quality does not decline across unemployment durations; the decline in firm fixed effects is therefore not caused by low-wage workers sorting into low- wage firms.

Instead, one likely explanation of why ﬁrm ﬁxed effects decrease strongly over the unemployment spell is that workers update beliefs about acceptable wages (or acceptable employers) with increasing unemployment durations [Le Barbanchon, Rathelot and Roulet

2019; Mueller, Spinnewijn and Topa 2018]. Alternatively, employers may use observable unemployment durations as a screening device for worker-level productivity, and high- wage employers are reluctant to higher workers with long unemployment durations (sig- naling). Both explanations potentially change reemployment wages of the same worker exiting at different durations and therefore could describe patterns of duration dependence.

However, as we cannot provide an identiﬁcation of the causal effect of unemployment duration on reemployment wages, the answer to the question whether duration dependence indeed requires ﬁrm level explanations is open for future research. Our results provide an interesting starting point for doing so.

The second column of Figure 4.8 is informative on how average wage components change under treatment conditional on unemployment duration, and on how these changes contribute to the LATE of the PBD extension on the respective wage components. In the absence of dynamic selection and sample selection this amounts to the dynamic treatment effects. Once again, the results for the firm fixed effects are strongest in terms of magnitude. The LATE accumulates very quickly during the first three months. The exhaustion points have the expected signs, i.e. binding exhaustion points imply considerably lower

firm fixed effects, but they are symmetric and cancel each other out. A causal interpretation, i.e. assuming that Figure 4.8 Panel B isolates the dynamic treatment effect, seems 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES132 very unlikely. This would require some workers to accept lower wages early on, because they have longer PBD. Unless firms overcompensate them with non-monetary benefits, this seems misaligned with basic economic intuition.

Instead, just as with the worker fixed effects, the difference could be driven by exits from our sample. In this case some workers in the treatment group with high worker fixed effects that would have moved to high-paying firms under short PBD do not reappear in our data. If so, the “wage loss” stems from some individuals finding options they prefer to employment subject to social security, e.g. they may have become self-employed or went abroad (Section 4.5.2). Also note that the duration effect of the PBD extension that pushes some workers from the first to the second exhaustion point, does not translate into a negative wage effect running through the firm fixed effect. That is, postponing unemployment by another six months does not lower the average firm quality upon reemployment. Taken together, the negative LATE of the PBD extension on firm fixed effects appears to be driven by sample selection.

The contribution of the momentary differences in the job fixed effects is much smaller in magnitude than for the firm fixed effects, but shows a very similar, though mirror-inverted, pattern. That is, workers with long PBD move to high-paying jobs when leaving unemployment at short durations smaller than three months. While such an increase in match quality can be well explained by search effects of unemployment insurance, the pattern is too similar to the momentary differences of the firm fixed effects to be caused by a different explanation. We therefore interpret the positive momentary differences on the job

fixed effect to be caused by high-wage workers, who would have moved to low-wage jobs in high-wage firms under short PBD, not coming back to employment subject to social security contributions. Interestingly, there are no spikes in average job fixed effects at the exhaustion points. As expected, the residuals contribute negatively to reemployment wages at the respectively binding exhaustion points. However, the momentary differences at the exhaustion points cancel out perfectly. Overall, the residuals therefore do no contribute to the LATE of the PBD extension on reemployment wages, neither through wage effects conditional on duration nor through a duration effect at the exhaustion points. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES133

In summary, our results imply that the steep decline of wages across unemployment durations is caused by a decline in the firm fixed effect and the residual across unemployment durations, and not by a decrease in average “worker quality” as commonly assumed. The duration effect of the PBD extension that induces some workers to prolong unemployment from the first to the second exhaustion point does not contribute to the negative LATE on reemployment wages. Instead, the LATE is determined by a decline in average firm fixed effects of workers leaving unemployment during the first three months of unemployment.

As a causal interpretation seems unlikely, the effect may be due to sample selection.

4.7. Conclusion

Our analysis focuses on how treatment effect heterogeneity with respect to unemployment durations and wages affects the interpretation of LATEs of potential benefit duration (PBD) extensions on reemployment wages. We make use of the same unemployment insurance regime and identification strategy as Schmieder, von Wachter and Bender[2016]. Hence, we estimate the impact of a change in PBD from 12 to 18 months at the threshold age of 42 for workers with a high labour market attachment on reemployment outcomes in a regression discontinuity design (RDD) framework. We use a linked employer-employee data set, the LIAB Mover Model, which allows us to decompose wages into person, firm and job fixed effects.

Combining the idea of heterogeneous treatment effects with a decomposition of reemployment wages into observables and worker, job and ﬁrm ﬁxed effects allows us to provide tools to purge wage effects from dynamic selection effects. The remaining variation is more informative about dynamic and static treatment effects. These tools also allow us to study the pattern of duration effects, dynamic selection and duration dependence, which is of interest on its own.

Schmieder, von Wachter and Bender[2016] conclude that the LATE of −0.01 log wage points operates completely via the duration effect of PBD. That is, compliers of the PBD extension prolong unemployment, thereby receiving lower wages upon reemployment due 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES134 to duration dependence. While we are able to replicate their point estimate of the static

LATE, our detailed analysis of dynamic selection and dynamic treatment effects net of dynamic selection leads to a different conclusion. Our results suggest that the PBD extension affects dynamic selection in two ways.

First, low-wage workers with long experience systematically prolong unemployment from the old exhaustion point of 12 months to the new exhaustion point of 18 months. However, the dynamic treatment effects at the two exhaustion points are symmetric, cancel each other out and therefore do not contribute to the LATE. Consequently, duration dependence does not lead to lower reemployment wages when prolonging unemployment from 12 to 18 months.

Second, we provide tentative evidence that the PBD extension induced some high-wage workers with short work experience to leave dependent employment subject to social security contributions, which corresponds to our sample. Such sample selection through the reemployment rate appears to be the only plausible explanation for the large and negative dynamic treatment effects that we ﬁnd at short durations. Indeed, the LATE on the ﬁrm

ﬁxed effects accumulates within the ﬁrst three months and remains constant from there on.

Further evidence for the sample selection bias is provided by the drop in average worker

ﬁxed effects at the age discontinuity, implying that some high-wage workers, who would have moved to high-wage ﬁrms under short PBD, decide to move to self-employment, go abroad, or move to inactivity under long PBD. In summary, we conclude that the LATE of the PBD extension on reemployment wages of −0.01 log wage points may be driven by sample selection bias. In this case, even the sign of the underlying average treatment effect is unknown.

An important caveat is that our linked RDD sample is small, so many substantive results are noisy and our conclusions remain suggestive. Future research could use larger samples to gain further insights. Still, our results clearly emphasize that it is crucial to understand treatment effect heterogeneity and dynamic selection, because they document substantial heterogeneity of both duration and wage effects. We provide evidence on the nature of dynamic selection, heterogeneity and what can and cannot be learned in the presence of 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES135 these problems. A better understanding of dynamic selection can shed light on the causes of unemployment duration. Understanding dynamic selection can also help us to isolate treatment effects and their causes, such as job search and sorting into ﬁrms. We show that considering heterogeneity makes it necessary to reinterpret established results of the effect of PBD on wages, alters how we interpret estimates of the impact of unemployment duration on wages, and can reconcile seemingly contradictory prior ﬁndings [Nekoei and

Weber 2017; Schmieder, von Wachter and Bender 2016]. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES136

4.A. Identiﬁcation of UI Eligibility in the Data

To identify individuals eligible for PBD we have to derive the age at the beginning of the unemployment spell and the labour market experience within the qualifying period of seven years. We can infer the age of the claimant from the start of unemployment which is reported at daily precision and the claimant’s date of birth date which is provided in months. Concerning the labour market experience, we either calculate the number of days in employment directly for workers who have not changed jobs within the qualifying period or we construct a lower bound measure of days in employment for the other claimants by assuming every day of unemployment recorded before 1993 as taking place during the qualifying period. Because our empirical research design relies on individuals with the maximal required days in employment in their age bracket, we can be sure to identify the actual PBD treatment correctly in the data. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES137

4.B. Wage Decomposition

We apply the methods of Mittag[2016] to estimate the model including three ﬁxed effects

(worker, ﬁrm and job). We use the entire sample of full-time employees in West Germany

(including workers of all ages and females) for the decomposition into person, firm and job effects. Otherwise the firm fixed effect may depend on the sample used in the outcome model. Table 4.B.1 reports our estimates of β, which are as expected and significant at any conventional level. The fixed effects explain a substantial share of the variation in wages, underlining the importance of time-invariant unobservables in wage determination. We then use these estimated fixed effects as dependent variables in Equation (4.1). The dependent variable in Equation (4.2) is log wages, so the units of the fixed effects are also log wages. Thus, treatment effects on these wage components estimated from Equation (4.1) can be interpreted as changes in percent of the daily wage of the individual.

Table 4.B.1.: Wage decompostition Coefﬁcients SE Experience 0.0075549*** 0.0000692 Experience Sq -0.00000321*** 0.0000000455 1993 0.3290116*** 0.005341 1994 0.2833377*** 0.0045276 1995 0.2521413*** 0.0038334 1996 0.1980015*** 0.0030971 1997 0.1435327*** 0.00242 1998 0.0912137*** 0.0016614 1999 0.0504677*** 0.0009083 2000 Reference Year 2001 -0.0437623*** 0.0009027 2002 -0.0966004*** 0.0015459 2003 -0.1313081*** 0.002236 2004 -0.1898366*** 0.0030394 2005 -0.2472489*** 0.0038093 2006 -0.3014595*** 0.0046257 2007 -0.3477242*** 0.0053302 2008 -0.3920538*** 0.0060598 N 31990529 Adjusted R-squared 0.8186 F-Test Coefﬁcients 3046.83 F-Test Fixed Effects 20.27

Calculated using the algorithm described in Mittag[2016]. Source: LIAB Mover Model, own calculations. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES138

The true worker fixed effect is the wage component determined by all (observable and unobservable) time-invariant characteristics. By definition, this can only vary with unemployment duration due to dynamic selection, because unemployment duration is a time- varying characteristic. However, this argument does not trivially extend to the estimated worker fixed effect we study. The estimated worker fixed effects may be biased in the sample we analyze, because we only analyze the subpopulation for which we observe an unemployment spell. If this unemployment spell has a meaningful effect on subsequent employment spell(s), then we are implicitly selecting a sample with at least one unusually low-wage employment spell in the sampled window. This low-wage spell will drive down the average wage we observe for this worker and thus negatively affect estimation error in the worker fixed effect. Thus, our estimated worker fixed effects for those with unemployment spells may suffer from a downward bias in finite samples where the number of time periods and matches does not go to infinity for each worker.

There are good conceptual and empirical reasons to believe this problem is negligible in our analysis. An important difference to the standard two-way ﬁxed effects model that mitigates the potential bias is that we include a job ﬁxed effect in Equation (4.2). This job

fixed effect captures the average deviation of wages in a specific spell from the long-run average that is not explained by X or the firm fixed effect. Thereby, it should capture the effect of unemployment by allowing matches after unemployment to differ from those before unemployment. While this does not imply that there is no bias in the individual fixed effects, we argue that if there is any bias, it is too small to drive the patterns we document. Strong empirical evidence for this claim comes from the fact that as long as workers spent more time in our sample before becoming unemployed than after the unemploye- ment spell, any pattern of bias we find mechanically has to have a stronger (sign-reversed) effect on the job fixed effect. Yet, we find no effect on average job quality, which allows us to bound the average bias on the individual fixed effect to be very close to zero. We also do not find a strong relation of the job fixed effect with unemployment duration, again ruling out meaningful effects on the derivative of θˆ with respect to d. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES139

One may be worried about bias in the average worker fixed effects causing bias in the difference between average worker fixed effects in the treatment and control group. Such a bias may generate a spurious “treatment effect” on the individual fixed effect and thereby bias estimates of the effect on other components. We compare two groups of workers with an unemployment spell, so any bias from unemployment affects both the treatment and the control group. The difference between the groups is only biased to the extent that treatment increases or reduces the bias. Note first that this would bias the difference between average treatment and control group worker fixed effects downward and would thereby bias any other estimated treatment effects toward zero.

One should still be weary of such a bias, but we do not find evidence of its existence. The average estimated worker fixed effect in the treatment group is indeed slightly lower than in the control group (Figure 4.3), but this difference appears to be caused by a slight reduction in the reemployment rate under treatment. The difference disappears when looking at all workers who become unemployed and not only those who find a job again. Specif- ically, Figure 4.4 clearly shows that the there is no reduction in the worker fixed effect for the treatment group in the overall sample of those entering unemployment independent of the subsequent reemployment probability.23

Figure 4.4 shows further that the negative difference in our analysis sample is offset by a higher overall worker fixed effect among workers in treatment with no subsequent employment spell (i.e. workers moving to self-employment, inactivity, going abroad, or becoming civil servants). Neither difference is statistically significant, but in addition to providing evidence against bias in the worker fixed effects, this finding suggests revisiting whether PBD affects the type or probability of employment after job loss.

23Note that the difference should still become smaller if bias was the cause, but not close entirely. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES140

4.C. RDD Validity and Prior Results

This section provides a full discussion of the RDD validity checks and our replication of the basic results of Schmieder, von Wachter and Bender[2016]. First, we find no evidence for bunching around the threshold or for any structural gaps in pre-unemployment characteristics around the threshold that would suggest strategic sorting of workers, thereby ren- dering the RDD invalid. Specifically, Figure 4.C.2a provides a frequency plot of age at the start of the unemployment spell around the age cutoff. There is no evidence that individuals postpone the start of unemployment to increase their PBD. 24 Figure 4.C.2b provides further evidence of the validity of the regression discontinuity design by confirming that, as in Schmieder, von Wachter and Bender[2016], there is no effect on pre-unemployment wages.

Figure 4.C.1.: RDD validity around the age threshold (a) Frequency of Observations (b) Log Pre-Unemployment Wages 450 4.45 400 4.4 350 4.35 Number of Spells 300 4.3 Log Pre−Unemployment Wage 250 4.25 40 41 42 43 44 40 41 42 43 44 Age at Start of Unemployment Age at Start of Unemployment

Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Panel A: Round marker at age 41 and 11 months corresponds to the number of observations without the conservative age correction. Source: LIAB Mover Model, own calculations.

24For those, who start unemployment at age 41 years and 11 months two values are displayed. One showing the number of observations with an age adjustment where we include workers born in the calendar month in which unemployment starts, given that the unemployment spell starts on the 1st or the 2nd of the month. Here we assume that these workers are born after the 2nd. This will include a small number of treated individuals, roughly 1/15th of those who start UI in the month of their birthday, in the control group and slightly attenuate our results. The other excludes these workers and makes the frequency plot continuous. In fact, the frequency of new unemployment spells slightly drops after the cutoff for the unadjusted case. Thus, if there is any threshold manipulation, it is into shorter PBD. We will assess the sensitivity of our results to this convention in future analyses. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES141

Next, we conﬁrm that estimates of the effects of PBD from our data are in line with the results of Schmieder, von Wachter and Bender[2016]. Figure 4.C.2 visualizes the change in unemployment duration, non-employment duration and reemployment wages at the discontinuity where PBD jumps from 12 to 18 months. Table 4.1 provides the corresponding RDD estimates. We follow Calonico, Cattaneo and Titiunik[2014] in providing bias- corrected estimates with robust standard errors. Note that this bias correction makes our point estimates differ from the effect sizes suggested by the graphs. The results are remarkably similar to those of Schmieder, von Wachter and Bender[2016]. Longer PBD increases time spent in unemployment and non-employment as one would expect:25 The six months

PBD extension leads to an increase of 1.51 months of unemployment duration, which is close to the effect of 1.77 month reported by Schmieder, von Wachter and Bender[2016].

We ﬁnd the PBD extension to lead to an additional 2.17 months of non-employment, which is larger than the 0.95 months in Schmieder, von Wachter and Bender[2016]. However, before the bias correction that they do not apply, our estimate is close to theirs at slightly less than one additional month of nonemployment.

Figure 4.C.2.: Effect of PBD on unemployment and nonemployment duration and reemployment wages (b) Nonemployment Dura- (c) Log Re-employment (a) Unemployment Duration tion Wages 10 4.35 16 4.3 8 14 4.25 Months Months Log Wage 6 12 4.2 4 10 4.15 40 41 42 43 44 40 41 42 43 44 40 41 42 43 44 Age at Start of Unemployment Age at Start of Unemployment Age at Start of Unemployment Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Source: LIAB Mover Model, own calculations.

25The differentiation between unemployment, defined as the duration of benefit receipt, and nonemployment, defined as the time gap between two full-time employment spells without any intermediate dependent employment is not trivial in this context and the differences between these two measures of “not working” require further analyses. We focus on unemployment in this preliminary draft, but will analyze and discuss non-employment further in future revisions. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES142

While the positive effect of a PBD extension on the duration of un- and nonemployment is well established, its effect on reemployment wages is debated more. Figure 4.C.3c shows that reemployment wages decrease by 0.05 log points (5 percent) at the age cutoff where

PBD increases from 12 to 18 months. The bias-corrected estimate in Table 4.1 is smaller at

−0.01 log points and not statistically signiﬁcant at the 10 percent level. The conventional

RDD point estimate is −0.02 log points. Both estimates are similar to the effect of 0.0078 log points in Schmieder, von Wachter and Bender[2016], but not signiﬁcantly different from zero. The estimate of Schmieder, von Wachter and Bender[2016] is signiﬁcant at the

5 percent level using the entire population of unemployed workers instead of our much smaller sample. Thus, the true effect is likely negative in both samples and our estimate only suffers from a lack of precision. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES143

4.D. Additional Figures and Tables

Figure 4.D.1.: Reemployment paths: Education level (a) Low (b) Medium (c) High .004 .004 .004 .003 .003 .003 .002 .002 .002 Share Low Educated Share High Educated Share Medium Educated .001 .001 .001 0 0 0 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 Duration in Months Duration in Months Duration in Months

Age below 42 Age above 42 Age below 42 Age above 42 Age below 42 Age above 42 Non-parametric local polynomial regression of low, medium and high level education on unemployment duration in days. Vertical lines mark PBD exhaustion at 12 and 18 months for the control and treatment group, respectively. Source: LIAB Mover Model, own calculations.

