How does a higher Full Retirement Age affect Careers?

Transcription

1 How does a higher Full Retirement Age affect Careers? Evidence from an Increase in the Female Full Retirement Age in Belgium Bart Cockx 1,2,3,4, Muriel Dejemeppe 2, Corinna Ghirelli 1,2,5, Yannick Thuy 1, and Bruno Van der Linden 2,3,4,6 1 SHERPPA - Ghent University (Ghent, Belgium) 2 IRES - Université Catholique de Louvain (Louvain-La-Neuve, Belgium) 3 IZA (Bonn, Germany) 4 CESIfo (Munich, Germany) 5 European Commission, DG JRC 6 FNRS (Belgium) February 2016 Preliminary version, please do not quote Abstract We study the impact of an increase in the female full retirement age in a country with an abundance of early retirement possibilities. The reform was introduced in 2006 and only for specific birth cohorts. We use the resulting quasi-experimental setting to implement a Regression Discontinuity design based on womens birth dates. Our data is drawn from administrative registers of the diverse Social Security institutions and of the National Register containing all Belgian inhabitants. Our preliminary results suggest that while affected women did take up their pension later, they did not leave employment later than unaffected women. One possible explanation for this is that instead of working longer, affected women enter into early retirement schemes at the usual age and simply remain there longer, as the reform was only partial and did not involve the early retirement schemes. Moreover, while the policy was announced ten years before being implemented, we do not find evidence for an anticipation of the policy. Further analysis is being done to more closely determine what treated women do in the extra year before full retirement.

2 1 Introduction Many developed countries face enormous pressure on their Social Security provisions due to ageing populations. An important contributing factor to this problem is the low labour force participation of older workers, especially women, and the corresponding low effective retirement ages. In the EU-15, the average effective retirement age over the period was 62.6 for men and 61.7 for women, while most countries have their normal full retirement age set at 65 (OECD, 2015). In Belgium, the country analysed in this paper, the difference between the full retirement age and the average effective retirement age is even bigger. Since the 1980s the average effective retirement age has not been above 60, and was in for men and 59.1 for women (OECD, 2015). Part of the explanation for this can be found in the abundance of early retirement options available to older workers in Belgium. One strategy that has been introduced by governments to cope with this challenge posed by population ageing has been to prolong the working career by means of an increase in the official retirement age. The idea behind this is that if people have to work more years before obtaining full pension rights, they would do so in order to keep the same pension allowances (OECD, 2006). As workers in most countries do not have many options to retire early, this is a strong incentive. However, as Belgian workers do have these options, the efficiency of an increase in just the normal full retirement age might not be sufficient. In this paper we focus on an increase in the full retirement age (FRA) for Belgian women in the private sector or in self-employment. This reform, starting on January 1, 2006, brought the female FRA from 63 to 64 and the necessary full career from 43 to 44. This increase was part of a bigger pension reform between 1997 and 2009, which brought the female full retirement from 60 to 65 with increases every two to three years. This reform aimed both at increasing the average effective retirement age of women and at removing gender discrimination in the Belgian pension system for private sector and self-employed workers, as the FRA for men in these sectors was already at Note that this reform did not alter any of the early retirement schemes. A growing literature conducts ex post evaluations of pension system reforms on the labour supply decision of older workers. We identify two strands of literature, one type of literature studies pension system reforms in a quasi-experimental setting, using an exogenous change in the eligibility conditions to (early) retirement to identify the effects of the pension treatment. The other type of studies studies the effects by means of structural models and simulation techniques. Our study belongs to the first group, as we exploit a quasi-experimental setting due to the exogenous increase in the retirement age of Belgian women from 63 to 64 in The results found in the literature are heterogeneous. However, empirical effects on a similar Swiss reform studied in Lalive and Staubli (2014) and Hanel and Riphahn (2012) suggests that there is a positive impact of the reform on 1 For public sector workers there were no changes as the female FRA was already at 65. 2

