The Labor Supply Effects of Unemployment Insurance. for Older Workers

Transcription

1 The Labor Supply Effects of Unemployment Insurance for Older Workers Matthew Gudgeon Johannes F. Schmieder Simon Trenkle Han Ye Boston University Boston University, Institute for Employment Boston University NBER, IZA, Research (IAB) and CESIfo PRELIMINARY DRAFT - PLEASE DO NOT CITE OR CIRCULATE - October 2016 Abstract Extending Unemployment Insurance (UI) benefits can affect labor supply along two margins: it can lengthen the unemployment duration of an individual who is entering UI the intensive margin and it can alter the inflows into UI the extensive margin. We study the labor supply effects of UI for older workers using German Social Security data and policy variation over 3 decades. We present evidence of extensive margin responses in the form of sharp bunching in UI inflows at various age discontinuities in UI eligibility for workers in their 50s, who may use UI as a pathway into early retirement. Using a combination of regression discontinuity designs and bunching techniques we quantify the magnitude of these responses exploiting a variety of thresholds, kinks and notches induced by the UI and retirement institutions. To provide estimates of the combined effects of UI extensions on time out of work through extensive and intensive margin responses, we use two approaches. First we employ a simple model based on two types that uses a static retirement model to model extensive margin responses. While this is very intuitive it has various shortcomings, in particular the types are unlikely to be well differentiated in practice and hard to identify. Our second approach therefore estimates a dynamic life-cycle model of labor supply where individuals face retirement and search intensity decisions. Preliminary results suggest that a 6 month UI extension for men increases non-employment durations by 1.46 months, almost twice as large as the 0.84 months coming from intensive margin responses alone. We are grateful to Kevin Lang, Daniele Paserman, Regina Riphahn, Till von Wachter, and seminar participants in Nuremberg and Boston for many helpful comments. All errors are our own. mgudgeon@bu.edu, johannes@bu.edu, Simon.Trenkle@iab.de, yehan@bu.edu

2 1 Introduction Unemployment Insurance (UI) benefits are an important policy tool to help workers smooth their consumption after job-loss. A large literature has studied the effects of UI extensions on labor supply using quasi experimental methods (see Schmieder and von Wachter, 2016, for a review). This literature has typically found that UI extensions have sizable effects on the duration of unemployment of individuals who become unemployed the intensive margin, while not having an effect on the inflow rates into unemployment the extensive margin. This can be most clearly seen in papers based on regression discontinuity designs around age or experience thresholds, where a standard validity check is to show that the density of inflows into UI does not change at the threshold (e.g. Card et al., 2007; Centeno and Novo, 2009; Schmieder et al., 2012; Lalive et al., forthcoming). However, this literature is largely based on relatively young workers in their 30s, 40s and early 50s, who are highly attached to the labor force. Older workers, in their late 50s and onwards, are much closer to retirement and may use UI as a stepping stone into retirement. This may be reinforced by firms that seek to reduce employment in response to a negative shock, by laying off or even buying out workers with relatively high outside options, thanks to the possibility of going into early retirement via an intermittent spell of UI. Understanding the labor supply behavior of older workers is particularly important given the common goal of extending the work life of the elderly and reducing the burden on the social security system. In this paper, we study the labor supply effects of UI extensions for older workers in Germany using the universe of social security data from 1975 to Numerous reforms to Germany s UI and retirement system over this period altered both the payoffs to entering UI at different age thresholds and the search incentives of the unemployed. Workers in their late 50s responded sharply to these policy changes. We see a clear increase in inflows to UI at various age thresholds where maximum UI duration eligibility increases, as well as sharp bunching at kinks in the budget set generated by the interaction of UI and retirement rules. To quantify the overall effect of UI extensions for older workers, we first develop a simple model that accounts for both intensive, that is conditional on becoming unemployed, and extensive, that is affecting the inflow into unemployment, margin responses. In the model, a share of individuals considers the option of entering UI as a pathway into retirement and chooses the age of entering UI based on the kinked, lifetime budget set induced by the UI system. We estimate the parameters of this model using a combination of non-parametric regression discontinuity designs and bunching estimators. We then use the structure of the model to back out the aggregate effect of a UI extension for older workers. We then built on this simple model by developing a dynamic labor supply model which allows us to estimate the underlying parameters without the assumption of two distinct types and estimate this 1