Figure 4.D.2.: Wages in previous employment by PBD and transition paths

Wages (a) “All” Transitions from Employment (b) Closing the Gap in Our Sample 4.7 4.7 4.6 4.6 4.5 4.5 Log Wage Log Wage 4.4 4.4 4.3 4.3 4.2 4.2 40 41 42 43 44 40 41 42 43 44 Age at Start of Unemployment Age at Start of Unemployment

All All minus EUE minus EN EN EUE plus EN EUE

Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Observations meet the UI eligibility criteria. Panel A: Sample All includes all transitions from employment from individuals regardless of reemployment. Panel B: Subsample EUE corresponds to the RDD-baseline sample. Subsample EN includes transitions out of the data (e.g. self- employment or going abroad). Source: LIAB Mover Model, own calculations. 4. POTENTIAL BENEFIT DURATION EFFECT ON REEMPLOYMENT TIMINGAND WAGES144

Figure 4.D.3.: Effect of PBD on wage components: Graphical RDD evidence (a) Firm Fixed Effect (b) Job Fixed Effect (c) Residuals .2 .1 .2 .15 .05 .15 0 .1 .1 Residual Wage Firm Fixed Effect Match Fixed Effect .05 .05 −.05 0 0 −.1 40 41 42 43 44 40 41 42 43 44 40 41 42 43 44 Age at Start of Unemployment Age at Start of Unemployment Age at Start of Unemployment Estimated using the wage decomposition (Equation 4.2). Linear ﬁt, bandwidth of 2 age years and monthly bins. Vertical line marks age threshold at age 42. Source: LIAB Mover Model, own calculations. 5. Labour Market Participation and Atypical Employment over the Life Cycle - A Cohort Analysis for Germany*

Abstract: We use data from the adult cohort of the National Education Panel Study to analyse the changes in the employment histories of cohorts born after World War II and the role of atypical employment in this context. Younger cohorts are characterised by acquiring more education, by entering into employment at a higher age, and by experiencing atypical employment more often. The latter is associated with much higher employment of women for younger cohorts. A sequence analysis of employment trajectories illustrates the opportunities and risks of atypical employment: The proportion of individuals whose entry into the labour market is almost exclusively characterised by atypical employment rises signiﬁcantly across the cohorts. Moreover, a substantial part of the increase in atypical employment is due to the increased participation of women, with part-time jobs or mini-jobs playing an important role in re-entering the labour market after career breaks.

*This chapter is co-authored with Ronald Bachmann and Marcus Tamm. It contains a minor revised version of: Bachmann, Ronald, Felder, Rahel and Marcus Tamm. 2018. “Labour Market Participation and Atyp- ical Employment over the Life Cycle - A Cohort Analysis for Germany.” Ruhr Economic Paper No. 786. It is an extended and translated version of: Bachmann, Ronald, Felder, Rahel and Marcus Tamm. 2017. “Er- werbstatigkeit¨ und atypische Beschaftigung¨ im Lebenszyklus - Ein Kohortenvergleich fur¨ Deutschland.” Per- spektiven der Wirtschaftspolitik 18(3):263-285. We thank the editors of that journal and three anonymous referees. All remaining errors are our own. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 146

5.1. Atypical Employment in Germany: Extent and Importance

The recent thriving employment performance of the German labour market and its reasons have attracted considerable attention by policymakers and academics alike [Dust- mann et al. 2014; Carrillo-Tudela, Launov and Robin 2018]. The increase of employment rates, however, went together with a considerable rise in wage inequality during the last decades [Card, Heining and Kline 2013]. Both the growth of employment and the rise in wage inequality are linked to the growing importance of atypical employment2 for the

German labour market. On the one hand, atypical employment has been an important driver of overall employment growth in Germany. The share of atypical employment in total employment rose from 12.8 per cent in 1991 to 20.8 per cent in 2015 [Statistisches

Bundesamt 2016a]. On the other hand, atypical employment is associated with a higher risk of unemployment and signiﬁcant disadvantages in pay, and has contributed to rising wage inequality in Germany [Biewen, Fitzenberger and de Lazzer 2018; Brehmer and

Seifert 2008; Gebel 2010; Giesecke and Groß 2002; Kvasnicka and Werwatz 2002; Paul 2016;

RWI 2016].

The welfare consequences of atypical employment depend crucially on its impact both on the labour market as a whole and at the individual level. We therefore examine the following research questions: (1) How has labour market participation evolved over the life cycle in recent decades? (2) What is the role of atypical employment, i.e. to what extent do workers pursue atypical employment over the course of their working lives and how have the corresponding employment proﬁles by age changed over time? (3) Which types of employment trajectories can be identiﬁed at the individual level?

2The definition of atypical employment differs between studies and often depends on the underlying data source. Atypical employment is usually defined as employment in fixed-term, part-time, mini-job (marginal employment) or temporary employment. In this study, freelance work is also included (see Section 5.2). There are also differences between studies as to whether jobs in “large part-time” jobs, which are characterized by above-average working hours, are classified as atypical forms of employment. Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung[2008] discusses that a classification can be arbitrary to some extent and that “in reality, the dividing-line [runs] somewhere within the group of open-ended part- time contracts and [... is] probably floating and [depends] not only on the number of hours worked but also on other employment characteristics such as the level of pay”. The present study follows the procedure in Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung[2008] and considers large part-time work as atypical employment and thus differs from the procedure of the Federal Statistical Office, which does not include large part-time work. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 147

The answer to the ﬁrst question is the backbone for the following analyses. The importance of atypical employment in recent times can only be assessed if it is clear how labour force participation developed over the life cycle when atypical employment played a smaller role than it does nowadays. Hence, our main empirical approach is to provide descriptive evidence on the life-cycle proﬁle of different birth cohorts in this context (see also Section

5.2). This approach does not allow us to establish a counterfactual situation in the sense of a causal analysis and we thus cannot interpret it in the sense of a causal cohort effect. How- ever, we are able to provide an encompassing portrait for different birth cohorts, which also makes it possible to make a comparison with the current situation.

The answer to the second question illustrates the importance of atypical employment for the labour market as a whole, and which subgroups of the population (by age and gender) are affected most. Combined with the ﬁrst research question, we can relate general labour market participation to atypical employment, both over the life cycle and over time. This is particularly important for understanding the strong increase in female participation over time.

To answer the third question, we focus on individual employment trajectories and typical sequences of different employment forms (regular employment, atypical employment, unemployment, etc.) over the life cycle. Since periods of employment may have long-term effects on an individual’s future employment trajectory, individual trajectories may deviate greatly from average behaviour. Atypical employment may therefore only be concentrated amongst certain groups of workers. We analyse these heterogeneities by depicting typical employment trajectories. This also allows us to make statements about the role and motives of atypical employment at specific points in individuals’ employment histories. Atyp- ical employment may, for example, facilitate access to the labour market at the first entry into the labour market or after career breaks, be a stepping stone to regular employment or represent the beginning of a permanent period of such employment. Which mechanism applies is controversially discussed in the literature and depends on the type of atypical employment. According to the Bundesagentur fur¨ Arbeit[2013], temporary agency work facilitates the access to the labour market for unemployed individuals. However, like fixed- 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 148 term employment, it is associated with lock-in effects [Kvasnicka 2009; Brehmer and Seifert

2008]. Similar lock-in effects are observed in part-time employment and marginal employment [Brehmer and Seifert 2008], although this is largely at the request of employees in order to better combine personal and family obligations with work or to gradually retire

[RWI 2013; Wolf 2010].

Our analyses provide important implications for economic policy. For example, if involvement in atypical types of employment at early career stages is accompanied by a significantly reduced probability of ever achieving a stable full-time employment relationship, regulating atypical employment more strictly will be justiﬁed. Considering distributional aspects, if the increase in atypical employment is concentrated on relatively few people throughout their whole working lives, an economic policy response may be warranted.

This will not be the case, if the increase is due to short periods of atypical employment for relatively many people.

We use a data set that links the survey data of the National Education Panel (NEPS) with administrative data (see Section 5.2) to answer the research questions. This allows us to study employment dynamics over the life cycle of individual workers. Hence, we can illustrate time spent in atypical employment throughout labour-market careers. Further- more, we observe a wide range of birth cohorts (birth cohorts 1944 to 1986) and compare employment histories between these birth cohorts. Therefore, we provide descriptive evidence on the changing importance and role of atypical employment over time. Since we expect stark differences in employment histories of men and women, we carry out all analyses separately by gender. In order to capture longer-term developments, we mainly focus on the West German labour market.

A distinction between cohort, age and time effects is naturally difﬁcult due to their linear dependence [e.g. Fitzenberger, Schnabel and Wunderlich 2004]. For example, differences in the employment behaviour of cohorts can be an indication of structural changes, such as reforms of labour market institutions (time effects), but they can also be based on fundamental differences between cohorts, such as changes in educational attainment (cohort 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 149 effects). This article does not seek to break down developments according to their individual causes (effects), but rather to describe and compare the different trajectories.

The most important contribution of this article to the existing literature thus lies in the life course perspective taken for different birth cohorts. This approach also offers a potential explanation for the rise in intragenerational inequality in lifetime earnings which has been observed in Germany already since the early 1960s [Bonke,¨ Corneo and Luthen¨ 2015]. In a related vein,B onke,¨ Giesecke and Luthen¨ [2015] analyse the evolution of earnings volatility over the life course. They ﬁnd an increase of earnings volatility (both permanent and transitory) across different cohorts in West Germany after 1960, particularly at labour market entry. This increase in earnings volatility is furthermore not restricted to low-income individuals, but is also observable for workers with higher earnings. The increased prevalence of atypical employment across cohorts, particularly (but not only) at labour market entry, which we document in this paper is a likely explanation for their ﬁndings.

From a methodological point of view, our analysis provides insights into the use of atypical employment in different periods of the working life and how this has changed over the past decades. We therefore go beyond studies using (repeated) cross-sectional data, such as in Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung[2012].

In comparison to causal analyses, for example those available for speciﬁc atypical employment forms such as ﬁxed-term employment [Boockmann and Hagen 2008], temporary agency work [Kvasnicka 2009] and marginal employment, i.e. mini-jobs [Caliendo,

Kunn¨ and Uhlendorff 2016], this article concentrates on stages of the life cycle at which certain forms of atypical employment are dominant. In this context, the long-term analysis of employment trajectories allows us to make statements about the role these forms of employment play in the labour force participation of individual employees over the entire life cycle. To our knowledge, the only article that carries out a comparable analysis isB ohnke,¨ Zeh and Link[2015]. The authors use the SOEP waves 2002 to 2011, so that different cohorts are observed in different phases of their labour-market career. However, cohorts overlap only partially, so that comparisons between the cohorts are not possible.

This article, which is partly based on a research report for the BMAS for the preparation of 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 150 the 5th Poverty and Wealth Report of the Federal Government [RWI 2016], is thus the ﬁrst study to analyse the signiﬁcance of atypical employment over the life cycle with a reliable comparison of different birth cohorts.

5.2. Data and Deﬁnition of Atypical Employment

For our analyses, we use the survey data ”NEPS-SC6-ADIAB”. It comprises the starting cohort 6 ”Adults” of the National Education Panel (NEPS)3, which includes individuals born between 1944 and 1986. The survey contains, among other things, a range of information on the current living conditions of these individuals as well as their entire employment and education history [Blossfeld and Von Maurice 2011]. The most recent wave that we consider for the analyses was conducted in 2012/13. A total of 17,137 individuals took part in the survey up to the corresponding wave. We link the survey data to administrative information from the Institute for Employment Research (IAB) given that the respondents agreed to a record linkage and could be identified in the Federal Labour Office’s administrative data [Antoni and Eberle 2015]. Overall, administrative records are available for 74 per cent of the survey respondents. The administrative data of the Federal Labour Office come from the weakly anonymous sample of integrated employment biographies of the

IAB (version 1975 - 2012).4 These biographies contain, among other things, information on jobs subject to social security contributions (in West Germany from 1975, in East Germany from 1991) and cover a period up to December 31, 2012 at the latest. For the analyses, the survey and administrative data are combined, in order to distinguish atypical employment as comprehensively as possible and information on the type of employment is incomplete if the two data sources are analysed separately.5

3The NEPS data (doi:10.5157/NEPS:SC6:5.0.0) were collected from 2008 to 2013 as part of the Framework Programme for the Promotion of Empirical Educational Research funded by the Federal Ministry of Education and Research (BMBF). Since 2014 NEPS has been funded by the Leibniz Institute for Educational Courses e.V. (LIfBi) at the Otto-Friedrich-University Bamberg in cooperation with a Germany-wide network. 4Access to the data was given via a guest stay at the Research Data Centre of the Federal Employment Agency at the Institute for Employment Research (FDZ) and subsequently via controlled remote data process- ing at the FDZ (project number: fdz872). 5For the analysis, survey and administrative data had to be linked at the level of individual employment episodes. After detailed examination, a method was chosen which uses the survey data as the primary source 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 151

The analysis takes into account the following forms of atypical employment: fixed-term employment, part-time employment (i.e. working less than 31 hours per week)6, marginal employment7, temporary agency work8 and freelance work. Regular employment relationships are defined as dependent employment relationships with indefinite duration

(employment subject to social security contributions or in civil service) on a full-time or close to full-time basis (i.e. at least 31 hours per week) outside the temporary agency employment sector. In addition to the types of atypical employment and the regular employment relationship mentioned above, ”other” employment episodes constitute a residual group. The latter mainly include different types of self-employment (except for freelance work).

In order to show long-term changes in employment behaviour and the extent to which individuals are affected by atypical employment, four birth cohorts are distinguished: the birth cohorts 1944-53, 1954-63, 1964-73 and 1974-86. At the time of the last survey wave in

2012, individuals of the oldest birth cohort had in some cases already exceeded the age of

64 years, whereas individuals of the youngest birth cohort are no older than 38 years. The cohort comparison focuses primarily on individuals living in West Germany in January

1989, since the consequences of reunification for East Germans greatly impair the validity of cohort comparisons of employment profiles. Thus, we only conduct a comparison be- and adds administrative information on the employment type (regular/marginal) and on the economic sector of the company. Further explanations can be found in [RWI 2016] Among other things, the survey data were used as a primary source because they cover a wider range of employment episodes (e.g. self-employment, civil service and pre-1975 episodes) and a wider range of inactivity (e.g. training, other inactivity) and the chronological order of episodes was already checked for inconsistencies in the survey to minimise possible recall bias. It can be assumed that recall bias increases with the temporal distance to the interview. If this is the case and (short-term) atypical employment episodes are more likely to be affected by recall bias (and are thus under-recorded), this leads to an overestimation of the differences in cohort comparison. 6Since in the retrospective survey data of the NEPS the number of working hours is only recorded at the beginning of each job spell, changes in working hours cannot be identified. At the same time, changes in marginal employment cannot be identified with sufficient precision by linking them to administrative data. Consequently, it is assumed for the analyses that these job characteristics do not change during the duration of an employment episode. The latter leads to some underestimation of transitions from full-time to part-time and from part-time to full-time. 7The data do not allow for the identification of marginal employment relationships prior to 1999 and it is likely that in most cases these are classified as part-time before that date. 8As in the previous literature for Germany, temporary agency employment is defined by the economic sector of the enterprise [cf. Kvasnicka 2009]. Accordingly, a job is defined as temporary agency employment if the enterprise, according to the classification WZ 2008, belongs to the 3-digit (WZ group) ”782 Temporary provision of labour” (or WZ group 745 according to WZ 2003/WZ 1993 or WZ group 865 according to WZ 1973, respectively) or, if this is missing, to the 2-digit (WZ department) ”78 Placement and provision of labour”. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 152 tween East and West German individuals for the youngest cohort. East Germans refer to individuals who lived in East Germany (including Berlin) in January 1989.

5.3. Life Cycle Employment Proﬁles by Birth Cohort and Gender

In this section, we ﬁrst examine, separately by gender and the four birth cohorts, periods in which individuals are strongly and rarely affected by atypical employment over their labour-market career. Also, we investigate which type of atypical employment occurs most frequently at different points of life. The following ﬁgures show the educational and labour force participation for individuals of the four birth cohorts from age 16 up to the current age of the interviewees (or at most up to age 64). We differentiate between the states ”employment”, ”unemployment”, ”education and training” and ”(other) inactivity”. Thus, the state ”other inactivity” does not include training periods. The state of

”employment” is broken down into three types: ”regular employment”, ”atypical employment” and ”other work” (mainly self-employment). Atypical jobs are further divided into

”ﬁxed-term employment”, ”large part-time”, ”small part-time”, ”marginal employment”,

”temporary agency employment” and ”freelance work”.9 To keep the figures readable, no confidence intervals are shown. Given the sample sizes,10 however, it is evident that not every difference between cohorts is significantly different from zero. The confidence intervals generally comprise a range of no more than 3 percentage points for probabilities

9If two or more of these states overlap within a month, the person is assigned a dominant status. The following hierarchy is used as the dominance rule: a) education and training, b) employment, c) unemployment, d) other inactivity. To the extent that individual employment relationships can be attributed to several of the atypical forms of employment, e.g. if a fixed-term part-time employment relationship exists, the forms of employment are each assigned to one of the atypical forms of employment according to a dominance rule. The following hierarchy is used for this purpose: a) freelance work, b) marginal employment, c) temporary agency employment, d) part-time work, e) fixed-term employment. If several employment relationships are performed simultaneously, a dominant employment relationship is determined on the basis of the following dominance rule: a) regular employment, b) full-time or large part-time self-employment, c) fixed-term full- time employment, d) open-ended part-time employment, e) fixed-term part-time employment, f) temporary agency employment, g) freelance work or small part-time self-employment, h) minor employment. 10For the individual cohorts, the data include the following numbers of observation: a) 1944-53: 1458 men (West) and 1253 women (West), b) 1954-63: 1925 men (West) and 2073 women (West), c) 1964-73: 1618 men (West) and 1738 women (West), and d) 1974-86: 1307 men (West), 1283 women (West), 439 men (East), and 376 women (East). 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 153 close to 10 per cent (and 90 per cent by analogy) and no more than 5 percentage points for probabilities close to 50 per cent.

