2.1 Retirement insurance and pathways to retirement

The German retirement insurance operates on a pay as you go basis. It is funded mostly by mandatory contributions of employers and employees. Regulated at the federal level, it covers more than 80% of the population, excluding only civil servants and the self-employed. It offers old-age, disability, and survivor benefits. Given the limited relevance of private pensions in Germany, the mandatory retirement system provides the main source of income for most elderly households (Frommert, Reference Frommert2010).Footnote ³

Generally, benefit amounts depend on amounts and years of contribution. More specifically, the benefit formula considers the sum of annual ‘earning points’. The earning points for 1 year represent the ratio of the individual's earnings and the mean earnings of all insured persons, i.e., the relative individual annual contribution. An individual who contributed to the retirement insurance over 45 years based on the annual average earnings would receive a gross benefit that amounts to about 48% of gross earnings (BMAS, 2019). Benefit eligibility is regulated along pathways to retirement. Examples for such pathways are retirement ‘due to unemployment’, ‘for women’, ‘after long term employment’, or ‘disability retirement’. Pathways can be used with an NRA or an ERA. With the ERA, the minimum retirement entry age and subsequent benefit amounts are lower than that with the NRA; specifically, each month of early benefit receipt prior to the NRA generates a permanent benefit cut by 0.3% (i.e., 3.6% for each full year).

We exploit two reforms of the entry age for the unemployment pathway into retirement. The unemployment pathway into retirement has been available for blue and white collar workers since 1957. The utilization of the pathway increased over successive birth cohorts. It was used by 10% of all retired men born in 1925, 24% of those born in 1935, and 30.8% of those born in 1940. For subsequent birth cohorts the shares declined, e.g., to 19.7% for the 1945 cohort (see DRV, 2018, p. 93).Footnote ⁴

The pathway was used particularly by individuals with low reemployment probabilities, who had lost their jobs in the wake of structural changes in firms and industries and whose occupational experience was no longer demanded in regional labor markets (Mika and Krickl, Reference Mika and Krickl2020). However, at the same time these individuals had accumulated substantial claims against the retirement insurance. Their past earnings and retirement benefits often exceeded those of the average retiree (DRV, 2018).

The first reform increased the NRA stepwise from age 60 to 65 starting with individuals born in January 1937 and ending with those born in December 1941 (see column 1 of Table 1 and Figure A.1 in the Appendix). At the same time, early retirement with benefit deductions became newly available at age 60 for all cohorts. The second reform increased the ERA stepwise from age 60 to 63 starting with individuals born in January 1946 and ending with those born in December 1948.Footnote ⁵ Both reforms contained regulations to protect the ‘legitimate expectation’ (Vertrauensschutz) of individuals.Footnote ⁶

Table 1.

Age at retirement by pathway and birth cohort

Notes: For a more complete description see Table A.1. The entry ‘rising to’ means that the retirement age increased by 1 month for each month of birth in the relevant birth cohort: in column 1 individuals born, e.g., in January of 1937 had an NRA of 60 and 1 month, those born in February of 1937 had an NRA of 60 and 2 months and so on, reaching an NRA of 61 years for those born in December of 1937.

Source: SGB VI, BMAS (2017), Steffen (Reference Steffen2018), and own calculations.

Column 2 of Table 1 depicts the retirement pathway for women (see also Figure A.1).Footnote ⁷ Here, the reform for birth cohorts 1940 and after is very similar to the first reform of the unemployment pathway. In addition to these two pathways, the ‘regular old-age’ retirement pathway has traditionally been available at age 65 and does not offer an ERA; ‘retirement after long-term employment’ is available for individuals with 35 insurance years and offers an ERA of 63 and an NRA of 65. A special pathway is available for individuals with severe handicaps. Finally, the disability pathway used to provide retirement benefits independent of age (Hanel, Reference Hanel2012). It is available until the ‘regular old-age’ retirement age is reached.Footnote ⁸

As retirement insurance contributions paid in East Germany were in part treated differently from those paid in West Germany, we follow Ye (Reference Ye2019) and use only individuals who contributed in West Germany. In robustness tests, we include individuals with labor market experience in East Germany.