3 model using a minimum distance estimator (in progress). Germany provides a particularly interesting context for studying UI extensions for older workers, since there has been a tremendous amount of policy variation over the past decades. In the early 1980s, maximum potential benefit durations (PBD) were capped at 12 months regardless of age. Throughout the 1980s maximum PBD were increased dramatically for older workers at a number of age thresholds, reaching up to 32 months of UI benefits for the oldest group. Between 1999 and 2007, Germany reversed track. Maximum benefit durations were reduced for older workers and Germany began the process of eliminating early retirement at age 60 after an unemployment spell. This increase and later decrease in UI generosity is matched by a sharp increase in the unemployment rate among older workers. Previous authors, such as Buchholz et al. (2013), have attributed this to a variety of policy changes aimed to reduce labor supply of older workers, but these papers have not attempted to isolate the impact of UI. While Germany provides many compelling advantages for studying the effects of UI for workers it also offers a number of challenges. The main complication is that in addition to UI there are a large number of other policies that changed over the past decades and that may affect inflows into UI and unemployment durations. Some of these changes are about regular and early retirement rules and relatively easy to understand, but there are also many rules based on collective labor agreements (CLAs) that are on the sectoral level or even specific to individual firms. Such CLAs may themselves take policy induced age discontinuities into account, for example by encouraging workers to exit firms at those age thresholds with severance packages. In this case one can view CLAs mainly as a mechanism of how age discontinuities lead to extensive margin responses. On the other hand, CLAs may also lead to bunching at age thresholds that are not directly related to retirement or UI institutions. This somewhat complicates our setting and we consider a variety of approaches to obtain meaningful estimates in light of such confounding. 1 Our setting also raises some interesting methodological issues. While several papers have estimated regression discontinuity designs in the presence of manipulation of the forcing variable (see for example Card and Giuliano, 2014; Gerard et al., 2015; Barreca et al., 2016; Hoxby and Bulman, 2016), this manipulation has typically been treated as a nuisance, with researchers attempting to avoid bias using techniques like excluding observations close to the threshold (donut-hole regressions). However, whether and when to enter UI is itself an important outcome and in practice individuals (together with firms) can influence this decision. When UI is used as a pathway to retirement, it essentially constitutes a labor supply decision in the face of a budget set defined by wage rates, the UI system and retirement rules. The 1 In ongoing work, we consider alternate samples of industry and sub-industries, and we scale down bunching estimates by observed bunching at unexplained thresholds. 2

4 UI system create kinks in this budget set and individuals choosing to enter UI as a step towards retirement should bunch at these kink points. We can thus use bunching techniques (Saez, 2010; Kleven, 2016) to back out labor supply elasticities for these workers, based on the amount of bunching around such kinks. 2 While bunching can help recover extensive margin decisions, it complicates identification of intensive margin effects. Ideally, we would use the discrete changes in potential benefit duration at the age thresholds to estimate intensive margin responses. Yet, extensive margin responses at or around these thresholds lead to direct violation of the RD assumption that there is no manipulation of the running variable and individuals on both sides of the cutoff are therefore comparable. We use three approaches to circumvent this challenge and obtain plausible estimates of intensive margin responses: First, we use donut-hole regressions to exclude the range where most of the bunching occurs. This is most credible when the bunching is not too extreme and there does not appear to be an overall shift in the density outside of a sharp window around the threshold. Second, we estimate intensive margin responses at slightly younger age thresholds, where bunching is less of an issue, in particular a threshold at age 54 during the 1990s. Third we estimate intensive margin responses on a sample of individuals who later return to the labor market, which likely obtains, as we argue below, a lower bound of the intensive margin response for these workers. In ongoing work, we follow Gerard et al. (2015) who explicitly provide a framework to estimate bounds in RD settings in the presence of sorting. Our paper is related to a large literature on retirement decisions. Several methodolocially related papers have analyzed bunching in retirement age to derive labor supply elasticities, for example Brown (2013) who looks at bunching at the regular retirement age for teachers and Manoli and Weber (2014) who analyze permanent exits from the labor force around tenure thresholds in Austria that lead to discrete increases in severance payments. Unlike these papers we look specifically at entry into UI, rather than exits from the labor force. Our paper also adds to a smaller literature looking at the interactions of the UI system with retirement decisions. For example, Lalive (2008) analyzes the effect of UI extensions for older workers around a discontinuity at age 50 in the Austrian UI system as well as a border discontinuity and finds relatively large disincentive effects, especially for women. He also shows that women seem to respond on the extensive margin to the change in UI generosity. Using partially the same variation as Lalive, Inderbitzin et al. (2016) show that much of this was due to early retirement responses. Similarly, Kyyrae and Wilke (2007), show that increasing the age 2 Note that not all bunching around UI age discontinuities is necessarily related to early retirement. It may also be that firms postpone lay-offs or workers postpone claiming UI benefits until they reach the threshold. This is likely to be most important at ages further away from the retierment age, such as the age threshold at age 54 in the 1990s and the threshold at age 55 in the 2000s. This is something that we plan to address more explicitly. 3