The inﬂuence of the educational expansion in Germany during the 1960s and 1970s becomes visible in Figure 5.1, which shows the proportion of each cohort in education over its life cycle. Younger cohorts acquire considerably more and longer (formal) education and are therefore, in contrast to older cohorts, still signiﬁcantly more likely to be in education in their mid-20s. This trend towards more education is somewhat more pronounced among women than among men. Acquiring more education implies that individuals enter the labour market later (see Figure 5.2). The proportion of employed men aged 25, for example, in the birth cohort 1974-86 is only 51 per cent compared to 66 per cent, 69 per cent and 75 per cent in the birth cohorts 1964-73, 1954-63 and 1944-53, respectively. At the age of 30 years and above, however, there are hardly any differences in the employment rates of men between the cohorts. Only the youngest cohort has a lower employment rate at the age of 30 years and older. The youngest cohort is more affected by unemployment than previous cohorts, in addition to the continuing higher rate of individuals in education.11

In the case of women, apart from the fact that younger cohorts enter the labour market later, it is particularly noticeable that (apart from the 1974-86 cohort) the employment rates of the younger cohorts are continuously above those of the older cohorts when looking at age 30 and above. For the birth cohorts 1974-86 such a development is also observable. Overall, women in younger cohorts are signiﬁcantly more likely to be employed than women in earlier cohorts. The traditional role of a housewife, in which the woman takes care exclusively of the household and children without participating on the labour market, is increasingly rare amongst young women. Nevertheless, women are much less likely than men to be employed, especially in their mid-20s and mid-30s.

11In contrast, the proportion of inactive individuals who are not in education does not differ between the youngest and older cohorts of men [RWI 2016]. Compared to other European countries the extent of youth unemployment in Germany is low overall, even among younger cohorts [Mascherini et al. 2012]. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 154

Figure 5.1.: Share of persons in education by age, year of birth and sex

100 Education 90 (Men West) 80

70 60 50 40 30 20 10 0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64

100 Share of of Share thepopulation % in 90 Education (Women West) 80

0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 Age cohort 1944-1953 cohort 1954-1963 cohort 1964-1973 cohort 1974-1986

Source: NEPS-SC6-ADIAB, own calculations.

The importance of regular and atypical employment is displayed in Figure 5.3. It shows the ratio of regular employment and atypical employment to total employment.12 Men and women in the youngest cohort (born 1974-86) are much less likely to be in regular

12See RWI[2016] for the share of other employees (i.e. in particular the self-employed) Self-employment usually increases with age and is more pronounced among men than among women. However, there are no major systematic differences between the cohorts. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 155 employment at a young age and therefore more likely to be in atypical employment than individuals in the older cohorts.

Figure 5.2.: Employment rates by age, birth cohorts and sex 100

80 Employment 70 (Men West) 60

0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64

100 Employment 90 (Women West) 80

Shareof thepopulation % in 70

0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 Age cohort 1944-1953 cohort 1954-1963 cohort 1964-1973 cohort 1974-1986

Source: NEPS-SC6-ADIAB, own calculations. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 156

For example, while 87 per cent of men in employment born in 1944-53 and 1954-63 were in regular employment at the age of 25, this applies only to 68 per cent of men in employment born in 1974-86. For young workers, regular employment is no longer the standard, and atypical employment is no longer the exception. However, not only the youngest cohort of West German men is less likely to be in regular employment and correspondingly more likely to be in atypical employment than earlier cohorts. The cohort of 1964-73 already shows an average five to six percentage points lower rate of regular employment and three percentage points higher rate of atypical employment than the 1944-53 and 1954-63 cohorts. Atypical employment for men is no longer concentrated exclusively on the first few years after entering the labour market, but is also increasing for middle ages. However, the vast majority of middle-aged men continue to work in regular employment and not in atypical jobs. Figure 5.3 also shows for men born in 1944-53 that the proportion of the atypically employed persons increases significantly towards the end of their working lives.

For women, the proportion of regular employees also fell across the cohorts, while the proportion of atypically employed persons rose (Figure 5.3). The decline in the ratio of regular employment to total employment across cohorts appears to be greater for women than for men. However, it should be noted that the employment rate of women has increased signiﬁcantly in recent years (Figure 5.2). It can therefore be assumed that at least part of this development can be attributed to the increased re-entry of mothers into the labour market after childbirth, which often occurs part-time, while mothers of previous cohorts have stayed away from the labour market for longer periods or even permanently. Find- ings from Figure 5.A.1 in the Appendix are consistent with this interpretation. It depicts employment rates relative to the total population instead of only employees. In relation to the total number of women, the drop in the share of regular employment among the middle-aged is signiﬁcantly lower, and in some cases even an increase across cohorts can be observed. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 157 64 64 62 62 60 60 58 58 56 56 54 54 52 52 50 50 48 48 46 46 44 44 42 42 40 40 38 38 cohort 1974-1986 (Women West) (Women 36 36 Regular Employment Regular (Women West) (Women 34 34 Atypical Employment Atypical 32 32 30 30 28 28 26 26 24 24 22 22 20 20 cohort 1964-1973 18 18 16 16 64 64 62 62 60 60 58 58 56 56 54 54 52 cohort 1954-1963 52 50 50 48 48 46 46 44 44 42 42 40 40 38 38 (Men West)(Men 36 36 (Men West) (Men cohort 1944-1953 Regular Employment Regular 34 34 Atypical Employment Atypical 32 32 30 30 28 28 26 26 24 24 22 22 20 20 Due to a small numberSource: of NEPS-SC6-ADIAB, cases, own some calculations. details are anonymised. The data series are interrupted at these points. 18 18 16 16 0 0

90 80 70 60 50 40 30 20 10 90 80 70 60 50 40 30 20 10

100 100

in % in workforce the of Share Figure 5.3. : Share of regular and atypical employees as a proportion of all employed persons by age, birth cohorts and sex 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 158

At the age of around 30, however, it is the youngest cohort that shows the lowest proportion of regular employees, as is the case for men. The share of atypically employed individuals in relation to the total population has increased signiﬁcantly across all cohorts.

To draw a more complete picture of atypical employment, Figures 5.4 and 5.5 break down the rate of atypical employment according to its speciﬁc type. The reference for the calculation of the percentages is again the total number of employees.13 For both men and women,

fixed-term jobs are mainly performed at a young age and the proportion of fixed-term employees decreases continuously with age. Freelance work is also carried out mainly at the beginning of a career, but (by men) also particularly in the last few years before retirement. Furthermore, there are no striking differences across age in temporary agency work, with the share of the total workforce in temporary agency work being rather small in the data. Significant differences between the sexes can be observed in the proportion of part- time or marginal employees. For men, the proportion of part-time or marginal employees increases only marginally with age, whereas for women, the proportion of part-time or marginal employees increases continuously from the age of 20 onwards.

Apart from the overall increase in atypical employment among younger cohorts, no changes in the type of atypical employment are apparent across cohorts. For example, all forms of atypical employment are more common among men born between 1974 and 1986 than for the cohort born between 1964 and 1973. However, older cohorts do not record marginal employment until later in life, but younger cohorts do so at a younger age. This is partly a statistical artefact, as marginal employment can only be identiﬁed in the data from 1999 onwards and probably was recorded as small part-time jobs in the years before 1999.

13For specific ages some specific forms of atypical employment cannot be reported, because the number of cases is too small. In these cases, the corresponding types of atypical employment are summarised and the difference to the total rate of atypical employment is shown (white area at the bottom of the figure). Due to too many such anonymizations at young ages, Figures 5.4 and 5.5 are restricted to age 19 onwards. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 159 63 63 61 61 59 59 57 57 55 55 53 53 51 51 49 49 fixed-term employment 47 47 45 45 1953 1963 - - 43 43 Age 41 41 39 39 (Men West) (Men (Men West) (Men Cohort 1944 Cohort 37 37 Cohort 1954 Cohort large part-time 35 35 33 33 31 31 29 29 27 27 25 25 small part-time 23 23 21 21 19 19 63 63 61 61 59 59 57 57 55 55 marginal employment 53 53 51 51 49 49 47 47 45 45 1986 1986 1973 - - 43 43 Age 41 41 39 39 (Men West)(Men (Men West) (Men temporary agencyworktemporary 37 37 Cohort 1974 Cohort Cohort 1964Cohort 35 35 Figure 5.4. : Type of atypical employment by age and cohort (men) 33 33 31 31 29 29 27 27 freelance work 25 25 23 23 Due to a small numberSource: of NEPS-SC6-ADIAB, cases, own some calculations. details are anonymised. The data series are interrupted at these points. 21 21 19 19

0 0

70 60 50 40 30 20 10 70 60 50 40 30 20 10 Share of the workforce in % in workforce the of Share 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 160 63 63 61 61 59 59 57 57 55 55 53 53 51 51 49 49 47 47 45 45 1953 fixed-term employment 1963 - - 43 43 Age 41 41 39 39 (Women (Women West) (Women (Women West) 37 37 Cohort 1944 Cohort 1954 35 35 33 33 large part-time 31 31 29 29 27 27 25 25 23 23 21 21 small part-time 19 19 63 63 61 61 59 59 57 57 55 55 53 53 marginal employment 51 51 49 49 47 47 1986 45 45 - 1973 - 43 43 Age 41 41 39 39 (Women (Women West) (Women (Women West) Cohort 1974 37 37 Cohort 1964 temporary agencywork 35 35 Figure 5.5. : Type of atypical employment by age and cohort (women) 33 33 31 31 29 29 27 27 25 25 freelance work 23 23 Due to a small numberSource: of NEPS-SC6-ADIAB, cases, own some calculations. details are anonymised. The data series are interrupted at these points. 21 21 19 19

0 0

70 60 50 40 30 20 10 70 60 50 40 30 20 10 Share of the workforce in % in workforce the of Share 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 161

The comparison of the distribution of the types of atypical employment between men and women clearly shows that the higher prevalence of atypical employment for women is rooted in significantly higher proportion of part-time and marginal employees. For women, the proportion of freelancers is also somewhat higher than for men, while the proportion of those employed in temporary agency work is lower on average. Further- more, the share of temporary jobs for women seems to be somewhat lower than for men, but this is (at least partly) due to the fact that fixed-term part-time jobs are classified in the corresponding part-time categories and not in the temporary category (see Footnote 8).

For the youngest cohort, i.e. the birth cohorts 1974-86, the employment trajectories have largely taken place in a labour market of a market-based economy for East Germans, too, so that meaningful comparisons can be made with the corresponding Western cohort. Nev- ertheless, it should be noted that the situation in the labour market, especially in the ﬁrst decade after the reuniﬁcation, differed markedly between East and West Germany. This might have idiosyncratically shaped the employment trajectories of the cohorts. Also, the experience of their parents and their environment may lead to disparities.

For men, the main differences lie in the duration of education and unemployment. The average proportion of men aged 20 to 30 in unemployment is almost 11 per cent in East

Germany, which is more than one third higher than in the West (less than 7 per cent). On average, East German men are also less frequently in education or training, or for a shorter time period. By contrast, the differences in the proportion of employed individuals and in the proportion of regular and atypical employees among those employed are very small.

Overall, women in East and West Germany display very different employment histories.

On the one hand, the employment rate in the East in the age range from the mid-20s to the mid-30s lies consistently above that in the West (on average by 5.4 percentage points).

For example, 72 per cent of women in the East and only 67 per cent of women in the West are employed in their early to mid-30s (see Figure 5.A.2 in the Appendix). West German women are much more likely to be inactive. On the other hand, employed women in East

Germany in their mid-20s and early 30s are initially somewhat less likely to be in regular employment than employed women in West Germany. From the age of 30 onwards, 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 162 however, they are more often regularly employed than employed women in West Ger- many (see Figure 5.A.3 in the Appendix).14 At the same time, employed women in West

Germany aged 30 and older are much more often in atypical employment and, in particular, more often in marginal employment or small part-time jobs (i.e. less than 21 hours per week) than women in East Germany, who are more likely to work in large part-time jobs (see Figure 5.A.4 in Appendix). The differences in the share of ﬁxed-term employees between East and West German women are rather small.

5.4. Typical Employment Trajectories: Results of a Sequence

Analysis

The previous analysis of age employment profiles illustrated the average behaviour of the cohorts. Education and employment trajectories can deviate greatly from this average and may be very heterogeneous in terms of their sequence. This is of particular interest for labour market policy: a situation in which many people go through atypical employment at the beginning of their career, and then switch to stable full-time employment, is to be assessed differently from a situation in which atypical employment is concentrated on a certain group of individuals and has a negative influence on their entire employment histories. The aim of the following analyses is therefore to descriptively show potential heterogeneities. For this purpose, we analyse the entire employment trajectory of individuals aged 16 and above. Hence, the starting ages and duration of an individual’s periods of education, atypical employment, regular employment, other gainful employment, unemployment or (other) inactivity are taken into account. From the large number of possible sequences of such episodes, similar patterns in employment trajectories are identified using the method of sequence analysis. From this, we derive types with similar employment trajectories, and identify differences between birth cohorts and men and women.

14One reason for the higher proportion of atypically employed women among East German women in their mid-20s to early-30s could be that they are on average 1.5 to 2 years younger at the birth of their ﬁrst child than West German women [cf. Statistisches Bundesamt 2016b]. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 163

In the sequence analysis, we compare individual employment sequences with each other using the optimal matching method. First of all, we assign one of the six employment statuses listed above to each year of age at the individual level, starting from the age of 16.15

Subsequently, the distances between all individual employment trajectories are determined

(so-called Levenshtein distance). The distance between two sequences will be higher if the two individuals experience different types of employment and if the timing of similar employment spells differs over the life course of two individuals. Graphically, this can be thought of as two sequences being placed next to each other and elements of the sequences being exchanged or deleted to make the sequences identical. The fewer adjustments are necessary, the more similar are the sequences and the smaller is their distance. Since there is a multitude of comparison possibilities, we apply the Needleman-Wunsch algorithm to arrange the sequences in such a way that the distances between all sequences are mini- mized [cf. Brzinsky-Fay, Kohler and Luniak 2006]. The resulting distances are divided into groups via a cluster analysis. Following WZB[2009] and RWI[2014 a], we use the Ward algorithm to group similar sequences. The Ward algorithm is a hierarchical-agglomerative cluster method that generates homogeneous clusters [cf.B ohnke,¨ Zeh and Link 2015].

To compare the cohorts as accurately as possible, we carry out sequence analyses separately for different age ranges. We distinguish a total of two age ranges. The age range from 16 to

30 can be examined for all four West German birth cohorts and the youngest East German cohort, and mainly represents the period of education and training and the ﬁrst entry into employment. The age range of 16 to 40 years can be considered for the three West German cohorts 1944-53, 1954-63 and 1964-73. In addition to the ﬁrst entry into employment, it also represents the early part of the main employment phase. The youngest birth cohort (1974-

86) cannot be taken into account here, as the majority of individuals in that cohort are still younger than 30 years at the time of the last survey or are only slightly older. For reasons

15In life years in which a person has several (employment) states, a dominant state is deﬁned. First, the criterion of maximum state duration in one year of life applies. If several statuses per year have the same duration, a rule of dominance similar to that in Section 5.3 is applied (see Footnote 8): education dominates all other statuses, followed by regular employment, atypical employment, other employment, unemployment and ﬁnally (other) inactivity. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 164 of space, no further age ranges are presented, as they show only very few developments that are not already observed in the younger age ranges.

We determine the typical career paths jointly for men and women. For the subsequent analysis of the frequencies of the employment trajectory types and their change across the cohorts, we portray results separately by gender. The number of different types with similar sequences is determined such that the types created are as different as possible and the results can be interpreted as meaningful as possible. First, we present the results for the age range 16 to 30.

5.4.1. Education and First Job Entry

For this stage of life up to age 30, we distinguish eight types of employment sequences.

Figure 5.6 shows the average number of years spent per employment state for each type.

Since the order in which the average durations are presented does not have to correspond to the sequences actually realized, Figure 5.7 shows for each type the most prevalent employment state at each year of life. This is determined using the modal value. This representation also makes it possible to illustrate differences in the sequences. Finally, separately by gender, Figure 5.8 shows the proportion of the different types of employment trajectories in each cohort. As already noted in Section 5.3, we point out that not every difference in the proportions between cohorts is signiﬁcantly different from zero. For the available sample sizes, the conﬁdence intervals typically include a range of no more than 1.5 percentage points for shares close to 2 per cent, no more than 3 percentage points for shares close to 10 per cent, and no more than 5 percentage points for shares close to 50 per cent.