Since anticipation behavior can affect the estimation of treatment effects, it is important to consider reform announcements: in 1989, a reform law was passed which stipulated that starting in 2001 retirement entry ages should start to increase toward age 65 beginning with the 1941 birth cohort (Steffen, Reference Steffen2018). However, in 1996 and 1997, reforms accelerated the increase in the retirement age and brought the starting date of these retirement entry age adjustments forward from 2001 to 1997. This newly affected individuals of the 1937 (instead of 1941) birth cohort (reform 1), who had little time for adjustment as they turned age 59 in 1996.

The ERA had been available since 1997 for birth cohorts 1937 and after. In 2004, a law was passed which mandated that the minimum age for early retirement on the unemployment pathway increased starting in 2006 (reform 2) (Steffen, Reference Steffen2018). As both reforms were passed into law very briefly before they became effective, we expect to observe short-term behavioral adjustments that are not yet attenuated through anticipatory behavior changes.

2.2 Related institutions: unemployment as a bridge to retirement

As we focus on retirement after unemployment, it is important to describe unemployment insurance (UI) rules. Historically, two types of unemployment benefits were provided: an insurance-based payment (UB I) that covered about 60% of last net earnings and a means tested system which provided a more modest minimum income support (UB II). The duration of UB I benefit payment depends on the number of UI contribution years and the age of the unemployed individual and was reformed several times (see Table A.3).

Since the 1990s, it was common practice to use the long unemployment benefit duration as a bridge to retirement: retirement benefits after unemployment were available starting at age 60 without benefit deductions. So workers who were laid off starting e.g., at the age of 57 and 4 months, received UI benefits for up to 32 months, and then entered retirement after unemployment. In our analysis, we follow Engels et al. (Reference Engels, Geyer and Haan2017) and account for changes in the duration of unemployment benefit payout by controlling for the individual and age-specific maximum entitlement length.

2.3 Expected reform effects and hypotheses

The first reform increased the NRA in a stepwise fashion and introduced an ERA. The second reform increased the ERA stepwise and thus disallowed retirement at early ages. The first reform implied a negative shock to individuals' social security wealth and the second reform restricted individuals' choice sets by making early retirement unavailable.

Under the first reform, it remained possible to retire starting at age 60. However, the reform introduced a benefit reduction after early retirement which increased over subsequent birth cohorts. To retire at age 60, the 1936 birth cohort did not suffer benefit reductions, cohort 1937 had to forgo up to 3.6%, and cohort 1942 and later lost 18% of their benefits.

The literature offers different theoretical frameworks to derive hypotheses with respect to the reforms' effects. The first is based on an intertemporal consumption model. In this view, the reforms generated income and substitution effects: the income effect consists of the decline in ‘pathway-specific pension wealth’, e.g., at age 60. The newly introduced additional reduction in pension benefits that follows each month by which retirement takes place prior to the new NRA increases the price of leisure. This generates the substitution effect in the choice of entering the unemployment pathway (for a more formal description, see, e.g., Hanel and Riphahn, Reference Hanel and Riphahn2012). If leisure is a normal good, both effects reduce the demand for leisure and increase labor force attachment and incentives to postpone retirement entry. We therefore expect prolonged employment and unemployment. In particular, workers who would have used unemployment as a bridge into retirement before age 60 without the reform may postpone unemployment to later ages: their unemployment may decline before age 60 and increase afterward. Also, we expect delayed retirement entry after the first reform. Since the reduction in pension benefits vary by cohort and by age, the income and substitution effects should differ across birth cohorts and for every cohort by age.

Alternative pathways of retirement, e.g., for those with severe handicaps, continued to offer retirement entry at age 60 without benefit deductions for the birth cohorts 1937–39. In addition, disability retirement was available without any age restrictions at full benefits for the 1937 cohort. For the birth cohorts 1938 and after, disability retirement at full benefits was available until December 31, 1999. For subsequent entries, benefit discounts were introduced if the retiree had not yet reached age 63 (see Table A.2). Individuals may have used these alternative pathways as substitute exit routes out of the labor force.