3 labour market participation. Both studies find that a delayed labour market exit. The specific institutional context of Belgium, with its many early retirement schemes, might lead to different results, although the policy reform itself is very similar as for example the Swiss one. Existing non-experimental studies for Belgium, such as Lefebvre and Orsini (2011) and Jousten, Sigismondi, et al. (2012), suggest that partial reforms that only modify the access to some retirement routes have limited effects. We exploit very rich administrative data on labour market histories (for private sector employment from as early as 1957), sick leave, exit destinations out of employment, and some essential firm and household characteristics. Based on these data we can estimate the impact on the employment rate, (early) retirement rate and the age of exit from employment. The pension reform creates a sharp discontinuity in the pension treatment: women born before December 1, 1942 are not faced with the exogenous change in the eligibility conditions to full retirement as they reach the old retirement age before In contrast, women born afterwards are affected and have to wait an additional year before they can claim their full pension. Similar to the study of Lalive and Staubli (2014) we identify the treatment effect with a regression discontinuity design. 2 The treatment effects in this framework are obtained by comparing the outcomes of women born just before and after the birth date threshold of December 1, We argue that this is a credible identification strategy leading to unbiased estimates of the treatment effect as the identification assumptions seem to be satisfied: we do not observe any manipulation of the assignment variable (the birth date), the women in our sample have similar observable characteristics and there are no other relevant policy changes around the cut-off. Any difference in outcome can thus be attributed to policy. We contribute to the literature in the following ways: first, to the best of our knowledge, we are the first to provide evidence on the effects of an increase in the legal retirement age for a country with a wide range of early retirement options. Previous studies focused on countries where workers have the incentive to remain in the labour market until the full retirement age. As these workers have few options to retire early, their employment rate is quite high. Instead, we focus on Belgium, where workers can benefit from a generous early retirement system. This results in a low employment rate at older ages and a very low, especially female, retirement age. Given these differences, we expect the impact of the reform to be different and, in fact, smaller in Belgium than in the countries mentioned above because the incentives created by the policy change are counteracted by the availability of the early exit options. Second, we are the first to study the increased female full retirement age for Belgium using a counterfactual approach. Third, there is a relatively long period between the moment that the policy was first announced in 1996 and the moment it was implemented in As a result, affected women had time to anticipate this change and adjust their labour supply if necessary. Moreover, the policy was announced while 2 Montizaan et al. (2010) and de Grip, Lindeboom, et al. (2012) used the same approach to evaluate the impact of increasing the retirement age in the Netherlands. 3

4 most affected women could not yet enter in the early retirement schemes. Finally, we intend to combine the wild bootstrap method to deal with a low number of clusters with our endogenous sampling weights. Our sample consists of women born one year before and after the birth date threshold. Since we group our sample by birth month, inference hinges on 24 months (clusters). With this low number of clusters we risk to underestimate the standard errors as empiricists tend to agree that at least 50 clusters are needed when clusters have roughly the same size, and that a higher number of clusters is required when clusters are unbalanced (Cameron, Gelbach, et al., 2008; MacKinnon and Webb, 2014). This problem can be tacked with wild bootstrap, which allows for intra-cluster correlation and heteroskedasticity (Cameron, Gelbach, et al., 2008). However, given our endogenous sampling procedure we need to adjust the procedure for this. We are currently still researching whether this adjustment is possible and feasible. Our findings can be summarized in two broad findings. First, it appears that, in general, treated women simply substitute the period in full retirement they could take before the policy change with a period of early retirement after the policy change. In other words, they do not remain employed longer. Second, as a whole, the population of treated Belgian women did not anticipate the policy change. We do not find any effects on employment or early retirement before the policy change of January This lack of forward looking behaviour leaves less room for the policy to work once it takes place. For example, it will be harder to reach women who have left employment before the policy change as they are perhaps less inclined to return to the labour market, as they would need to search for a new job. The paper is structured as follows. We start with a literature review in Section 2. In Section 3 we describe the institutional context, while in Section 4 we describe the sampling scheme and the data we used. Section 5 outlines the empirical strategy (identification and estimation). In Section 6 we present our empirical findings and in the final section we summarise the results and conclude. 2 Literature Overview There is a large literature of the effects of pension system reforms on the labour supply decision of older workers. We first discuss two strands of international literature on pension reforms. One type of literature studies pension system reforms in a quasi-experimental setting, exploiting exogenous changes in the eligibility conditions to (early) retirement. The other type of studies has attempted to quantify the effects by means of structural models and simulation techniques. Our study belongs to the first strand, as we exploit a quasi-experimental setting due to the exogenous increase in the retirement age. Next to the international literature we also discuss the existing empirical evidence for Belgium in this section. So far there has been no counterfactual impact study on the increase in the female full retirement age. 4