Three types can be identiﬁed that enter regular employment relatively quickly after the

ﬁrst entry into the labour market, and whose entry into the labour market can therefore be described as unproblematic (types A1, A2 and A3). The main difference between the types is in the duration of the preceding education and training. The shift in the importance of the respective types across the cohorts that is due to the educational expansion becomes visible in this picture, particularly for men. For example, while 42 per cent of the West 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 165

German men in the 1944-53 birth cohort still had a relatively short period in education and a rapid transition to regular employment (type A1), only 8 per cent of the West German men in the 1974-86 cohort still show this pattern. Type A2 with a somewhat longer period in education shows a small decline across the male birth cohorts from 18 per cent to 14 per cent, too. On the other hand, the share of the type with a long (presumably mostly academic) education and a rapid transition to regular employment (type A3) increased from 14 per cent to 21 per cent in the corresponding cohorts. In addition, one type can be identiﬁed which is almost exclusively in the training system until the age of 30 (type A4) and whose share in the oldest cohort was 12 per cent of men in West Germany, compared to 36 per cent in the youngest cohort. Among West German women, the share of type A4 has risen from 3 per cent in the oldest to 34 per cent in the youngest cohort, too. Although starting from a lower initial value, but to a similar extent as for men, the proportion of West

German women of type A1 fell from 26 per cent to 6 per cent.

Figure 5.6.: Average duration of employment state by employment type (age range 16-30)

Type A1 3 0 12

Type A2 5 1 8 1

Type A3 9 0 5 1

Type A4 11 1 1 1

Type A5 4 6 2 1 0

Type A6 3 1 4 6 0

Type A7 4 1 2 1 5

Type A8 5 0 2 7 1

0 2 4 6 8 10 12 14 16 Duration in Years (After Age 16)

education atypical employment regular employment other work other inactivity unemployment

Source: NEPS-SC6-ADIAB, own calculations. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 166

Figure 5.7.: Dominant employment state by age and type of employment (age range 16-30)

Type A1

Type A2

Type A3

Type A4

Type A5

Type A6

Type A7

Type A8

16 18 20 22 24 26 28 30 Age

education atypical employment regular employment other work other inactivity unemployment

Source: NEPS-SC6-ADIAB, own calculations.

Figure 5.8.: Share of employment types by cohort and sex (age range 16-30)

12 Men '44-'53 West 42 18 14 4 2 2 6 14 Men '54-'63 West 33 18 21 4 12 7 16 Men '64-'73 West 26 19 24 5 12 7 36 Men '74-'86 West 8 14 21 9 1 4 6 28 Men '74-'86 East 8 19 20 13 2 7 3

3 Women '44-'53 West 26 15 3 8 36 4 5 6 Women '54-'63 West 27 18 7 9 23 5 5 11 Women '64-'73 West 19 23 15 9 16 3 3 34 Women '74-'86 West 6 13 16 14 11 5 2 34 Women '74-'86 East 5 16 14 16 7 4 3

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

type A1 type A2 type A3 type A4 type A5 type A6 type A7 type A8

Source: NEPS-SC6-ADIAB, own calculations.

Individuals with longer periods of atypical employment in the labour market entry phase are depicted in type A5. Their share is signiﬁcantly higher among women than among 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 167 men and has increased in both groups over time. In the youngest cohort of West German men, 9 per cent are of this type, in the corresponding cohort of West German women 14 per cent show this employment pattern. For the cohort born three decades earlier, the share of this type of employment trajectory with longer periods in atypical employment was only about half as high.

Type A6 has an employment pattern that already includes long periods of inactivity up to age 30. Between education and inactivity there are only short periods of employment, but these are usually regular instead of atypical. Type A6 occurs almost exclusively among women and represents the traditional housewife. In a comparison of cohorts, their proportion fell from 36 per cent among women born in 1944-53 to 11 per cent among women born in 1974-86. On the one hand, this decline can be attributed to the lower proportion of women who withdraw completely from the labour market over a longer period after the birth of a child, but on the other hand, in view of the comparatively short age range up to

30 years, to a shift in age at ﬁrst birth [cf.P otzsch¨ and Emmerling 2008]. Many women, especially of the youngest cohort, are already older than 30 years at the time of their ﬁrst birth.

Another type of employment history (A7) mainly comprises individuals with very difﬁcult entry into the labour market. Although representatives of the group have several years of employment, unemployment is the primary state after the relatively short period in education. The share of this type is rather low overall. Among West German men, the proportion has increased across the cohorts from about 2 per cent in the oldest cohort to about 4 per cent in the youngest cohort. Among West German women, the corresponding proportion is somewhat higher, but it has increased less strongly across the cohorts.

Finally, type A8 represents the labour market entry of self-employed individuals. For men, the proportion is about 6 to 7 per cent and there is no clear trend across the cohorts; for women, the proportion is somewhat lower and falls across the cohorts from about 5 to only 2 per cent in the youngest cohort.

A comparison of East and West for the youngest cohort shows that self-employed careers

(type A8) are pursued only half as often by East German men up to 30 years of age and that 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 168 problematic entry into the labour market is more frequent than in the West. The proportion of type A7 with long unemployment periods is almost twice as high for East German men as for West German men, and type A5 with longer periods of atypical employment is also much more common for men in the East than in the West. Longer periods in atypical employment (type A5) are also slightly more common among East German women than among West German women. On the other hand, East German women are much less likely to show an employment pattern with a long period of inactivity (type A6). Extremely long periods in education without labour force participation before the age of 30 (type A4) are also much rarer among East German men than among West German men.

In summary, for the first period of entry into employment, changes in educational attainment and the general increase in the proportion of working women have led to significant shifts in employment patterns. At the same time, especially for type A5, it becomes apparent that it is increasingly difficult for men and women in the younger cohorts to enter the labour market and that they work in atypical employment for longer periods of time. If instead of using 8 types of trajectories we differentiate a larger number of types, it is not only for type A5 that there is a marked increase in the proportion of individuals who are severely affected by atypical employment. A subgroup of type A4, i.e. individuals with a long phase in education, also shows a strong involvement in atypical employment when entering the labour market. The importance of this subgroup also increases significantly across the cohorts, for example from 4 per cent in the oldest to 9 per cent in the youngest cohort for men and from 1 to 13 per cent for women, respectively.

5.4.2. Early Main Employment Period

For the age range from 16 to 40, which covers not only the ﬁrst labour market entry but also the early main employment period, we distinguish between eight types of employment trajectories. The average number of years per employment state is shown in Figure 5.9, the most prevalent employment state at each year of age is shown in Figure 5.10, and the 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 169 proportion of the different types of employment trajectories in each cohort are shown in

Figure 5.11.

In the age range from 16 to 40, two types of employment can be observed which after leaving education are mainly active in regular employment and which differ primarily in the duration of education and training (types B1 and B2). In the case of men, the educational expansion and the associated shift from type B1 to type B2 is visible. In addition, the sum of the shares of the two types with long periods in regular employment decreases slightly over the birth cohorts. In the oldest cohort these two types of trajectories account for 73 per cent, in the cohort 1954-63 about 74 per cent and in the cohort 1964-73 about 69 per cent. In the case of women, however, the proportion of the two types with long periods in regular employment increases from a total of about 28 per cent in the 1944-53 cohorts, over 32 per cent in the 1954-63 cohort, to about 35 per cent in the 1964-73 cohort.

Substantially longer periods in education with quite heterogeneous entry into employment can be observed for type B3. Periods in atypical and regular employment as well as in self- employment are distinctive of this highly educated group. Compared with employment types B1 and B2, however, entry into the labour market is clearly less smooth. The share of type B3 is signiﬁcantly higher for both men and women for younger cohorts than for older cohorts.

Employment trajectory type B4 is characterised by long periods spent in regular employment and shorter periods in atypical employment, self-employment or inactivity. Figure

5.10 suggests that atypical employment follows rather than precedes regular employment, suggesting late parenthood with rapid re-entry into the labour market (initially via atypical or self-employment). However, the proportion of type B4 among men, especially in the oldest cohort, is no smaller than among women (6 per cent each). Comparing the cohorts, the share of this type of employment is constant among men, but it has almost doubled among women until cohort 1964-73 (to 11 per cent).

Type B5 enters atypical employment much earlier and remains there longer, and is very prevalent among women in particular. Representatives of this employment pattern up to the age of 40 spend an average of 13 years in atypical employment. These are mainly 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 170 mothers who take only short breaks from work after childbirth and then return quickly to the labour market, mainly part-time or in mini-jobs. In the oldest cohort of West German women (born 1944-53), this type of employment trajectory accounts for 19 per cent; in the 1964-73 cohort, a signiﬁcant share of 26 per cent of women return to work quickly.

In contrast, the proportion of women who are not employed in the long term (type B6) decreases across the cohorts from 33 per cent in the 1944-53 birth cohort to 14 per cent in the 1964-73 birth cohort.16

Figure 5.9.: Average duration of employment state by type of employment (age range 16- 40)

Type B1 4 0 20 1

Type B2 10 1 12 1

Type B3 14 2 2 4 1 1

Type B4 5 4 12 2 2

Type B5 5 13 4 2 0

Type B6 3 2 6 13 0

Type B7 4 3 5 1 7

Type B8 4 1 3 14 2

0 2 4 6 8 10 12 14 16 18 20 22 24 26 Duration in Years (After Age 16)

education atypical employment regular employment other work other inactivity unemployment

Source: NEPS-SC6-ADIAB, own calculations.

16Looking at the age range from 16 to 50, it can be seen that the group of women with a long period of inactivity after the birth of a child (type B6) appears to be divided into two subtypes. One of these types remains largely absent from the labour market even in the ﬁfth decade of life. The other type returns to the labour market primarily via atypical employment during the ﬁfth decade of life after a long period of inactivity which on average lasts 12 years. Here, too, a change towards a stronger labour market orientation of women can be observed for subsequent birth cohorts; in the oldest cohort, 16 per cent leave the labour market permanently, while in the subsequent cohort it is only half that share. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 171

Figure 5.10.: Dominant employment state by age and employment type (age range 16-40)

Type B1

Type B2

Type B3

Type B4

Type B5

Type B6

Type B7

Type B8

16 18 20 22 24 26 28 30 32 34 36 38 40 Age education atypical employment regular employment other work other inactivity unemployment

Source: NEPS-SC6-ADIAB, own calculations.

Figure 5.11.: Share of employment types by cohort and sex (age range 16-40)

Source: NEPS-SC6-ADIAB, own calculations.

For the proportion of self-employed (type B8) and the proportion of individuals affected by permanent unemployment (type B7) only minor changes can be observed between the cohorts. Type B7 has relatively short periods of regular employment (most likely at the 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 172 beginning of employment) and longer periods of unemployment, sometimes alternating with shorter periods of atypical employment. Moreover, for some life stages, no concrete information is available on the employment state, which is why the average durations shown in Figure 5.9 are lower in total than for the other types of employment trajectories.

5.5. Summary and Conclusion

The recent growth in employment in Germany, but also the increase in atypical employment and income inequality has attracted considerable attention recently [e.g. Card, Hein- ing and Kline 2013; Dustmann et al. 2014]. Taking a life-cycle perspective, we therefore examine three closely related issues which are highly relevant in this context: the development of labour force participation over the life cycle, the role of atypical employment in this context and, ﬁnally, the identiﬁcation of typical employment trajectories at the individual level. All three topics deal with differences between birth cohorts, so that developments over time become visible, which are informative about welfare consequences at the individual level and about related economic policies. Our results can be summarized as follows.

First, employment patterns across the cohorts born since WWII have changed signiﬁcantly.

To start with, younger cohorts are characterized by a longer period in education and a corresponding later entry into the labour market, which can be attributed to the educational expansion. There has also been a fundamental change in the employment behaviour of women. After the birth of a child, women are less likely to leave the labour market for extended periods of time and are more likely to return to work relatively quickly.

Second, atypical employment plays a crucial role in this context. In general, the importance of atypical employment is increasing across cohorts. For women, the growing prevalence of atypical employment is closely linked to the development of labour market participation, as much of the increase in employment takes place in the form of atypical employment, mainly part-time and mini-jobs. The proportion of women in permanent and pre- 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 173 dominantly regular employment in the age range 25 to 40 is rising in part across cohorts, but appears to have changed little in the age group 40 and above.

Third, looking at individual employment trajectories over the life cycle, various distinct patterns can be identified. A comparison of the different cohorts for women shows a relatively strong departure from the model of the traditional housewife and thus a stronger focus on labour market participation, albeit with a reduced number of hours. As the sequence analysis of the period in education and training and the first entry into employment shows, entry into the labour market is increasingly more difficult for men from younger than from older cohorts and is characterised by longer periods of atypical employment.

This phenomenon can also be observed among formally well-trained persons.

Overall, atypical employment has increased in Germany, especially in the form of two patterns across cohorts; ﬁrstly as an opportunity for women to enter the labour market despite childcare obligations, and secondly in the form of increasingly insecure entry into employment. The latter pattern could be a potential explanation for the rise in both the intragenerational earnings inequality and the rise in earnings volatility over the life cycle which have been documented for Germany since the 1960s [Bonke,¨ Corneo and Luthen¨

2015;B onke,¨ Giesecke and Luthen¨ 2015]. It should however be pointed out that the present paper, due to its primarily descriptive approach, cannot clarify whether insecure entry into employment serves as a stepping stone or rather leads to a dead end. The causal literature partly questions the stepping-stone function or attests it only to a small extent [Caliendo,

Kunn¨ and Uhlendorff 2016; Kvasnicka 2009].

Given the descriptive nature of this analysis, economic policy conclusions should be drawn with caution. Nevertheless, some relevant evidence emerges in this context. First, the close link between the increase in women’s employment and the importance of atypical employment - especially as a way of re-entering the labour market after a career break - shows that increased regulation of atypical employment could have negative side effects on women’s participation in the labour market. In a labour market characterised by a labour surplus, for example, this would be the case when the attractiveness of part-time work for employers decreases. On the other hand, the stricter regulation could also lead to an increase in 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 174 labour supply, which could mean higher labour market participation in a labour market characterised by skills shortages. Secondly, the importance of atypical employment for women indicates that the reconciliation of family and working life remains problematic, at least if a quick return to full-time employment is considered desirable. Further efforts should therefore be made in this area, in particular with regard to the availability of childcare [e.g. Schober and Spieß 2014], so that families seeking full-time employment for both partners can achieve it. Thirdly, the development of atypical employment and its effects should be studied further in the light of previous ﬁndings on negative wage effects, employment probabilities, and uncertainty about a possible stepping-stone function, which are also relevant for men, for whom the corresponding employment histories now occur frequently, too.

In addition to the latter point, some open questions remain for future analysis. The cohort being very strongly affected by atypical employment when it ﬁrst enters the labour market has not yet entered the main employment period in a sufﬁcient number. The future must show whether a possible transition to regular employment only takes place later than was the case for earlier cohorts or whether careers in persistently insecure or otherwise (unintentionally) atypical employment relationships arise to a relevant extent. It also remains to be seen when and how the decline in atypical employment reported in

Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung[2015] in the period from 2010 to 2014 will be reﬂected in long-term employment patterns. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 175

5.A. Figures

Figure 5.A.1.: Share of regular and atypical employees as a proportion of the total population by age (women) 100 Regular Employment 90 in the Population (Women West) 80

in % in 0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64

100 Atypical Employment 90 in the Population (Women West) 80

70 Shareof thepopulation 60 50 40 30 20 10 0 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 Age

cohort 1944-1953 cohort 1954-1963 cohort 1964-1973 cohort 1974-1986

Figure 5.A.2.: Employment rates by age, sex and region (age group 1974-86) 100 Employment 90

30 Shareof thepopulation % in 20

0 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 Age

Men West Men East Women West Women East

Source: NEPS-SC6-ADIAB, own calculations. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 177

Figure 5.A.3.: Share of regular and atypical employees as a proportion of total employment by age, sex and region (born 1974-86) 100 Regular Employment 90

in% 0 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

100 Atypical Employment 90

80 Shareof theworkforce 70

0 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

Men West Men East Women West Women East

Due to the small number of observations, some details are anonymised. The data series are interrupted at the respective points. Source: NEPS-SC6-ADIAB, own calculations. 5. LABOUR MARKET PARTICIPATION AND ATYPICAL EMPLOYMENT 178 35 35 34 34 33 33 32 32 31 31 freelance work 30 30 29 29 Ost 28 28 East 27 27 Age 26 26 Frauen Frauen Women Women West Women 25 25 24 24 23 23 temporary agencywork 22 22 21 21 20 20 19 19

0 0

70 60 50 40 30 20 10 70 60 50 40 30 20 10

in % in workforce the of Share marginal employment Ost 35 35 Frauen Frauen 34 34 33 33 small part-time 32 32 31 31 30 30 29 29 28 28 large part-time East 27 27 Age 26 26 Men Men West Men 25 25 24 24 23 23 22 22 21 21 20 20 fixed-term employment 19 19

0 0

70 60 50 40 30 20 10 70 60 50 40 30 20 10 Share of the workforce in % in workforce the of Share Figure 5.A.4. : Type of atypical employment by age, region and sex (age group 1974-86) Due to the smallpoints. number of observations, someSource: details NEPS-SC6-ADIAB, own are calculations. anonymised. The data series are interrupted at the respective 6. Worker Motives for Multiple Job Holding in Germany*

Abstract: I investigate the recent strong increase of multiple job holding in Germany. From

2003 to 2014 the number of registered workers with more than one job rose from 1.0 to 2.2 million. Among individuals with multiple jobs a primary job subject to and a secondary job exempted from social security contributions and from taxes, a so-called minijob, is predominant. Examining employment biographies from administrative records, I identify worker motives for taking up an additional job by comparing worker and job characteristics as well as subsequent career paths of multiple and single job holders. I ﬁnd that immediate monetary intentions relate to multiple job holding. Multiple job holders exhibit an increased job, sector and task mobility as well as are more likely to receive a pay raise relative to single job holders. This suggests that an investment motive plays an important role for multiple job holding, too.

