Retirement age reforms and worker substitutability: Implications for employment of older workers

Badalyan, Sona Working Paper Retirement age reforms and worker substitutability: Implications for employment of older workers IAB-Discussion Paper, No. 14/2025 Provided in Cooperation with: Institute for Employment Research (IAB) Suggested Citation: Badalyan, Sona (2025) : Retirement age reforms and worker substitutability: Implications for employment of older workers, IAB-Discussion Paper, No. 14/2025, Institut für Arbeitsmarktund Berufsforschung (IAB), Nürnberg, https://doi.org/10.48720/IAB.DP.2514 This Version is available at: https://hdl.handle.net/10419/330590 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-sa/4.0/deed.de IAB-DISCUSSION PAPER Articles on labour market issues 14|2025 Retirement Age Reforms and Worker Substitutability: Implications for Employment of Older Workers Sona Badalyan ISSN 2195‑2663 Retirement Age Reforms and Worker Substitutability: Implications for Employment of Older Workers Sona Badalyan (Institut für Arbeitsmarkt‑ und Berufsforschung (IAB) & CERGE‑EI) Mit der Reihe „IAB‑Discussion Paper“ will das Forschungsinstitut der Bundesagentur für Arbeit den Dialog mit der externen Wissenschaft intensivieren. Durch die rasche Verbreitung von Forschungsergebnissen über das Internet soll noch vor Drucklegung Kritik angeregt und Qualität gesichert werden. The “IAB‑Discussion Paper” is published by the research institute of the German Federal Employment Agency in order to intensify the dialogue with the scientific community. The prompt publication of the latest research results via the internet intends to stimulate criticism and to ensure research quality at an early stage before printing. Contents 1. Introduction............................................................................... 6 2. Institutional setting and conceptual framework..................................... 12 2.1. Institutional setting..................................................................... 13 2.2. Conceptual framework and implications ............................................ 15 2.2.1. Firm’s problem of employment decisions .................................. 15 2.2.2. Wage determination under Nash bargaining............................... 17 3. Data ..................................................................................... 19 3.1. The Sample of Integrated Employer‑Employee Data .............................. 19 3.2. Sample construction for analyses .................................................... 20 4. Identification............................................................................ 21 4.1. Regression discontinuity design ...................................................... 21 4.2. Descriptive evidence on the presence of discontinuity ........................... 24 5. Results .................................................................................. 24 5.1. The effect of the rise in ERA on employment states .............................. 25 5.2. Robustness and sensitivity checks for the baseline RDD results ................ 26 5.3. The effect of the rise in ERA on wages.............................................. 28 6. Labor demand mechanisms: replacement costs..................................... 29 6.1. The role of job‑specific skills ......................................................... 29 6.2. The role of internal and external substitutability .................................. 33 6.3. The effect of raised ERA on wages by replacement costs ........................ 39 7. Conclusion .............................................................................. 41 References...................................................................................... 42 Appendix .................................................................................... 48 A1. Appendix figures ........................................................................ 48 B1. Appendix tables ......................................................................... 63 IAB‑Discussion Paper 14|2025 3 Abstract This paper studies how labor demand factors—specifically worker substitutability and job‑specific skills—shape employment responses to a rise in the early retirement age. Using a regression discontinuity design, I exploit a 1999 German reform that eliminated the option for women to retire at age 60. Before the reform, older workers could exit voluntarily, thereby imposing turnover costs on firms. Afterward, firms were better able to retain less substitutable workers for whom turnover costs are higher. At the same time, the loss of early pension eligibility reduced workers’ outside options, allowing firms to offer lower wages, often through partial retirement. Zusammenfassung Dieses Papier untersucht, wie arbeitsnachfrageseitige Faktoren – insbesondere die Ersetzbarkeit von Arbeitskräften und berufsspezifische Fähigkeiten – die Beschäftigungsreaktionen auf eine Anhebung des frühestmöglichen Rentenalters beeinflussen. Mithilfe eines Regression‑Discontinuity‑Designs analysiere ich eine Reform in Deutschland im Jahr 1999, die die Möglichkeit für Frauen abschaffte, bereits mit 60 Jahren in Rente zu gehen. Vor der Reform konnten ältere Beschäftigte freiwillig aus dem Erwerbsleben ausscheiden, was den Unternehmen Fluktuationskosten verursachte. Nach der Reform waren Betriebe besser in der Lage, schwer ersetzbare Arbeitskräfte mit höheren Austrittskosten zu halten. Gleichzeitig verschlechterte sich durch den Wegfall des vorgezogenen Rentenzugangs die Verhandlungsposition der Beschäftigten, was es den Unternehmen ermöglichte, niedrigere Löhne durchzusetzen – häufig in Form von Altersteilzeit. JEL H32, H55, J21, J24, J26 Keywords aging, raise in the retirement age, internal labor markets, human capital, worker substitutability IAB‑Discussion Paper 14|2025 4 Acknowledgments I thank Dan Black, Wolfgang Dauth, Randall Filer, Štěpán Jurajda, Andreas Mense, Nikolas Mittag, and Paolo Zacchia for their feedback; Dan Black for inviting me to the University of Chicago, where parts of this paper were written; Michael Moritz, Wolfgang Dauth, and the “Regional Labour Markets” Department at the Institute for Employment Research (IAB) for their belief in this project; Philipp vom Berge and Katja Wolf for help with understanding the data; Thomas Zwick for sharing a programming file that computes pension eligibility; Deborah Nováková for language editing. This paper benefited from presentations at HUN REN 2025; ifo Institute 2025; EWMES 2024; EALE 2024 & 2023; IAB Brown Bag and Regio Flash Talks 2024; ESPE 2024; IZA Summer School 2024; Dutch National Bank 2024; Young Economists Seminar (Croatian National Bank) 2024; SITES 2023; AIEL 2023 & 2022; Student Workshop at Harris School of Public Policy at UChicago 2023; BSE Summer School 2022; Czech Economic Society Biennial Conference 2022; Armenian Economic Association Annual Meeting 2022; CERGE‑EI Brown Bag 2023, Applied Microeconometrics Reading Group 2022, DW 2022 & DPW 2020 Seminars. This study was supported by Charles University, GAUK project No. 333221. This paper is part of a project that has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska‑Curie grant agreement No. 870245. IAB‑Discussion Paper 14|2025 5 1 Introduction The dynamics of labor markets are profoundly influenced by the interplay between worker substitutability and firm‑specific human capital. The ease with which workers can be replaced affects various labor supply decisions, including absences due to temporary illness (Hensvik/Rosenqvist, 2019), the duration of actual parental leave in reaction to extension of parental leave duration (Ginja/Karimi/Xiao, 2023) and increase of paid parental leave eligibility coverage (Huebener et al., 2024), and labor supply following a coworker’s death (Jäger/Heining, 2022). Worker substitutability has also been associated with wage losses after job displacement (Jacobson/LaLonde/Sullivan, 1993), as workers with more specific skills, such as those tied to a particular industry or occupation, face greater difficulty finding comparable jobs in the external labor market. However, the role of worker substitutability in the context of retirement, a significant driver of workforce turnover, remains underexplored. While substantial literature examines how statutory retirement age reforms impact labor supply (Atalay/Barrett, 2015; Brinch/Vestad/Zweimüller, 2015; Geyer/Welteke, 2021; Hanel/Riphahn, 2012; Hernæs et al., 2016; Lalive/Staubli, 2015; Lalive/Magesan/Staubli, 2023; Manoli/Weber, 2016; Mastrobuoni, 2009; Staubli/Zweimüller, 2013; Vestad, 2013), there is limited understanding of how labor demand mechanisms, such as job‑specific skills and worker substitutability, shape employment responses to such reforms because these papers often assume that labor demand is perfectly elastic at the relevant margins. In contrast, my paper argues that labor demand is not uniformly elastic and highlights the role of worker substitutability in shaping firms’ retention decisions. This paper aims to bridge this gap in the retirement literature by integrating insights from studies on worker substitutability with research on employment reactions to retirement reforms. Understanding this mechanism is crucial, as it offers deeper insights into how worker substitutability influences labor supply adjustments to retirement reforms and the coping strategies adopted by workers and firms. This challenges the standard assumption of uniformly elastic labor demand and offers new insights into the incidence and efficiency of retirement reforms. The seminal study by Becker (1962) posits that firm‑specific human capital renders incumbent workers less substitutable by external hires. In the context of reforms that raise the retirement age, this theory suggests that employment responses by older workers may exhibit substantial heterogeneity based on their substitutability and the specificity of the human capital required for their roles. A pertinent question arises: When early retirement options are curtailed, do firms respond uniformly across worker types, or do employment gains disproportionately accrue to those with more specific skills and lower substitutability? IAB‑Discussion Paper 14|2025 6 Such differences may reflect how firms and workers coordinate—depending on their turnover costs—in response to extended employment horizons. The demand for workers rises due to firm‑ or job‑specific human capital, or challenges in finding suitable replacements internally or externally. However, in the presence of outside options in the form of pensions, firms may have difficulties retaining such workers. Reforms raising the retirement age could help firms to retain such workers.1 Employment decisions at older ages are affected by many factors, including health, ability, income, and flexibility of contracts and firms; hence, in the absence of exogenous drivers, such decisions are likely endogenous at the individual level. Moreover, given an option to retire and receive a pension, workers may opt to exit the workplace and instead prioritize personal benefits (such as health, leisure time with family, etc) over firm factors (such as their substitutability and costs of replacement) in deciding to retire. A reform that raises the retirement age shifts the employment dynamics of those affected. I overcome this endogeneity challenge by studying the effects of a reform in Germany that abolished the women’s pathway to early retirement by making the statutory retirement ages gender neutral. This reform resulted in a sharp rise of at least three years (from 60 to 63) in the Early Retirement Age (ERA), the earliest age women could begin to claim a pension. This discontinuous policy change, which impacted women born from 1952 onward, provides a natural experiment for causally identifying the effect of raising the retirement age on employment and wages using a Regression Discontinuity Design (RDD), and exploring the relationship of worker substitutability with a large labor supply increase. The German labor market, characterized by substantial variation in worker substitutability2 and strong dismissal protections, offers a suitable setting for investigating whether workers delay retirement based on their skills and substitutability. The availability of comprehensive German establishment data, which encompasses entire workforces and employment histories, together with job cell data (3‑digit occupation groups within the establishments), enables analysis of internal markets, measurement of the availability of internal substitutes (workers sharing the same 3‑digit occupation), and a study of personnel practices employed by the establishments. To examine how employment responses to the rise in retirement age interact with worker substitutability, I start by sketching a simple model of the interplay between the reform that raises the age of the option to receive pensions, turnover costs, and employment decisions 1 Stole/Zwiebel (1996a) and Stole/Zwiebel (1996b) provided theoretical discussions of intra‑firm bargaining and its relation to firm‑specific human capital, while Lazear (2009) and Cahuc/Marque/Wasmer (2008) extended the discussion by arguing that, similar to firm‑specific human capital, the ease with which a firm can find a suitable replacement could affect the wages of workers. However, having lower bargaining power after removal of the option to receive pensions, firms may be in a stronger position than workers. 