5 2.1 Results from Impact Evaluations A growing literature conducts ex post counterfactual evaluations of pension system reforms on the labour supply decision of older workers. Lalive and Staubli (2014) studied the impact of a reform of the Swiss pension system that raised the female retirement age from 62 to 63 in 2001 and from 63 to 64 in 2005 by means of a regression discontinuity design (RDD) based on administrative data. They explored the impact of raising the female retirement age on the pension claiming age and on the participation in other social security programs (e.g. disability or unemployment insurance) of older women. They find that a higher full retirement age of one year delays labour market exit by 7.9 months and claiming of retirement benefits by 6.6 months. In our empirical analysis we use a similar approach as the one used in this study. The same reform has been evaluated by Hanel and Riphahn (2012) by means of a difference-in-difference (DiD) approach. Using Labor Force Surveys data, the impact of these policies is evaluated by comparing the entry into retirement before and after each policy change between the age group affected by the policy and the age group not affected by the policy. They find that a permanent reduction of retirement benefits by 3.4% induces a decline in the age-specific annual retirement probability by over 50%. Based on data from the UK s Labour Force Survey data, Cribb et al. (2013) investigate the impact impact of increasing the UK state pension age from 60 to 61 between April 2010 and March 2011 on a number of labour market outcomes of older women and of their partners, by means of a DiD approach. As for Austria, Staubli and Zweimüller (2013) evaluate the impact of increasing the early retirement age from 60 to 62 for men and from 55 to for women between 2001 and 2010 on the employment rate of older workers and on the participation to disability or unemployment schemes. Atalay and G. Barrett (2015) evaluate the impact of increasing the female retirement age in Australia between July 1995 and July 2013 on the retirement status of female older workers and on the participation in alternative government programs, e.g. the disability benefits. As for Norway, Hernæs et al. (2015) exploit a comprehensive restructuring of the early retirement system in 2011 to examine labour supply responses to alternative pension reform strategies (e.g. abolishment of the early retirement earnings and improved work incentives). The analysis is based on a comparison of the first birth cohorts who were potentially affected by the reform and the last cohorts who were not affected by the reform in terms of a number of labour market outcomes. A number of studies have evaluated the impact of reforms of the US Social Security System pension System. First, Song and Manchester (2007) examine the impact of the removal of the earnings test to receive Social Security benefits, which concerns individuals who have attained the full retirement age (FRA) as from Second, Pingle (2006) examined the impact of increasing the Delayed Retirement Credit (DRC) from one percent to three percent as from 1982 on the labour supply of older men - an incentive that was meant to promote delayed entry into retirement. Third, Mastrobuoni (2009) evaluates the impact of increasing the age eligibility conditions to retirement as 5

6 from 2000 on the retirement behaviour of older workers based on a DiD approach. Similarly, Behaghel and Blau (2012) investigate the effect of the same reform on the hazard of exiting employment and the hazard of claiming the retirement benefits. Duggan et al. (2007) investigate the impact of this reform on the increase in the participation in the Social Security Disability Insurance. Lastly, Manchester and Song (2007) evaluated the impact of two of the aforementioned reforms of the US Social Security System on the retirement behaviour of older workers: the removal of the retirement earnings test and the increase in the full retirement age. 3 A number of studies have focused on a recent reform of the pension system of the Dutch public sector that substantially lowers pension benefits and creates incentives to postpone retirement as from January 2006 for everyone born after December 31, de Grip, Lindeboom, et al. (2012) assess the mental health effects of the Dutch pension reform using matched survey data and administrative data for male full-time employees in the public sector who were born in 1949 or The analysis is based on a sharp RDD in which the mental health of workers under the old pension system (born in 1949) are compared with the mental health of workers under the new pension system (born in 1950) in 2008, i.e. two years after the policy change. Following a similar strategy, Montizaan et al. (2010) examined whether the increase in the expected retirement age induced by the Dutch pension reform increased the training participation of public sector workers affected by the reform. de Grip, Fouarge, et al. (2013) analysed the effects of an announced increase in the eligibility age for the Dutch public old age pension on individuals retirement expectations. 4 The causal ex-ante impact of the announcement is identified by means of a sharp RDD using matched administrative and survey data. 5 Finally, Elsayed et al. (2015) study the impact of a gradual retirement scheme on the expected retirement age and total labour supply of Dutch public sector workers based on a stated preference approach. 6 They find that the replacement of full-time retirement schemes with a gradual retirement scheme stimulates workers to postpone retirement, but that gradual retirement is not a preferred option among workers. 2.2 Results from Structural Models Based on administrative data, Asch et al. (2005) study the financial incentives on retirement decisions focusing on workers employed at the US Department of Defense and provide evidence that the 3 Based on a synthetic panel for the period , Blau and Goodstein (2010) investigated the impact of the main changes in the US Social Security system (e.g.delayed retirement credit, increase in the full retirement age, removal of the retirement earnings test) on the trends of the labour force participation of older male workers in the US. 4 The announcement was done by the Dutch government in June 2010 and entailed the increase the statutory pension age in 2020 from 65 to 66 years for all inhabitants born after Based on survey data from the Irish Longitudinal Study on Ageing, A. Barrett and Mosca (2013) examined whether the expected age of retirement of older workers is affected by the announcement of retirement age increase (made by the Irish government in March of 2010) or by bad economic news ( Black Thursday in September 2010). 6 It consists of six vignettes in which respondents get hypothetical retirement scenarios and are asked to indicate at what age they would retire and to compare and rank the preferred choices. 6