5 threshold of early retirement via UI benefits from 53 to 57 reduced unemployment durations significantly. Bennmarker et al. (2013) exploit a Swedish reform which reduced UI duration from 90 to 60 weeks for unemployed aged 55 56, finding that transitions to employment increased and were concentrated around the time of required inflow into active labor market programs. Hairault et al. (2010) provides some evidence based on French survey data, that for the unemployed job search behavior depends on the distance to retirement age. 3 Several papers analyze the interaction between various retirement rules and labor supply in Germany (see Giesecke and Kind (2013); Boersch-Supan et al. (2004); Boersch-Supan and Hendrik (2011)among others). This paper proceeds in four steps. We first provide a very general decomposition of the effect of UI extensions on time out of work in the presence of intensive and extensive margin labor supply responses, which highlights the importance of imposing additional structure to fully estimate these responses. In section 3, we present the institutional background and briefly describe the core features of the German unemployment insurance and retirement institutions. In Section 4, we present reduced form evidence of both intensive and extensive margin responses. To do so, we estimate RDs at all the older age cutoffs available to us, using various approaches to handle sorting at these cutoffs. We also estimate bunching elasticities at all the notches and kinks in the budget set for older workers entering UI, as if every worker were indeed choosing their exit age strategically. This evidence provides a first glimpse at the range of magnitudes of intensive and extensive margin effects. Section 5 presents a simple model that allows us to quantify the overall effects of UI extensions for older workers. It proposes a model that splits workers into two types using information on whether or not the worker permanently exits the labor force. While this is an imperfect solution, it allows us to obtain a first estimate of the effect of UI extensions on total time out of work. Finally, Section 6 develops a dynamic model that nests the simple two type model, allowing for more credible estimates of the overall effects of UI extensions. We estimate this structural model using the moments in the data. 2 The Effect of UI Extensions on Total Time out of Work We develop a general framework to describe how potential UI benefit duration affects time out of work in the presence of extensive margin responses. There is a mass of workers N, who enter the workforce at age 1 and reach a mandatory retirement age at T R. P =(P 1, P 2,..., P T R) is the vector of potential UI durations, which may depend on age. In each period t (meaning 3 While our focus is to quantify the overall effect of UI extensions on labor supply of older workers, rather than discussing optimal policy, our analysis can be viewed as an important input into welfare computations. For papers on the optimal design of UI for older workers, see for example: Hairault et al. (2012), Michelacci and Ruffo (2015), and Inderbitzin et al. (2016). 4

6 at each age), a worker is either working or not working. We do not distinguish between unemployment and nonemployment and hence use the terms interchangeably. The fraction of workers entering unemployment at age t is denoted as g t (P). If an individual becomes unemployed, the duration of nonemployment is defined as the time between entering unemployment and either starting a job again or when the individual retires. We denote the expected nonemployment duration of individuals becoming unemployed at age t as D t (P). The expected total time out of work (T u ) for an individual is given by T u (P) = T R t=1 g t(p)d t (P). Without specifying micro-foundations of this labor market, we can write the relationship between inflows and durations on benefit durations as reduced form functions and decompose the effects of an increase in potential benefit into intensive and extensive margin components. For simplicity, we focus on the case where P is constant for all exit ages (P = (P, P,..P ) ). A change in P can thus be decomposed into: dt u = T R t=1 g t D t P }{{} Intensive Margin + T R t=1 D t g t P }{{} Extensive Margin (1) The first term represents the standard intensive margin effect of UI extensions on nonemployment durations that most of the UI literature has estimated. The second term represents the changes in inflows into unemployment. The central question of this paper is how to credibly estimate this total effect for older workers. To see why this is difficult using only reduced form techniques it is instructive to consider two benchmark cases. For the first case, take the younger workers that have been the focus of much of the UI literature. Schmieder et al. (2012) show that younger workers do not significantly alter their entry probabilities into UI in response to changes in P. In this case gt P = 0 and we can focus on estimating the intensive margin effect, which can be credibly estimated using RD designs at age cutoffs in P as in Schmieder et al. (2012). Note that even in this case there is a challenge to cleanly estimating dt u, as any increase in P above some age threshold would change the pool of people at risk of becoming unemployed. So if P increases D, employment falls, and hence future g t might decrease, violating gt P can get a reasonable first order estimate of dt u response to changes in P using estimates from the prior literature. 4 = 0. Nevertheless, we in the case of no strategic entry into UI in Now consider a case at the other extreme. Suppose older workers only use UI as a bridge to retirement, and never become unemployed except by their own choice. Once they exit 4 Note also that if one is interested in more complicated changes in potential benefit durations (that is not just an increase at a single age level k), then it is still relatively straightforward to estimate the intensive margin effect by aggregating estimates of D k P k at different age levels. 5

7 they stay non-employed until pension, so at each age, non-employment duration is fixed and D t P = 0. These workers time their exit date optimally to maximize lifetime utility over consumption and leisure subject to their budget constraints. Depending on the institutional setting, transitioning to retirement via unemployment may be appealing, as we will show is the case for Germany where there is a kink in the budget set P years before the retirement age. Extending P moves this kink, and hence moves UI exit mass (reducing it at the prior kink point and shifting to the new kink). If this were the correct model, we could calibrate its key parameters using bunching techniques in a manner similar to Brown (2013) and then simulate dt u. In practice, neither of these cases fully captures the complexity of reality. Older workers are likely to still have strong intensive margin responses to changes in P, and some older workers, even at later ages, will find themselves unemployed not by their own choice. Below, we tackle this complexity, first with a simple two-type model (Section 5, and later with a dynamic life-cycle model that nests the two type model. 3 Institutional Background and Data 3.1 Unemployment insurance The German unemployment insurance system provides income replacement to eligible workers who lose their job. Prior to 1985, eligible workers were entitled to at most 12 months of potential benefit durations (PBD). Beginning in 1985, numerous reforms changed PBDs in a manner that tied the maximum PBD to recipients exact age at the beginning of their UI spell. 5 Maximum PBDs were increased for workers above age 42 via several reforms between 1985 and The most generous PBD was available after the 1987 reform for workers older than age 54 who were eligible for up to 32 months of UI benefits. Reforms in 1999 and 2006 gradually decreased the generosity of the system. In 1999, age thresholds were increased, and then, beginning 2006, maximum PBD was reduced from 32 to 18 months for workers above age 55, while everyone else could only receive 12 months. There was a modest reversal of this trend in 2008 when workers above 58 could attain PBD of 24 months. Figure 1 shows the evolution of maximum PBD for older workers by age of UI entry over this time period. We use the same 6 distinct periods below in the empirical part. We omit the 1985 rules, both because this is the shortest period and since it appears that some individuals who entered UI in 1985 retroactively benefited from the UI extensions in later years. 6 5 See Hunt (1995) and Fitzenberger and Wilke (2010) for an analysis and discussion of these reforms. 6 Also note that the 1999 reform was enacted in 1997 but individuals in our sample with high labor force 6