2 Previous literature for Germany has shown that frictions in replacing workers are important (Jäger/Heining, 2022; Huebener et al., 2024). IAB‑Discussion Paper 14|2025 7 at 60‑62. I also outline a Nash bargaining model with implications for the effects of the reform and of substitutability on wages conditional on employment at 60‑62. To test these implications empirically, I first construct several proxies for worker substitutability (and therefore turnover costs). First, I examine whether workers with specific skills are more likely to be retained at older ages. Specific human capital and managerial roles are key determinants of worker substitutability, as external replacements for these skills are often scarce (Baker/Gibbs/Holmstrom, 1994)3. Consistent with theories of firm‑ and job‑specific human capital (Becker, 1962), Bertheau (2021) shows that jobs requiring teamwork and training with senior workers are more often filled internally. In the context of retirement reform, this suggests that establishments where older workers’ positions rely on job‑ or firm‑specific human capital may benefit most from the extended retention of older workers. Next, I explore internal (coworkers in the same occupation) and external (potential hires in a commuting zone for a given occupation or industry) labor market thickness. According to Topel/Ward (1992), both internal and external labor markets affect workers’ life‑cycle labor market outcomes. In thin labor markets, finding suitable replacements is more challenging, making worker turnover costly for firms (Lazear, 1979). Automation can substitute for some types of labor, leading to reduced employment and wages, particularly in economies with aging populations like Germany (Acemoglu/Restrepo, 2022). Hence, I test whether the substitutability matters beyond the worker level, by dividing occupations by routineness, a proxy for substitution by automation. Finally, I consider the tradability of industries as another dimension of worker substitutability. Firms in tradable industries can replace workers not only locally but also by outsourcing tasks globally, increasing substitutability (Drenik et al., 2023). While characteristics such as managerial status or skill specificity may reflect both firm‑side costs and worker‑side preferences, I interpret heterogeneity in the reform’s effects primarily through the lens of firms’ retention incentives — that is, the labor demand channel. My findings confirm the implications of the model and indicate that the reform increased employment among women aged 60–62 by 17.3 percentage points (a 22% increase relative to the control mean of workers who were eligible to retire at 60). These results are robust to variations in model specification. To gauge the potential scale of the reform’s impact, I conduct a back‑of‑the‑envelope calculation. This treatment effect would translate into roughly 540,000 additional women remaining employed at ages 60–62 due to the reform.4 Conditional on employment, the workers whose retirement age rose by the reform are less likely to bargain for higher wages at ages 60–62, compared to those previously eligible for pension benefits. The reform removed access to early retirement, weakening outside 3 See also Bartel et al. (2014), Friedrich/Hackmann (2021), Jäger/Heining (2022), Jaravel/Petkova/Bell (2018) 4 This is a rough calculation based on local treatment effects for women born in 1951–1952, who were employed continuously at 58‑59 years old. The estimate assumes that the sample is nationally representative and that the effect generalizes across cohorts affected by the reform. It does not adjust for compositional differences or cohort trends and should be interpreted as illustrative. IAB‑Discussion Paper 14|2025 8 2.2 Conceptual framework and implications Firms operating in imperfect labor markets face frictions in replacing experienced workers, particularly those with occupation‑ or firm‑specific skills. As these workers approach retirement age, firms risk productivity losses and incur hiring costs due to turnover. Early retirement eligibility grants workers considerable autonomy in deciding when to exit the labor force. This paper studies how a policy reform that raised the early retirement age (ERA) from 60 to at least 63 alters the interaction between worker substitutability and retirement behavior. The model builds on the idea that retirement is not only a worker’s choice, but also reflects the relative bargaining power of workers and firms. When early retirement is an option, workers with valuable skills may leverage this as a bargaining chip in wage negotiations. When the option is removed, firms can retain even valuable workers without raising wages. To understand how firms respond to a rise in early retirement age, I first develop a static model in which the firm’s decision to retain a worker depends explicitly on the worker’s substitutability and the policy environment. I then extend the framework with a Nash bargaining model, allowing wages to be endogenously determined. This yields testable implications conditional on employment.10 2.2.1. Firm's problem of employment decisions Setup. Consider a firm employing worker i, who is approaching retirement age. Continued employment at ages 60–62 depends on whether the match between the worker and the firm remains viable. I model this using a latent retention condition, in which both the firm and the worker must benefit from continued employment. While the firm’s willingness to accommodate employment reflects the economic value of the match, the worker’s outside option plays a key role in the joint decision. I do not model active dismissal, consistent with strong employment protections. Instead, I interpret “retention” as the match continuing when both parties find it preferable to separation. The reform removes a key voluntary exit channel (early retirement), extending employment among older workers, especially those with high specificity. 10 The theoretical framework presented in this section builds on Nash bargaining models of labor market frictions (e.g., Pissarides (2000)), adapting them to retirement contexts by incorporating outside options shaped by policy. It also draws on Acemoglu/Pischke (1999) in the implications of firm‑specific skills for wage setting and turnover, and from Gruber/Wise (2008) the responsiveness of retirement to institutional incentives. Lastly, the interaction between substitutability and tax incidence in determining the incidence of adjustment costs is conceptually linked to Gruber (1997), and is applied here to changes in retirement age. IAB‑Discussion Paper 14|2025 15 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎧ ⎫ ⎪ ⎪ ⎨ ⎬ max πi = di · (y − o(Ri, si)) +(1 − di) · (−c(si)) (1) di∈{0,1} | {z } | {z } ⎪ ⎪ ⎩ surplus if cost if ⎭ match continues match ends (∂c(si) > 0) ∂si where: yi is the worker’s output. I abstract from heterogeneity in output across workers, as substitutable workers may be either more or less productive depending on job fit and skill specificity.11 The model uses a single specificity parameter si to capture employer‑side turnover costs. These costs increase when workers are more difficult to replace (due to specialized knowledge or task‑specific skills) and when they have limited outside options (due to thin external markets for their skills). In this sense, specificity si represents a reduced‑form measure encompassing both skill specificity and substitutability. o(Ri, si) is the outside option, shaped by the policy reform Ri and worker specificity si. It is decreasing in Ri (∂o(Ri,si) < 0) ∂Ri because pension eligibility is delayed post‑reform, and decreasing in si because workers with specific skills face thinner external labor markets, especially after age 60 (∂o(Ri,si) < 0) ∂si . Such specificity could, for example, decrease the likelihood of leaving the social insurance for other pathways than retirement (move to another country, start self‑employment, etc.).12 c(si) denotes the replacement cost of an employee, increasing in specificity, and reflecting dismissal, severance, and other compensation payments, as well as hiring, training, and productivity ramp‑up costs that rise with the degree of human capital specificity . Solution. The match continues if the joint surplus from continuing exceeds the cost of separation: y − o(Ri, si) + c(si) > 0 (2) Interpretation. While the equation is modeled as a firm‑side optimization problem, the 11 I thank Wolfgang Dauth for this discussion. 12 Such argument is also in line with the finding of literature on displaced workers: those unemployed who switch to another industry or occupation experience much larger declines in earnings (e.g. Neal (1995); Addison/Portugal (1989)). IAB‑Discussion Paper 14|2025 16 ∂ do(Ri, si) dc(si) (y − o(Ri, si) + c(si)) = − + > 0 (4) ∂si dsi dsi | {z } | {z } <0 >0 ∂ do(Ri, si) (y − o(Ri, si) + c(si)) = − > 0 (3) ∂Ri dRi | {z } <0 interpretation reflects a joint agreement between the firm and the worker: continued employment occurs when both benefit relative to separation. The reform shifts this condition by removing early retirement as a fallback, thus altering the outside option o(Ri, si) and increasing the likelihood of match continuation, especially for workers with high specificity. Implication 1: A higher ERA rises employment of older workers. Delaying pension eligibility reduces o(Ri, si), making continued employment more attractive to the firm and less avoidable for the worker, thereby increasing employment. Implication 2: Workers with higher specificity (si) are more likely to remain employed. Less substitutable workers are more likely to remain in employment due to both weaker outside options and higher replacement costs. As a result, the joint surplus of continued employment is larger, sustaining the match. 2.2.2. Wage determination under Nash bargaining Conditional on retention (di = 1), the firm and the worker bargain over the wage wi based on the total surplus generated by employment: Si = y − o(Ri, si) (5) With worker bargaining power β ∈ (0, 1), the Nash wage splits the surplus between the worker and the firm. The wage thus depends on both the worker’s productivity y and their outside option o(Ri, si). The general form of the Nash‑bargained wage is: IAB‑Discussion Paper 14|2025 17 wi = β · y + (1 − β) · o(Ri, si) (6) |{z} | {z } productivity‑based reward outside‑option fallback ∂wi do(Ri, si) = (1 − β) · < 0 (7) ∂Ri dRi | {z } <0 Intuition. The worker’s wage is a weighted average of what they contribute to the firm’s output (through y) and what they could earn elsewhere (via o(Ri, si)). Workers with high productivity naturally command higher wages, all else equal. However, their outside option—such as retirement income or alternative employment—also determines their bargaining position. If their fallback option weakens, the firm can offer a lower wage even if the worker is productive. To understand how wages change due to the reform and differences in substitutability, I examine how wi responds to changes in Ri and si. Implication 3: The reform lowers wages via weaker outside options. When the policy raises the early retirement age (i.e., increases Ri), the outside option o(Ri, si) declines. This reduces the worker’s fallback position in wage negotiations, shifting surplus toward the firm. The wage falls even though the worker remains employed. This mechanism is stronger when the worker has low bargaining power β, and when the reduction in outside options is large. Implication 4: The effect of specificity on wages is ambiguous. ∂wi dy(si) do(Ri, si) = β · + (1 − β) · (8) ∂si dsi dsi Higher specificity si affects both productivity and outside options. If more specific workers are more productive (dy(si) > 0 dsi ), then wages may increase through the first term. However, specificity also reduces outside options (do(Ri,si) < 0 dsi ), which lowers wages through the second term. If outside options deteriorate faster than productivity improves, or if the worker has low bargaining power, the overall effect on wages may be negative. IAB‑Discussion Paper 14|2025 18 Intuition. Even if the worker is valuable to the firm (due to difficult‑to‑replace skills), the firm may exploit their lack of external alternatives. The reform amplifies this asymmetry by removing early retirement as a viable fallback, especially for workers in thin external labor markets (e.g., managers, specialists). This is a form of monopsony power, where the employer’s ability to set wages below marginal product is strengthened by the worker’s limited exit options. Summary. Implications three and four jointly imply that the wage response to the reform depends on the interaction between substitutability and bargaining frictions. For workers with low specificity (who are easy to replace), wages fall mostly due to the loss of retirement options. For workers with high specificity, the story is more nuanced: while their productivity makes them costly to replace (increasing employment), their weakened fallback position gives the firm the leverage to suppress wages. Thus, employment may rise while wages fall or stagnate, despite high skill specificity, due to increased employer monopsony power post‑reform. 