7 retirement decision of these workers is not affected by social norms. Burtless (1986) evaluate the impact of unanticipated increases in US Social Security benefit (voted in 1969 and 1972) based on data from Longitudinal Retirement History Survey. Based on administrative data, Bernal and Vermeulen (2014) estimate a structural model of retirement and simulate the impact of increasing the retirement age on the effective retirement age of older single individuals in the Netherlands - a reform approved on July Fehr et al. (2012) evaluate the impact of increasing the normal retirement age from 65 to 67 as from 2007 in Germany by means of a general equilibrium model with overlapping generations. Based on the administrative data, Manoli and Weber (2014) exploit the exogenous variation from mandated discontinuous increases in retirement benefits of the Austrian pension system to estimate the inter-temporal labour supply elasticities. Their estimates are quite low, which suggests that the disutility of labour supply rises relatively quickly with additional years of work. 7 In addition, Manoli, Mullen, et al. (2015) exploit variation from five pension reforms in Austria (between 1984 and 2003) to identify and estimate income and price elasticities based on administrative data. They use these elasticities to study the labour supply and welfare consequences of a number of hypothetical pension reforms based on a structural model of retirement decision. Costrell and McGee (2010) analyse the retirement incentives embedded in Arkansas teacher pension scheme. Based on administrative data, they simulate the effect of eliminating the early retirement provision, and raising the normal retirement condition. 2.3 Results from Micro-Simulations on Belgium Finally, a number of studies have focused on the Belgian pension system. Most of these contributions entailed micro-simulations that tested the impact of a number of hypothetical reforms. Lefebvre and Orsini (2011) builds up a structural model of the labour supply of older workers in Belgium which is used to simulate the impact of a number of policy reforms. 8 They find that reforms that modify the access to every retirement route have more potential effects on the probability of exit than those simply targeting the generosity of the retirement scheme. Jousten, Sigismondi, et al. (2012) simulate different hypothetical reforms of the Belgian pension system using a micro-simulation approach. Based on administrative data, they evaluate the impact of the reform scenarios for the individuals as well as the public budgets. They show that partial reforms have limited effects, both 7 Based on administrative data, Brown (2013) estimates the price elasticity of lifetime labour supply of Californian public school teachers exploiting the reform of Californian teachers pension system that explicitly altered the financial return to an additional year of work. 8 Data are derived from the micro-simulation model MIMOSIS for a static simulation model based on administrative information that allows the reconstruction of the individual inter-temporal budget constraint of wage-earners. 7

8 in distributional and in fiscal terms. 9 Lastly, a number of studies examines specific aspects of the Belgian pension system. Jousten et al. (2014) investigates the potential link between health status, disability insurance and retirement old age wage earners in Belgium. Dellis et al. (2004) studies the incentives of early retirement in Belgium. Jousten et al. (2011) evaluate whether there is a systematic link between indicators of population health and the participation in the Disability Insurance, and shed light on the role of the Disability Insurance system in the Belgian Social Security. Capéau and Decoster (2015) study the impact of ageing on job choice behaviour, using a structural discrete choice model. They investigates the causes of the low labour market participation of older workers in Belgium, i.e. whether it is a matter of preferences or whether less jobs are available for the elderly. 3 Institutional Context In this section we discuss the institutional setting in which the 2006 reform of the Belgian female full retirement age (FRA) took place. This increase brought the FRA from 63 to 64 for women in self-employment and in salaried private sector employment. We start by introducing the pension reform of 2006 and situate it in the broader Belgian pension system around that time. Then we look at other policies for older workers that existed in that period, such as the early exit possibilities or the measures implemented to prevent them. It is important to note that the reform of the female FRA was announced several years before it was actually implemented. 3.1 Belgium s Pension System Until 1997 the female full retirement age for self-employed and salaried private sector workers was 60, as opposed to 65 for men. To obtain a full pension, men had to work 45 years whereas women only had to work 40 years. As this was deemed discriminatory according to the European Court and the European Commission, the Belgian government decided to increase the female FRA to the level of men. From July 1997 until January 2009 the female FRA was increased stepwise every two to three years from 60 to 65, as can be seen in Table 1. Note that for public sector workers the female FRA was already at 65 so there were no changes in that sector. In this paper we focus on the reform of January 1, 2006, which brought the female FRA from 63 to 64 and the necessary full career from 43 to 44. Note that these changes were all announced in 1997 and thus several years before they actually took place, with the exception of the reform of July At each moment in the left column of Table 1, the increase in the FRA determined an exogenous change in eligibility to old age pension benefits which affected specific birth cohorts. For the 2006 reform women born December 1942 are the youngest to be affected. While older women could still 9 Additional simulation analyses on the impact of potential pension reforms in Belgium are Jousten, Desmet, et al. (2007) and De Vil et al. (2015). 8