8 Replacement rates for UI were relatively stable over the period (67-68% for an individual with children and 63-60% for an individual without children). Individuals who exhausted UI benefits prior to 2005 and whose net liquid wealth fell below a certain threshold were eligible for unemployment assistance (UA) benefits with an effective average replacement rate of around 30%. 7 From 2005 on, UA was replaced by unemployment insurance benefits 2 (UIB 2), a completely means tested program. Both UA and UIB 2 are unlimited in duration. Appendix Table A.1 documents all the institutional changes in benefit durations, replacement rates and age thresholds that have occurred since These changes provide highly useful empirical variation both at the age thresholds and by changing incentives on when to enter unemployment if using unemployment as a bridge to retirement, as we elaborate on in the next section. 3.2 Pension system and early retirement via unemployment Germany has a generous pay-as-you-go public pension insurance with high effective replacement rates. Participation is mandatory, with the exception of civil servants and the selfemployed. The pension system is financed with contribution payments, which are shared equally by employers and employees. Pension benefits depend on a workers earnings, years of contribution, an adjustment factor, and the type of pension claimed. The benefits are roughly proportional to lifetime income at an average replacement rate of 50% (Deutsche Rentenversicherung, 2016). The statutory retirement age (SRA) for a regular old age pension remained at 65 throughout our sample period, with the only prerequisite being 5 years of contributions. Several alternate pathways make retiring before 65 an option. The five main pathways to retirement are regular old-age pensions, old-age pensions for long-term insured, old-age pensions for women, old-age pensions due to unemployment (and, later, part-time work) and old-age pensions for severely disabled persons, see for example Boersch-Supan and Wilke (2005). Appendix Table A.2 documents the earliest possible retirement age for each of these pathways over the past 4 decades. We focus primarily on the pathway into retirement via unemployment. 8 The unemployment pathway (UI pathway) provided eligible workers an option to retire at the age of 60. The eligibility requirements for this pathway are 1) at least 15 years of contributions at least 8 of which must have occurred in the past 10 years and 2) being unattachment were eligible to the more generous durations until In principle, replacement rates were between 50% and 57% but lower in practice due to deductions like spousal income. See Schmieder et al. (2012) for a discussion of this. 8 While early retirement due to disability is also a quantitative important, Riphahn (1997) argues that in practice they are not close substitutes and that retirement due to disability is in fact usually associated with a health shock. 7

9 employed for at least 1 year after the age of 58 years and 6 months. Both the generosity of UI benefits and the lenient job search requirement for older workers make old-age pensions due to unemployment attractive. Workers 58 and older could receive unemployment benefits without actively looking for a job or other obligations. 9 In the first 3 periods of our sample, the unpenalized/normal retirement age (NRA) as well as the earliest possible retirement age (ERA) via the UI pathway was age 60. This means persons satisfying the unemployment requirements could retire at 60 with no penalty other than the loss of additional years of pension contributions. The ERA via the UI pathway stayed at 60 during the first four periods of our sample (see Appendix Table A.2). Beginning with cohorts born after 1945, this age increased in monthly steps from 60 to 63, ending with cohorts born in The NRA via the UI pathway stayed at 60 until 1997 (almost until the end of period 3). Beginning with cohorts born in January 1937 (turning 60 in 1997), the SRA increased in monthly steps up to 65, ending with cohorts born December If these people chose to retire early at the ERA (60 up until 2006), they faced a 0.3% pension reduction per each month they retired in advance of the NRA. For cohorts born after 1952 (after our sample) this pathway into retirement was entirely abolished. The possibility of using UI as a bridge to retirement introduces a kink in a lifetime budget constraint relating lifetime income to year of exit into UI. 10 Individuals retiring before 60-P are forced to spend time reliant on a spouse or on UA/UIB 2 before their pension, whereas individuals who leave at or after 60-P can take the full UI duration and transfer directly into pensions. This reduces the value of an extra year of work after the kink, decreasing the slope of the budget constraint. In general, the size of the kink is exacerbated by the generosity of the UI system, the size of the drop comparing UI to UA/UIB 2, and how generously time on UI is counted towards pension contributions 11 Figure 2 plots the evolution of stylized lifetime budget constraints over the periods. The jumps (notches) in the budget constraint are generated by the discontinuous increases in PBD that result if one becomes unemployed right above the age threshold. Note that if a lifetime labor supply model is the correct model for individuals, they would also have incentives to bunch at these notch points in their transition to retirement. See the Appendix for a detailed descriptions of how these budget sets are constructed. 9 This so-called 58er-Regelung was introduced end of 1985 and in place until end of Note that a worker may be sanctioned if he or she quits a job voluntarily. These sanctions take the form of losing the first few weeks of benefits and vary from a 4-12 week penalty over the study period. These sanctions, which are not always applied, are insufficient to offset the appeal of using UI as a pathway into retirement. For now we ignore the sanctions, but given that they make the UI pathway slightly less attractive this whould lead to a downward bias in estimated labor supply elasticities. 11 In practice, unemployment counts as an 80% contribution year calculated on pre-unemployment wages. 8