The second reform abolished the option to use ERA on the unemployment pathway prior to age 63 in a stepwise fashion for birth cohorts 1946 and after. This reform may have no lifecycle income effects if prior benefit discounts after early retirement were actuarially fair. Nevertheless, behavioral adjustments will follow from changes to age-specific budget constraints: if the pathway to early retirement is unavailable other labor market options are more attractive as they provide an income. The effect of increasing ERA has been looked at for different countries before: see, e.g., Cribb et al. (Reference Cribb, Emmerson and Tetlow2016) for the UK, Staubli and Zweimüller (Reference Staubli and Zweimüller2013) for Austria, Vestad (Reference Vestad2013) studied the effect of a reduced ERA for Norway, and Geyer and Welteke (Reference Geyer and Welteke2021) and Geyer et al. (Reference Geyer, Haan, Hammerschmid and Peters2020) studied different pathways in Germany. These studies generally find that such a reform (a) affects employment and unemployment, and (b) incentivizes program substitution.

A second framework to explain retirement behavior has recently been developed by Seibold (Reference Seibold2021). He uses the example of 644 pension benefit continuities in Germany to establish convincing evidence that financial incentives alone cannot explain retirement patterns. Instead, he suggests that NRA and ERA become reference points for individual retirement decisions. Under this framework, the two reforms considered here shifted the reference points for retirement entry after unemployment which should then yield similar adjustment patterns.

Overall, we expect that after both reforms individuals stay in employment longer (H1). Because unemployment as a bridge to retirement becomes less attractive, after reform 1 (NRA), unemployment may decline prior to age 60 and increase at age 60 and after. In contrast, after reform 2 (ERA), unemployment may decline at age 60 and after because retirement without deductions was available only at age 65. In sum, we expect that unemployment is postponed to later ages (H2). We expect that both reforms contributed to a delay in the utilization of the unemployment pathway to retirement because it became either more expensive (reform 1) or inaccessible (reform 2) (H3). Also, we expect that the demand for alternative pathways, such as disability retirement or retirement for the severely handicapped, increased (H4).

In addition, we compare the responses of two reforms. Staubli and Zweimüller (Reference Staubli and Zweimüller2013) suggest that responses to changes in the ERA are larger than responses to changes in the NRA: while individuals can ignore changes in the NRA by accepting reduced benefit payments upon retirement postponements of the ERA must be heeded. We compare the responses to reforms 1 and 2, which affect the same pathway but affect different birth cohorts.

3.1 Data

We use administrative data offered by the German Pension Insurance. The Versicherungskontenstichprobe (VSKT) provides roughly a 1% random sample of persons aged 15–67 and covered by the German Statutory Pension Insurance. The data are available annually since 2002.Footnote ⁹ Each wave provides information on demographic and pension relevant characteristics such as year and month of birth, nationality, information on monthly labor market status and earning points (Himmelreicher and Stegmann, Reference Himmelreicher and Stegmann2008; Stegmann, Reference Stegmann2008).

We use different samples for the two reforms. They consist of men and women born between 1935–39 and 1945–48. Due to special pension rules, we do not consider civil servants, self-employed, miners, and persons with pension entitlements according to the law on foreign pensions (FRG).Footnote ¹⁰

We analyze the reform effects by comparing the labor market behavior of treated and non-treated individuals. We compare the birth cohorts 1937–39 and 1946–48 who are affected by the reforms to the cohorts 1935–36 and 1945 who are not affected. In addition, we take advantage of the fact that females who are eligible for the female retirement pathway are basically not affected by the reform of the unemployment pathway.Footnote ¹¹ Therefore, they constitute a control group in our analyses. The female retirement pathway provides an NRA of 60 for birth cohorts potentially affected by our first reform (1935–39) and an ERA of 60 for those birth cohorts potentially affected by our second reform (1945–48). Since the female birth cohorts 1940–44 were affected by the NRA reform for female retirement (see Table 1), we do not use these cohorts and thus limit our post-reform cohorts for reform 1 to 1937–39 and our pre-reform cohorts for reform 2 to 1945.Footnote ¹² For reform 1 (NRA), we therefore consider individuals born 1935–39 and for reform 2 (ERA) individuals born 1945–948.