*This chapter is single-authored. It is currently a mimeo. I am grateful to Ronald Bachmann, Thomas K. Bauer, Matthias Giesecke, Corinna Hentschker, Sandra Schaffner and Christina Vonnahme as well as to participants at seminars at RWI for helpful comments. All remaining errors are my own. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 180

6.1. Introduction

From 2003 to 2014 the number of employed workers in Germany having more than one job has more than doubled. In 2014, every 15th dependent employed worker was a multiple job holder [Klinger and Weber 2017]. Thus, multiple job holding2 has become an important labour market state. During the same period, the German labour market underwent a series of institutional changes, the Hartz reforms, which increased its ﬂexibility and led to a strong job expansion with a employment growth rate of about 10 percent [Schneider,

Rinne et al. 2017].

In this paper, I analyse multiple job holding focusing on the worker side. I exploit the richness of administrative employment biography records to differentiate types of multiple job holding, to investigate a wide range of socio-economic information of multiple job holders and to study their labour market transitions. Thus, I identify worker motives for multiple job holding by examining its determinants as well as its consequences for career progression. This provides answers for the policy relevant questions of who decides to hold more than one job, why and with which implications.

The economic literature has elaborated four main reasons why individuals work in more than one job [among others Conway and Kimmel 1998; Shishko and Rostker 1976]. The hours constraint and hedging motives are related to immediate monetary intentions. In particular, the hours constraint motive reﬂects that workers take up an additional job, because they are willing to work more hours, but cannot do so in their primary job. The hedging motive arises from precautionary behaviour. Workers hold multiple jobs to diver- sify the risk of income loss. In contrast, the utility and investment motives root in expected pay-offs from job heterogeneity. While the utility motive suggests satisfaction gains from job heterogeneity as main reason, the investment motive highlights the potential for a ben- eﬁcial career progression.

Early empirical studies primarily focus on immediate monetary intentions. In general, they provide evidence for the hours constraint motive [e.g Krishnan 1990; Kimmel and Powell

2The literature uses the terms multiple job holding and moonlighting interchangeably. I deviate from this convention by not using the expression moonlighting to avoid framing this employment state as illegal. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 181

1999] and against the hedging motive [e.g Bell, Hart and Wright 1997; Wu, Baimbridge and

Zhu 2009]. More recent research conﬁrms both heterogeneous job intentions, the utility

[Dickey, Watson and Zangelidis 2011] as well as the investment motive [Panos, Pouliakas and Zangelidis 2014]. For Germany, Heineck[2009] analyses worker motives for multiple holding with survey data from the Socio-economic Panel (SOEP). He shows that immediate monetary intentions prevail stronger than heterogeneous job intentions. Klinger and We- ber[2017] support this view using a cross sectional sample of the Integrated Employment

Biographies (IEB) for the year 2014.

The investigation of worker motives for multiple job holding poses two main challenges.

First, an individual’s decision to work in an additional job is selective causing endogeneity in analyses. Therefore, such studies are generally of descriptive nature and compare the populations of multiple to single job holders. In accordance with the literature I follow this approach. Second, longitudinal data is necessary to comprehensively evaluate worker motives. In particular, examining the investment motive requires reliable information on labour market transitions. However, to my knowledge all previous studies are conducted with survey data which are prone to measurement error.3 The latter stems from recall bias, caused by respondents’ wrong recollection of events, as well as from time aggregation bias, due to the gaps between interviews. Thus, these measures for labour market mobility may be biased. Furthermore, survey data generally provide small sample sizes and include only limited information about different types of multiple job holding. On the basis of administrative records I can tackle the second challenge.

My contribution to the literature is twofold. I am the ﬁrst to study multiple job holding using longitudinal administrative records. Therefore, I can study worker motives more rigorous. Moreover, I complement research on multiple job holding by focusing on the recent increase in this employment type in Germany, which is exceptional and has not been studied in detail yet.

3Even though Klinger and Weber[2017] use administrative data, they perform a cross sectional study for the year 2014 which precludes an investigation of labour market transitions. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 182

I ﬁnd that multiple job holding in Germany actually corresponds to dual job holding. In particular, after the Hartz II reform in 2003 it is of the type primary job subject to social security contributions with a secondary subsidized minijob. A minijob is a marginal part- time employment for which the worker does not have to pay social security contributions and taxes. Furthermore, the results suggest that individuals hold multiple jobs due to immediate monetary intentions. Multiple job holding is more prevalent among primary part-time employed, females, low educated and routine workers indicating the need to increase earnings with an additional subsidized job. At the same time, multiple job holding relates to career progression. Multiple job holders are more likely to switch jobs, sectors and tasks as well as to receive a wage increase relative to single job holders, which supports an investment motive. Moreover, there exists strong unobserved heterogeneities in both worker mobility as well as its determinants between multiple and single job holders.

The paper proceeds as follows. The next section introduces the data. Section 6.3 documents the increase in multiple job holding in Germany and reviews the implemented labour market reforms during this period. Section 6.4 identiﬁes worker motives for multiple job holding by examining its determinants as well as its consequences for career progression. The

ﬁnal section concludes.

6.2. Data

I use the weakly anonymous Sample of Integrated Labour Market Biographies (SIAB) to study multiple job holding.4 The data are a 2 percent representative sample of the German population which is either in dependent employment or registered unemployment. For each individual the administrative data set includes the universe of employment and unemployment spells at daily precision from 1975 to 2014. The spell notiﬁcations determine workers’ social security contributions as well as unemployment beneﬁts for individuals.

Hence, the data are of high accuracy. Furthermore, the SIAB provide a range of socio- demographic characteristics, such as sex, age, and education, as well as important ﬁrm

4Data access was provided via on-site use at the Research Data Centre (FDZ) of the German Federal Em- ployment Agency (BA) at the Institute for Employment Research (IAB) and subsequently remote data access. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 183 information, such as size, location and industrial sector. Wages are recorded as earnings per day.

From 1999 onwards the data distinguish full-, from part-time and marginal work. Employ- ment of less than 19 hours a week is considered as part-time work. Marginal employment is a type of part-time work, which is additionally capped at a monthly income level. In

Germany, marginal employment is subsidized: Workers do not have to pay social security contributions and in most cases minijobs are exempted from income taxes.5

Multiple job holding occurs if an individual has two or more employment notiﬁcations in the data at the same time. I observe in total six forms of dual job holding. That refers to every possible combination of two spells in full-, part-time and marginal employment.6

The number of parallel spells is unlimited in the SIAB. Therefore, I can examine every existing degree of multiple job holding.

The key advantage of studying multiple job holding with the SIAB is that they precisely record transitions and job durations. In the administrative data, every job generates a new employment spell, which allows me to track all employment state transitions, i.e. from a single to two jobs, as well as changes in industrial sector, occupation and the wage.

Hence, I can study workers’ career progression observing exact job durations as well as job, sector, task and wage mobility which is important for investigating the investment motive behind multiple job holding. In contrast, the corresponding measures in survey data are potentially subject to measurement error. Per-se, they are affected by recall and time aggregation bias arising from respondents’ wrong recollection of the events and the gaps between the interviews, respectively. Generally, these biases cause unobserved labour market transitions [Bachmann and Schaffner 2009], which distort the examination.

However, studying multiple job holding with the SIAB is subject to two limitations. First, the data neither include employment spells in self-employment nor in civil service. Al- though Alden[1971] and Panos, Pouliakas and Zangelidis[2014] report that self-employment is the predominant secondary job at least in the UK, most studies exclude this form of mul-

5Only holding multiple minijobs with total earnings above the cap constitutes the exception. 6Hence, two jobs in full-time are also observed in the data. This can for example occur if a new and old job overlap. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 184 tiple job holding. The main reasons are missing information for wages and working hours

[Hirsch, Husain and Winters 2016] and distinct motives for self-employed to hold an additional job, because they are per se not working hours constrained [Boheim¨ and Taylor

2004].7 Also, civil servants constitute an idiosyncratic sample of workers as holding an additional job is restricted by law.8 For this group multiple job holding must be approved by the employer, is strictly hours constrained and capped in earnings.

Second, the data do not provide details on the household context. Thus, I can evaluate worker motives for multiple job holding only at the individual level. To my knowledge, only Krishnan[1990] considers the household context and ﬁnds evidence for the hours constraint motive. She shows that the higher a wife’s contribution to the household income, the lower is the husband’s likelihood to hold an additional job.

6.3. Aggregate Trends and Institutional Background

I start by presenting aggregate trends in multiple job holding in Germany and then move to differentiate types of multiple job holding. Workers aged below 18 and above 64 as well as apprentices in vocational training are excluded from the data. In the following analyses, the annual labour market state for each individual is determined by the cut-off date June

30th.

Figure 6.1 plots the number of individuals employed and the share of multiple hob holders between 1999 and 2014. Over the period, the number of employed rises non-monotonically from 28 to 30 millions with a decrease to 27 million between 2001 and 2005. The share of multiple job holders rises strongly among employed individuals. From 1999 to 2014 the corresponding share more than doubles from 3.5 to 7.3 percent. These numbers align well with the IAB report by Klinger and Weber[2017], who use a larger sample from the same data source.9

7A small and separate literature studies these multiple job holders named hybrid entrepreneurs [e.g Schulz, Urbig and Procher 2017]. 8Refer for details to: Bundesbeamtengesetz, Paragraph 99. 9Klinger and Weber[2017] have access to a 10 percent representative sample of dependent employed from the Integrated Employment Biographies (IEB). 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 185

Figure 6.1.: Number of employed and the share of multiple job holders , 1999-2014

31 18

14 Multiple Job Holders (in %)

29 12

27 6

4 Number of Workers (in mio.)

25 0 2000 2003 2006 2009 2012 Year

Workers Total Share MJH

Annually, cut-off date June 30th. MJH refers to multiple job holders. Numbers are extrapolated from the 2 percent sample to the full population by multiplication with 50. The vertical line marks the Hartz II reform in 2003. Source: SIAB, own calculations.

Most of the increase in the share of multiple job holding occurs with the implementation of the Hartz reforms in the early 2000s. Foremost, these reforms aimed at enhancing the

flexibility of the German labour market. Concerning multiple job holding, the Hartz II legislation, which was enforced in the first months of 2003,10 reorganized marginal employment. In Germany, marginal employment is a type of part-time work capped at a monthly income level and subsidized. Employees are not obliged to pay social security contributions and in most cases minijobs are exempted from income taxes. Marginal employed exhibit similar worker protections as regular employed such as protection against unfair dismissal or in case of sickness,11 as well as entitlements, such as holidays and financial premia. Generally, a job in marginal employment is referred to as a minijob and a person holding such a job as a minijobber, respectively.

Hartz II extended the monthly earnings cap of 325 to 400 Euro and abolished the working hours restriction. Furthermore, the administrative costs for employers to hire a minijobber were lowered by establishing a state agency for marginal employment, the Minijob-

Zentrale. Also, a reduced social security contribution rate for private household employers has been introduced: A ﬁrm’s rate amounts to 23 percent, instead a private household pays 10 percent. Note that the establishment of the Minijob-Zentrale as well as the reduced

10Refer for details to: Bundesgesetzblatt Year 2002, Part I, No. 87. 11That is, sick days are work days. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 186 contributions for private households potentially led to a transformation of previous un- registered in registered jobs which goes hand in hand with the reform’s goal of reducing moonlighting. Moreover, Hartz II considerably increased the incentives to hold multiple jobs. Workers with a primary job subject to social security contributions were now allowed to take up a secondary subsidized minijob.12

The legislation for marginal employment has been modiﬁed again in 2006 and 2013. How- ever, without speciﬁc changes for multiple job holders. In 2006, the employer payment rate for social security contribution rose from 23 to 28 percent.13 In 2013, the monthly earnings cap increased from 400 to 450 Euro and the rule for a worker’s contribution to the pension insurance changed from exemption to opt-out.14

Figure 6.2 reﬂects a steep rise in minijobs after 2003. Galassi[2018] studies Hartz II as a anti-poverty policy in terms of wages and transitions to better jobs for targeted workers.

Using the Socio-economic Panel (SOEP) she ﬁnds no positive wage effects for low-earners.

However, the reform increased the probability for inactive and unemployed people to transit to employment. Furthermore, Caliendo and Wrohlich[2010] analyse short-run effects of Hartz II on marginal employment. In the short-run, the increase in minijobs cannot be causally related to Hartz II. Nevertheless, they ﬁnd that single men are more likely to take up a secondary minijob.

Figure 6.3 supports this evidence by showing how strong multiple job holding reacted in absolute terms in the aggregate and by disaggregated employment types. It displays the total number of multiple job holders, the number of dual job holders, the number of dual job holders of the type primary job subject to social contributions and a secondary minijob and the number of dual job holders of the type primary job full-time employment and a secondary minijob over the period 1999 to 2014. The difference between the latter two corresponds to the sub-type primary part-time job.

12Between 1999 and 2002 this was precluded by law (see Bundesgesetzblatt Year 1999, Part I, No. 14). 13Refer for details to: Bundesgesetzblatt Year 2006, Part I, No. 30. Collischon, Cygan-Rehm and Riphahn [2018] exploit this variation to show that marginal employment crowds out regular employment in small ﬁrms. 14Refer for details to: Bundesgesetzblatt Year 2012, Part I, No. 58. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 187

Figure 6.2.: Number of minijobbers, 1999-2014 6 5.5 5 4.5 4 Number of Workers ( in mio. ) 3.5 3 2000 2003 2006 2009 2012 Years Annually, cut-off date June 30th. Numbers are extrapolated from the 2 percent sample to the full population by multiplication with 50. The vertical line marks the Hartz II reform in 2003. Source: SIAB, own calculations.

Between 1999 and 2002, before the reform, Germany had around 0.95 million multiple job holders. From 2002 to 2004, starting with the reform, this number substantially rose to

1.4 million. After 2004 the increase continues, although less steep, to 2.2 millions in 2014 implying a growth rate of 129 percent in 15 years.

Figure 6.3.: Types of multiple job holding, 1999-2014 3 2.5 2 1.5 1 Number of Workers ( in mio. ) .5 0 2000 2003 2006 2009 2012 Years

MJH 2 Jobs 2 Jobs: SSC and Mini 2 Jobs: FT and Mini

Annually, cut-off date June 30th. MJH refers to all multiple job holders, 2 Jobs to dual job holders, 2 Jobs: SSC and Mini to dual job holders of the type primary job subject to social contributions and a secondary minijob and 2 Jobs: FT and Mini to dual job holders of the type primary job full-time employment and a secondary minijob. Numbers are extrapolated from the 2 percent sample to the full population by multiplication with 50. The vertical line marks the Hartz II reform in 2003. Source: SIAB, own calculations.

Figure 6.3 illustrates that multiple job holding in Germany actually corresponds to dual job holding. Over the period, less than ﬁve percent hold more than two jobs. With the change in incentives, as expected, the number of multiple job holders with a primary job subject to social security contributions and a secondary minijob grows rapidly after the reform: 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 188 from half a million in 2002 to 1.1 million in 2004. In 2004, they represent 80 percent of all multiple job holders. This prevalence persists until 2014.

Descriptive evidence from Figure 6.3 further suggests that the reforms in 2006 and 2013 had no visible impact on multiple job holding since the pre-reform do not differ from post- reform growth rates.

6.4. Worker Motives

I examine the worker motives for multiple job holding in Germany over the period 2003 to 2014 by studying its determinants and labour market consequences. Comparing worker and job characteristics of multiple with single job holders yields insights on the determinants for multiple job holding. Analysing labour market mobility separately for multiple and single job holders provides evidence about the labour market consequences of multiple job holding.

I investigate worker motives for multiple job holding focusing on the type primary job subject to social security contributions, which may be in full- or part-time, with a secondary minijob. As illustrated in Section 6.3 this type marks the development in multiple job holding in Germany after the Hartz II reform in 2003 and represents more than 80 percent of all multiple job holders.

In the following analyses, annual employment states as well as worker and job characteristics are determined by the cut-off date June 30th. This procedure induces the same time aggregation bias in measuring worker mobility as in survey data. The bias occurs if an individual changes more than once the labour market state between the annual cut-off dates. Alternatively, the annual employment state can be determined by the longest state within a year. However, this creates the same bias. Nevertheless, compared to survey data the procedure does not involve a recall bias, because the SIAB are based on administrative records. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 189

6.4.1. Determinants

Table 6.1 displays worker and job characteristics of multiple and single job holders both aggregated and disaggregated by sex. Given the high number of observations, all reported differences are statistically signiﬁcant at the one percent level. On the aggregate level, the comparison (Column 1 minus Column 2) shows that multiple job holders are on average

0.6 years younger and have acquired 3.17 years more labour market experience than single job holders. Together, this suggests that multiple job holders are less educated than single job holders. In fact, Table 6.1 shows that they possess higher shares of low education and lower shares of higher education. Multiple job holding is relatively less prevalent for workers with a German citizenship. Moreover, in East Germany multiple job holding is less common than in West Germany. Six percent of multiple job holders work in East

Germany (Column 1), whereas the corresponding share for single job holders is 16 percent

(Column 2).