3 Data This section consists of two parts. First, I describe the data I utilize, its sampling procedure, and its suitability to my research question. Second, I describe how I constructed my sample, the reasoning behind each restriction, and the resulting sample size. 3.1 The Sample of Integrated Employer-Employee Data I use the Sample of Integrated Employer‑Employee Data (SIEED7518), a random 1.5% sample of all establishments in Germany. The establishment identifiers are fixed by industry, ownership, and location at the municipality level; hence, an establishment is not equivalent to a firm in all cases. Nevertheless, I use the terms firms and establishments interchangeably. Employers are obliged to report data on all of their employees subject to social security contributions. Self‑employed and civil servants are not covered by the data. At the end of each year, employers report the start and end date of employment, wages, and other occupational, educational, and demographic indicators of all of their workers. Typically, the data is a snapshot of the employment state as of June 30th of each year. Employers are also obliged to report changes in employment contracts.13 13 One of the data limitations is the lack of working hours; hence, I am limited to the analyses of only the extensive margin of employment. IAB‑Discussion Paper 14|2025 19 For each of these establishments, the entire employment biographies of all employees are included over the observation period 1975‑2018 for West Germany and 1992‑2018 for East Germany. Hence, the data also include the establishments that did not constitute the random 1.5% of the establishments originally sampled, in case the workers from the establishments originally sampled were ever employed elsewhere. Observing the entire workforce of the sampled establishments is critical for my analyses, because I study substitutability mechanisms behind employment reactions to the raise in retirement age, which requires observing all coworkers of a given establishment. Schmidtlein/Seth/Vom Berge (2020) describe the data sampling in more detail. 3.2 Sample construction for analyses To construct the final sample for my analysis, I keep only women born in 1951, the control group, i.e., women who were potentially eligible for wfor the women’s pathway to early retirement, if they accumulated enough years of social security contributions in later life; and 1952, the treatment group, i.e., women who experienced the rise in the women’s ERA. I drop women who were ever employed as miners and sailors (for clarity) because their retirement rules differ from those in other occupations.14 To address the issue of parallel spells in the data, which is possible, for example, due to dual earners (employed at several establishments simultaneously), I keep the spells in the randomly selected 1.5% establishments. If both spells come from randomly sampled establishments, I keep the spells where the worker accumulated more tenure. In cases where the employee works in two randomly selected establishments and has accumulated an equal amount of tenure in each of them, I keep the job with the highest wage. Dropping parallel spells allows me to construct Panel data and study the firm mechanisms for only the establishments to which the dual workers are more attached. The final data consists of person‑age entries (in age‑month), where I observe women from the age of 42 (age‑month 504) until 66 (age‑month 792). The choice of this time frame is driven by the fact that the first affected cohort was 47 years old at the time of the reform announcement in 1999, and in some of my analyses I want to observe employment (1) before the reform announcement, (2) between the reform announcement and its inaction at 14 The seminal work by Geyer/Welteke (2021) on labor supply responses to the 1999 reform makes a restriction of keeping only women who are eligible for the women’s pathway to retirement at the age of 60. I make restrictions that proxy for eligibility, following Lorenz et al. (2018). I do not explicitly make sample restrictions that keep the women eligible for the women’s pathway (e.g., 15 years of contributions in total and ten years after 40 years old, etc), because I do not observe the unemployment spells that also contribute to the contribution years. Because unemployment spells still count towards the contributions to social security, not making this restriction results in smaller treatment effects in my sample, compared to that of Geyer/Welteke (2021) IAB‑Discussion Paper 14|2025 20 60, (3) and workers who continue working beyond both the ERA (60 or at least 63) and NRA (65 or 65.5). First, studying employment before the reform announcement shows whether the treatment and control groups had different labor supply frequencies before the reform announcement. Second, studying employment between 47‑60 can show whether the rise in ERA leads to different employment choices during middle age, in expectation of a longer employment period. Finally, studying the effects beyond the new ERA shows how the effect of raising ERA also spills over to post‑ERA employment, which could show indirect employment effects beyond the age targeted by the reform, further increasing its effectiveness in keeping workers in employment longer. I keep workers who are continuously (in each age‑month) employed at 58 and 59. To make such restriction plausible, I have to assume that the employment at 58‑59 is not is unaffected by the reform. Geyer/Welteke (2021) show that there are no employment effects before the age of 60, therefore, such restriction is not likely to lead to a selection bias. Because most of the main heterogeneity variables are constructed at the establishment level, this restriction helps me to obtain a sample of workers with sufficient attachment to their establishments. The final data consists of 32,770 workers, and 9,036,582 worker‑age months (Table B1 records the number of workers after each restriction). Out of these workers, 15,640 are in the control group (born in 1951), and 17,130 are in the treatment group (born in 1952). 4 Identification First, I describe the identification strategy based on reform discontinuity in birth dates, and then I provide some descriptive results that confirm the presence of discontinuity in the data. 4.1 Regression discontinuity design I follow Geyer/Welteke (2021) to locally identify the effect of the reform that raised the ERA on employment, τm, in an RDD framework15: 15 There are several differences from the identification in Geyer/Welteke (2021). First, I do not control for the presence of children in my RDD regression as I do not observe such variables in the data. Second, because the most recent year observed in my data is 2018, the data allow me to pool all the age months corresponding to 60‑62 years of age in the baseline regression and beyond 63 in the supplementary analyzes, while Geyer/Welteke (2021) pooled only 60‑62 due to their right‑censored data in 2016. Finally, I use the mean square‑based optimal bandwidth, while they use a 12‑month ad‑hoc bandwidth selection procedure. IAB‑Discussion Paper 14|2025 21 yim = αm + τm 1 {b i ≥ b ∗ }+ (9) ′ + β0m 1 {b i < b ∗ }(bi − b ∗ ) + β1m 1 {b i ≥ b ∗ }(bi − b ∗ ) + Xiβm + ϵim where yim ‑ is employment state, recorded for each woman i at every age‑months m; bi is the birth cohort of the individual i; 1 {b i ≥ b∗} is an indicator showing that i was born after the cutoff b∗ (January 1952), i.e., experienced the rise in the ERA (treatment group); while 1 b i < b∗}{ includes the individuals who are below the cutoff (control group). I use a local linear regression, and by interacting the running variable (bi − b∗) with the treatment indicator, I allow for different slopes in treatment and control groups. Figure 1 shows that a linear trend in the running variable is a plausible assumption, and there is a clear discontinuity that is unlikely to be attributed to a wrong functional form of polynomials. To compute the RDD estimates, I use a triangular kernel function and the optimal bandwidth choice based on mean square error (Imbens/Kalyanaraman, 2012). As a result, I calculate the bias‑corrected RDD estimates with a robust variance estimator. I also control for calendar month, a dummy for Western German residence, wages at the age of 46, and two education categories (out of 3), because previous literature confirms that education is an important determinant of employment at an older age (Geyer et al., 2022). I cluster the standard errors at the birth month level to account for the potential correlation of standard errors ϵim for the women belonging to the same birth cohort.16 In robustness and sensitivity checks, I re‑run the regressions, altering all the specification parameters‑ the procedures for estimating the parameters and covariance matrices, polynomial order, kernel weights, bandwidth choice, included covariates, and clustering level. The baseline regressions pool the 60‑62 age (720‑756 age months) together, because this is the age frame that was affected by the ERA reform. This identification results in a local average treatment effect of higher ERA on employment outcomes at ages 60‑62 (coefficient τm in equation Equation 9).17 Identification assumptions. This identification relies on two main assumptions. (1) Smoothness in density. This assumption requires continuity of the running variable (birth cohort) around the cutoff, which eliminates the possibility of strategic bunching (manipulation of the treatment status) at the cutoff. This assumption holds by construction 16 Clustering at the level of birth dates aligns with literature suggesting clustering the standard errors at the treatment level. 17 Because I cannot claim that all the women included in my sample were eligible for women’s pathway to early retirement, the coefficient could also capture the Intention‑to‑Treat (ITT) effect. However, Lorenz et al. (2018) show which sample restrictions are likely to lead to eligibility imputations, and because most of my restrictions match their proposed restrictions, my sample likely captures most of women eligible for the women’s pathway to early retirement. IAB‑Discussion Paper 14|2025 22 because it is impossible to change one’s own birth date.18 Nevertheless, in the sensitivity tests, I re‑estimate the main regressions by omitting the observations close to the cutoff and confirm the robustness of the results. (2) Smoothness in covariates. This assumption requires continuity of the distribution of the observed and unobserved variables around the threshold, showing that the assignment of the treatment around the cutoff is as good as random. Table B2 shows that there is no sizeable significant discontinuity in pre‑determined variables. In particular, I choose a variable showing whether a woman has Western origin (proxied by the place of living according to the first biographical spell) and nationality, as these variables are fixed over time and hence are pre‑determined. Main outcome variables. In terms of outcome variables, at each age month, I create three mutually exclusive main labor market categories ‑ employment, nonemployment, and retirement. I further disentangle the employment into three groups‑ employees liable to social security, marginal part‑time employment, and partial retirement. Nonemployment stands for a gap in the employment age‑month spells. I proxy retirement with the last labor market activity of a worker. Figure A2 displays the evolution of the three main employment states over age by treatment status, i.e., the gap in employment and retirement statuses at 60‑62. In addition to these employment state categories, I also define wages, because I am also interested in wages conditional on employment.19 Wages are created at the detailed monthly level, and are non‑zero only if the worker is employed. Effect heterogeneity. To study the mechanisms behind these effects, I perform subsample analysis using several categories of variables, which show turnover costs associated with retirement in the next section. Because the research question relates to the labor demand factors influencing employment at ages 60‑62, I define these variables at the age of 58, just before the pre‑reform retirement age of 60. 