9 Table 1: Exogenous Changes in the Full Retirement Age for Women working in the private sector or as self-employed FRA Career Affected requirement birth cohorts Until June 30, Born before June 1, 1937 From July 1, Born June 1, 1937 or later From January 1, Born December 1, 1938 or later From January 1, Born December 1, 1940 or later From January 1, Born December 1, 1942 or later From January 1, Born December 1, 1944 or later Pensions can only be claimed one month after reaching the FRA. Therefore women born in (for example) December 1942 become 63 in December As their full retirement can thus only start in January 2006, they are affected by the policy and need to wait until January 2007 to claim their full retirement at the new FRA of 64. This table does not contain the changes to this scheme introduced since 2012, such as the decision to increase the FRA further to 66 in 2025 and 67 in retire at the (previously increased) FRA of 63, these younger women had to wait until 64. This translates to retiring one calendar year later, from January 2007 under the new rules instead of January 2006 under the old rules. The only cohorts not affected by any of these reforms are women born before June 1937, as they reached the FRA of 60 in June 1997, i.e. before the first increase in the pension age. The FRA mentioned above refers to the age at which women obtain full pension rights in the normal old age pension component. This is one type of pensions in the first pillar of Belgium s three pillar pension system (see below). Until the 2006 reform women could claim a full pension in the normal old-age pension component either the month after reaching age 63 or after reaching a career duration of 43 years. The calculation of pension benefits in this scheme is shown in the formula 1 below; it shows that for shorter contribution histories the pension is proportionally reduced. However, annual benefits cannot go below a certain minimum, to allow women with a sufficiently long contribution period (15 years) but very low earnings to have a sufficiently high pension. p a=1 Benefit t = (Earn a Z a ) Full Career Length RR t I t, (1) with ˆ t calendar time after retirement; ˆ a years with pension contributions, 1 a p < t and p the last year with pension contributions; ˆ benefit t the pension benefits received in year t; 9

10 ˆ earn a the labour earnings 10 in year a, subject to certain minima and maxima; ˆ Z a a reweighting factor that adjusts the yearly earnings earn a for inflation and welfare increases between year a and year p; ˆ the Full Career Length set at 43 years before the reform; ˆ RR t the replacement rate, 60% for single women and 75% for women with a dependent spouse; ˆ I t a reweighting factor that allows for further inflation and welfare adjustments once the pension has been determined in year p. The reform of 2006 adjusts Equation 1 by increasing the Full Career Length from 43 to 44. This implies that the same career before and after the reform (i.e. the same number of years with pension contributions) results in a lower monthly pension benefit after the reform, as the lifetime earnings are divided by a higher number. When in later sections we refer to the pension benefits of a woman, we refer to this specific scheme. In what follows we will give a short overview of Belgium s three pillar system, limited to the regulations available to private sector and self-employed workers. Unless mentioned differently, the pension system described below refers to the situation right before the reform of January 1, Pillar I is the state pension, available to all workers and organised as a pay-as-you-go system. This means that today s workers finance today s pensions via social security contributions on their wages. The height of a pension depends on career duration and earned income. The first pillar is further composed of three types of pension, the normal old age pension component, the survivor pension and the guaranteed minimum income for older persons. The first type, the normal old age pension, has been discussed above as this is the scheme affected by the pension reform. The second component, the survivor pension, is granted under certain conditions to the spouse of a deceased person. The allowance is equal to 80% of the pension paid or accrued to the deceased. Finally, the third component, the guaranteed minimum income for older persons, is a means-tested safety-net income financed by general taxation instead of contributions, making it a social assistance benefit and not a security benefit. It covers elderly people above the full retirement age who have no pension rights based on a professional activity or whose pension rights are very low. The allowance is indexed to the health index. Pillars II and III instead provide complementary pensions on a voluntarily basis at the firm or individual level. Instead of being organised as a pay-as-you-go scheme such as Pillar I, they are fully-funded schemes, where today s workers finance their own future pension benefits. More specifically, Pillar II is a supplementary pension for private sector workers organized at the company or sector level. Self-employed workers can also set up a second pillar pension. It is financed by contributions paid by employers, self-employed or employees to a pension institution. Sectoral 10 Some types of inactivity, such as sickness or unemployment are assimilated to employment. 10