Turning to job characteristics, multiple job holders remain on average 3.1 years in the same job.15 Job duration16 for single job holders is 5.3 years longer. Also, multiple job holders have a eleven percent points lower share of full-time employment in the primary job. In line with lower levels of education, they are more likely to perform routine and non-routine manual tasks in their primary jobs than single job holders. The opposite holds true for non- routine cognitive tasks.17

The empirical literature documents that women and men overall exhibit distinct labour market behaviour. Female labour supply is more elastic over the life-cycle as well as with regard to wages and non-wage income [among others Evers, De Mooij and Van Vuuren

2008]. Hence, Table 6.1 disaggregates the comparison between multiple and single job holders by sex. Contrasting Column 3 to Column 5, the female and male multiple job holders, reveals that slightly more women hold multiple jobs with a share of 55 percent.18

15The same job for multiple job holders implies both remaining in the identical primary job and minijob. 16This refers to completed job duration observed in the spell data before applying the cut-off date. 17I classify tasks according to a time-ﬁxed variant of Cortes[2016]. The code was provided by Ronald Bachmann. 18The share is calculated from the numbers of observation of female and male multiple job holders. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 190

Table 6.1.: Worker and job characteristics of multiple and single job holders by sex, 2003- 2014

All Women Men Delta

Unit MJH SJH MJH SJH MJH SJH All Women Men (1) (2) (3) (4) (5) (6) (1)-(2) (3)-(4) (5)-(6)

(a) Worker Characteristics Age Yr. 41.65 42.24 41.85 42.33 41.41 42.18 -0.59 -0.48 -0.77 Experience Yr. 18.65 15.48 17.84 14.72 19.65 16.08 3.17 3.12 3.57 Education Low % 11.84 8.57 11.13 8.58 12.72 8.56 3.27 2.55 4.16 Medium % 55.04 58.72 54.82 60.18 55.32 57.55 -3.68 -5.36 -2.23 High % 4.94 11.63 4.63 9.88 5.33 13.01 -6.69 -5.25 -7.68 German % 88.17 93.11 90.31 94.46 85.54 92.04 -4.94 -4.15 -6.50 East % 6.14 15.80 6.98 17.59 5.1 14.39 -9.66 -10.61 -9.29

(b) Primary Job Characteristics Duration Yr. 3.13 8.93 3.12 8.46 3.14 9.31 -5.8 -5.34 -6.17 Full-Time % 67.82 79.05 49.43 60.58 90.56 93.72 -11.23 -11.15 -3.16 Task R % 45.17 42.36 40.72 35.70 50.68 47.65 2.81 5.02 3.03 NRM % 27.79 21.91 32.00 29.02 22.58 16.27 5.88 2.98 6.31 NRC % 22.92 32.14 23.96 32.45 21.63 31.90 -9.22 -8.49 -10.27 Observations 361,866 5,579,029 200,005 2,468,631 161,861 3,110,398 MJH refers to multiple and SJH to single job holders. Delta are differences between MJH and SJH. R refers to routine, NRM to non-routine manual and NRC to non-routine cognitive tasks, respectively. The shares of education as well as task groups do not add up to 100 percent due to missing values in the data. Source: SIAB (1975-2014), own calculations.

Female and male multiple job holders are remarkably similar in their individual characteristics and with regard to job duration. However, every second woman is employed in part-time in her primary job compared to only 9 percent among men. Also, the tasks performed in the primary job differ strongly. A higher share of women perform non-routine manual and cognitive tasks than men. The opposite holds true for routine tasks.

Concerning differences between multiple and single job holders by sex, that is for women

Column 3 minus Column 4 and for men Column 5 minus Column 6, both display overall analogous ﬁndings as the aggregate examination (Column 1 minus Column 2). The differences are in all worker and job characteristics of the same direction as in the aggregate. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 191

Next, I run logistic regressions for the probability of holding multiple jobs. The dependent variable takes the value one for multiple and the value zero for single job holders, respectively. The regressions include individual controls for labour market experience, education, nationality, region as well as whether the primary job is full-time and which type of task is performed.

Table 6.2.: Determinants of multiple job holding by sex, 2003-2014

All Women Men

M.E. S.E. M.E. S.E. M.E. S.E. (1) (2) (3) (4) (5) (6)

(a) Worker Characteristics Experience 0.002*** 0.00001 0.002*** 0.00002 0.001*** 0.00001 Education Low 0.005*** 0.00030 0.002*** 0.00051 0.006*** 0.00036 Medium Reference Group High -0.033*** 0.00045 -0.033*** 0.00075 -0.028*** 0.00054 German -0.032*** 0.00032 -0.042*** 0.00057 -0.027*** 0.00036 East -0.047*** 0.00040 -0.056*** 0.00062 -0.040*** 0.00054

(b) Primary Job Characteristics Full-Time -0.029*** 0.00021 -0.022*** 0.00032 -0.019*** 0.00041 Task R -0.008*** 0.00023 -0.001*** 0.00037 -0.012*** 0.00029 NRM Reference Group NRC -0.022*** 0.00028 -0.019*** 0.00043 -0.023*** 0.00037 Observations 5,725,255 2,591,502 3,133,753 M.E. are the marginal effects for the probability to hold multiple jobs from a logit model. S.E. are the corresponding standard errors. The dependent variable takes the value one for multiple and the value zero for single job holders, respectively. The controls include all variables displayed in the table. R refers to routine, NRM to non-routine manual and NRC to non-routine cognitive tasks, respectively. */ **/ *** refers to α = 0.1/0.05/0.01. Source: SIAB (1975-2014), own calculations.

Table 6.2 shows the corresponding estimation results aggregated and disaggregated by sex.

The marginal effects for the aggregate (Column 1) as well as for women (Column 3) and men (Column 5) confirm that multiple job holding is more prevalent among experienced and low educated workers. Furthermore, it is less prevalent among highly educated in- 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 192 dividuals as well as individuals holding a German nationality, working in East Germany, employed in full-time as well as performing routine and non-routine cognitive tasks. All estimates are statistically significantly different from zero at the 1 percent level. The estimated coefficients for the aggregate and disaggregated by sex are of similar size for each variable.

The presented results for the determinants of multiple job holding are in line with the general consensus in the empirical literature [e.g Kimmel and Powell 1999; Heineck 2009;

Klinger and Weber 2017]. They suggest that workers hold an additional job predominantly due to monetary intentions since multiple job holding relates more to women, to lower educated and part-time employed.

6.4.2. Consequences

The consequences of multiple job holding have received less attention than analyses of the determinants of multiple job holding. Only Panos, Pouliakas and Zangelidis[2014] investigate the former using survey data for British men. They ﬁnd that multiple job holders have a higher probability to change jobs than single job holders. Showing that multiple job holders are also more likely with a job-to-job transitions to switch the industrial sector and to receive a pay raise, Panos, Pouliakas and Zangelidis[2014] provide evidence for an investment motive behind multiple job holding. The investment motive suggests that workers hold more than one job to obtain new skills and experience as well as a stepping-stone for a new career path.

In the following, I analyse labour market transitions of multiple job holders in Germany.

Table 6.3 compares annual transition rates of multiple and single job holders between 2003 and 2013 both aggregated and disaggregated by sex.19 The labour market states in t + 1 include being not registered, unemployed and employed in a different job.20 Moreover, the table illustrates differences in job-to-job transitions concerning the primary job with regard to switches in three-digit industrial sectors and job tasks as well as with regard to the rate to

19For 2014, I cannot calculate transition rates since the data ends in 2014. 20Notice that having a different job for multiple job holders refers to job changes in each of the two jobs. Staying in the same job is deﬁned as holding in both periods the identical primary job and minijob. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 193 receive a wage increase. The latter takes the value one if the job-to-job transitions includes a pay raise and zero otherwise. All underlying transition rates for multiple and single job holders are calculated relative to the respective group sizes in t.

Table 6.3.: Differences in labour market mobility between multiple and single job holders by sex, 2003-2013

All Women Men

State Next Year Delta Exp. N Delta Exp. N Delta Exp. N (1) (2) (3) (4) (5) (6) (7) (8) (9) Not Registered -1.17 19.17 4,307,331 -1.60 58.31 1,921,926 -1.15 -10.36 2,385,405 Unemployment 1.55 -31.15 4,133,241 1.72 -34.69 1,822,528 1.32 -29.95 2,310,713 Same Job Comparison Group Different Job 27.15 -2.75 4,736,637 27.06 -4.08 2,121,622 26.83 -2.26 2,615,015 with Sector Change 8.24 -13.44 4,288,233 9.24 -10.75 1,877,605 7.38 -14.62 2,410,628 with Task Change 7.72 -6.03 4,144,558 8.27 -7.00 1,827,964 7.07 -4.48 2,316,594 with Wage Increase(1) 22.74 -3.89 3,526,043 25.56 -2.85 1,188,314 20.74 -4.92 2,337,729 Delta are differences in transition rates to each state between MJH and SJH in percentage points. Exp. refers to the explained part of these differences, in percent, from a Blinder-Oaxaca decomposition proposed by Sinning, Hahn and Bauer[2008] using a logit model. The dependent variable takes for each transition the value one for workers moving to the respective state. In all logit regressions the dependent variable takes the value zero for worker staying in the same job. That is, workers staying in the same job constitute the comparison group. The control variables include all variables displayed in the Table 6.2. Source: SIAB (1975-2014), own calculations.

Column 1 of Table 6.3 displays percentage points differences in the transition rates to the corresponding labour market state between multiple and single job holders for the aggregate. Accordingly, multiple job holders have a 1.17 percentage points lower transition rate to not being administratively registered next year than single job holders. The unemployment inﬂow rate is higher, by 1.55 percentage points.

In contrast, differences in job mobility between multiple and single job holders are striking. The job-to-job transition rate for multiple job holders is 27.15 percentage points above the corresponding rate for single job holders. Also, multiple job holders are more likely with a primary job change to switch sectors, by 8.24 percentage points, and job tasks, by

7.72 percentage points. Since they exhibit a higher rate to receive a wage increase in their primary job with a job change, by 22.74 percentage points, too, the comparisons provide 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 194 tentative evidence for an investment motive. These raw differences suggest that multiple job holding relates to career progression.

Job mobility patterns are sex speciﬁc. The general understanding is that women are less mobile [among others Royalty 1998; Theodossiou 2002]. Therefore, Column 4 and Column

7 contrast the labour market transitions for women and men, respectively. Comparing their mobility behaviour with the aggregate (Column 1), both reveal very similar differences between multiple and single job holders.

In the next step, I investigate whether these gaps in labour market transitions between multiple and single job holders can be explained by differences in observable characteristics. Therefore, I perform a Blinder-Oaxaca Decomposition [Blinder 1973; Oaxaca 1973] for nonlinear regression models as proposed by Sinning, Hahn and Bauer[2008]. I run for each transition logistic regressions in which the dependent variable takes the value one for workers moving to the respective labour market state and zero for worker staying in the same job. Hence, I follow Panos, Pouliakas and Zangelidis[2014] by using the group of workers staying in the same job as the comparison group. The logistic regressions control for the same explanatory variables as before. Thus, the explanatory variables include labour market experience, education, nationality, region as well as whether the primary job is full-time and which type of task is performed.

Column 2, Column 5 and Column 8 of Table 6.3 show how much of the raw differences in worker mobility can be explained by differences in the means of the included observable characteristics between multiple and single job holders in the aggregate, disaggregated for women and for men, respectively. The corresponding shares are displayed in percent.

Refer for the complete Blinder-Oaxaca Decomposition to 6.A.1. For most transitions, the shares are negative implying that differences in observable characteristics widen the mobility gap between multiple and single job holders. The exceptions are the transitions to being not registered for the aggregate and women as their values are positive. Generally, these ﬁndings suggest strong unobserved heterogeneities in both worker mobility as well as its determinants between multiple and single job holders. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 195

6.5. Conclusion

In this paper, I analyse the recent increase in multiple job holding in Germany focusing on labour market supply. I exploit the richness of the weakly anonymous Sample of Integrated

Labour Market Biographies (SIAB) to investigate worker motives. This administrative data allow me to analyse different types of multiple job holding, to investigate a wide range of socio-economic information of multiple job holders and to study their labour market transitions. Hence, I identify worker motives for multiple job holding by examining its determinants as well as its consequences for career progression. In particular, with the

SIAB I can study labour market mobility with a higher accuracy compared to survey data, which is commonly used to analyse multiple job holding.

I ﬁnd that in Germany multiple job holding actually is dual job holding. The Hartz II reform in 2003 increased incentives to hold an secondary subsidized minijob. Between 2003 and 2014 multiple job holding rose from 1.0 to 2.2 million. Since the reform the type primary job subject to social security contributions with a minijob is most prevalent. A minijob is marginal part-time employment which is for the worker exempted from social security contributions and from taxes. This type represents over 80 percent of all multiple job holders. The analyses implies that workers hold an additional job partly due to monetary intentions. Multiple job holding is more prevalent among primary part-time employed, females, low educated and low skilled. This may imply that individuals are in need to increase earnings with an additional subsidized job. Alongside, multiple job holding positively relates to career progression which provides support the investment motive. Multi- ple job holders are more likely to switch jobs, sectors and tasks as well as to receive a wage increase relative to single job holders. The results prevail when separating the analyses by sex. Moreover, I ﬁnd strong unobserved heterogeneities in both worker mobility as well as its determinants between multiple and single job holders.

The study suggests that the recent increase in multiple job holding in Germany relates to career progression. Therefore, multiple job holding can be assessed from a policy perspective as beneﬁcial. However, holding a secondary minijob instead of increasing earnings in 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 196 the primary job would imply ceteris paribus lower aggregate social security contributions and taxes. 6. WORKER MOTIVESFOR MULTIPLE JOB HOLDINGIN GERMANY 197

6.A. Tables

Table 6.A.1.: Differences in labour market mobility between multiple and single job holders by sex, 2003-2013, complete decomposition

All Women Men

State Next Year Delta Exp. Unexp. N Delta Exp. Unexp. N Delta Exp. Unexp. N (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Not Registered -1.17 19.17 80.83 4,307,331 -1.60 58.31 41.69 1,921,926 -1.15 -10.36 110.36 2,385,405 Unemployment 1.55 -31.15 131.15 4,133,241 1.72 -34.69 134.69 1,822,528 1.32 -29.95 129.95 2,310,713 Same Job Comparison Group Different Job 27.15 -2.75 102.75 4,736,637 27.06 -4.08 104.08 2,121,622 26.83 -2.26 102.26 2,615,015 with Sector Change 8.24 -13.44 113.44 4,288,233 9.24 -10.75 110.75 1,877,605 7.38 -14.62 114.62 2,410,628 with Task Change 7.72 -6.03 106.03 4,144,558 8.27 -7.00 107.00 1,827,964 7.07 -4.48 104.48 2,316,594 with Wage Increase(1) 22.74 -3.89 103.89 3,526,043 25.56 -2.85 102.85 1,188,314 20.74 -4.92 104.92 2,337,729 Delta are differences in transition rates to each state between MJH and SJH in percentage points. Exp. refers to the explained part of these differences, in percent, from a Blinder-Oaxaca decomposition proposed by Sinning, Hahn and Bauer[2008] using a logit model. Unexp. refers to the unexplained part of these differences, in percent, respectively. The dependent variable takes for each transition the value one for workers moving to the respective state. In all logit regressions the dependent variable takes the value zero for worker staying in the same job. That is, workers staying in the same job constitute the comparison group. The control variables include all variables displayed in the Table 6.2. Source: SIAB (1975-2014), own calculations. Bibliography

Abowd, John M., Francis Kramarz and Simon Woodcock. 2008. Econometric Analyses of Linked Employer–Employee Data. In The Econometrics of Panel Data, ed. Laszl´ o´ Maty´ as´ and Patrick Sevestre. Heidelberg: Springer chapter 22, pp. 727–760. Abowd, John M., Robert H. Creecy, Francis Kramarz et al. 2002. Computing Person and Firm Effects Using Linked Longitudinal Employer–Employee Data. CES Technical Paper TP-2002-06. US Census Bureau. Abraham, Katharine G. and James L. Medoff. 1984. “Length of Service and Layoffs in Union and Nonunion Work Groups.” Industrial and Labor Relations Review 38(1):87–97. Alden, Jeremy. 1971. “Double-Jobholding: A Regional Analysis of Scotland.” Scottish Jour- nal of Political Economy 18(1):99–112. Andrews, Martyn J., Len Gill, Thorsten Schank and Richard Upward. 2008. “High Wage Workers and Low Wage Firms: Negative Assortative Matching or Limited Mobility Bias?” Journal of the Royal Statistical Society: Series A (Statistics in Society) 171(3):673–697. Angrist, Joshua D., Guido W. Imbens and Donald B. Rubin. 1996. “Identification of Causal Effects Using Instrumental Variables.” Journal of the American Statistical Associ- ation 91(434):444–455. Antoni, M. and J. Eberle. 2015. Kurzdokumentation NEPS-SC6-ADIAB. Technical report Nurnberg,¨ IAB. Auer, Peter and Sandrine Cazes. 2000. “The Resilience of the Long-Term Employment Relationship: Evidence from the Industrialized Countries.” International Labour Review 139(4):379–408. Bachmann, Ronald. 2005. Labour Market Dynamics in Germany: Hirings, Separations, and Job-to-Job Transitions Over the Business Cycle. SFB 649 Discussion Paper 45. Son- derforschungsbereich 649: Okonomisches¨ Risiko, Humboldt-Universitat¨ Berlin. Bachmann, Ronald, Miriam Bomer,¨ Hanna Frings, Wolfgang Durig,¨ Lisa Sofie Hockel¨ and Fernanda Martinez Flores. 2016. “Nachfolgestudie zur Analyse der Geringfugigen¨ Beschaftigungsverh¨ altnisse¨ (Minijobs) sowie den Auswirkungen des Gesetzlichen Min- destlohns.” Research Report for the Ministerium fürArbeit, Integration und Soziales des Landes Nordrhein-Westfalen . Bachmann, Ronald, Peggy Bechara, Anica Kramer and Sylvi Rzepka. 2015. “Labour Mar- ket Dynamics and Worker Heterogeneity during the Great Recession - Evidence from Europe.” IZA Journal of European Labor Studies 4(19):1–29. Bachmann, Ronald and Rahel Felder. 2018. “Job Stability in Europe over the Cycle.” Inter- national Labour Review 157(3):481–518. BIBLIOGRAPHY 199