10 Throughout the rest of this paper, we focus on the notches and the first kink induced by using UI as a bridge to retirement. In practice, agents might also use UI as a bridge to the long-term contribution retirement age of 63 or the regular retirement age of Since we cannot credibly calculate whether or not a person is eligible for the long-term contribution rate, examining bunching at these kinks is problematic. In further work we aim to examine these in additional depth. Note also, that changes in other pathways may create alternative substitutes for workers aiming to retire early. Appendix Table A.3 summarizes the reforms for all of the different pathways over our study period. Our sample consists only of men, so the pathway for women is not relevant. The only other possible confounder at these early exits could be changes disability pension and old-age pension for disabled workers. 3.3 Data We use rich administrative data from the Integrated Employment Biographies (IEB) in Germany. This data contains information on all social security reliable employment periods and periods of UI receipt between the years 1975 and The employment information covers daily information on approximately 80% of the regular workforce, with the self-employed and civil servants being the most common exceptions (see Web Appendix for further details). The data on UI receipt stems from administrative UI records and entails information on the exact duration of UI-receipt and the amount of daily benefits. For now we focus on men only, since early retirement rules differ slightly between men and women. We plan to also do the full analysis for women in the near future. We select all UI-entries between 1980 and 2010, which qualify based on their working history for their age-specific maximum PBD. This leaves a five year window before the first period and a three year window after the last, allowing us to calculate UI eligibility for all individuals and unemployment durations for up to three years after UI entry. For the selected individuals, we construct detailed biographical information such as experience tenure or past exposure to unemployment. Since some of the requirements for maximum PBD eligibility such as the duration over which claims could be accumulated changed over the study period, the restrictions set on this duration differs slightly between periods. We summarize these restrictions in Appendix Table A.4. Additionally, we exclude mining and steel construction from our analysis, since both sectors are known to have specific and strong early-retirement rules for at least some of the periods. For other specific subgroups which face some, but less clear or pronounced early retirement rules we do not exclude cases a priori, but address them throughout the analysis. 12 Individuals cannot receive UI past 65, and cannot receive UI and pensions simultaneously. 9

11 4 Reduced Form Evidence This section displays and analyzes the behavior of older individuals entering UI over three decades. To get a reduced form sense of the magnitude of intensive and extensive margin responses to UI extensions, we estimate regression discontinuities at each of the age cutoffs for PBD increases and we estimate bunching responses at each of the retirement-via-ui kinks (which change as P is increased). 4.1 Graphical Evidence Figure 3 shows the number of individuals entering UI benefits by age for different periods in our sample. The periods correspond to thedifferent UI regimes depicted in Figure 1. Red vertical lines correspond to age thresholds in PBD and blue vertical lines correspond to the kink in the lifetime budget set due to the bridge to retirement via UI (as depicted in Figure 2). As stated in 3.2, early retirement via UI was possible without penalties at age 60 in the first 3 periods, and with cohort-varying penalties in period 4. In 1980 to 1984, all individuals are eligible to 12 months of UI in this period, regardless of age. However there is a kink in the budget set (see Figure 2 (a)) at age 59, since individuals who enter UI at 59 can bridge the entire 12 months of unemployment with UI benefits and then enter retirement via UI at 60. Thus by entering UI after age 59, individuals are covered by UI or pension benefits after unemployment without any gaps of no income, while individuals who enter prior to age 59 will spend some time without either of them and would have to cover that period out of their savings, from their spouses or by receiving other transfers (such as UA). Figure 3 a) shows the inflows into UI in this period. There is very sharp spike in inflows just to the right of age 59 and an increased density for several months afterwards. This is highly consistent with individuals choosing entry into UI strategically at age 59 as a pathway into retirement. Throughout the different periods, the changes in PBD move the kink in the budget set coming from using UI as a pathway into retirement without intermittent uncovered periods. The panels in Figure 3 generally show a spike in the UI inflows at each of the kink points: 59 in , 58 in 1986-June 1987, at 57 and 4 months from There continues to be some bunching at 57 and 4 months in , but consistent with the new cohortspecific penalties on retiring at age 60, overall bunching is reduced (and spread out) over the period. Once retirement at 60 (even with penalties) is eliminated in 2006, we no longer see this bunching. Figures 4 and 5 complement Figure 3 by showing mean UI benefit duration as well as mean non-employment duration by exit age in each period. We cap non-employment durations at 10