To obtain comparable subsamples, we follow Engels et al. (Reference Engels, Geyer and Haan2017) and consider only those men and women in our sample who fulfill the eligibility criteria for the female retirement pathway when they are at age 55.Footnote ¹³ These eligibility criteria demand a waiting period of 15 years and a compulsory contribution period of at least 10 years after reaching age 40. Clearly, we cannot condition on the eligibility criteria of the unemployment pathway since this involves the potentially endogenous response of anticipatory unemployment.

For every birth cohort, we use one wave of the VSKT to avoid duplicate observations.Footnote ¹⁴ Based on the biographical information, we code monthly observations for each individual from age 60 plus 0 months to age 62 plus 11 months. As discussed above some individuals are not affected by the reform because of regulations to protect their ‘legitimate expectation’. We code the treatment status of those individuals, about 2% of the sample, as not treated. Overall, we use 11,240 different individuals and 404,640 person-month observations for the first and 8,566 individuals and 308,376 person-month observations for the second reform.Footnote ¹⁵

Our dependent variables characterize five different states. They describe for every age month whether an observation is in ‘employment’, ‘unemployment’, ‘old-age retirement’, ‘severely handicapped retirement’, or ‘disability retirement’. ‘Old-age retirement’ combines retirement entries through all pathways except for ‘severely handicapped retirement’, or ‘disability retirement’.Footnote ¹⁶

Table 2 describes our dependent variables. In the sample for reform 1, 21% of the full sample are in employment, about 4% in unemployment, 57% of the person-month observations of 60–62 years olds are already in old-age or severely handicapped retirement, and about 13% in disability retirement.Footnote ¹⁷ The reform 2 sample has higher shares in employment and unemployment and a lower share in retirement.

Table 2.

Descriptive statistics of dependent variables

Notes: All observations are in the age range 60–62. Old-age retirement combines retirement after all pathways except for severely handicapped retirement and disability retirement; given our age restrictions this comprises only the two pathways of retirement due to unemployment and retirement for women (see Table A.1). Tables A.5–A.8 describe all variables and provide further descriptive statistics. The number of observations in the columns entitled ‘All’ exceed the sum of male and female observations because sampling restrictions with respect to the protection of ‘legitimate expectations’ are not yet imposed in the columns ‘All’ (see Sections 2.1 and 3.1 in the text). The affected observations are dropped from the male subsamples described in the columns entitled ‘Men’.

Source: SUFVSKT2002_FAU_Schrader-SUFVSKT2013_FAU_Schrader, own calculations.

Figures 1 and 2 depict cohort- and gender-specific employment, unemployment, and old-age retirement behavior by age for both reforms. Figure 1a shows employment rates by age for men and women of cohorts 1935–39 (we dropped some birth cohorts to enhance clarity). After age 60, employment drops for men and women. However, males stay in employment longer than females. As expected, the treated male cohorts 1937 and 1939 have higher employment rates than the pre-reform cohort 1935 with little difference across cohorts for females. The developments for the non-affected cohorts are similar for males and females supporting our strategy to use females as a control group. Figure 1b presents unemployment rates by cohort, gender, and age. While these rates are rather similar for females across birth cohorts, the male rates clearly reflect the reform incentives: the unemployment rate of post-reform cohorts exceeds that of pre-reform cohorts at almost all age months. The effect is most visible at the early ages. The male and female developments for the pre-reform cohorts are parallel. Figure 1c depicts old-age retirement starting at age 60. Generally, males' retirement rates are lower than females'. While male cohorts affected by the reform have lower old-age retirement rates than the pre-reform cohorts this does not hold for females, matching expectations. As before, females and males in the cohort not affected by the reform (1935) have similar age patterns.

Figure 1.

(a) Reform 1 (NRA) – Employment rate by age, cohort, and gender. (b) Reform 1 (NRA) – unemployment rate by age, cohort, and gender. (c) Reform 1 (NRA) – old-age retirement rate by age, cohort, and gender.