Bachmann, Ronald and Sandra Schaffner. 2009. Biases in the Measurement of Labour Mar- ket Dynamics. Technical Report 12. TU Dortmund. Balleer, Almut, Britta Gehrke, Wolfgang Lechthaler and Christian Merkl. 2016. “Does Short–Time Work Save Jobs? A Business Cycle Analysis.” European Economic Review 84:99–122. Bassanini, Andrea and Andrea Garnero. 2013. “Dismissal Protection and Worker Flows in OECD Countries: Evidence from Cross-Country/Cross-Industry Data.” Labour Eco- nomics 21:25–41. Bassanini, Andrea and Romain Duval. 2006. “The Determinants of Unemployment Across OECD Countries: Reassessing the Role of Policies and Institutions.” OECD Economic Studies 42(1):7–86. Bedoya, Guadalupe, Luca Bittarello, Jonathan M. Davis and Nikolas Mittag. 2017. Dis- tributional Impact Analysis: Toolkit and Illustrations of Impacts Beyond the Average Treatment Effect. Policy Research Working Paper 8139. The World Bank. Bell, David, Robert A. Hart and Robert E. Wright. 1997. Multiple Job Holding as a ’Hedge’ Against Unemployment. CEPR Discussion Papers 1626. Belot, Michele and Jan C. Van Ours. 2004. “Does the Recent Success of some OECD Coun- tries in Lowering their Unemployment Rates Lie in the Clever Design of their Labor Market Reforms?” Oxford Economic Papers 56(4):621–642. Bentolila, Samuel and Giuseppe Bertola. 1990. “Firing Costs and Labour Demand: How Bad Is Eurosclerosis?” The Review of Economic Studies 57(3):381–402. Bentolila, Samuel, Pierre Cahuc, Juan J. Dolado and Thomas Le Barbanchon. 2012. “Two- Tier Labour Markets in the Great Recession: France Versus Spain.” Economic Journal 122(562):F155–F187. Bertola, Giuseppe. 1990. “Job Security, Employment and Wages.” European Economic Review 34(4):851–879. Bertola, Giuseppe. 2017. “European Unemployment Revisited: Shocks, Institutions, Inte- gration.” Research in Economics 3(71):588–612. Bertola, Giuseppe and Richard Rogerson. 1997. “Institutions and Labor Reallocation.” Eu- ropean Economic Review 41(6):1147–1171. Biewen, Martin, Bernd Fitzenberger and Jakob de Lazzer. 2018. “The Role of Employment Interruptions and Part-Time Work for the Rise in Wage Inequality.” IZA Journal of Labor Economics 7(1):10. Blanchard, Olivier. 1999. “European Unemployment: The Role of Shocks and Institutions.” Bafﬁ Lecture Rome. Blanchard, Olivier Jean and Peter Diamond. 1994. “Ranking, Unemployment Duration, and Wages.” The Review of Economic Studies 61(3):417–434. BIBLIOGRAPHY 200

Blanchard, Olivier and Justin Wolfers. 2000. “The Role of Shocks and Institutions in the Rise of European Unemployment: The Aggregate Evidence.” The Economic Journal 110(462):1– 33. Blanchard, Olivier and Pedro Portugal. 2001. “What Hides Behind an Unemployment Rate: Comparing Portuguese and US Labor Markets.” American Economic Review 91(1):187– 207. Blinder, Alan S. 1973. “Wage Discrimination: Reduced Form and Structural Estimates.” Journal of Human resources pp. 436–455. Blossfeld, Hans-Peter and Jutta Von Maurice. 2011. “Education As a Lifelong Process.” Zeitschrift fürErziehungswissenschaft 14(2):19–34. BLS (Bureau of Labor Statistics). 2014. Employee Tenure in 2014. Working Paper USDL-14- 1714. BLS News Release Washington, DC, US Department of Labor. Boeri, Tito. 1999. “Enforcement of Employment Security Regulations, On-the-Job Search and Unemployment Duration.” European Economic Review 43(1):65–89. Boeri, Tito. 2011. Institutional Reforms and Dualism in European Labor Markets. In Hand- book of labor economics, ed. Orley Ashenfelter and David Card. Vol. 4B. Amsterdam, North Holland, pp. 1173–1236. Boeri, Tito and Jan Van Ours. 2013. The Economics of Imperfect Labor Markets. Princeton University Press. Boheim,¨ Rene´ and Mark Taylor. 2004. And in the Evening She’s a Singer with the Band- Second Jobs, Plight or Pleasure? IZA Discussion Paper Series 1081. IZA. Bohnke,¨ Petra, Janina Zeh and Sebastian Link. 2015. “Atypische Beschaftigung¨ im Er- werbsverlauf: Verlaufstypen als Ausdruck Sozialer Spaltung?” Zeitschrift fürSoziologie 44(4):234–252. Bonke,¨ Timm, Giacomo Corneo and Holger Luthen.¨ 2015. “Lifetime Earnings Inequality in Germany.” Journal of Labor Economics 33(1):171–208. Bonke,¨ Timm, Matthias Giesecke and Holger Luthen.¨ 2015. The Dynamics of Earnings in Germany: Evidence from Social Security Records. Ruhr Economic Paper 582. RWI. Boockmann, Bernhard and Susanne Steffes. 2010. “Workers, Firms, or Institutions: What Determines Job Duration for Male Employees in Germany?” Industrial and Labor Relations Review 64(1):109–127. Boockmann, Bernhard and Tobias Hagen. 2008. “Fixed-Term Contracts As Sorting Mecha- nisms: Evidence from Job Durations in West Germany.” Labour Economics 15(5):984–1005. Brehmer, Wolfram and Hartmut Seifert. 2008. “Sind Atypische Beschaftigungsverh¨ altnisse¨ Prekar?¨ eine Empirische Analyse Sozialer Risiken.” Zeitschrift fürArbeitsmarktForschung– Journal for Labour Market Research 41(4):501–531. Brzinsky-Fay, Christian, Ulrich Kohler and Magdalena Luniak. 2006. “Sequence Analysis with Stata.” The Stata Journal 6(4):435–460. BIBLIOGRAPHY 201

Bundesagentur fur¨ Arbeit. 2013. Zeitarbeit in Deutschland–Aktuelle Entwicklungen. Stand januar 2013. Burda, Michael C. and Jennifer Hunt. 2011. “What Explains the German Labor Market Miracle in the Great Recession?” Brookings Papers on Economic Activity 2011(1):273–319. Burda, Michael C. and Mark Weder. 2016. “Payroll Taxes, Social Insurance, and Business Cycles.” Journal of the European Economic Association 14(2):438–467. Burgess, Simon. 1999. Reallocation of Labour: An International Comparison Using Job Tenure Data. CEPR Discussion Paper 416. London, Centre for Economic Performance, London School of Economics and Political Science. Cahuc, Pierre, Olivier Charlot and Franck Malherbet. 2016. “Explaining the Spread of Tem- porary Jobs and Its Impact on Labor Turnover.” International Economic Review 57(2):533– 572. Caliendo, Marco and Katharina Wrohlich. 2010. “Evaluating the German ’Mini-Job’ Reform Using a Natural Experiment.” Applied Economics 42(19):2475–2489. Caliendo, Marco, Konstantinos Tatsiramos and Arne Uhlendorff. 2013. “Beneﬁt Dura- tion, Unemployment Duration and Job Match Quality: A Regression-Discontinuity Ap- proach.” Journal of Applied Econometrics 28(4):604–627. Caliendo, Marco, Steffen Kunn¨ and Arne Uhlendorff. 2016. “Earnings Exemptions for Un- employed Workers: The Relationship between Marginal Employment, Unemployment Duration and Job Quality.” Labour Economics 42:177–193. Calonico, Sebastian, Matias D. Cattaneo and Rocio Titiunik. 2014. “Robust Nonparamet- ric Conﬁdence Intervals for Regression-Discontinuity Designs.” Econometrica 82(6):2295– 2326. Card, David, Jochen Kluve and Andrea Weber. 2010. “Active Labour Market Policy Evalu- ations: A Meta-Analysis.” The Economic Journal 120(548, Features November 2010):F452– F477. Card, David, Jochen Kluve and Andrea Weber. 2018. “What Works? A Meta Analysis of Recent Active Labor Market Program Evaluations.” Journal of the European Economic Association 16(3):894–931. Card, David, Jorg¨ Heining and Patrick Kline. 2013. “Workplace Heterogeneity and the Rise of West German Wage Inequality.” The Quarterly Journal of Economics 128(3):967–1015. Card, David, Raj Chetty and Andrea Weber. 2007. “Cash-On-Hand and Competing Models of Intertemporal Behavior: New Evidence from the Labor Market.” The Quarterly Journal of Economics 122(4):1511–1560. Carrillo-Tudela, Carlos, Andrey Launov and Jean-Marc Robin. 2018. The Fall in German Unemployment: A Flow Analysis. CEPR Discussion Paper DP12846. Cazes, Sandrine and Mirco Tonin. 2010. “Employment Protection Legislation and Job Sta- bility: A European Cross-Country Analysis.” International Labour Review 149(3):261–285. BIBLIOGRAPHY 202

Clauwaert, Stefan and Isabelle Schomann.¨ 2012. The Crisis and National Labour Law Re- forms: A Mapping Exercise. ETUI Working Paper 2012.04. Brussels, European Trade Union Institute. Collischon, Matthias, Kamila Cygan-Rehm and Regina T. Riphahn. 2018. Employment Effects of Payroll Tax Subsidies. CESifo Working Papers 7111. Conway, Karen Smith and Jean Kimmel. 1998. “Male Labor Supply Estimates and the Decision to Moonlight.” Labour Economics 5(2):135–166. Cortes, Guido M. 2016. “Where Have the Middle–Wage Workers Gone? A Study of Polar- ization using Panel Data.” Journal of Labor Economics 34(1):63–105. de Serres, Alain and Fabrice Murtin. 2013. Do Policies that Reduce Unemployment Raise its Volatility? OECD Economics Department Working Paper 1020. Deaton, Angus. 2009. “Instruments of Development: Randomisation in the Tropics, and the Search for the Elusive Keys to Economic Development.” Proceedings of the British Academy 162:123–160. Dickey, Heather, Verity Watson and Alexandros Zangelidis. 2011. “Is It All about Money? An Examination of the Motives behind Moonlighting.” Applied Economics 43(26):3767– 3774. Dustmann, Christian, Bernd Fitzenberger, Uta Schonberg¨ and Alexandra Spitz-Oener. 2014. “From Sick Man of Europe to Economic Superstar: Germany’s Resurgent Econ- omy.” Journal of Economic Perspectives 28(1):167–88. Eamets, Raul, Epp Kallaste, Jaan Masso and Marit Room. 2003. “How Flexible Are the Labour Markets in CEE Countries? A Macro Level Approach.” Transfer: European Review of Labour and Research 9(1):101–121. Ebell, Monique. 2011. “On the Cyclicality of Unemployment: Resurrecting the Participa- tion Margin.” Labour Economics 18(6):822–836. Ehrlinghagen, Marcel and Matthias Knuth. 2002. Kein Turbo-Arbeitsmarkt in Sicht. IAT Report 4. Gelsenkirchen, Institut Arbeit und Technik. Eichhorst, Werner, Michael Feil and Paul Marx. 2010. “Crisis, What Crisis? Pat- terns of Adaptation in European Labor Markets.” Applied Economics Quarterly 61(Supplement):29–64. Elsby, M., B. Hobijn and A. Sahin. 2013. “Unemployment Dynamics in the OECD.” Review of Economics and Statistics 95(2):530–548. Elsby, Michael W., Ryan Michaels and Gary Solon. 2009. “The Ins and Outs of Cyclical Unemployment.” American Economic Journal: Macroeconomics 1(1):84–100. Eurofound. 2014. European Restructuring Monitor. Luxembourg, Publications Ofﬁce of the European Union. European Central Bank. 2012. Euro Area Labour Markets and the Crisis. ECB Occasional Paper Series 138. BIBLIOGRAPHY 203

European Commission. 2001. Employment in Europe 2001: Recent Trends and Prospects. Luxembourg, Publications Office of the European Union. Eurostat. 2014a. EU Labour Force Survey Database User Guide. Luxembourg. Eurostat. 2014b. “Harmonized Unemployment Rate.”. http://epp.eurostat.ec.europa. eu/tgm/table.do?tab=table&language=en&pcode=teilm020, accessed 03/02/2014. Evers, Michiel, Ruud De Mooij and Daniel Van Vuuren. 2008. “The Wage Elasticity of Labour Supply: A Synthesis of Empirical Estimates.” De Economist 156(1):25–43. Farber, Henry S. 2010. Job Loss and the Decline in Job Security in the United States. In Labor in the New Economy, ed. James Spletzer Katherine G. Abraham and Michael Harper. University of Chicago Press pp. 223–262. Fitzenberger, Bernd, Reinhold Schnabel and Gaby Wunderlich. 2004. “The Gender Gap in Labor Market Participation and Employment: A Cohort Analysis for West Germany.” Journal of Population Economics 17(1):83–116. Flaig, Gebhard and Horst Rottmann. 2013. “Labour Market Institutions and Unemploy- ment: An International Panel Data Analysis.” Empirica 40(4):635–654. Flood, Sarah, Miriam King, Steven Ruggles and J. Robert Warren. 2015. Integrated Pub- lic Use Microdata Series, Current Population Survey: Version 4.0. [Machine-Readable Database]. Technical Report. Minneapolis, University of Minnesota. Freeman, Richard B. 1978. A Fixed Effect Logit Model of the Impact of Unionism on Quits. NBER Working Paper 280. Fujita, Shigeru and Garey Ramey. 2009. “The Cyclicality of Separation and Job Finding Rates.” International Economic Review 50(2):415–430. Gal, Peter and Adam Theising. 2015. The Macroeconomic Impact of Policies on Labour Market Outcomes in OECD Countries. OECD Economics Department Working Paper 1271. Galassi, Gabriela. 2018. The German Mini-Job Reform: Intended and Unintended Conse- quences. Unpublished. European University Institute. Gebel, Michael. 2010. “Early Career Consequences of Temporary Employment in Germany and the UK.” Work, Employment and Society 24(4):641–660. Gielen, Anne C. and Trudie Schils. 2014. Non-Standard Employment Patterns in the Netherlands. Policy Paper 77. IZA. Giesecke, Johannes and Martin Groß. 2002. “Befristete Beschaftigung:¨ Chance oder Risiko?” KZfSS KölnerZeitschrift fürSoziologie und Sozialpsychologie 54(1):85–108. Gnocchi, Stefano, Andresa Lagerborg and Evi Pappa. 2015. “Do Labor Market Institutions Matter for Business Cycles?” Journal of Economic Dynamics and Control 51:299–317. Goerke, Laszlo and Markus Pannenberg. 2011. “Trade Union Membership and Dis- missals.” Labour Economics 18(6):810–821. BIBLIOGRAPHY 204

Goette, Lorenz, Uwe Sunde and Thomas K. Bauer. 2007. “Wage Rigidity: Measurement, Causes and Consequences.” The Economic Journal 117(524):F499–F507. Grogger, Jeffrey and Conny Wunsch. 2012. “Unemployment Insurance and Departures from Employment: Evidence from a German Reform.” Unpublished Manuscript. Hall, Robert E. 1982. “The Importance of Lifetime Jobs in the US Economy.” American Economic Review 72(4):716–724. Haltiwanger, John, Henry Hyatt, Lisa B. Kahn and Erika McEntarfer. 2017. Cyclical Job Ladders by Firm Size and Firm Wage. Working Paper 23485. National Bureau of Eco- nomic Research. Ham, John C. and Robert J. LaLonde. 1996. “The Effect of Sample Selection and Initial Conditions in Duration Models: Evidence from Experimental Data on Training.” Econo- metrica 64(1):175–205. Hanushek, Eric A., Guido Schwerdt, Ludger Woessmann and Lei Zhang. 2017. “General Education, Vocational Education, and Labour-Market Outcomes over the Life-Cycle.” Journal of Human Resources 52(1):48–87. Heckman, James J. 2008. “Econometric Causality.” International Statistical Review 76(1):1–27. Heckman, James J. and Edward J. Vytlacil. 2000. “The Relationship between Treatment Parameters within a Latent Variable Framework.” Economics Letters 66(1):33–39. Heckman, James J. and Edward J. Vytlacil. 2001. “Policy-Relevant Treatment Effects.” American Economic Review 91(2):107–111. Heckman, James J. and Edward J. Vytlacil. 2007. Econometric Evaluation of Social Pro- grams, Part II: Using the Marginal Treatment Effect to Organize Alternative Econometric Estimators to Evaluate Social Programs, and to Forecast Their Effects in New Environ- ments. In Handbook of Econometrics, ed. James J. Heckman and Edward Leamer. Vol. 6b Amsterdam: chapter 71, pp. 4875–5143. Heckman, James J. and Sergio Urzua. 2010. “Comparing IV with Structural Models: What Simple IV Can and Cannot Identify.” Journal of Econometrics 156(1):27–37. Heckman, James J., Sergio Urzua and Edward J. Vytlacil. 2006. “Understanding Instru- mental Variables in Models with Essential Heterogeneity.” The Review of Economics and Statistics 88(3):389–432. Heineck, Guido. 2009. “The Determinants of Secondary Jobholding in Germany and the UK.” Zeitschrift f¨urArbeitsmarktforschung 42(2):107. Hijzen, Alexander and Pedro S. Martins. 2016. No Extension without Representation? Evi- dence from a Natural Experiment in Collective Bargaining. International Monetary Fund Working Paper 16/143. Hirsch, Barry T., Muhammad M. Husain and John V. Winters. 2016. “Multiple Job Holding, Local Labor Markets, and the Business Cycle.” IZA Journal of Labor Economics 5(1):4. BIBLIOGRAPHY 205