11 pensions systems are built up by collective agreements between the management and the workforce of a sector of activity. Pillar III consists of personal pension savings which qualify for tax reductions. 3.2 Other policies aimed at older workers Next to the three pillar pension system discussed above, there exists a wide range of early retirement programs for older workers. In Table 4 in the Appendix we give a summary of the most important programs existing around the moment of the policy change in 2006, such as the anticipated pension where workers can retire early by taking up their pension before the FRA or the conventional early retirement scheme (the so-called bridge pension scheme), where workers pass through unemployment with a supplementary allowance from their last employer. 11 In general, these schemes provide generous allowances to participants and, except for the anticipated pension, keep accumulating (assimilated) career years for the pension calculation. This results in high participation rates in these schemes and a low average effective retirement age. Next to the policies that facilitate early retirement, the government also has several mechanisms that aim to keep workers employed until the FRA. Table 3 in the Appendix gives a summary of the most important programs around the moment of the policy change in 2006, such as the pension bonus that grants workers who stay employed longer a lump sum increase in their pension; a structural wage subsidy that lowers employer social security contributions for older workers, thereby making these older workers more employable and an outplacement mechanism designed to give laid-off older workers better employability prospects. In general the incentives provided by these mechanisms are not very strong. 12 In more recent years, beyond the scope of our analysis, the government has tried to make early retirement less appealing and working longer more so. In anticipation of our estimation strategy, where we compare women just (not) affected by the 2006 reform, we should ensure that the increase in the FRA is the only factor that plays a role. It is important to rule out the possibility that other factors or other policies do not simultaneously affect the treated group differently than the non-treated group. In the absence of such policies, any difference in the outcomes between the women treated by the policy and the women not treated by the policy can be attributed to the increase in the FRA. As Tables 3 and 4 in the Appendix show, there were not many other reforms around The most important reform was the Pension Bonus starting in January This reform is aimed at the same women as those affected by the reform, as it only concerns pensions started in January 2007 or later and refers to periods worked since January However, the incentives given by this policy are rather weak, as working one day extra after age 62 results in e2 extra pension benefits per year. Other reforms tried to discourage 11 Given that by 2007 everybody in our sample will reach the full retirement age, we do not focus on early retirement policies after this moment. 12 Smith et al. (2014) shows that the pension bonus has no impact before When we study the 2009 pension reform in the future this mechanism should be taken into account more carefully though. 11

12 early exits through levying social security contributions on employer-paid allowances in the (pseudo) early retirement and time credit schemes. All these reforms aim at keeping older workers employed longer, either by promoting employment or by making early labour market exits less profitable. Although most of these reforms are small in terms of incentives, they could have a minor impact on the effects, favouring positive employment effects. 4 Data and Sample Selection 4.1 Database We use rich individual data that were obtained by merging administrative registers of the diverse Social Security institutions and of the National Register containing all Belgian inhabitants. The database became more comprehensive over time. From as early as 1957 until 1998 we have for employees in the private sector yearly information on earnings, the number of working days and hours (in case of part-time work) and the worker type (blue or white collar). From 1998 onwards this information is available on a quarterly basis (measurement at the end of each quarter), not only for employees in the private sector, but also in the public sector. In addition, from then onwards it also contains on limited firm information (size and sector), the industrial committee to which the worker belongs to 13, the timing of self-employment spells, on UB receipt, as well as on participation in the Career Break, TC schemes and early retirement schemes. Finally, since 2003 the data have been complemented by information on receipt of statutory (possibly early) retirement benefits, on sick leave and replacement income in case of disability, occupational diseases or accidents. Finally, since 1998 the National Register provides yearly information on December 31 on individual and household characteristics, such as age, gender, nationality, district of residence, household size (by age group) and type (single or couple, with or without children). The observation period in this study ends in the last quarter of Endogenous Sampling Mechanism The sample on which we base our analysis was drawn with the purpose to evaluate the effect of a wage cost subsidy for employees in the private sector aged 58 years or more. The Belgian government introduced this subsidy in 2002 to enhance the employment of older workers (Albanese and Cockx, 2015). To that end a representative sample was drawn of 243,655 individuals born between the 1st of April 1941 and the 31st of March 1950, i.e. aged between 52 and 61 in Because in Belgium many individuals are already inactive in that age bracket, the sample was not only stratified 13 Industrial committees are organized for each type of worker at the sectoral level. In these committees trade unions and employer organizations negotiate the collective agreements. These agreements are binding for all workers belonging to this industrial committee, irrespectively of whether they are unionised and, hence, represented in the negotiation. 12