12 36 months and set all non-employment durations above that at 36 months, including censored spells where we do not observe individuals returning to employment. Overall, these outcomes consistently match the PBD reforms (with the exception of some early applications of the rules in 1984). For example Figure 4a) supports that individuals in the first period use UI as a pathway to retirement, since individuals who enter UI right at age 59 use almost all of their UI eligibility (of 12 months). 13 Similarly, in Figure 5a) it is apparent that past age 59 unemployment durations are very long (the average is almost at the cap of 36 months) and in practice almost no one who enters unemployment at that age returns to work. Moreover, we can see very clearly in the first 3 periods that people exiting at the kink point for early retirement via UI take the maximum PBD. The amount of UI taken then reduces with age up to 1 year, the minimum required for entry into pension via UI. The duration of non-employment after these kinks is also extremely close to the cap. This paints a consistent picture: the bunching observed at and after the kink in the first 3 periods is due to people using UI as a bridge to retirement. Beginning in period 4, when the cohort-specific penalties on entering pension via UI at age 60 were enacted, we see that UI after the kink remains much flatter and above 1 year and a half on average, suggesting that people are responding to the penalties and retiring at their new NRA. Nonemployment durations also decrease as more people, lacking the very early retirement option, transition back to work. Finally, note that UI durations drops to 0 at age 65 as UI cannot be taken past the SRA of 65. We also see some decreases in UI approaching the age 63 retirement, which is only an option for a subset of our sample. In addition to bunching at the kink point for early retirement via UI, there is clear bunching at many of the age thresholds at which PBD for UI is extended. This can be very clearly seen in Figure 3 at the age 55 threshold in periods 5 and 6, but also at the age 57 threshold from April 1999 to January 2006, and the age 54 threshold from July 1987 to March Interestingly, there is a positive relationship that the higher the age of a UI threshold, the more bunching there is. While, younger age thresholds, such as the age 49 threshold in or the 52 threshold from 1999 to 2006 show almost no bunching. This makes sense if bunching is mainly due to individuals entering UI as a pathway into early retirement, which is simply not an attractive option (no matter the UI durations) for individuals who are too far away from the retirement age. Figure 3 also shows that there is clear bunching at various other threshold. For example, the first period shows some bunching exactly at age 55, 57, 58, 60, and 62. The small amount of bunching at 62 can be explained by a similar kink into retirement at age 63 for the subset of the population with very long term contributions. However, the bunchings at 55, 57 and 13 Note also that entries after age 60 still take 1 year of UI (see Figure 4 (a)), consistent with the requirement that entries into the pension via unemployment scheme necessitate 1 year of UI. 11

13 58 and 60 are not driven by kinks or notches in our budget set. While some of this could be round number bunching or bunching at reference points, much of this is driven by specific collective labor agreements at the firm or sectoral level that specified retirement packages and ages. Indeed, this type of bunching is almost entirely absent in the years leading up to and including 1982 (see Figure 8 (a) for the period), consistent with the timing of the first major CLAs specifying retirement ages (see Trampusch et al. (2010)). Our sample drops the sectors mining and steel which have clearly defined CLAs, but inevitably picks up other sectors and firms with CLAs. For example in period 3, there is sharp bunching at exactly age 55 and 56, as well as a shift upwards in the density after age 55. During this period many firms reduced employment through CLAs that bought out older workers. Age 55, and to a lesser extent age 56, was a common cutoff used in these CLAs. The importance of these CLAs fades throughout the late 90s and early 2000s. In robustness exercises, we consider alternate samples and ways to address any confounding, but we note here that the bunching at explained kinks generally dwarfs bunching at these alternate thresholds. The following two sections present reduced form estimates on the effect of PBD extensions. First, we estimate regression discontinuities at all the PBD age thresholds, that correspond to the standard intensive margin effects of extending P on nonemployment duration for older workers in Equation 1. We address the challenges faced by sorting at some of these thresholds. Second, we estimate the labor supply elasticity in a standard lifecycle model using bunching at the kink induced by the UI pathway into retirement. Since this kink moves with PBD, it is one way to quantify the magnitude of an extensive margin effect i.e. workers timing their entry into UI strategically as a function of P. The amount of bunching (as opposed to the elasticity) can itself be viewed as a reduced form parameter. 4.2 Estimating Intensive Margin Responses using Regression Discontinuity Estimators This section presents estimates of the effect of PBD on nonemployment duration using a regression discontinuity design (RDD). To this end, we use discontinuous changes in PBD at the various cutoffs for older worker. In particular, we use all cutoffs at ages 49 or larger since 1987, resulting in 8 different cutoffs. 14 We estimate variants of the following specification: y i = δ 1(a i A) P BD + f(a i ) + X i β + ε i (2) 14 We omit the period due to it being a transition period. Figure 4 shows clearly that there is no first stage in this period. 12