Hunt, Jennifer. 1995. “The Effect of Unemployment Compensation on Unemployment Du- ration in Germany.” Journal of Labor Economics 13(1):88–120. ILO. 1988. Current International Recommendations on Labour Statistics, 1988 Edition. Interna- tional Labour Organisation, Geneva. Imbens, Guido W. 2010. “Better LATE Than Nothing: Some Comments on Deaton (2009) and Heckman and Urzua (2009).” Journal of Economic Literature 48(2):399–423. Imbens, Guido W. and Karthik Kalyanaraman. 2012. “Optimal Bandwidth Choice for the Regression Discontinuity Estimator.” The Review of Economic Studies 79(3):933–959. Ivlevs, Artjoms and Michail Veliziotis. 2017. “What Do Unions Do in Times of Economic Crisis? Evidence from Central and Eastern Europe.” European Journal of Industrial Rela- tions 81-96(23):1. Jackson, C. Kirabo. 2013. “Match Quality, Worker Productivity, and Worker Mobility: Di- rect Evidence from Teachers.” Review of Economics and Statistics 95(4):1096–1116. Jacobi, Lena and Jochen Kluve. 2007. “Before and After the Hartz Reforms: The Perfor- mance of Active Labour Market Policy in Germany.” Journal for Labour Market Research 40(1):45–64. Jovanovic, Boyan. 1979. “Firm-Speciﬁc Capital and Turnover.” Journal of Political Economy 87(6):1246–1260. Kimmel, Jean and Lisa M. Powell. 1999. “Moonlighting Trends and Related Policy Issues in Canada and the United States.” Canadian Public Policy/Analyse de Politiques pp. 207–231. Klinger, Sabine and Enzo Weber. 2017. Zweitbeschaftigungen¨ in Deutschland: Immer mehr Menschen haben einen Nebenjob. IAB-Kurzbericht 22/2017. Krishnan, Pramila. 1990. “The Economics of Moonlighting: A Double Self-Selection Model.” The Review of Economics and Statistics pp. 361–367. Kvasnicka, Michael. 2009. Does Temporary Help Work Provide a Stepping Stone to Regu- lar Employment? In Studies of Labor Market Intermediation. University of Chicago Press pp. 335–372. Kvasnicka, Michael and Axel Werwatz. 2002. “Lohneffekte der Zeitarbeit.” DIW Wochen- bericht 69(49):847–854. Lalive, Rafael. 2007. “Unemployment Beneﬁts, Unemployment Duration, and Post- Unemployment Jobs: A Regression Discontinuity Approach.” American Economic Review 97(2):108–112. Layard, P. Richard G., Stephen J. Nickell and Richard Jackman. 1991. Unemployment: Macroeconomic Performance and the Labour Market. Oxford University Press. Layard, P. Richard G., Stephen J. Nickell and Richard Jackman. 2005. Unemployment: Macroeconomic Performance and the Labour Market. Oxford University Press. BIBLIOGRAPHY 206

Le Barbanchon, Thomas, Roland Rathelot and Alexandra Roulet. 2019. “Unemployment Insurance and Rservation Wages: Evidence from Administrative Data.” Journal of Public Economics 171:1 – 17. Martin, John P. and Stefano Scarpetta. 2012. “Setting It Right: Employment Protection, Labour Reallocation and Productivity.” De Economist 160(2):89–116. Mascherini, Massimiliano, Lidia Salvatore, Anja Meierkord and Jean-Marie Jungblut. 2012. NEETS–Young People Not in Employment, Education or Training: Characteristics, Costs and Policy Responses in Europe. Luxembourg, Publications Ofﬁce of the European Union. Masso, Jaan and Kerly Krillo. 2011. Labour Markets in the Baltic States during the Crisis 2008/2009: The Effect on Different Labour Market Groups. Working Paper 79. Faculty of Economics and Business Administration, University of Tartu. McDonald, Ian M. and Robert M. Solow. 1981. “Wage Bargaining and Employment.” The American Economic Review 71(5):896–908. Medoff, James L. 1979. “Layoffs and Alternatives under Trade Unions in US Manufactur- ing.” The American Economic Review 69(3):380–395. Messina, Julian´ and Giovanna Vallanti. 2007. “Job Flow Dynamics and Firing Restrictions: Evidence from Europe.” The Economic Journal 117(521):F279–F301. Meyer, Bruce D. 1990. “Unemployment Insurance and Unemployment Spells.” Economet- rica 58(4):757–782. Mincer, Jacob. 1974. Schooling, Experience, and Earnings. New York: Human Behavior & Social Institutions No. 2. Columbia University Press. Mittag, Nikolas. 2016. A Simple Method to Estimate Large Fixed Effects Models Applied to Wage Determinants and Matching. Discussion Paper 10447. IZA. Mittag, Nikolas. 2019. “A Nonparametric k-Sample Test of Conditional Independence.” Unpublished Manuscript. Mortensen, Dale T. and Christopher A. Pissarides. 1994. “Job Creation and Job Destruction in the Theory of Unemployment.” The Review of Economic Studies 61(3):397–415. Mortensen, Dale T. and Christopher A. Pissarides. 1999a. New Developments in Models of Search in the Labor Market. Handbook of Labor Economics, in: O. Ashenfelter & D. Card (ed.), Edition 1, Chapter 39, 2567–2627, Elsevier. Mortensen, Dale T. and Christopher A. Pissarides. 1999b. “Unemployment Responses to ’Skillbiased’ Technology Shocks: The Role of Labour Market Policy.” Economic Journal 109(455):242–265. Mueller, Andreas I, Johannes Spinnewijn and Giorgio Topa. 2018. Job Seekers’ Perceptions and Employment Prospects: Heterogeneity, Duration Dependence and Bias. Working Paper 25294. National Bureau of Economic Research. Nekoei, Arash and Andrea Weber. 2017. “Does Extending Unemployment Beneﬁts Im- prove Job Quality?” The American Economic Review 107(2):527–561. BIBLIOGRAPHY 207

Neumark, David. 2000. Changes in Job Stability and Job Security: A Collective Effort to Untangle, Reconcile, and Interpret the Evidence. NBER Working Paper 72. Cambridge, MA, National Bureau of Economic Research. Nickell, Stephen. 1997. “Unemployment and Labor Market Rigidities: Europe versus North America.” The Journal of Economic Perspectives 11(3):55–74. Nickell, Stephen J. and Martyn Andrews. 1983. “Unions, Real Wages and Employment in Britain 1951-79.” Oxford Economic Papers 35(Supplement):183–206. Nickell, Stephen, Luca Nunziata and Wolfgang Ochel. 2005. “Unemployment in the OECD since the 1960s. What Do We Know?” The Economic Journal 115(500):1–27. Nunziata, Luca. 2002. Unemployment, Labour Market Institutions and Shocks. Economics Series Working Papers 2002-w16. Oaxaca, Ronald. 1973. “Male-Female Wage Differentials in Urban Labor Markets.” Interna- tional Economic Review pp. 693–709. OECD. 2013. OECD Employment Outlook 2013. OECD Publishing. OECD (Organisation for Economic Co-operation and Development). 2004. OECD Employ- ment Outlook. IATReport Gelsenkirchen . Institut Arbeit und Technik. Ohanian, Lee E. and Andrea Raffo. 2012. “Aggregate Hours Worked in OECD Countries: New Measurement and Implications for Business Cycles.” Journal of Monetary Economics 59(1):40–56. Orlandi, Fabrice. 2012. Structural Unemployment and its Determinants in the EU Coun- tries. European Economy - Economic Papers 455. Directorate General Economic and Monetary Affairs (DG ECFIN), European Commission. Panos, Georgios A., Konstantinos Pouliakas and Alexandros Zangelidis. 2014. “Multiple Job Holding, Skill Diversiﬁcation, and Mobility.” Industrial Relations: A Journal of Economy and Society 53(2):223–272. Paul, Marie. 2016. “Is There a Causal Effect of Working Part-Time on Current and Future Wages?” The Scandinavian Journal of Economics 118(3):494–523. Perman, Roger, Gaetan Stephan and Christophe Tavera.´ 2015. “Okun’s Law - A Meta- Analysis.” The Manchester School 83(1):101–126. Petrongolo, Barbara and Christopher A. Pissarides. 2008. “The Ins and Outs of European Unemployment.” The American Economic Review 98(2):256–262. Pierse, Tom and John McHale. 2015. “Unions and Involuntary Job Separations.” Human Resourse Management Journal 25(4):496–515. Potzsch,¨ Olga and Dieter Emmerling. 2008. Geburten und Kinderlosigkeit in Deutsch- land: Bericht uber¨ die Sondererhebung 2006 ”Geburten in Deutschland”. Technical report Statistisches Bundesamt. Rinne, Ulf and Klaus F. Zimmermann. 2013. “Is Germany the North Star of Labor Market Policy?” IMF Economic Review 61(4):702–729. BIBLIOGRAPHY 208

Royalty, Anne Beeson. 1998. “Job–to–Job and Job–to–Nonemployment Turnover by Gender and Education Level.” Journal of Labor Economics 16(2):392–433. RWI. 2013. Studie zur Analyse der geringfugigen¨ Beschaftigungsverh¨ altnisse,¨ Endbericht zum Forschungsvorhaben des Ministeriums fur¨ Arbeit, Integration und Soziales des Landes Nordrhein-Westfalen. RWI Projektberichte. RWI. RWI. 2014a. Labour Market Transitions in Turbulent Times Task B. Final Report for Euro- found. RWI. RWI. 2014b. Personen, die nicht am Erwerbsleben teilnehmen: Analyse sozio- okonomischer¨ Merkmale unter besonderer Berucksichtigung¨ des Haushaltskontextes und Bestimmung des Arbeitskraftepotenzials.¨ RWI Projektberichte. RWI. RWI. 2016. Risiken atypischer Beschaftigungsformen¨ fur¨ die beruﬂiche Entwicklung und Erwerbseinkommen im Lebensverlauf, Endbericht zum Forschungsvorhaben des BMAS. RWI Projektberichte. RWI. Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung. 2008. Die Finanzkrise Meistern–Wachstumskrafte¨ Starken.¨ Jahresgutachten 2008/2009. Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung. 2012. Stabile Architektur fur¨ Europa–Handlungsbedarf im Inland. Jahresgutachten 2012/2013. Sachverstandigenrat¨ zur Begutachtung der Gesamtwirtschaftlichen Entwicklung. 2015. Zukunftsfahigkeit¨ in den Mittelpunkt. Jahresgutachten 2015/2016. Scarpetta, Stefano. 1996. “Assessing the Role of Labour Market Policies and Institutional Settings on Unemployment: A Cross-Country Study.” OECD Economic Studies 26(1):43– 98. Schmieder, Johannes F., Till von Wachter and Stefan Bender. 2012. “The Effects of Extended Unemployment Insurance over the Business Cycle: Evidence from Regression Disconti- nuity Estimates over 20 Years.” The Quarterly Journal of Economics 127(2):701–752. Schmieder, Johannes F., Till von Wachter and Stefan Bender. 2016. “The Effect of Unem- ployment Beneﬁts and Nonemployment Durations on Wages.” The American Economic Review 106(3):739–777. Schneider, Hilmar, Ulf Rinne et al. 2017. “The Labor Market in Germany, 2000–2016.” IZA World of Labor pp. 379–379. Schober, Pia S. and C. Katharina Spieß. 2014. “Die Kita-Qualitat¨ ist fur¨ das Erwerbsver- halten von Muttern¨ mit Kleinkindern Relevant: Zusammenhang Eindeutiger in Ost- deutschland.” DIW-Wochenbericht 81(21):463–471. Schulz, Matthias, Diemo Urbig and Vivien Procher. 2017. “The Role of Hybrid En- trepreneurship in Explaining Multiple Job Holders’ Earnings Structure.” Journal of Busi- ness Venturing Insights 7:9–14. Shimer, Robert. 2005. “The Cyclicality of Hires, Separations, and Job-to-Job Transitions.” Federal Reserve Bank of St. Louis Review 87(4):493–507. BIBLIOGRAPHY 209

Shishko, Robert and Bernard Rostker. 1976. “The Economics of Multiple Job Holding.” The American Economic Review 66(3):298–308. Sinning, Mathias, Markus Hahn and Thomas K. Bauer. 2008. “The Blinder-Oaxaca Decom- position for Nonlinear Regression Models.” The Stata Journal 8(4):480–492. Statistisches Bundesamt. 2016a. Atypische Beschaftigung–Kernerwerbst¨ atige¨ nach Einzelnen Erwerbsformen. https://www.destatis.de/DE/ ZahlenFakten/GesamtwirtschaftUmwelt/Arbeitsmarkt/Erwerbstaetigkeit/ TabellenArbeitskraefteerhebung/AtypKernerwerbErwerbsformZR.html. Down- loaded on 18.12.2016. Statistisches Bundesamt. 2016b. Bevolkerung¨ und Erwerbstatigkeit.¨ Fachserie 1 reihe 1.1. Wiesbaden, Statistisches Bundesamt. Tatsiramos, Konstantinos. 2009. “Unemployment Insurance in Europe: Unemployment Duration and Subsequent Employment Stability.” Journal of the European Economic Asso- ciation 7(6):1225–1260. Theodossiou, Ioannis. 2002. “Factors Affecting the Job–to–Joblessness Turnover and Gen- der.” LABOUR 16(4):729–746. Traxler, Franz and Bernhard Kittel. 2000. “The Bargaining System and Performance: A Comparison of 18 OECD countries.” Comparative Political Studies 33(9):1154–1190. Van Ours, Jan C. and Milan Vodopivec. 2008. “Does Reducing Unemployment Insurance Generosity Reduce Job Match Quality?” Journal of Public Economics 92(3):684–695. Wolf, Elke. 2010. Lohndifferenziale zwischen Vollzeit-und Teilzeitbeschaftigten¨ in Ost-und Westdeutschland. WSI-Diskussionspapier. WSI. Wolfers, Justin. 2003. “Is Business Cycle Volatility Costly? Evidence from Surveys of Sub- jective Well-Being.” International Finance 6(1):1–26. Woodcock, Simon D. 2015. “Match Effects.” Research in Economics 69(1):100–121. Wu, Zhongmin, Mark Baimbridge and Yu Zhu. 2009. “Multiple Job Holding in the United Kingdom: Evidence from the British Household Panel Survey.” Applied Economics 41(21):2751–2766. WZB. 2009. Erwerbsverlaufe¨ und Weiterbildungsbeteiligung von Wiedereinsteigerinnen, Forschungsprojekt fur¨ das BMFSFJ. Technical report WZB. Yashiv, Eran. 2008. “U.S. Labor Market Dynamics Revisited.” Scandinavian Journal of Eco- nomics 109(4):779–806. Acknowledgements

I would like to express my gratitude and appreciation for the support I have received in writing this dissertation. I thank my supervisor, Thomas K. Bauer, for providing academic guidance and concise feedback throughout my doctoral research. I am also grateful to my co-supervisor and co-author, Ronald Bachmann, for sharpening my academic thinking, the great amount of time he spent supervising me and his trust. My co-authors, Hanna Frings, Nikolas Mittag and Marcus Tamm, have signiﬁcantly contributed to this dissertation. My doctoral research has tremendously beneﬁted from con- structive discussions and from the productive team effort. Moreover, I am thankful to my RWI colleagues, especially to colleagues of my research group Labor Markets, Education and Population for a very supportive and encouraging work environment. Generally, the RWI provided an ideal research setting for writing the dissertation. Finally, I am indebted to my family. In particular, I am grateful to my husband, Wolf, for his support in writing this dissertation and in life. Curriculum Vitae

Rahel Felder

Professional Experience 2014 – present RWI, Researcher, Research Division ”Labor Markets, Education, Popula- tion”, Essen 2013 – 2014 ZEW, Intern and student research assistant, Research Division ”Labour Markets, Human Resources and Social Policy”, Mannheim 2012 Research assistant, Prof. Dr. Carlos Carrilo-Tudela, University of Essex 2010 RWI, Intern, Research Division ”Labor Markets, Education, Population”, Essen

Education 2014 – 2019 Doctoral student, Ruhr-University Bochum 2012 – 2014 MSc in Economics, University of Mannheim 2007 – 2012 BSc in Economics, University of Konstanz 2009 Semester abroad, Macquarie University, Sydney

Peer Reviewed Publications Bachmann, R. and R. Felder (2018), Job Stability in Europe over the Cycle. International Labour Review 157 (3): 481-516. Bachmann, R., R. Felder and M. Tamm (2017), Erwerbstatigkeit¨ und atypische Beschaftigung¨ im Lebenszyklus - Ein Kohortenvergleich fur¨ Deutschland. Perspektiven der Wirtschaftspoli- tik 18 (3): 263-285.

Discussion Papers Bachmann, R., R. Felder and M. Tamm (2018), Labour Market Participation and Atypical Employment over the Life Cycle - A Cohort Analysis for Germany. Ruhr Economic Papers #786. RUB, RWI. Bachmann, R., R. Felder, S. Schaffner and M. Tamm (2018), Some (Maybe) Unpleasant Arithmetic in Minimum Wage Evaluations - the Role of Power, Signiﬁcance and Sample Size. Ruhr Economic Papers #772. RUB, RWI. Bachmann, R. and R. Felder (2017), Labour Market Transitions, Shocks and Institutions in Turbulent Times: A Cross-Country Analysis. Ruhr Economic Papers #709. RUB, RWI.