13 according to gender, but also into 9 birth cohorts c (=1,2,,9) and 5 strata r (=1,2,,5). These strata were defined according to employment status in the private sector and the earned wage in the period around the 2002 reform. This stratification aimed at over-representing groups that are relatively rare in that age bracket and more responsive to the labour market policy reform: low-wage employees in the private sector and individuals transiting in and out of employment during this period. In Appendix A.2 more details on this stratification can be found. Because the stratification involves outcome variables of interest, it is endogenous and it is well known that consistent estimation then requires to appropriately weigh the data in these strata (e.g. Manski and Lerman, 1977; Cameron and Trivedi, 2005). If we denote the sampling weight for individual observation i belonging to birth cohort c and to substratum r by W cr,i, then W cr,i = N cr N n n cr, (2) where N cr denotes the size of the population in substratum cr 14, n cr the corresponding sample size, N 9 c=1 5 r=1 N cr the total population size and n the corresponding sample size. As to avoid cumbersome notation, gender is not explicitly referred to. The weighting formula comes from a double re-weighing, within and between cohorts Final Sample & Outcome Variables In this study we evaluate the impact of an increase in the legal retirement age for women from 63 to 64 in January 1, The pension reform creates a quasi-experimental setting to which we apply a Regression Discontinuity (RD) design (Lalive and Staubli, 2014; Montizaan et al., 2010; de Grip, Lindeboom, et al., 2012) exploiting the sharp discontinuity in the eligibility conditions to full retirement; for women eligible to a pension before January 1, 2006 nothing changes, while the other women face a change in the eligibility criteria. We then obtain our results by comparing women just below and above this eligibility threshold. We select women born up to one year before and after the cut-off birth date of December , which is from December 1, 1941 up to November 30, Women born before the cut-off birth date form our control group, whereas women born afterwards form the treated group. This results in an initial sample of women, representing women in the Belgian population. There are non-treated and treated women, who represent and women in the population. In population terms 45.9% of the women in our sample are in the control group and the remaining 54.1% in the treated group. 14 We have information on the population sizes in each substratum, i.e. on N cr. 15 First, to restore the representativeness within the cohorts we reweigh the units within each cohort by W c cr = Ncr N c nc n cr (where N c and n c are the size of the cohort in the population and in the sample). To make the cohorts in the sample representative for the population, we weigh each cohort a second time: SW cr = W c cr Nc N 13 n n c, so that W cr = Ncr N n n cr.

14 With this baseline sample we can calculate the effects for our most important outcome variables, such as the pension take-up rate, the early retirement rate and the employment rate. So far, we have finished our baseline estimations, the results can be found in Section 6. We are currently performing a more in-depth analysis on sub-populations of our initial sample, after which we plan to run several robustness checks. A more in-depth analysis will allow us to look at among others the average age of exit from employment and the average age of retirement. For this type of outcomes we cannot calculate the outcome variable for women who have not made the necessary transitions (e.g. leave employment or receive a pension). To be able to nonetheless calculate these outcomes we need to restrict our sample. First, to make sure we cover a similar window of observation for all women in the sample we look at the period in which they are aged Second, to look at for example the age of exit from employment we would need to restrict our sample to women who are observed in employment at least during one quarter between ages and who are not in employment any more at age 68. This follows the approach of Lalive and Staubli (2014) in their evaluation of a similar Swiss reform. Observe that this sample restriction will have to be tested, to make sure that we still represent the same population. The outcome variables discussed in this paper are defined as follows: (1) The pension take-up rate (PR) is defined as the share of women receiving a full pension compared to the total number of women in our sample. Note that, as explained in Section 3 on the institutional context of Belgium, in principle women can claim this pension before the full retirement age, albeit a lower pension than they would have had at the full retirement age. A full pension is defined as either a pension in the private or self-employment pension regime. 17 We thus focus on self-earned pensions and not on pensions received through the spouse. Moreover, by focussing on pensions in the affected regimes we ensure that we consider people potentially affected by the policy. The pension rate is defined by treatment group, so the denominator for the treated (controls) contains the total number of women in our treated (control) group in our sample. The entire PR analysis is based on the definition of the PR by calendar quarter. (2) The employment rate (ER) is defined as the share of women in salaried or self-employment compared to the total number of women in our sample. In later analyses we will look at the sector-specific employment rates to see if there are different effects among them. Like the pension rate, the employment rate is defined by treatment group and calendar quarter. (3) The early retirement rate (ERR) is defined as the share of women in bridge pension or unemployment compared to the total number of women in our sample. 18 We combine bridge pension 16 Women born in December 1941 (the oldest cohort) become 56.5 in the second quarter of 1998, this is the first quarter for which all necessary quarterly data are available. Women born in November 1943 (the youngest cohort) become 68 in the fourth quarter of 2011, the last quarter for which we have data. 17 Pension data is only available since 2003 in our database, so this outcome variable can only be studied from this moment onwards. 18 The unemployment can be with or without job search requirement, but given that our sample consists of older 14