18 One could argue that the reform cohorts could be chosen by policy‑makers in a way that violates the assumption, for example, by the cohort of baby‑boomers, etc. However, because I compare cohorts born around the cutoff, and the cutoff does not appear in any other reforms, policies, or characteristics (both of these cohorts are typically classified in the baby‑boomer generation) that would make the 1951 cohort different from the 1952 cohort, there is no reason to believe that the assumption is likely to be violated. 19 Although wages are top‑coded in the social security data, this data feature is unlikely to constitute an issue for the analyses as women are less likely to cross the threshold for wage censoring. IAB‑Discussion Paper 14|2025 23 4.2 Descriptive evidence on the presence of discontinuity The abolishment of women’s pathway to early retirement led to a large increase in employment rates at 60‑62, as shown in the right Panel of Figure 1. While overall there is an upward‑sloping employment trend at 60‑62 over the birth cohorts, there is also a clear discontinuity around the 1952 cohort. Only around 75% of women born in 1946‑1951 were employed at 60‑6220. However, the employment rate jumped to approximately 90 percent starting with the 1952 cohort, the reform cutoff. Figure A3 extends the analyses to display employment rates by treatment status at all age months (corresponding to the ages between 42 and 66), and confirms the presence of a discontinuity in employment rates at 60‑62 (due to the 1999 reform that I study) and to a smaller magnitude of discontinuity at the ages 65‑65.5 (due to the 2007 reform). Estimating the treatment effects of the 2007 reform is beyond the scope of this paper; hence, in the next section, I causally quantify the largest employment discontinuity that happens due to the 1999 reform, i.e., at 60‑62. 5 Results In this section, I first focus on the effect of the 1999 reform on employment, confirming the results of prior studies on this reform (Geyer/Welteke, 2021). I show the effects of retirement on employment trajectories before studying the labor demand mechanisms of employment, because I want to provide a general picture of the labor supply behavior overall before zooming in on the total employment mechanisms. Based on the theoretical framework, I expect that the rise in the ERA should extend employment among affected workers, particularly those whose exit would impose high turnover costs on firms. These costs are likely higher for workers with specific skills or those employed in occupations with limited internal or external substitutes. Therefore, I expect the employment effects of the reform to be stronger for such workers. On wages, the model predicts ambiguous effects depending on workers’ outside options and replacement difficulty: lower bargaining power due to the loss of pension eligibility may lead to wage decreases, while high replacement costs for specific or non‑substitutable workers could result in wage premiums to incentivize retention. 20 This control mean is higher than that in existing literature studying the labor supply response of this reform (Geyer/Welteke, 2021), likely because the sampling of SIEED and my sample restriction (employment at the ages 58‑59) results in a sample of workers who are more attached to the labor force. IAB‑Discussion Paper 14|2025 24 Figure 3.: The effect of the rise in ERA on employment at ages 60‑62 by return to experience in a given occupation and occupational hierarchy level Notes: Coefficient plots for RDD regressions around the 1952 cutoff. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and the highest education. The subsample analysis in the left Panel is performed by the human capital specificity of occupation. The right Panel stands for managerial status. The vertical lines indicate 95% confidence intervals based on robust stan‑ dard errors clustered at the birth month level. The control means (on the x‑axis) show the employment share at the ages of 60‑62 in the corresponding subsample over the control group (born in 1951). A corresponding table with more details can be found in Table B12. Managerial status. Managerial occupations often entail a higher degree of firm‑specific and occupation‑specific human capital, due to their reliance on accumulated institutional knowledge, leadership skills, and relationship‑specific investments within the firm. Managers are typically more difficult to replace than are non‑managers, particularly at older ages when experience and firm‑specific knowledge peak. Therefore, distinguishing between managers and non‑managers offers a meaningful way to capture heterogeneity in turnover costs and the value of worker retention following a rise in the early retirement age. Even if two workers have the same returns to experience, managerial roles may imply extra firm‑specific value. In a related study, Jäger/Heining (2022) find that the death of a manager or a worker in a specialized occupation results in more negative effects on the coworkers in other occupations. In my setting, if a worker is a manager, she likely has many coworkers under her hierarchy, communicates with them more, and has more information, making her less substitutable, and thus making the extension of her working life more valuable. I create a variable showing managerial or supervisory status based on the last two digits of the 5‑digit occupations. I pool the supervisors and managers into the dummy variable manager. 28 Depending on occupation type, some occupational hierarchies have managers, while others have supervisors as the highest occupation level in a hierarchy. I thank Philipp vom Berge for the help with the data. IAB‑Discussion Paper 14|2025 28 31 The right Panel of Figure 3 shows that workers in managerial positions are significantly more likely to remain employed at older ages in reaction to the reform. The workers in managerial positions extend their employment by 43.1 p.p. (55.3% relative to the control mean‑ the managers whose retirement age was not altered by the reform), while the non‑managers raise their retirement ages by 17.1 p.p. (22.1% increase relative to the control mean). The difference in point estimates (26 p.p.) suggests that the employment response to the treatment is substantially larger for workers performing managerial occupations. Alternative measures of skills and specificity. To test whether the results presented above are sensitive to approximating worker skills, I explore alternative proxies for worker skill specificity. The baseline analysis relies on hierarchical job positions as indicators of skill‑specific roles. As an alternative, I use an occupational classification by Blossfeld (1985). This classification groups occupations into ten categories and shows the occupational split by required skills‑ simple vs professional.29 Figure A6 shows that across all occupational groups, workers in skilled (i.e., professional) categories exhibit greater employment gains after the reform than those in corresponding simple roles. Managers and professionals are particularly likely to remain employed longer, reinforcing the idea that skills and job specificity drive retention. Although one might suspect that this is driven by longer job tenure, Figure A7 shows that employment gains do not increase monotonically with tenure. This suggests that hierarchical position captures more than tenure alone.30 I further explore whether employment responses differ across occupational task types, offering another dimension of worker substitutability. Following Dengler/Matthes/Paulus (2014), I categorize jobs along two dimensions: (i) skill content—analytical, interactive, cognitive, or manual tasks—and (ii) routineness—routine vs non‑routine.31 High‑skilled workers typically perform analytical or interactive tasks, while low‑skilled workers perform manual non‑routine tasks. Figure A14 shows that workers in high‑skill task occupations experience the largest post‑reform employment extension. In contrast, workers in routine occupations—often more replaceable by automation—do not exhibit systematically different employment responses. This suggests that, in the studied period, task routineness and the potential for automation play a lesser role in driving employment effects than overall skill specificity. While automation may become a more relevant channel in the future, the evidence here points primarily to skill‑related substitutability as the key mechanism. 29 I use the codes from material published by Schmieder/von Wachter/Bender (2016) to implement this classification. Education level is not a suitable candidate for skill differentiation in this context, as it is directly controlled for in the baseline specification due to institutional reasons (chapter 2 and chapter 4). The results by education level are shown in Panel C of Table B22 and exhibit no meaningful differences. 30 Tenure is an imperfect proxy for skills in this context because eligibility for retirement at age 63 depends on tenure. Thus, its use conflates eligibility rules with substitutability. 31 This classification is matched to my main data using the 3‑digit occupation identifier. Task types include analytical non‑routine, interactive non‑routine, cognitive routine, manual routine, and manual non‑routine. IAB‑Discussion Paper 14|2025 32 6.2 The role of internal and external substitutability The next group of variables showing the turnover costs and substitutability of workers is based on the markets‑ internal (by availability of coworkers in the same job cell as an older woman) and external (potential hires in the local labor market). The main motivation for studying internal labor market thickness is that the scarcer the job performed is, the more difficult it is for the employer to replace potential retirees with coworkers, thus leading to higher employment responses to the retirement reform. Internal substitutability is particularly important, as internal workers are imperfect substitutes for external workers (Jäger/Heining, 2022); hence, often the internal substitutes weigh more than the external substitutes. When fewer workers are working in the specific occupation of an older woman in the commuting zone, the less substitutable such a woman is. Similarly, when fewer workers are working in the specific industry of an establishment in local labor markets, the less substitutable the older women of such establishments are by external hires. Availability of internal substitutes. To capture internal substitutability, I use the number of available coworkers in the same 3‑digit occupation as women born around the reform cutoff. I count only workers in employment positions subject to social security. Following Huebener et al. (2024), I define three categories of such variables by the availability of coworkers in the same 3‑digit occupation as the affected women: 0, 1‑4, and 5 or more internal substitutes. I perform the analyses for establishments with fewer than 100 workers, as the levels of substitutability will be less dependent on establishment size (such restriction also closely follows Huebener et al. (2024) definitions). The left Panel of Figure 4 shows that when there are no coworkers who perform the same job as the older workers, the older workers are more likely to remain employed at 60‑62 following a retirement reform. The group with more than five substitutes has significantly lower employment responses than those with 0 coworkers in the given job cell. Workers who have no internal replacements respond to the reform by extending their employment by 26.8 p.p. (35% increase relative to the control mean of workers who were allowed to retire at 60 and were employed in non‑substitutable establishments). While the effects are insignificant for the group of workers who have between one and four coworkers, the workers who have more than five coworkers in the same occupation extend their employment by 6.5 p.p. (7.9% increase relative to the control mean). The difference in point estimates (20.3 p.p.) suggests that the employment response to the reform that raised the ERA is substantially larger for workers who have no internal substitutes, relative to those