Notes: The NRA for birth cohorts 1935 was 60; the NRA for birth cohorts 1937 (1939) increased in monthly steps from 60 to 61 (62 to 63) for those born from January through December. The ERA did not exist for birth cohort 1935; for birth cohorts 1937 and after it was at 60. For the figures, we deleted male, post-reform cohort observations who due to protection of legitimate expectation (measured as of age 60) were not treated by the reforms. We show the gender-specific number of individuals of a given birth cohort and age in employment relative to the number of all individuals in that gender, age, and birth cohort cell in our sample. Source: SUFVSKT2002_FAU_Schrader-SUFVSKT2013_FAU_Schrader, own calculations.

Figure 2.

(a) Reform 2 (ERA) – employment rate by age, cohort, and gender. (b) Reform 2 (ERA) – unemployment rate by age, cohort, and gender. (c) Reform 2 (ERA) – old-age retirement rate by age, cohort, and gender.

Notes: The ERA for birth cohort 1945 was 60; the ERA for birth cohorts 1946 (1948) increased in monthly steps from 60 to 61 (62 to 63) for those born from January through December. The NRA was 65 for all birth cohorts. For the figures, we deleted male, post-reform cohort observations who due to protection of legitimate expectation (measured as of age 60) were not treated by the reforms. We show the gender-specific number of individuals of a given birth cohort and age in employment relative to the number of all individuals in that gender, age, and birth cohort cell in our sample. Source: SUFVSKT2002_FAU_Schrader-SUFVSKT2013_FAU_Schrader, own calculations.

Figure 2 present the same outcomes with respect to reform 2, for cohorts 1945–48 (cohort 1947 omitted to enhance clarity). Figure 2a shows similar employment-age patterns for males and females pre-reform (i.e., birth cohort 1945). For the cohorts affected by the reform, we find substantial changes in employment rates for men, but not for women. The unemployment rates in Figure 2b are unclear which is likely related to reforms of unemployment benefit duration in 2006 and 2008 for which we do not control here.Footnote ¹⁸ However, the male pre-reform cohort stands out. Also, we find a drop in unemployment rates at the early ages and an increase in later ages for cohort 1948, as expected. Finally, Figure 2c shows retirement rates and confirms cohort and gender differences: the retirement of the treated male cohorts declines stepwise by age and birth cohort which follows exactly the incentives generated by reform 2. As in Figure 1c, we observe a similar age pattern for the male and female pre-reform cohort.

Figures 1 and 2 show that affected and non-affected male birth cohorts differ in their employment, unemployment, and retirement behavior in agreement with the expected reform effects. Female birth cohorts are suitable controls as they show similar pre-reform patterns.

3.2 Empirical methods

We aim to identify the causal effect of two reforms on retirement behavior and labor force participation choices of older workers. We exploit the fact that both reforms affected specific birth cohorts at specific ages, e.g., reform 1 modified the NRA of individuals age 60 and born in 1937 or later and reform 2 modified the ERA of individuals age 60 and born in 1946 or later. Based on the combination of birth cohort and observation period, we can identify the causal reform effect if we assume that the reform is the only determinant of possible behavior changes across birth cohorts at a given age. This identification strategy is widely applied in the literature and requires that there are no changes in age-specific cohort trends absent the reform (e.g., Mastrobuoni, Reference Mastrobuoni2009; Hanel and Riphahn, Reference Hanel and Riphahn2012; Staubli and Zweimüller, Reference Staubli and Zweimüller2013).

We can go beyond this standard approach and allow for age-specific cohort trends. We additionally compare groups of men treated by the reform and women not treated by the reforms, both, at ages when the reform was effective or not, and for birth cohorts which were and were not affected. This DIDID setting is possible because the change in the unemployment pathway to retirement was not accompanied by a similar and simultaneous change in the retirement pathway for females and because the changes were implemented stepwise by cohort.

Under the requirements of the female retirement pathway, women continued to be able to retire at an NRA of 60 after our reform 1 (NRA) and similarly to use an ERA of 60 after our reform 2 (ERA) (see Table 1). Thus, we distinguish men who are affected by the reform (treatment group) from women who are not affected (control group) and compare their behavioral choices at ages which were affected and ages that were not affected by the reforms for birth cohorts affected (post-reform) and not affected (pre-reform) by the reforms. We calculate the reform effect as the difference in changes between men and women across ages and cohorts. This identifies causal effects if the behavioral adjustments across cohort groups and genders would have been identical for different ages without the reform. This is a weaker identifying assumption than that required for a DID model: we can allow for changes in the labor market affecting men and women differently across birth cohorts as long as they are not age-specific.