14 y i is the outcome variable for individual i (typically non-employment duration), a i is the age at time of UI entry (measured on the daily level) and 1(a i A) is an indicator function which equals one when individuals age is above the cutoff A where benefits are extended discontinuously by P BD months. In this specification, δ measures the effect of a one month increase in PBD. The function f(a i ) is set to be a linear function with different slopes on each side of the cutoff in the baseline specification. X i is a vector of additional controls. We use a local polynomial regression with rectangular kernel and cluster standard errors on the daily level. We set the bandwidth to two years, but restrict it for the 49 and 54 years cutoff to one year on the right side, due to other policy discontinuities that are present at 50 and 55 in the earlier periods. One general concern in using a RDD at these older ages in our setting is the presence of sorting/bunching at the cutoffs. Figure 3 shows that the number of UI entries is not smooth at most cutoffs where UI is extended. The degree of sorting seems to vary between cutoffs and is usually most pronounced within the first 1 to 2 months around the cutoff. We first apply an impartial solution to address this concern by excluding 2 months on each side of the cutoff in all our regressions. Second, we add a specification with detailed individual controls such as education, tenure and other pre-unemployment characteristics (see regression table for details) that addresses this selection concern directly. Third, in ongoing work, we apply a method due to Gerard et al. (2015), that allows explicitly to apply RDDs in situations where the running variable is manipulated. 15 Column 1 and 2 of Table 1 show the estimates of a one month increase of PBD on nonemployment duration (capped at 36 months). There is a strong and significant effect at most cutoffs, varying between 0.11 and 0.47 in the specification without controls. Picking the age 54 specification with controls, we note that, on average, older workers experience an additional 0.14 months of unemployment for each addition month of potential unemployment duration. Note that this estimate is similar to the estimates in Schmieder et al. (2012) at younger age cut-offs. However, unlike in that paper we restrict our sample to men only who tend to be less responsive to UI and also have slightly different sample restrictions (e.g. omitting some industries). When we replicate the estimates at the younger cutoffs as in Schmieder et al. (2012) for our sample, we find significantly smaller disincentive effects of UI. This suggests that the disincentive effect of extending UI is higher for older workers. Moreover, even in our main sample, the effects of PBD on nonemployment duration increase with age. For the most recent period for example, the point estimate increases from 0.14 at age 50 up to 0.3 for the oldest cutoff at age 58. While the inclusion with controls lowers 15 Columns 3 and 4 of Table 1 estimate the same RDs on the restricted sample of people who re-enter the labor market at some later date. This sample exhibits less sorting at the discontinuities (as can be seen in Figure 6) and thus can be viewed as a reasonable lower bound on the true effects. 13

15 the point estimates in all specifications, this pattern is robust. Of course, sorting is also more severe at the higher age cutoffs, and to the extent that this is inadequately addressed that could be in part driving the pattern. While these patterns alone are of interest, the policy relevance of these results is compounded by the presence of extensive margin effects. 4.3 Estimating Extensive Margin Responses using Bunching Estimators This section estimates bunching responses at the retirement-via-ui kinks. The amount of bunching can be viewed as a reduced form parameter, but converting this into an elasticity requires a model. To this end, we assume all older workers entering UI can be modeled using a standard life time labor supply model, as in Brown (2013). That is workers choose their entry date strategically to maximize utility over lifetime consumption and leisure subject to a lifetime budget set. As long as workers abilities are drawn from a continuous distribution, the distribution of UI entries will be smooth. The introduction of a kink in the budget set results in bunching at the kink point, and the amount of bunching allows estimation of a labor supply elasticity, using by now standard techniques (See e.g. Saez (2010); Kleven (2016)). We refer the interested reader to the appendix for details on the model and for specifics on how we construct the budget set for the average individual in our data. Table 2 presents the elasticity estimation using bunching at the retirement-via-ui kink in the first 4 periods in the data. Note that beginning in 1997 (near the end of period 3) the NRA for exit via UI was already being gradually increased in a cohort-specific fashion. We fit a 7th degree polynomial to the data excluding the bunching bins, which are determined visually. This poses little challenge in the first two periods, the third already demonstrates some more diffuse bunching, not just at 57 and 4 months but also at 58 and 59. We include all of this in the bunching region, given that Figure 4 indicates these workers are taking exactly the right amount of UI to bring them to early retirement. Doing so produces stable elasticities in the first three periods of These elasticities are relatively similar to those in Brown (2013), whose preferred estimate is She examines the retirement behavior of Californian teachers whose normal retirement age is 60 and whose average pension replacement rate is 59%. It is clear that people are willing to adjust their entry into unemployment in response to both the duration of P and changes in financial incentives. The estimated elasticity is much lower in the 4th period, as expected given the gradual dissolution of the kink at 60 resulting from changes in the NRA. 16 Standard error calculations using bootstrapping are in progress. 14