IAB‑Discussion Paper 14|2025 33 who have at least five internal substitutes, in line with the prediction that firms retain workers who are more difficult to replace.32 Figure 4.: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by num‑ ber of internal and external substitutes for the given occupation Notes: Coefficient plots from RDD regressions around the January 1952 cutoff. The estimates are obtained us‑ ing local linear regressions with first‑order polynomials, a triangular kernel, and mean square error–optimal bandwidth selection. Controls include calendar month of birth, Western German residence, wages at age 46, and education. Subsample analyses are conducted by internal substitutability in the left Panel and external substitutability in the right Panel. Internal substitutability is measured by the number of coworkers in the same 3‑digit occupation as the old worker, restricting the sample to establishments with fewer than 100 workers. The right Panel shows external labor market thickness (ELMT), based on the commuting zone at most half as concen‑ trated in a given occupation relative to the country‑level (ELMT < 0.5), or at least half as concentrated but less concentrated than the country‑level (0.5 < ELMT < 1), and at least as concentrated as the country‑level concentration (ELMT > 1). Vertical lines represent 95% confidence intervals based on robust standard errors clustered at the birth‑month level. Control means (on the x‑axis) refer to the average employment rate at ages 60–62 among the control group within each group’s optimal bandwidth. The corresponding detailed tables are reported in Table B13 and Table B17. External labor market thickness (ELMT). I define ELMT in two steps. First, I create 141 local labor markets based on high within‑region and low between‑region commuting for work, following Kropp/Schwengler (2011). Next, I create an index ELMT kc , showing the local labor market share of 3‑digit occupation (or industry) employment (E kc /E c ) over the national share of occupation (or industry) employment (E k /E). I count only workers between 18 and 64 years old who are either in employment subject to social security contributions or trainee workers. 32 Figure A9 repeats the analyses for all establishments, regardless of size. When at least five coworkers perform the same job as a woman, the effects are still large, despite being slightly smaller (but not significantly smaller) than those of women with no internal worker substitutes. This pattern could be driven by the variation in treatment effects by establishment size. Indeed, in larger establishments, women are more likely to work longer in reaction to the reform than those in smaller establishments (Figure A8); hence, when analyzing internal substitutability, it is important to account for the establishment size by restricting the sample to those with at most 100 workers. Even in large establishments, if there are no internal substitutes, the effects are quite large, which highlights that, although in large firms workers stay in employment longer, those who have no substitutes still work longer regardless of the establishment size. IAB‑Discussion Paper 14|2025 34 Ekc/Ec ELMTkc = (10) Ek/E where k is a 3‑digit occupation (or industry), and c is a commuting zone, Ekc shows the number of workers employed in the occupation (or industry) k, and in the commuting zone c, Ec is the number of workers employed in the commuting zone c and all the occupations (or industries) together, Ek is the number of workers employed in the occupation (or industry) k in all the commuting zones together, while E is the number of workers employed in all the occupations (or industries) and all the commuting zones together (i.e., country).33 Figure 5 displays an example of this index construction for the nursing occupation and hospital activities industry. While Passau has many workers employed in these industries relative to the national level, Leipzig does not. This means that for an establishment located in Leipzig, an older worker in a given occupation and industry is more valuable (i.e., such a worker is associated with higher turnover costs) than for an establishment located in Leipzig. I call an external labor market thick if this index is over 1, i.e., if the thickness of an occupation (or industry) in a given commuting zone is denser than the thickness at the national level. Additionally, I define a group where the index ELMTkc is below 0.5 (i.e., the commuting zone at most half as concentrated in a given occupation or industry as the country‑level), between 0.5 and one (at least half as concentrated but less concentrated than the country‑level). 33 All of these variables are defined based on my SIEED data, but becuase the sample is representative of all German establishments in the country (and the random sampling provides representativeness of workforce subject to social security at the commuting zone level), I expect these indices to proxy the country‑level index well. IAB‑Discussion Paper 14|2025 35 Figure 5.: External labor market thickness by German industry and occupation in 2010 Notes: This map shows the computed external labor market thicknesses (ELMT) for each of the 141 local labor markets based on the Kropp/Schwengler (2011) classifications, which are constructed based on high within‑ region and low between‑region commuting. I compute ELMT based on Equation 10 for the industry and occu‑ pation largest share of female employees: “Hospital activities industry” (left Panel) and “Nursing occupation” (right Panel). I plot the ELMT indexes (Equation 10) on the map based on the ten deciles presented in the left corner of each graph. IAB‑Discussion Paper 14|2025 36 The right Panel of Figure 4 displays the RDD results split by external labor market thicknesses of occupations. If women are employed in a commuting zone at most half as concentrated in a given occupation as the country‑level, they extend their employment by 41.3 p.p. (58.4% increase relative to the control group). I find that if a woman is employed in a commuting zone at least half as concentrated but less concentrated than the country‑level, the employment increase is 15.6 p.p. (19.9% increase relative to the control mean). Finally, in the commuting zones in which a given occupation is more represented than at the national level, the reform leads to an 11.7 p.p. increase in employment at ages 60‑ 62 (15.3 % increase relative to the control mean). This increase in employment is 29.6 p.p. lower than in commuting zones at most half as concentrated in a given occupation as the country‑level. This result indicates that the response to the reform that raised the retirement age is higher for workers in occupations with thin external labor markets, where they are less substitutable than in thicker markets. I examine heterogeneity in employment effects along external labor market thickness at the industry level (Figure A11). Unlike the baseline occupation‑based results, which showed clear differences by substitutability, I find no significant heterogeneity in responses across industries with different levels of labor market thickness. One potential explanation is that industry‑level measures are too broad to capture substitutability for specific skills or tasks. Additionally, larger firms, which are included in the full sample, may be less affected by external labor market conditions because they can rely more on internal replacement options. To account for this concern, I re‑estimate the analysis for a subsample of establishments with fewer than 100 employees, where firms are less likely to rely on internal labor markets. In this subsample, the effects of the reform do differ significantly by industry‑level labor market thickness: I find that workers are more likely to remain employed in industries in which the external labor markets are thin. (Figure A12). This suggests that external substitutability matters more when firms face tighter external constraints and cannot rely on internal hires. Overall, the occupation‑based measure of external substitutability remains more informative than the industry‑based measure, because thick industry labor markets may reflect broader agglomeration patterns rather than job‑level substitutability. Moreover, industry thickness may not map well onto the specific skills that firms need to replace. Gender‑specific substitutability. The main results rely on gender‑neutral measures of external labor market thickness (ELMT), pooling employment densities of both men and women. However, Germany exhibits pronounced occupational and industry segregation by gender, and the reform exclusively affected women. If women face limited competition or IAB‑Discussion Paper 14|2025 37 hiring barriers in male‑dominated fields, their effective substitutability may depend on the gender composition within occupations and establishments.34 To explore this, I construct a gender‑specific version of the ELMT index using only female employment densities (a modification of Equation 10) and re‑estimate the main analysis across the previously defined three ELMT categories. As expected, the variation in the female‑specific ELMT is smaller, and the results become statistically insignificant (Figure A13). One possible explanation is that employers do not confine their replacement pool to women and may consider male hires instead. In such cases, a gender‑neutral ELMT measure may better reflect the labor supply elasticity firms actually face. However, this interpretation is not definitive. Relying solely on female data reduces statistical power, and the resulting ELMT measure may be noisier and less correlated with true substitutability. Therefore, I cannot fully assess gender‑specific substitutability with precision in this setting. Nonetheless, to test whether gender segregation interacts with employment responses, I perform additional subsample analyses by the gender dominance of occupations and establishments.35 I find no significant differences in employment responses between male‑ vs female‑dominated contexts. One possible interpretation is that, conditional on occupation and firm size, women and men are generally substitutable from the firm’s perspective, and the substitutability measures used in the baseline are robust to gender composition. Does the external substitutability matter beyond the local level? Tradability of industries. The results above show that workers employed in less substitutable occupations in a given local labor market are more likely to extend their employment in response to the reform. I analyze the broad industry groups and discuss the results in terms of the conventional classification of industries by tradability to test whether the workers in tradable industries are more likely to respond to the raised retirement age. Such analyses allow me to test whether external substitutability matters beyond the local level. In tradable industries, firms can replace workers not only locally but also by outsourcing tasks globally, increasing substitutability (Drenik et al., 2023). I classify the industries by tradability following Gregory/Salomons/Zierahn (2022).36 Figure A15 shows no difference between tradable and 34 For example, Illing/Schwank/Tô (2024) find gender gaps in wages at the hiring stage for vacancies created by worker deaths in Germany. 35 I follow Tophoven et al. (2015) and define gender‑integrated occupations or establishments as those in which the proportion of men and women ranges from 21% to 79%. Gender‑dominated occupations or establishments are those in which the share of one gender exceeds 80%. 36 Tradable industries are: Mining (WZ08: B), Manufacturing (WZ08: C), Electricity, water supply (WZ08: D, E), Transport, storage (WZ08: H), Financial services (WZ08: K), Real estate (WZ08: L), Agriculture (WZ08: A), Information and communication (WZ08: J), Scientific and technical services (WZ08: M). Non‑tradable industries are Construction (WZ08: F), Wholesale and retail trade (WZ08: G), Hotels, restaurant (WZ08: I), IAB‑Discussion Paper 14|2025 38 untradable sectors. The result implies that substitutability does not matter beyond the local level when it comes to the effects of the reform on remaining in employment after 60.37 To conclude, I find that job‑specific skills and low internal and external substitutability are associated with a stronger increase in employment at ages 60–62 following the reform. While the analysis captures equilibrium effects — that is, match‑specific attributes shaped by both worker and firm — the pronounced retention of managers and specific workers is consistent with higher replacement costs, pointing to an important role for labor demand frictions. 