The recent literature intensely discusses spillover effects in retirement choices of couples. There is substantial evidence showing coordination in spouses' retirement choices (e.g., Hospido, Reference Hospido2015; Atalay et al., Reference Atalay, Barrett and Siminsky2019). Clearly, such coordination behavior can modify the policy reforms' effects as changes of the retirement entry age for men affects the behavior of women and vice versa. If spillovers are neglected in the analysis of reform effects, the results can be biased downward (see Coile, Reference Coile2004). Besides the direct reform effects of increasing retirement entry ages on the husbands' retirement, the reform could induce indirect reform effects by increasing the wives' retirement ages, as well. Since we compare male and female labor market outcomes and as our data do not allow us to control for the spouses' labor market choices our overall impact assessment of the reform may be biased downward.

Let men indicate whether the individual is male (men = 1) or female and belongs to the control group (men = 0). C contains month and year of birth fixed effects and post is equal to one for cohorts affected and equal to zero for cohorts not affected by the reform. A represents the individuals' monthly age in a given monthly observation a. Some ages are affected by the reform and others are not; age (valued 0 or 1) indicates whether an individual is affected by the reform in a given month. X are individual level control variables; we use measures of past earnings, health, tenure, and insurance group indicators (blue collar, white collar, other) as of age 55, i.e., prior to the observation window. Finally, to account for unemployment and disability benefit reforms (see Sections 2.1 and 2.2), we also control for the maximum length of (potential) unemployment benefit receipt and the month- and person-specific potential discount after disability retirement.Footnote ¹⁹ We use the following linear regression model:

(1)$$\eqalign{Y_{i, a} =\; & \beta _0 + \beta _1men_i + \beta _2C_i + \beta _3A_{ia} + \beta _4C_i\cdot A_{ia} + \beta _5men_i\cdot C_i \cr & \quad + \beta _6men_i\cdot A_{ia} + \gamma ( {men_i\cdot post_i\cdot age_{ia}} ) + X\theta + \varepsilon _{i, a}.} $$

We estimate the parameters β, γ, and θ using least-squares regression, where γ represents the causal effect of interest. ɛ is a random error. Note that the triple interaction term is identical to using an indicator for being affected by the reform versus not (see, e.g., the specification by Staubli and Zweimüller, Reference Staubli and Zweimüller2013). We consider variations of this specification where we control for a post reform indicator instead of the detailed date of birth fixed effects, C.

In addition to the interaction model 1, we apply specifications that control for the reform intensity for given individuals (see Duggan et al., Reference Duggan, Singleton and Song2007, for a related approach). Appendix Tables A.9 and A.10 describe the number of months of benefit discounts after reform 1 (NRA) and the number of months by which an early retirement must be postponed after reform 2 (ERA). We use these indicators of reform intensity (I) that vary at the cohort, gender, and age level in a separate specification and estimate the following model for both reforms:Footnote ²⁰

(2)$$\eqalign{Y_{i, a} =\; & \beta _0 + \beta _1men_i + \beta _2C_i + \beta _3A_{ia} + \beta _4C_i\cdot A_{ia} \cr & \quad + \beta _5men_i\cdot C_i + \beta _6men_i\cdot A_{ia} + \gamma I_{ia} + X\theta + \varepsilon _{i, a}.} $$

The estimate for γ yields the causal reform effect. As before, we apply alternative specifications which use a simple post reform indicator instead of monthly date of birth fixed effects, C. Generally, γ is identified by the reform which changed the incentives jointly by age and cohort. In the case of reform 1 (NRA), intensity represents the magnitude of the benefit discount that follows upon immediate retirement for reform 2 (ERA) it represents the waiting time to ERA. As we are not conditioning our sample on meeting the unemployment pathway criteria, our estimates represent intention to treat (ITT) effects. We report robust standard errors that are clustered at the individual level.