16 The very same lifetime labor supply model would predict bunching at the notches in the budget sets depicted in Figure 2. Using similar bunching techniques, developed in this case by Kleven and Waseem (2013), we can estimate labor supply elasticity at the notches. We fit a 7th degree polynomial to the (upward) notch excluding the bunching mass to the right of the notch visually, and gradually exclude mass to the left of the notch until the missing mass (distance between the counterfactual density and the data on the left of the notch) equals the bunching mass. We do this at all the possible notches. Elasticity results are in Table 3. We note immediately that the estimated elasticities are much lower and not consistent with the estimates. The estimated elasticities are largest in the last two periods. Some of the bunching at age 50 in is possibly confounded with CLAs and firm specific retirement practices. In the last period the largest estimated elasticity is at age 58. Put another way, we would severely over-predict bunching at these notches if we applied the model estimated at the retirement-via-ui kinks to the notches. In part, and especially at younger ages, this is likely a result of the lifetime labor supply model being unsuitable to exits at these early ages. The last two panels of Figure 5 shows clearly that even at age 58, there are a large share of workers returning to the labor force (mean duration of non employment is still relatively far from its cap of 36). Overall, Tables 1 3 support the idea that UI extensions can have relevant intensive and extensive margin effects: both parts of Equation 1 matter for older workers. Nevertheless, it is not obvious how to combine these reduced form estimates into a credible estimate of dt u. This is the concern of the rest of the paper. In Section 5 we approach the problem by making the strong assumption that there are two types of agents ones that never exit voluntarily before the SRA and others that never get unemployed randomly and exit into UI according to our lifetime labor supply model. Section 6 develops a dynamic model that nests the preceding two type model but allows for more complex and realistic behavior. 5 Estimating the Nonemployment Effect of UI Extensions within the Two Type Framework. 5.1 A Two Type Model To provide a point estimate of the overall nonemployment effect of UI extensions, we now assume there are two types of people: retirement types (r) and lay-off types (l). Lay-off types retire only at the mandatory retirement age, but have an exogenous probability of losing their job at any age t, after which they search for a job according to a standard searching process. In contrast, retirement types never become unemployed exogenously. Retirement types choose their retirement age to maximize utility, taking a static lifetime budget set (determined in part 15

17 by P ) as given. Formally, there are N i individuals of each type i {r, l}. θ i = N i N is the share of individuals of type i. g i t(p )N i is now the number of individuals of type i entering UI at time t. Additionally, let the share of i-types entering UI at time t be given by: ω i t(p ) = gt(p i )N i j r,l gj t (P )N = gi t(p )N i j g t (P )N = gi t(p ) g t (P ) θ i The effect of a change in potential benefit durations on total time out of work is now given by: dt u T R = θ l gt l ddt l T R + t=1 T R +θ r t=1 g r t t=1 ddt r T R + Dt l dgt l t=1 D r t P dg r t We assume that one can reasonably approximate dgl t = 0. That is ltype inflows are unaffected by P. Furthermore, we will model r-types using the simple life-cycle model of retirement discussed in Section 4.3. Since retirees always have a nonemployment duration equal to T R t, this does not respond to changes in P, implying ddr t simplifies to: dt u = θ l T R t=1 g l t dd l t T R + θ r t=1 D r t dg r t = 0. Thus, our problem (3) This decomposition is of little use if we cannot credibly distinguish types. We propose one way to do so, that while flawed, serves as an instructive precursor to a fuller dynamic model. We will proxy for r-types using people who permanently exit employment and for l-types as persons who return to employment after a non-employment spell. Using this decomposition of types allows us to directly observe θ l, θ r, g l t, g r t, D l t and D r t. We for various k on the sample of non-permanent exiters using RDs at age cutoffs estimate ddl t k in potential benefit durations. We estimate the full set of changes in dgr t (2013) model to the data and simulating changes in P. by fitting a Brown To this end, we divide UI entries by permanent and non-permanent exiters in Figures 6 and 7. Somewhat reassuringly, there is far lass bunching in the non-permanent exit sample, particularly in the first two panels between While we see some indications of extensive margin activity in later periods, it is never as stark as in the permanent exit samples. 16

18 While we heavily caveat this two type approach, we find the exercise beneficial in that it places an intuitive and simple framework on the data and allows us to obtain a baseline calibration of the total effect of extending UI durations (P) that can later be compared with the predictions of a model that nests this two type specification and allows for more nuanced behavior. 5.2 Estimating dgr t and ddl t This section calculates the key components of the 2 type model by calculating ddl t and dgr t on the non-permanent and permanent exit samples respectively. We estimate ddl t using regression discontinuities at the age thresholds depicted in Figure 1 on the non-permanent exit sample. Table 1 columns 3 and 4 do this for all possible age thresholds on the non-permanent exit sample. Since the regression discontinuities only provide an estimate of the intensive margin response at a given age k, we need to extrapolate estimates across ages or across space to get the effects of extending P on all persons aged For now, we assume the estimates do not differ greatly across ages or time, and use only the estimate at age 54 in column (3) of We simulate dgr t on the sample of permanent exiters using our calibrated lifetime labor supply model, based on bunching estimates. Table 2 Column (2) presents elasticity estimations on the permanent exit sample in the first four periods. Not surprisingly, these are higher than the elasticities estimated on the full sample, ranging from Figure 8 (b) shows an example of this estimation for a subsample using only data (when P was 1). 17 The change in slope at the retirement-via-ui kink given by our budget sets is The figure shows the normalized bunching mass (using a 1 year bandwidth) is large at and translates into an elasticity of We then transform the counterfactual distribution into abilities via the first order condition of the model to give us all the necessary pieces to simulate the life time labor supply model and see how workers respond to changes in UI. To illustrate the usefulness of this model, we consider a policy simulation that takes the institutions as given in and extends P from 1 to 1.5 years. We calculate dt u persons aged Policy Simulation We calculate the effect of increasing P from 1 year to 1.5 years using Equation 3. For ddl t we use the RD results from the earliest (and longest) possible period in Table 1. For this period for the non-permanent exit sample, we estimate the effect of the 17 This sample serves to avoid bunching at other ages due to CLAs that started in for 17