6.3 The effect of raised ERA on wages by replacement costs The theoretical framework in chapter 2 predicts that raising the early retirement age weakens older workers’ outside options, most directly by removing the fallback of early pension access. Such elimination of outside options in the form of pensions reduces wages for affected workers on average. In this section, I analyze whether the effects on wages display heterogeneity by substitutability and job‑specific skills. There are two main opposite forces that display heterogeneity. On the one hand, the negative effect might be more pronounced for workers with high specificity (e.g., job‑specific skills or high‑level managerial roles), because their outside options may be especially limited. On the other hand, if such workers are more productive, firms may have incentives to offer wage premia to retain them, potentially offsetting the negative effect on their wages (see the derivations in chapter 2). Hence, the effect of the rise in ERA on wages by substitutability and the specificity of skills required to perform the given job may be both positive and negative. I test this implication by estimating RDD regressions with monthly wages as the outcome, focusing on subsamples that differ in job‑specificity and substitutability. Figure 6 presents the results. As expected, the overall wage effect is negative, consistent with reduced outside options weakening employee bargaining power, but effects vary across groups. Among the more replaceable workers, wages decline post‑reform. In contrast, managers and those in occupations that are difficult to replace externally sometimes experience wage gains after the reform, likely reflecting firms’ reluctance to lose strategically important employees. Public administration (WZ08: O), Education (WZ08: P), Health and social services (WZ08: Q), Cultural, social and personal services (WZ08: R, S), Household‑related services (WZ08: T), Other economic services (WZ08: N), Extraterritorial organizations (WZ08: U). I thank Duncan Roth for the help with the data. 37 In addition, the generalized categories of industries help me to test whether the external substitutability operates beyond the national level. I define industries by mapping based on the IAB establishment Panel, following the procedure described in Dauth/Eppelsheimer (2020). Figure A16 does not display significant differences by tradability. IAB‑Discussion Paper 14|2025 39 Figure 6.: Subsample analyses for the effect of the rise in ERA on wages at ages 60‑62 by substi‑ tutability measures Panel A: Human capital specificity Panel B: Hierarchical positions Panel C: Internal substitutability in the sample of small establishments Panel D: External substitutability (occupations) Notes: Coefficient plots from RDD regressions around the January 1952 cutoff. The estimates are obtained us‑ ing local linear regressions with first‑order polynomials, a triangular kernel, and mean square error–optimal bandwidth selection. Controls include calendar month of birth, Western residence, wages at age 46, and edu‑ cation. The vertical lines represent 95% confidence intervals based on robust standard errors clustered at the birth‑month level. Control means (on the x‑axis) refer to the average employment rate at ages 60–62 among the control group within each group’s optimal bandwidth. The corresponding detailed tables are reported in Table B23, Table B24, and Table B25. This result may reflect firm retention motives: when specific workers contribute more to firm profits, firms may offer wage premia despite weak outside options. However, selection into employment, whereby only the most productive or critical workers remain, may bias the upward wage effects observed in these groups. Overall, these findings highlight that the wage effects of the retirement reform are shaped by a complex interplay between retention needs and bargaining power, conditional on continued employment. IAB‑Discussion Paper 14|2025 40 Vestad, Ola Lotherington (2013): Labour Supply Effects of Early Retirement Provision. In: Labour Economics, Vol. 25, p. 98–109. Zweimüller, Josef; Winter‑Ebmer, Rudolf; Falkinger, Josef (1996): Retirement of Spouses and Social Security Reform. In: European Economic Review, Vol. 40, No. 2, p. 449–472. Zwick, Thomas; Bruns, Mona; Geyer, Johannes; Lorenz, Svenja (2022): Early Retirement of Employees in Demanding Jobs: Evidence from a German Pension Reform. In: The Journal of the Economics of Ageing, Vol. 22, p. 100 387. IAB‑Discussion Paper 14|2025 47 Appendix A1 Appendix figures Figure A1.: The assignment of normal retirement age by birth cohorts Notes: This figure depicts the assignment rule of normal retirement age by birth cohorts. Before the 1952 cohort, there was a women’s pathway to retirement (dashed line). The vertical dashed line at the January 1952 cohort indicates the birth cutoff from which the women’s pathway to early retirement was abolished. Starting from the 1952 cohort, the NRA for people eligible for the regular pathway to retirement is equal to the NRA for long‑term insured, which used to be 65, but increased by monthly increments per birth year starting from the 1947 cohort (black line). IAB‑Discussion Paper 14|2025 48 Figure A2.: Fraction of women employed, nonemployed, and retired at each age‑month by treat‑ ment and control group Notes: This figure displays the evolution of three main employment states (employment in black, nonemploy‑ ment in dark gray, and retirement in light gray‑ see chapter 4 for more details) over age by treatment status: (i) treated ‑ women born in 1952 (solid lines), and (ii) control‑ women born in 1951 (dashed lines). The first short‑ dashed vertical line (at age 47) corresponds to the age of the 1st treated cohort in 1999. The next two short dashed vertical lines show the age frame between the old ERA scheme (at age 60) and the new one (at least age 63) per the 1999 reform, while the last two short‑dashed vertical lines show the old NRA scheme (at age 65) and the new one (at age 65 years and six months) per the 2007 reform. IAB‑Discussion Paper 14|2025 49 Figure A3.: Fraction of women employed at each age‑month by treatment and control group Notes: This figure displays the fraction of women employed at each age month by two treatment statuses: treated (the 1952 birth cohort, in black) and control (the 1951 birth cohort, in gray). The period between the two dashed lines at 60 and 63 years old indicates the gaps between the two groups due to the 1999 reform un‑ der study. IAB‑Discussion Paper 14|2025 50 Figure A4.: The effect of the rise in ERA: RDD plot Notes: RDD regression of the share of employed at ages 60‑62 around the 1952 cutoff. For computing the RDD estimates, I use first‑order polynomials (upper graph) or automatic 4th order (lower graph), triangular kernel function, and mean square‑based optimal bandwidth selection procedure. The vertical line marks the birth cohort threshold 1952 (e.g., 0 corresponds to January 1952, ‑6 corresponds to people born six months before, in June 1951). IAB‑Discussion Paper 14|2025 51 Figure A5.: RDD by age in months Notes: Coefficient plots. Each vertical line corresponds to the RDD regression of the share of employed at a given age‑month. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. The points represent the estimated robust coefficients, and the bars represent the 95% confidence intervals, clustered at the birth month level. The red solid line rep‑ resents the control mean (with corresponding values displayed on the reversed y‑axis), while the red dashed lines represent the confidence intervals for the control means. IAB‑Discussion Paper 14|2025 52 Figure A6.: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by aggregate occupations Notes: Coefficient plots for RDD regressions around the 1952 cutoff. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. I perform subsample analyses by ten categories of occupations based on occupational classification. The vertical lines indicate 95% confidence intervals based on robust standard errors clustered at the birth month level. The con‑ trol means (on the x‑axis) show the employment share at the ages of 60‑62 in the corresponding subsample over the control group (born in 1951). A corresponding table with more details can be found in Table B15. IAB‑Discussion Paper 14|2025 53 Figure A7.: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by tenure Panel A: tenure measured at 46 years old (Me=4.5 Panel B: tenure measured at 58 years old (Me=7.7 years) years) Notes: Coefficient plots for RDD regressions around the 1952 cutoff. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. I perform subsample analyses by median split of tenure recorded at 46 years old (Panel A), and 58 years old (Panel B)‑ 4.5 and 7.7 years, respectively. The vertical lines indicate 95% confidence intervals based on robust standard errors clustered at the birth month level. The control means (on the x‑axis) show the employment share at the ages of 60‑62 in the corresponding subsample over the control group (born in 1951). A corresponding table with more details can be found in Table B14. IAB‑Discussion Paper 14|2025 54 Figure A8.: The effect of the rise in ERA on employment at ages 60‑62 by establishment size Notes: Coefficient plots for RDD regressions around the 1952 cutoff. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. I perform subsample analyses by establishment size categories. The three categories of establishment size are (1) up to 19, (2) 20‑249, (3) 250‑999, and (4) more than 1,000 workers employed at the establishment. The vertical lines indicate 95% confidence intervals based on robust standard errors clustered at the birth month level. The con‑ trol means (on the x‑axis) show the employment share at the ages of 60‑62 in the corresponding subsample over the control group (born in 1951). A corresponding table with more details can be found in Table B16. IAB‑Discussion Paper 14|2025 55 Figure A9.: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by number of internal substitutes Notes: Coefficient plots for RDD regressions around the 1952 cutoff. For computing the RDD estimates, I use local linear regressions, a triangular kernel function, and mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. I perform subsample analyses by the number of coworkers in the same 3‑digit occupation, restricting the sample to es‑ tablishments with fewer than 100 workers. The vertical lines indicate 95% confidence intervals based on robust standard errors clustered at the birth month level. The control means (on the x‑axis) show the employment share at the ages of 60‑62 in the corresponding subsample over the control group (born in 1951). A correspond‑ ing table with more details can be found in Table B13. IAB‑Discussion Paper 14|2025 56 B1 Appendix tables Table B1.: Baseline sample size after each restriction in German social security data N women, (birth cohort 1951) N women, (birth cohort 1952) N total unrestricted 34570 36776 71346 delete miners 34562 36771 71333 delete sailors 34560 36768 71328 delete parallel spells ‑ ‑ ‑ delete age‑months below 42 years old 32236 34166 66402 delete age‑months above 66 31988 33936 65924 delete repeating age‑months ‑ ‑ ‑ delete if not employed at 58‑59 15640 17130 32770 Notes: This table records the sample size after each of the restrictions in German social security data. The first column names the restrictions. The second and third columns list the sample size of treated and control groups, while the last column records the total sample size, i.e., the sum of the two pre‑ ceding columns. Table B2.: Balance check. The effect of the rise in ERA on covariates (1) West origin (2) non‑German The rise in ERA ‑0.007 0.013∗∗∗ (0.009) (0.005) Bandwidth 2.8 3.4 Observations 1179720 1179720 Notes: This table shows the effect of the rise in ERA on Western German origin (column 1) and non‑German nationality (column 2) (RDD regres‑ sion in Equation 9). The cutoff is January 1952, starting from which ERA was raised by at least 3 years. I pool all observations from the month af‑ ter a worker’s 60th birthday to their 63rd birthday. I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for West‑ ern residence, wages at the age of 46, and educa‑ tion. Robust standard errors in parentheses are clustered at the birth‑month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 63 Table B3.: The effect of the rise in ERA on employment outcomes at 60‑62 years old (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage The rise in ERA 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.027) (0.014) (0.016) (0.005) (0.006) (0.021) (23.368) Bandwidth 2.9 3.9 3.9 4.5 3.0 3.0 3.4 Control mean 0.774 0.455 0.232 0.079 0.050 0.179 1719.644 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Robust standard errors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Figure 2. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 64 Table B4.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by altering the estimation procedure (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage Panel A: bias‑corrected RD estimates with robust variance estimator (baseline) Robust 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.027) (0.014) (0.016) (0.005) (0.006) (0.021) (23.368) Panel B: conventional RD estimates with conventional variance estimator Conventional 0.166∗∗∗ 0.078∗∗∗ 0.003 0.051∗∗∗ ‑0.020∗∗∗ ‑0.144∗∗∗ ‑64.181∗∗∗ (0.002) (0.009) (0.014) (0.003) (0.002) (0.002) (21.622) Panel C: bias‑corrected RD estimates with conventional variance estimator Bias‑corrected 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.002) (0.009) (0.014) (0.003) (0.002) (0.002) (21.622) Bandwidth 2.9 3.9 3.9 4.5 3.0 3.0 3.4 Control mean 0.774 0.455 0.232 0.079 0.050 0.179 1719.644 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Panel A shows the bias‑corrected RD estimates with robust variance estimator, Panel B ‑conventional RD estimates with conventional variance estimator, Panel C ‑bias‑corrected RD estimates with conventional bias estimator. Standard errors in parentheses are clustered at the birth‑month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 65 Table B5.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by specified polynomial order (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage Panel A: polynomial function of order 1 (baseline) The rise in ERA 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.027) (0.014) (0.016) (0.005) (0.006) (0.021) (23.368) Bandwidth 2.9 3.9 3.9 4.5 3.0 3.0 3.4 Control mean 0.774 0.455 0.232 0.079 0.050 0.179 1719.644 Panel B: polynomial function of order 2 The rise in ERA 0.254∗∗∗ 0.063∗∗∗ 0.131∗∗∗ 0.056∗∗∗ ‑0.040∗∗∗ ‑0.215∗∗∗ ‑145.377∗∗∗ (0.047) (0.022) (0.023) (0.010) (0.014) (0.032) (33.774) Bandwidth 3.3 4.6 3.2 4.6 3.4 3.3 4.9 Control mean 0.769 0.458 0.232 0.079 0.050 0.181 1724.441 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I use a first‑order polynomial function in Panel A, and a second‑order polynomial in Panel B. I control for calendar month, a dummy for Western residence, wages at 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Robust standard errors in parentheses are clustered at the birth‑month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 66 Table B6.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the specified kernel function (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage Panel A: triangular weights (baseline) The rise in ERA 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.027) (0.014) (0.016) (0.005) (0.006) (0.021) (23.368) Bandwidth 2.9 3.9 3.9 4.5 3.0 3.0 3.4 Control mean 0.774 0.455 0.232 0.079 0.050 0.179 1719.644 Panel B: Epanechnikov kernel The rise in ERA 0.171∗∗∗ 0.071∗∗∗ 0.008 0.048∗∗∗ ‑0.020∗∗∗ ‑0.148∗∗∗ ‑99.628∗∗∗ (0.029) (0.016) (0.017) (0.006) (0.006) (0.022) (22.389) Bandwidth 2.9 3.8 4.0 4.3 3.0 3.0 3.5 Control mean 0.774 0.455 0.231 0.079 0.050 0.179 1719.644 Panel C: uniform kernel The rise in ERA 0.168∗∗∗ 0.076∗∗∗ 0.002 0.047∗∗∗ ‑0.023∗∗∗ ‑0.146∗∗∗ ‑133.632∗∗∗ (0.029) (0.021) (0.019) (0.004) (0.006) (0.024) (25.635) Bandwidth 2.7 3.3 3.1 2.8 2.7 2.9 2.6 Control mean 0.774 0.455 0.232 0.080 0.047 0.179 1723.688 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function in Panel A, Epanechnikov kernel in Panel B, and uniform weights in Panel C. I use a mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Robust standard errors in parentheses are clustered at the birth‑month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 67 Table B7.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by ad‑hoc bandwidth choices (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage Panel A: all the birth cohorts The rise in ERA 0.132∗∗∗ 0.091∗∗∗ ‑0.010 0.051∗∗∗ ‑0.013∗∗∗ ‑0.119∗∗∗ ‑4.547 (0.014) (0.014) (0.016) (0.005) (0.004) (0.011) (38.270) Bandwidth 12.0 12.0 12.0 12.0 12.0 12.0 12.0 Control mean 0.772 0.458 0.228 0.086 0.050 0.178 1744.540 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Panel B: excluding December 1951 and January 1952 birth cohorts The rise in ERA 0.115∗∗∗ 0.123∗∗∗ ‑0.055∗∗∗ 0.047∗∗∗ ‑0.006 ‑0.109∗∗∗ 97.799∗∗∗ (0.023) (0.018) (0.015) (0.007) (0.007) (0.017) (23.570) Bandwidth 12.0 12.0 12.0 12.0 12.0 12.0 12.0 Control mean 0.773 0.458 0.229 0.086 0.050 0.177 1744.764 Observations 1077408 1077408 1077408 1077408 1077408 1077408 895417 N workers 29928 29928 29928 29928 29928 29928 28662 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function and a 12‑month ad‑hoc bandwidth choice. Panel A displays the regressions with all cohorts born 1 year before or after the January 1952 cutoff, while Panel B removes the observations of women born 1 month around the cutoff. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Robust standard errors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Figure 2. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 68 Table B8.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the choice of covariates included (1) employment Panel A: baseline (month dummies, education, and western German residence) The rise in ERA 0.173∗∗∗ (0.027) Panel B: additionally controlling for regional origin and foreigner (non‑German) status The rise in ERA 0.173∗∗∗ (0.027) Panel C: no controls The rise in ERA 0.152∗∗∗ (0.024) Bandwidth 2.8 Control mean 0.772 Observations 1179720 N workers 32770 Notes: This table shows the effect of rise in the ERA on employment (RDD regression in Equa‑ tion 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. In Panel A, I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. In Panel B, I additionally control for western origin and foreigner status. I have no control variables in Panel C. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Ro‑ bust standard errors in parentheses are clustered at the birth‑month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 69 Table B9.: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the specified clustering method for standard errors (1) employ‑ ment (2) employees liable to social security (3) marginal part‑time employment (4) partial retirement (5) non‑ employ‑ ment (6) retire‑ ment (7) monthly wage Panel A: clustering at the birth date level (baseline) The rise in ERA 0.173∗∗∗ 0.070∗∗∗ 0.015 0.048∗∗∗ ‑0.021∗∗∗ ‑0.150∗∗∗ ‑116.522∗∗∗ (0.027) (0.014) (0.016) (0.005) (0.006) (0.021) (23.368) Bandwidth 2.9 3.9 3.9 4.5 3.0 3.0 3.4 Control mean 0.774 0.455 0.232 0.079 0.050 0.179 1719.644 Observations 1179720 1179720 1179720 1179720 1179720 1179720 1179720 N workers 32770 32770 32770 32770 32770 32770 32770 Panel B: clustering at the establishment level The rise in ERA 0.148∗∗∗ 0.070∗∗ 0.022 0.051∗∗ ‑0.017 ‑0.131∗∗∗ ‑136.181 (0.027) (0.035) (0.022) (0.021) (0.011) (0.026) (100.397) Bandwidth 3.3 3.2 3.3 3.6 3.3 3.4 2.9 Control mean 0.769 0.455 0.232 0.081 0.050 0.181 1723.688 Observations 1179720 1179720 1179720 1179720 1179720 1179720 980014 N workers 32770 32770 32770 32770 32770 32770 31346 Notes: These tables show the regression discontinuity design estimates around the cutoff of 1952, starting from which ERA rose by at least 3 years (Equation 9). I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). There are 3 mutually exclusive outcome variables: employment (column 1), nonemployment (column 5), and retirement (column 6). Employment can be further decomposed into columns 2‑4. I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I control for calendar month, a dummy for Western German residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951 (the control group). Robust standard errors in parentheses are clustered at the birth‑month level in Panel A and establishment level in Panel B. The corresponding coefficient plot can be found in Figure 2. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 70 Table B10.: Falsification test: RDD on employment at 60‑62 years old around placebo cutoffs (1) employment Panel A: 1948 cohort females Robust RDD ‑0.025 (0.101) Bandwidth 4.0 Observations 728892 N workers 20247 Panel B: 1949 cohort females Robust RDD 0.004 (0.784) Bandwidth 3.7 Observations 853812 N workers 23717 Panel C: 1950 cohort females Robust RDD ‑0.004 (0.438) Bandwidth 3.0 Observations 985104 N workers 27364 Panel D: 1951 cohort females Robust RDD 0.021 ∗ (0.062) Bandwidth 3.2 Observations 1083420 N workers 30095 Notes: This table shows the ef‑ fect of the rise in ERA on employ‑ ment (RDD regression in Equa‑ tion 9). Panel A performs RDD for the women born in 1947– 1948, around the January 1948 cutoff; Panel B ‑ born in 1948– 1949, around the January 1949 cutoff; Panel C ‑ born in 1949– 1950, around the January 1950 cutoff; and Panel D ‑ born in 1950–1951, around the January 1951 cutoff. I pool all obser‑ vations from the month after a worker’s 60th birthday to their 63rd birthday (age months corre‑ sponding to ages 60–62). I use a triangular kernel function and a mean square error‑based opti‑ mal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. Robust standard errors in parentheses are clustered at the birth month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 71 Table B11.: Falsification test: RDD on employment at 60‑62 years old around the reform cutoff for males (1) employment Robust RDD 0.051∗∗∗ (0.016) Bandwidth 3.2 Observations 1230624 N workers 34184 Notes: This table shows the ef‑ fect of the rise in ERA on employ‑ ment (RDD regression in Equa‑ tion 9) for males. The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birth‑ day (age months corresponding to ages 60–62). I use a tri‑ angular kernel function and a mean square error‑based opti‑ mal bandwidth choice. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to men born in 1951. Robust standard errors in parentheses are clustered at the birth month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 72 Table B18.: The effect of the rise in ERA on employment by task type employment (1) analytic non‑routine (2) interactive non‑routine (3) cognitive routine (4) manual routine (5) manual non‑routine The rise in ERA 0.248∗∗∗ 0.145∗∗∗ 0.246∗∗∗ 0.194∗∗ 0.027∗∗ (0.023) (0.018) (0.058) (0.077) (0.013) Bandwidth 4.2 3.1 2.9 2.9 3.7 Control mean 0.739 0.776 0.768 0.754 0.800 Observations 91152 218952 417384 88416 320724 N workers 2532 6082 11594 2456 8909 Notes: This table shows the effect of the rise in ERA on employment (RDD regression in Equation 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I perform subsample analyses in five task‑type categories. I control for calendar month, a dummy for Western resi‑ dence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard errors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Figure A14. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 79 Table B19.: The effect of the rise in ERA on employment at 60‑62 years old by industry by tradability employment (1) non‑tradable (2) tradable The rise in ERA 0.127∗∗∗ 0.227∗∗∗ (0.015) (0.057) Bandwidth 3.1 3.0 Control mean 0.784 0.737 Observations 838044 334044 N workers 23279 9279 Standard errors in parentheses ∗ ∗∗ ∗∗∗ p < 0.10, p < 0.05, p < 0.01 Notes: This table shows the effect of the rise in ERA on employment (RDD regression in Equa‑ tion 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I perform subsample analyses by tradability of sectors. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Ro‑ bust standard errors in parentheses are clus‑ tered at the birth‑month level. The corre‑ sponding coefficient plot can be found in Fig‑ ure A15. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 80 Table B20.: The effect of the rise in ERA on employment at 60‑62 years old by industry categories employment (1) Agriculture, hunting and forestry, fishing (2) Food and beverage (3) Manu‑ facture of consumer products (4) Manu‑ facture of industrial goods (5) Manufacture of capital and consu mer goods (6) Cons‑ truc‑ tion (7) Hotel and res‑ taurant (8) Trans‑ port, storage (9) Edu‑ cation The rise in ERA 0.124 0.158∗∗∗ 0.135∗∗∗ 0.157 0.145∗∗∗ 0.236∗∗∗ 0.164∗∗∗ 0.178∗∗∗ 0.124∗∗∗ (0.092) (0.050) (0.012) (0.128) (0.041) (0.018) (0.010) (0.033) (0.018) Bandwidth 3.2 4.3 3.2 2.7 4.8 3.5 3.5 3.4 2.9 Control mean 0.660 0.740 0.726 0.787 0.732 0.786 0.786 0.771 0.784 Observations 17136 34668 30960 41400 44748 21960 279036 252252 424080 N workers 476 963 860 1150 1243 610 7751 7007 11780 Notes: This table shows the effect of the rise in ERA on employment (RDD regression in Equation 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. I perform subsample analyses by industry categories. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard errors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Figure A16 ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 81 Table B21.: The effect of the rise in ERA on employment at 60‑62 years old by gender domination employment Panel A: gender domination in occupation gender‑integrated female‑dominated male‑dominated The rise in ERA 0.122∗∗∗ 0.288∗∗∗ 0.245∗∗∗ (0.018) (0.027) (0.029) Bandwidth 2.8 3.7 4.5 Control mean 0.736 0.778 0.724 Observations 174600 76752 20376 N workers 4850 2132 566 Panel B: gender domination in establishment gender‑integrated female‑dominated male‑dominated The rise in ERA 0.188∗∗∗ 0.207∗∗∗ 0.178∗∗ (0.015) (0.025) (0.072) Bandwidth 4.4 4.0 4.1 Control mean 0.741 0.782 0.681 Observations 144000 95184 19656 N workers 4000 2644 546 Notes: This table shows the effect of the rise in ERA on employment (RDD regres‑ sion in Equation 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangu‑ lar kernel function and a mean square error‑based optimal bandwidth choice. The subsample analyses are performed by gender dominance of occupations (Panel A) and establishments (Panel B). Gender‑integrated occupations and establishments are defined as those in which the proportion of men and women ranges from 21% to 79%. Gender‑dominated occupations/establishments are those in which the share of one of the genders exceeds 80%. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard errors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Figure A10 in the Appendix. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 82 Table B22.: The effect of the rise in ERA on employment at 60‑62 years old by demographic charac‑ teristics of employees employment Panel A: residence East West The rise in ERA 0.164∗∗∗ 0.174∗∗∗ (0.039) (0.024) Bandwidth 2.8 3.0 Observations 228168 949392 N workers 6338 26372 Panel B: residence of origin East West The rise in ERA 0.172∗∗∗ 0.171∗∗∗ (0.038) (0.025) Bandwidth 2.9 2.9 Observations 232776 945756 N workers 6466 26271 Panel C: education high school vocational university The rise in ERA 0.165∗∗∗ 0.183∗∗∗ 0.094∗∗∗ (0.029) (0.039) (0.024) Bandwidth 3.4 2.8 3.7 Observations 160740 897840 155340 N workers 4545 24940 4315 Notes: This table shows the effect of the rise in ERA on em‑ ployment (RDD regression in Equation 9). The cutoff is Jan‑ uary 1952, starting from which the ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. Panel A per‑ forms subsample analyses by the residence of the workers (dummy variable); Panel B divides the workers by Eastern and Western German origin, proxied by the place of residence of the first worker as observed in the employment biography; and Panel C divides the sample by educational categories. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. Robust standard errors in parentheses are clustered at the birth month level. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 83 Table B23.: The effect of the rise in ERA on monthly wages at 60‑62 years old by measures of worker skills monthly wages (1) (2) Panel A: human capital specificity of occupations low high The rise in ERA ‑233.398∗∗∗ ‑90.348∗∗∗ (64.777) (15.485) Bandwidth 2.7 3.9 Control mean 1694.602 1744.467 Observations 458872 520927 N workers 14619 16721 Panel B: by hierarchical position not a manager manager The rise in ERA ‑144.746∗∗∗ 1360.297∗∗∗ (25.361) (195.102) Bandwidth 3.4 3.4 Control mean 1704.381 3287.176 Observations 968243 11771 N workers 30976 370 Notes: This table shows the effect of the rise in ERA on monthly wages (RDD regression in Equation 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all observations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. Panel A is per‑ formed by ”HK specificity”‑ which stands for human capital specificity of occupation. It is based on the re‑ turn of experience in Mincer equations performed sep‑ arately for each of the 3‑digit occupations. Then, I cre‑ ate a dummy variable based on a median split across all the occupations. Panel B stands for managerial sta‑ tus, which is created as a dummy from the last 2 digits of the 5‑digit occupational variables. I control for cal‑ endar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard er‑ rors in parentheses are clustered at the birth‑month level. The corresponding coefficient plot can be found in Panel A and Panel B of Figure 6. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 84 Table B24.: The effect of the rise in ERA on monthly wages at 60‑62 years old by internal substi‑ tutability (number of coworkers in the same occupation) monthly wages (1) 0 (2) 1‑4 (3) at least 5 Panel A: all the establishment categories The rise in ERA ‑217.352 ‑960.250∗∗∗ ‑33.783∗∗∗ (327.504) (156.779) (6.537) Bandwidth 3.5 3.4 3.6 Control mean 1566.372 1349.531 1754.040 Observations 44454 33085 877485 N workers 1427 1054 28054 Panel B: establishments with fewer than 100 workers The rise in ERA ‑1104.690∗∗∗ ‑997.944∗∗∗ ‑474.748∗∗ (98.434) (160.777) (184.439) Bandwidth 3.7 3.3 2.8 Control mean 1818.764 1761.9 1792.3 Observations 18601 20033 47272 N workers 610 641 1503 Notes: This table shows the effect of the rise in ERA on monthly wages (RDD regression in Equation 9). The cutoff is January 1952, starting from which ERA rose by at least 3 years. I pool all obser‑ vations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a trian‑ gular kernel function and a mean square error‑based optimal band‑ width choice. I perform subsample analyses by 3 categories of in‑ ternal substitutes: 0, 1‑4, and at least 5 workers. The Panel A dis‑ plays the results for all the sizes of establishments, while Panel B zooms in on the smaller establishments with fewer than 100 work‑ ers. I control for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard errors in parentheses are clustered at the birth month level. The corresponding coefficient plot can be found in Panel C and Panel D of Figure 6. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 85 Table B25.: The effect of the rise in ERA on monthly wages at 60‑62 years old by external substi‑ tutability measures monthly wages (1) below 0.5 (2) 0.5‑1 (3) above 1 Panel A: external labor market thickness (industry) The rise in ERA ‑155.193 ‑81.246∗∗∗ ‑90.073∗∗∗ (214.400) (12.265) (21.603) Bandwidth 4.1 4.0 4.1 Control mean 1403.030 1518.215 1903.207 Observations 54534 352647 567063 N workers 1738 11245 18164 Panel B: external labor market thickness (occupation) The rise in ERA 335.624∗∗∗ ‑78.478∗∗∗ ‑178.565∗∗∗ (59.037) (22.142) (38.253) Bandwidth 4.1 4.0 3.3 Control mean 1320.551 1602.430 1872.761 Observations 38515 430331 505147 N workers 1264 13728 16148 Notes: This table shows the effect of the rise in ERA on monthly wages (RDD regression in Equation 9). The cutoff is January 1952, starting from which the ERA rose by at least 3 years. I pool all ob‑ servations from the month after a worker’s 60th birthday to their 63rd birthday (age months corresponding to ages 60–62). I use a triangular kernel function and a mean square error‑based optimal bandwidth choice. Panel A shows subsample analyses by exter‑ nal labor market thickness (ELMT) for a given occupation, based on the index taking values below 0.5, 0.5‑1, and above 1. Panel B shows subsample analyses by ELMT for a given industry. I con‑ trol for calendar month, a dummy for Western residence, wages at the age of 46, and education. The control means are the average values of the outcomes when I limit the sample to women born in 1951. Robust standard errors in parentheses are clustered at the birth month level. The corresponding coefficient plot can be found in Panel E and Panel F of Figure 6. ∗ (p < 0.10), ∗∗ (p < 0.05), ∗∗∗ (p < 0.01). IAB‑Discussion Paper 14|2025 86 List of Figures Figure 1: Discontinuity in birth cohorts..................................................... 14 Figure 2: The effect of the rise in ERA on the employment state (overall and from each category) ...................................................................... 25 Figure 3: The effect of the rise in ERA on employment at ages 60‑62 by return to experience in a given occupation and occupational hierarchy level ........ 31 Figure 4: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by number of internal and external substitutes for the given occupation ........................................................................... 34 Figure 5: External labor market thickness by German industry and occupation in 2010 .................................................................................. 36 Figure 6: Subsample analyses for the effect of the rise in ERA on wages at ages 60‑62 by substitutability measures............................................... 40 Figure A1: The assignment of normal retirement age by birth cohorts ................. 48 Figure A2: Fraction of women employed, nonemployed, and retired at each age‑month by treatment and control group.................................... 49 Figure A3: Fraction of women employed at each age‑month by treatment and control group ....................................................................... 50 Figure A4: The effect of the rise in ERA: RDD plot ......................................... 51 Figure A5: RDD by age in months............................................................. 52 Figure A6: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by aggregate occupations ........................................... 53 Figure A7: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by tenure Figure A8: The effect of the rise in ERA on employment at ages 60‑62 by establishment size Figure A9: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by number of internal substitutes Figure A10: The effect of the rise in ERA on employment at ages 60‑62 by gender‑ composition of occupations and establishments Figure A11: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by external substitutability of a given industry ............................................................... 54 .................................................................. 55 .................................. 56 .............................. 57 .................. 58 Figure A12: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by external substitutability, restricting to small establishments with at most 100 workers ...................................... 59 Figure A13: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by external substitutability, using only the female workforce for computations ................................................................... 59 IAB‑Discussion Paper 14|2025 87 Figure A14: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by task type ........................................................... 60 Figure A15: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by tradability of industries Figure A16: Subsample analyses for the effect of the rise in ERA on employment at ages 60‑62 by aggregate industry categories .......................................... 61 ................................... 62 List of Tables Table B1: Baseline sample size after each restriction in German social security data ................................................................................... 63 Table B2: Balance check. The effect of the rise in ERA on covariates .................. 63 Table B3: The effect of the rise in ERA on employment outcomes at 60‑62 years old ..................................................................................... 64 Table B4: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by altering the estimation procedure............................................................................. 65 Table B5: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by specified polynomial order .................................................................................. 66 Table B6: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the specified kernel function ............................................................................... 67 Table B7: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by ad‑hoc bandwidth choices ................................................................................ 68 Table B8: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the choice of covariates included .............................................................................. 69 Table B9: Robustness and sensitivity checks. The effect of the rise in ERA on employment outcomes at 60‑62 years old by the specified clustering method for standard errors ........................................................ 70 Table B10: Falsification test: RDD on employment at 60‑62 years old around placebo cutoffs ................................................................................ 71 Table B11: Falsification test: RDD on employment at 60‑62 years old around the reform cutoff for males............................................................. 72 IAB‑Discussion Paper 14|2025 88