15 and unemployment as this is, in se, the same thing for these women: both schemes allow women to prematurely leave the labour market without having to claim an anticipated pension. Therefore we talk about early retirement in the remainder of this paper when we talk about the bridge pension and unemployment combination. Like the pension rate, the early retirement rate is defined by treatment group and calendar quarter. 4.4 Descriptive analysis As described in the previous section, our sample is composed of women born within two years from each other. Consequently, we expect that these women are relatively homogeneous in their characteristics. We condition on four categorical covariates in our analysis. These variables contain information on the nationality, region of residence, number of children in the household and marital status. All covariates are observed yearly. For our analysis we take the earliest observation possible (i.e. 1998q4 (1999q4) for the control (treated) group), but impute missing values by the earliest non-missing observation up to 2011q4, such that no women have missing observations. 19 A summary of the descriptive statistics can be found in Table 2, where we distinguish between women in the control and treated group. As expected, there is little difference in the covariate values. Figure 1 gives a first idea on the possible effect of the higher retirement age on our baseline outcome variables: the pension take-up rate, the employment rate and the early retirement rate. First, in panel A we plot the pension take-up rate by treatment status and per calendar quarter, from 2003q1 until 2011q4. As mentioned in the previous section we have to restrict our window of observation as the pension data is only available since Before women in the control group (blue line) reach the full retirement age, pension take-up is relatively similar compared to the treated group (red line). The slightly higher pension rate for the non-treated women is because they are on average one year older and because not all treated women have reached the earliest age to claim an anticipated pension yet in 2003q1. When the non-treated women reach the full retirement age, their pension take-up rate goes up from 12% in 2004 to 56% from 2006 onwards. 20 Treated women can only claim their full pension two calendar years after non-treated women. The figure shows that treated women do not increase their anticipated pension take-up at the moment of the policy change in 2006q1, indicated by the vertical red line. While treated women start their full retirement on average later, this does not yet mean that they also leave employment later. Moreover, although the magnitude and nature of the effect suggests that this is caused by the policy (i.e. that there is a discontinuity at the cut-off), we still check this formally in the results section. Panels B and women, in practise they did not really have to meet this requirement. See Section 3 for more details on the institutional context of these labour market states. 19 We impute around 3.5% of the observations (corresponding to around 10% of the population due to oversampling of inactive individuals in our database - who have less information). 20 The rate does not increase to 100% because a some women in our sample were ever employed and could thus not claim a pension. 15

16 16 Table 2: Descriptive Statistics of Selected Treated and Control Groups (weighted by W cr,i ) Controls Treated Mean SD Min Max Mean SD Min Max Region of residence Brussels Flanders Wallonia Nationality Belgian Non-Belgian Marital status Spouse No spouse Children in household Children No children Sample Size (Unweighted) 6,095 8,380 Represented Population Size 45,650 53,803

17 C discussed next contain a first explanation on what happens to treated women in this extra year before full retirement. Next, in panel B we plot the employment rate by treatment status and per calendar quarter, from 1998q2 until 2011q4. The treated women, in red, have a higher employment rate in almost every quarter. This is not surprising as they are on average one year younger than non-treated women. In 2002 there is an important drop in the employment rate for non-treated women, followed by a similar drop one year later for the treated women. This corresponds to a drop around age 60, when most women can retire early. Given that there is this important shift in labour market participation, we will investigate whether this has been caused by (an anticipation of) the policy change. The jump in 2003q1, which is shared by both treatment groups, likely points to a positive business cycle effect after several quarters of low growth. At the moment of the policy change itself, indicated by the vertical red line in January 2006, the difference between the employment rates of the treated and non-treated women has increased to around 5 percentage points. In our results section we investigate whether this difference is due to a discontinuity or not. Given the evidence from panel A, it seems that treated women do not actually increase their labour supply in the extra year before full retirement. Finally, in panel C we plot the early retirement employment rate by treatment status and per calendar quarter, from 1998q2 until 2011q4. As all women (have to) leave this scheme when they reach the full retirement age, the rate drops to zero in 2006q1 for the non-treated women (blue line) and in 2008q1 for the treated women (red line). The two rates overlap until 2002, when non-treated women can start to claim an anticipated pension. This is followed by a similar drop in the early retirement rate for treated women one year later. At the moment of the policy change treated women continue to use the early retirement scheme. This, in combination with the evidence from panels A and B, implies that treated women mostly spend their time in inactivity during the extra year before they reach the full retirement age. 5 Empirical Strategy 5.1 Identification We estimate whether women adapt their labour market behaviour when their full retirement age (FRA) is increased. The source of identification is this quasi-experimental increase in female FRA; for women born December 1, 1942 or later the FRA increased from 63 to 64, while it remained at 63 for women born before this date. In our analysis we estimate the causal effects of interest by comparing women born within the year before this cut-off birth date of December 1, 1942 (control group) with women born within the year after the cut-off (treated group). Such a comparison results in unbiased estimates of the treatment effect if the women on both sides of the cut-off exhibit similar observable and unobservable characteristics (Lalive and Staubli, 2014, p ). A Regression 17